Protocol for applying a network-enabled gene discovery pipeline to non-model plant species Authors: Dae Kwan Ko, Federica Brandizzi Published: 2026 ABSTRACT SummaryIdentifying upstream regulators of key genes is essential for understanding gene regulatory mechanisms and translating these insights into functional targets. Here, we present a protocol for applying the network-enabled gene discovery pipeline (NEEDLE) to non-model plant species. We describe steps for environment setup, data preparation, computational analysis, expected outputs, and parameter considerations. NEEDLE integrates RNA sequencing (RNA-seq) processing, weighted gene co-expression analysis (WGCNA), Gene Network Inference with Ensemble of trees (GENIE3), and promoter conservation analysis to prioritize candidate transcriptional regulators.For complete details on the use and execution of this protocol, please refer to Ko and Brandizzi.1 Before you begin This protocol describes the application of NEEDLE to identify transcription factors (TFs) associated with gene expression dynamics during sorghum stem development. The workflow is demonstrated using an RNA sequencing (RNA-seq) dataset comprising 11 developmental samples, as reported in the associated research article.1 Although the protocol uses sorghum as an example, NEEDLE can be applied to other plant species when transcriptome datasets capture sufficient biological variation across time points, developmental stages, tissues, genotypes, or treatment conditions.Note: Based on our testing across diverse transcriptome datasets, six or more biologically meaningful samples are generally recommended for stable network modeling; however, this threshold should be considered a practical guideline rather than a strict requirement, as the optimal sample number may vary depending on dataset complexity and the biological question. Note: Based on our testing across diverse transcriptome datasets, six or more biologically meaningful samples are generally recommended for stable network modeling; however, this threshold should be considered a practical guideline rather than a strict requirement, as the optimal sample number may vary depending on dataset complexity and the biological question. NEEDLE consists of three major components: (i) RNA-seq data processing, (ii) WGCNA-based coexpression network construction followed by GENIE3-based gene regulatory network (GRN) inference, and (iii) validation of predicted TF–target relationships through in silico promoter scanning and conservation analysis, followed by optional in planta transient expression assays such as dual-luciferase reporter analysis (Figure 1). The present protocol focuses on the computational and in silico portion of NEEDLE, whereas detailed procedures for in planta validation are described in the associated research article1 and are not included here.Figure 1Overview of the NEEDLE workflowSchematic representation of the NEEDLE pipeline described in this protocol. The workflow begins with selection or generation of an appropriate RNA-seq dataset to address the biological question. Raw sequencing data are then processed to generate expression matrices for downstream analysis.The computational workflow proceeds through RNA-seq data processing, WGCNA-based coexpression network construction, GENIE3-based GRN inference, and promoter scanning and conservation analysis for in silico validation of predicted regulatory relationships. Corresponding protocol step numbers are shown next to each major analysis stage: RNA-seq processing, Steps 1–13; WGCNA-based coexpression network construction, Steps 14–24; GENIE3-based GRN construction, Steps 25–30; and promoter scanning and conservation analysis, Steps 31–33. Overview of the NEEDLE workflow Schematic representation of the NEEDLE pipeline described in this protocol. The workflow begins with selection or generation of an appropriate RNA-seq dataset to address the biological question. Raw sequencing data are then processed to generate expression matrices for downstream analysis. The computational workflow proceeds through RNA-seq data processing, WGCNA-based coexpression network construction, GENIE3-based GRN inference, and promoter scanning and conservation analysis for in silico validation of predicted regulatory relationships. Corresponding protocol step numbers are shown next to each major analysis stage: RNA-seq processing, Steps 1–13; WGCNA-based coexpression network construction, Steps 14–24; GENIE3-based GRN construction, Steps 25–30; and promoter scanning and conservation analysis, Steps 31–33. This protocol requires both command-line software and R packages. Command-line software is used mainly for downloading, quality-controlling, trimming, aligning, sorting, and quantifying RNA-seq reads, whereas R packages are used for differential expression analysis, coexpression network construction, GENIE3-based regulatory inference, and visualization.Note: These requirements are listed separately below and in the key resources table to improve clarity for users. The workflow was validated using the software and packages listed in the key resources table. Exact versions are not required for all tools, but users seeking full reproducibility should use the indicated versions when possible and record all software and package versions before analysis. Note: These requirements are listed separately below and in the key resources table to improve clarity for users. The workflow was validated using the software and packages listed in the key resources table. Exact versions are not required for all tools, but users seeking full reproducibility should use the indicated versions when possible and record all software and package versions before analysis. Innovation Most gene discovery pipelines aimed at identifying regulatory drivers of agronomic traits benefit from integrating multiple omics layers, such as epigenomics, proteomics, and chromatin accessibility data. However, these resources are often unavailable for crop and non-model plant species.2,3 NEEDLE addresses this limitation by providing a transcriptome-driven workflow for prioritizing candidate regulators using widely accessible RNA-seq datasets. Rather than introducing a new algorithm, NEEDLE integrates and standardizes established analytical components, including RNA-seq processing, WGCNA-based coexpression network construction, intramodular connectivity analysis, GENIE3-based GRN inference, and promoter conservation analysis, into a coherent pipeline optimized for dynamic transcriptome data. A key strength of NEEDLE is its structured prioritization strategy, which combines module membership, gene connectivity, and regulatory hierarchy metrics to systematically nominate candidate TFs associated with genes or pathways of interest. The broader NEEDLE workflow also supports downstream experimental validation of selected TF–target relationships using transient expression assays. In the present protocol, we focus on the computational and in silico components of NEEDLE, while experimental validation procedures are described in the associated research article.1 This protocol lowers the barrier to network-guided gene discovery in data-limited systems and expands the practical utility of systems biology approaches for crop improvement. Hardware and software preparation 1.Prepare a computational environment suitable for transcriptome analysis.a.Ensure access to a Linux- or macOS-based workstation or high-performance computing (HPC) system.b.Confirm that the system has sufficient memory and storage to process RNA-seq datasets. Note: NEEDLE relies exclusively on transcriptome data and does not require epigenomic or proteomic inputs. 2.Install or confirm access to the required command-line software for RNA-seq processing.a.Install or confirm access to SRA Toolkit, FastQC, Cutadapt, Bowtie2, TopHat2, SAMtools, and HTSeq.b.Install Cufflinks only if FPKM estimation is needed.c.Install Cytoscape only if network visualization will be performed locally.3.Install R and the required R packages.a.Install R version 4.0.3 or later.b.Install DESeq2, dplyr, RColorBrewer, gplots, WGCNA, flashClust, pheatmap, and GENIE3.c.Record R package version using sessionInfo().4.Download and organize transcriptome input data.a.Obtain raw FASTQ files or processed gene expression matrices for samples to be analyzed.b.Prepare a metadata table containing sample identifiers, developmental stage or treatment information, and replicate annotations.c.Organize raw data, processed expression matrices, metadata files, and analysis outputs into clearly labeled directories. Note: Use of an HPC system is recommended for processing large RNA-seq datasets, but it is not required for all steps. Users without access to an HPC system may run the workflow on a local Linux- or macOS-based workstation, provided sufficient memory and disk space are available. 1.Prepare a computational environment suitable for transcriptome analysis.a.Ensure access to a Linux- or macOS-based workstation or high-performance computing (HPC) system.b.Confirm that the system has sufficient memory and storage to process RNA-seq datasets. Prepare a computational environment suitable for transcriptome analysis.a.Ensure access to a Linux- or macOS-based workstation or high-performance computing (HPC) system.b.Confirm that the system has sufficient memory and storage to process RNA-seq datasets. a.Ensure access to a Linux- or macOS-based workstation or high-performance computing (HPC) system. Ensure access to a Linux- or macOS-based workstation or high-performance computing (HPC) system. b.Confirm that the system has sufficient memory and storage to process RNA-seq datasets. Confirm that the system has sufficient memory and storage to process RNA-seq datasets. Note: NEEDLE relies exclusively on transcriptome data and does not require epigenomic or proteomic inputs. 2.Install or confirm access to the required command-line software for RNA-seq processing.a.Install or confirm access to SRA Toolkit, FastQC, Cutadapt, Bowtie2, TopHat2, SAMtools, and HTSeq.b.Install Cufflinks only if FPKM estimation is needed.c.Install Cytoscape only if network visualization will be performed locally. Install or confirm access to the required command-line software for RNA-seq processing.a.Install or confirm access to SRA Toolkit, FastQC, Cutadapt, Bowtie2, TopHat2, SAMtools, and HTSeq.b.Install Cufflinks only if FPKM estimation is needed.c.Install Cytoscape only if network visualization will be performed locally. a.Install or confirm access to SRA Toolkit, FastQC, Cutadapt, Bowtie2, TopHat2, SAMtools, and HTSeq. Install or confirm access to SRA Toolkit, FastQC, Cutadapt, Bowtie2, TopHat2, SAMtools, and HTSeq. b.Install Cufflinks only if FPKM estimation is needed. Install Cufflinks only if FPKM estimation is needed. c.Install Cytoscape only if network visualization will be performed locally. Install Cytoscape only if network visualization will be performed locally. 3.Install R and the required R packages.a.Install R version 4.0.3 or later.b.Install DESeq2, dplyr, RColorBrewer, gplots, WGCNA, flashClust, pheatmap, and GENIE3.c.Record R package version using sessionInfo(). Install R and the required R packages.a.Install R version 4.0.3 or later.b.Install DESeq2, dplyr, RColorBrewer, gplots, WGCNA, flashClust, pheatmap, and GENIE3.c.Record R package version using sessionInfo(). a.Install R version 4.0.3 or later. Install R version 4.0.3 or later. b.Install DESeq2, dplyr, RColorBrewer, gplots, WGCNA, flashClust, pheatmap, and GENIE3. Install DESeq2, dplyr, RColorBrewer, gplots, WGCNA, flashClust, pheatmap, and GENIE3. c.Record R package version using sessionInfo(). Record R package version using sessionInfo(). 4.Download and organize transcriptome input data.a.Obtain raw FASTQ files or processed gene expression matrices for samples to be analyzed.b.Prepare a metadata table containing sample identifiers, developmental stage or treatment information, and replicate annotations.c.Organize raw data, processed expression matrices, metadata files, and analysis outputs into clearly labeled directories. Download and organize transcriptome input data.a.Obtain raw FASTQ files or processed gene expression matrices for samples to be analyzed.b.Prepare a metadata table containing sample identifiers, developmental stage or treatment information, and replicate annotations.c.Organize raw data, processed expression matrices, metadata files, and analysis outputs into clearly labeled directories. a.Obtain raw FASTQ files or processed gene expression matrices for samples to be analyzed. Obtain raw FASTQ files or processed gene expression matrices for samples to be analyzed. b.Prepare a metadata table containing sample identifiers, developmental stage or treatment information, and replicate annotations. Prepare a metadata table containing sample identifiers, developmental stage or treatment information, and replicate annotations. c.Organize raw data, processed expression matrices, metadata files, and analysis outputs into clearly labeled directories. Organize raw data, processed expression matrices, metadata files, and analysis outputs into clearly labeled directories. Note: Use of an HPC system is recommended for processing large RNA-seq datasets, but it is not required for all steps. Users without access to an HPC system may run the workflow on a local Linux- or macOS-based workstation, provided sufficient memory and disk space are available. Clone the script directory Timing: 30 min Timing: 30 min All scripts required to run this protocol are publicly available in the NEEDLE GitHub repository: https://github.com/DaeKwan-Ko/needle.5.Open a Linux terminal on a local workstation or High-Performance Computing (HPC) system.6.Clone the repository Bashgit clone https://github.com/DaeKwan-Ko/needle 5.Open a Linux terminal on a local workstation or High-Performance Computing (HPC) system. Open a Linux terminal on a local workstation or High-Performance Computing (HPC) system. 6.Clone the repository Clone the repository Bash git clone https://github.com/DaeKwan-Ko/needle This command downloads the full script directory into the current working directory.Note: The repository contains multiple top-level directories corresponding to different analyses and plant species. In this protocol, users should follow the scripts and example files associated with the sorghum dataset.CRITICAL: Record the commit version or release tag used for the analysis to ensure reproducibility. Note: The repository contains multiple top-level directories corresponding to different analyses and plant species. In this protocol, users should follow the scripts and example files associated with the sorghum dataset. CRITICAL: Record the commit version or release tag used for the analysis to ensure reproducibility. Install required command-line software and R packages Timing: 2 h, depending on system configurations Timing: 2 h, depending on system configurations This protocol requires command-line software for RNA-seq processing and R packages for differential expression analysis, coexpression network construction, regulatory inference, and visualization. These are listed separately below to clarify which tools are installed and executed outside R and which are installed within R.7.Confirm hardware availability.a.Ensure that computational resources meet the requirements described in Computational requirements and in the key resources table.8.Confirm operating systems and R installation.a.Use a Linux- or macOS-based operating system, either on a local workstation or an HPC system.b.Confirm that R version 4.0.3 or later is installed.9.Install or confirm access to the required command-line software.a.Install or confirm access to SRA Toolkit, FastQC, Cutadapt, Bowtie2, TopHat2, SAMtools, and HTSeq.b.Install Cufflinks if FPKM estimation is needed.c.Install Cytoscape if network visualization will be performed.d.If using an HPC system that uses environment modules, confirm availability of the required tools. Bashmodule avail bowtie2module avail samtoolsmodule avail sra-toolkit10.Install the required R packages. Rinstall.packages(c(“dplyr”, “RColorBrewer”, “gplots”, “WGCNA”, “flashClust”, “pheatmap”))if (!requireNamespace(“BiocManager”, quietly = TRUE)) { install.packages(“BiocManager”)}BiocManager::install(c(“DESeq2”, “GENIE3”))sessionInfo()CRITICAL: Exact versions are not required for all users, but users seeking to reproduce the outputs reported in this protocol should use the software and package versions listed in the key resources table when possible. Differences in package versions, random seed handling, or network parameters may affect module assignment, network topology, and ranked TF–target outputs.Optional: If SRA Toolkit or other command-line tools are not available through an HPC module system, users can install them locally, for example through Conda. Bashconda create -n needle_env -c conda-forge -c bioconda sra-tools fastqc cutadapt bowtie2 tophat samtools htseqconda activate needle_env 7.Confirm hardware availability.a.Ensure that computational resources meet the requirements described in Computational requirements and in the key resources table. Confirm hardware availability.a.Ensure that computational resources meet the requirements described in Computational requirements and in the key resources table. a.Ensure that computational resources meet the requirements described in Computational requirements and in the key resources table. Ensure that computational resources meet the requirements described in Computational requirements and in the key resources table. 8.Confirm operating systems and R installation.a.Use a Linux- or macOS-based operating system, either on a local workstation or an HPC system.b.Confirm that R version 4.0.3 or later is installed. Confirm operating systems and R installation.a.Use a Linux- or macOS-based operating system, either on a local workstation or an HPC system.b.Confirm that R version 4.0.3 or later is installed. a.Use a Linux- or macOS-based operating system, either on a local workstation or an HPC system. Use a Linux- or macOS-based operating system, either on a local workstation or an HPC system. b.Confirm that R version 4.0.3 or later is installed. Confirm that R version 4.0.3 or later is installed. 9.Install or confirm access to the required command-line software.a.Install or confirm access to SRA Toolkit, FastQC, Cutadapt, Bowtie2, TopHat2, SAMtools, and HTSeq.b.Install Cufflinks if FPKM estimation is needed.c.Install Cytoscape if network visualization will be performed.d.If using an HPC system that uses environment modules, confirm availability of the required tools. Install or confirm access to the required command-line software.a.Install or confirm access to SRA Toolkit, FastQC, Cutadapt, Bowtie2, TopHat2, SAMtools, and HTSeq.b.Install Cufflinks if FPKM estimation is needed.c.Install Cytoscape if network visualization will be performed.d.If using an HPC system that uses environment modules, confirm availability of the required tools. a.Install or confirm access to SRA Toolkit, FastQC, Cutadapt, Bowtie2, TopHat2, SAMtools, and HTSeq. Install or confirm access to SRA Toolkit, FastQC, Cutadapt, Bowtie2, TopHat2, SAMtools, and HTSeq. b.Install Cufflinks if FPKM estimation is needed. Install Cufflinks if FPKM estimation is needed. c.Install Cytoscape if network visualization will be performed. Install Cytoscape if network visualization will be performed. d.If using an HPC system that uses environment modules, confirm availability of the required tools. If using an HPC system that uses environment modules, confirm availability of the required tools. Bash module avail bowtie2 module avail samtools module avail sra-toolkit 10.Install the required R packages. Install the required R packages. R install.packages(c(“dplyr”, “RColorBrewer”, “gplots”, “WGCNA”, “flashClust”, “pheatmap”)) if (!requireNamespace(“BiocManager”, quietly = TRUE)) { install.packages(“BiocManager”) } BiocManager::install(c(“DESeq2”, “GENIE3”)) sessionInfo() CRITICAL: Exact versions are not required for all users, but users seeking to reproduce the outputs reported in this protocol should use the software and package versions listed in the key resources table when possible. Differences in package versions, random seed handling, or network parameters may affect module assignment, network topology, and ranked TF–target outputs. Optional: If SRA Toolkit or other command-line tools are not available through an HPC module system, users can install them locally, for example through Conda. Bash conda create -n needle_env -c conda-forge -c bioconda sra-tools fastqc cutadapt bowtie2 tophat samtools htseq conda activate needle_env Download input RNA-seq raw files Timing: ∼1 h, depending on network speed Timing: ∼1 h, depending on network speed This protocol uses the sorghum RNA-seq dataset described in Ko and Brandizzi, consisting of 11 developmental samples with three biological replicates.1 Users may either download the same dataset or substitute their own RNA-seq dataset.11.Load SRA Toolkit if using an HPC module system. Bashmodule load SRA-Toolkit/2.10.7-centos_linux6412.Install SRA Toolkit locally if it is not available as an HPC module. Bashconda create -n needle_sra -c conda-forge -c bioconda sra-toolsconda activate needle_sraNote: Alternatively, users can download precompiled SRA Toolkit binaries from NCBI or download FASTQ files directly through repository web interfaces when available.13.Download FASTQ files using prefetch and fasterq-dump. Bashprefetch FILEIDfasterq-dump FILEID --threads 4 --outdir fastq/ 11.Load SRA Toolkit if using an HPC module system. Load SRA Toolkit if using an HPC module system. Bash module load SRA-Toolkit/2.10.7-centos_linux64 12.Install SRA Toolkit locally if it is not available as an HPC module. Install SRA Toolkit locally if it is not available as an HPC module. Bash conda create -n needle_sra -c conda-forge -c bioconda sra-tools conda activate needle_sra Note: Alternatively, users can download precompiled SRA Toolkit binaries from NCBI or download FASTQ files directly through repository web interfaces when available. 13.Download FASTQ files using prefetch and fasterq-dump. Download FASTQ files using prefetch and fasterq-dump. Bash prefetch FILEID fasterq-dump FILEID --threads 4 --outdir fastq/ Replace FILEID with the corresponding SRA accession number for each sample.Note: If the dataset is paired-end, users can add --split-files to generate separate mate files. Bashfasterq-dump FILEID --split-files --threads 4 --outdir fastq/CRITICAL: Ensure sufficient disk space before downloading and converting SRA files. RNA-seq FASTQ files are typically large, often several GB per sample. For large datasets, consider using external or institutional storage. Note: If the dataset is paired-end, users can add --split-files to generate separate mate files. Bash fasterq-dump FILEID --split-files --threads 4 --outdir fastq/ CRITICAL: Ensure sufficient disk space before downloading and converting SRA files. RNA-seq FASTQ files are typically large, often several GB per sample. For large datasets, consider using external or institutional storage. Split combined paired-end reads, if required Timing: ∼1 h, depending on file size Timing: ∼1 h, depending on file size This step is required only when paired-end reads are interleaved, meaning that R1 and R2 reads are stored together in a single FASTQ file.Note: Skip this step if FASTQ files are already provided as separate R1 and R2 files or if fasterq-dump --split-files generated separate files ending in _1.fastq and _2.fastq.14.Inspect the FASTQ file structure to determine whether reads are interleaved.15.Split R1 reads from an interleaved FASTQ file. Bashsed -ne '1∼8p;2∼8p;3∼8p;4∼8p' INPUT.fastq > OUTPUT_R1.fastq16.Split R2 reads from an interleaved FASTQ file. Bashsed -ne '5∼8p;6∼8p;7∼8p;8∼8p' INPUT.fastq > OUTPUT_R2.fastq17.Verify that the split R1 and R2 FASTQ files contain the same number of reads. Bashwc -l OUTPUT_R1.fastqwc -l OUTPUT_R2.fastq Note: Skip this step if FASTQ files are already provided as separate R1 and R2 files or if fasterq-dump --split-files generated separate files ending in _1.fastq and _2.fastq. 14.Inspect the FASTQ file structure to determine whether reads are interleaved. Inspect the FASTQ file structure to determine whether reads are interleaved. 15.Split R1 reads from an interleaved FASTQ file. Split R1 reads from an interleaved FASTQ file. Bash sed -ne '1∼8p;2∼8p;3∼8p;4∼8p' INPUT.fastq > OUTPUT_R1.fastq 16.Split R2 reads from an interleaved FASTQ file. Split R2 reads from an interleaved FASTQ file. Bash sed -ne '5∼8p;6∼8p;7∼8p;8∼8p' INPUT.fastq > OUTPUT_R2.fastq 17.Verify that the split R1 and R2 FASTQ files contain the same number of reads. Verify that the split R1 and R2 FASTQ files contain the same number of reads. Bash wc -l OUTPUT_R1.fastq wc -l OUTPUT_R2.fastq Divide the line count for each file by 4 to obtain the number of reads. The R1 and R2 files should contain the same number of reads.CRITICAL: Incorrectly splitting paired-end reads will disrupt read pairing and reduce alignment quality. Confirm that R1 and R2 files have matching read counts before proceeding.Optional: Compress split FASTQ files to reduce storage requirements. Bashgzip OUTPUT_R1.fastqgzip OUTPUT_R2.fastq CRITICAL: Incorrectly splitting paired-end reads will disrupt read pairing and reduce alignment quality. Confirm that R1 and R2 files have matching read counts before proceeding. Optional: Compress split FASTQ files to reduce storage requirements. Bash gzip OUTPUT_R1.fastq gzip OUTPUT_R2.fastq Computational requirements For optimal performance, particularly during RNA-seq read mapping and quantification, access to an HPC system is recommended because raw sequencing files can be large and alignment steps are computationally intensive. However, an HPC system is not strictly required for all users or all steps. Small to moderate RNA-seq datasets can be processed on a Linux- or macOS-based workstation if sufficient memory, CPU capacity, and storage are available. For local execution, we recommend a modern workstation equipped with an 8-core or higher CPU, at least 16 GB of RAM, and a minimum of 100 GB of available storage. Increased RAM of 32 GB or more is recommended for larger transcriptome datasets, especially those containing more than 20 samples or a large number of expressed genes. Downstream analyses, including differential expression analysis, WGCNA-based coexpression network construction, module detection, connectivity analysis, and GENIE3-based GRN inference, can generally be performed on a modern personal workstation for moderate-sized datasets. For large datasets, users may reduce memory demands by prefiltering lowly expressed or low-variance genes, using WGCNA blockwise module detection, or running memory-intensive steps on an HPC node. Key resources table REAGENT or RESOURCESOURCEIDENTIFIERSoftware and algorithmsR statistical computing environmentR Core Teamhttps://www.R-project.org/; version 4.0.3 or later recommendedSRA ToolkitNCBIhttps://github.com/ncbi/sra-tools; version 2.10.7 used in this protocol; newer versions are acceptableFastQCBabraham Bioinformaticshttps://www.bioinformatics.babraham.ac.uk/projects/fastqc/CutadaptMartin et al.4https://cutadapt.readthedocs.io/en/stable/Bowtie2Langmead & Salzberg5https://bowtie-bio.sourceforge.net/bowtie2/index.shtmlTopHat2Kim et al.6https://ccb.jhu.edu/software/tophat/index.shtmlSAMtoolsLi et al.7https://samtools.sourceforge.net/HTSeqPutri et al.8https://htseq.readthedocs.io/en/latest/CufflinksTrapnell et al.9http://cole-trapnell-lab.github.io/cufflinks/; optional, required only for FPKM estimationmVISTA online platformFrazer et al.10https://genome.lbl.gov/vista/index.shtml; used as an online web toolCytoscapeShannon et al.11https://cytoscape.org/; optional, used for network visualizationDESeq2Love et al.12https://bioconductor.org/packages/release/bioc/html/DESeq2.htmldplyrCRANhttps://cran.r-project.org/package=dplyrRColorBrewerCRANhttps://cran.r-project.org/package=RColorBrewergplotsCRANhttps://cran.r-project.org/package=gplotsWGCNALangfelder & Horvath13https://cran.r-project.org/web/packages/WGCNA/index.htmlflashClustCRANhttps://cran.r-project.org/package=flashClustpheatmapCRANhttps://cran.r-project.org/package=pheatmapGENIE3Huynh-Thu et al.14https://bioconductor.org/packages/release/bioc/html/GENIE3.htmlNEEDLE scriptsKo and Brandizzi1https://github.com/DaeKwan-Ko/needle; https://doi.org/10.5281/zenodo.14560598GitSoftware Freedom Conservancyhttps://git-scm.com/; used to clone the NEEDLE repositoryConda or MinicondaAnaconda, Inc.https://docs.conda.io/; optional, used for local software installationDeposited dataSorghum RNA-seq dataset, stem developmentKo and Brandizzi1https://genome.jgi.doe.gov/portal/SorbicEProfiling_14/SorbicEProfiling_14.info.htmlSorghum genome v3.1.1Phytozome, JGIhttps://phytozome-next.jgi.doe.gov/Plant Transcription Factor Database (PlantTFDB)Jin et al.15https://planttfdb.gao-lab.org/OtherLocal workstationApple iMac, example configuration used in this study3.8 GHz 8-Core Intel Core i7; 16–32 GB RAMHPC clusterLinux-based HPC system≥32 GB RAM; ≥8 processing cores recommended for large RNA-seq datasets REAGENT or RESOURCESOURCEIDENTIFIERSoftware and algorithmsR statistical computing environmentR Core Teamhttps://www.R-project.org/; version 4.0.3 or later recommendedSRA ToolkitNCBIhttps://github.com/ncbi/sra-tools; version 2.10.7 used in this protocol; newer versions are acceptableFastQCBabraham Bioinformaticshttps://www.bioinformatics.babraham.ac.uk/projects/fastqc/CutadaptMartin et al.4https://cutadapt.readthedocs.io/en/stable/Bowtie2Langmead & Salzberg5https://bowtie-bio.sourceforge.net/bowtie2/index.shtmlTopHat2Kim et al.6https://ccb.jhu.edu/software/tophat/index.shtmlSAMtoolsLi et al.7https://samtools.sourceforge.net/HTSeqPutri et al.8https://htseq.readthedocs.io/en/latest/CufflinksTrapnell et al.9http://cole-trapnell-lab.github.io/cufflinks/; optional, required only for FPKM estimationmVISTA online platformFrazer et al.10https://genome.lbl.gov/vista/index.shtml; used as an online web toolCytoscapeShannon et al.11https://cytoscape.org/; optional, used for network visualizationDESeq2Love et al.12https://bioconductor.org/packages/release/bioc/html/DESeq2.htmldplyrCRANhttps://cran.r-project.org/package=dplyrRColorBrewerCRANhttps://cran.r-project.org/package=RColorBrewergplotsCRANhttps://cran.r-project.org/package=gplotsWGCNALangfelder & Horvath13https://cran.r-project.org/web/packages/WGCNA/index.htmlflashClustCRANhttps://cran.r-project.org/package=flashClustpheatmapCRANhttps://cran.r-project.org/package=pheatmapGENIE3Huynh-Thu et al.14https://bioconductor.org/packages/release/bioc/html/GENIE3.htmlNEEDLE scriptsKo and Brandizzi1https://github.com/DaeKwan-Ko/needle; https://doi.org/10.5281/zenodo.14560598GitSoftware Freedom Conservancyhttps://git-scm.com/; used to clone the NEEDLE repositoryConda or MinicondaAnaconda, Inc.https://docs.conda.io/; optional, used for local software installationDeposited dataSorghum RNA-seq dataset, stem developmentKo and Brandizzi1https://genome.jgi.doe.gov/portal/SorbicEProfiling_14/SorbicEProfiling_14.info.htmlSorghum genome v3.1.1Phytozome, JGIhttps://phytozome-next.jgi.doe.gov/Plant Transcription Factor Database (PlantTFDB)Jin et al.15https://planttfdb.gao-lab.org/OtherLocal workstationApple iMac, example configuration used in this study3.8 GHz 8-Core Intel Core i7; 16–32 GB RAMHPC clusterLinux-based HPC system≥32 GB RAM; ≥8 processing cores recommended for large RNA-seq datasets Materials and equipment • Recommended computing environment ◦A Linux- or macOS-based workstation or HPC system is recommended for running this protocol.◦HPC access is particularly useful for RNA-seq read mapping and quantification because raw sequencing files can be large and alignment steps are computationally intensive.◦An HPC system is not required for all users or all steps. Small to moderate RNA-seq datasets can be processed on a local workstation if sufficient memory, CPU capacity, and storage are available. • Minimum local workstation configuration ◦CPU: 8-core processor or higher◦Memory: At least 16 GB RAM◦Storage: At least 100 GB available disk space◦Operating system: Linux or macOS • Recommended configuration for larger datasets ◦Memory: 32 GB RAM or more◦Storage: More than 100 GB available disk space, depending on the number and size of FASTQ files◦For datasets containing more than 20 samples or a large number of expressed genes, additional memory may be needed for WGCNA and GENIE3 analyses. • Memory-reduction options ◦Users working with large transcriptome datasets may reduce memory demands by filtering lowly expressed genes before network construction.◦Users may also remove genes with low expression variance, use WGCNA blockwise module detection, or run memory-intensive steps on an HPC node. • Expected computational burden ◦RNA-seq read alignment and quantification are usually the most computationally intensive steps.◦Downstream analyses, including DESeq2 differential expression analysis, WGCNA-based coexpression network construction, module detection, connectivity analysis, and GENIE3-based gene regulatory network inference, can generally be performed on a modern workstation for moderate-sized datasets. • Recommended computing environment ◦A Linux- or macOS-based workstation or HPC system is recommended for running this protocol.◦HPC access is particularly useful for RNA-seq read mapping and quantification because raw sequencing files can be large and alignment steps are computationally intensive.◦An HPC system is not required for all users or all steps. Small to moderate RNA-seq datasets can be processed on a local workstation if sufficient memory, CPU capacity, and storage are available. Recommended computing environment ◦A Linux- or macOS-based workstation or HPC system is recommended for running this protocol.◦HPC access is particularly useful for RNA-seq read mapping and quantification because raw sequencing files can be large and alignment steps are computationally intensive.◦An HPC system is not required for all users or all steps. Small to moderate RNA-seq datasets can be processed on a local workstation if sufficient memory, CPU capacity, and storage are available. ◦A Linux- or macOS-based workstation or HPC system is recommended for running this protocol. A Linux- or macOS-based workstation or HPC system is recommended for running this protocol. ◦HPC access is particularly useful for RNA-seq read mapping and quantification because raw sequencing files can be large and alignment steps are computationally intensive. HPC access is particularly useful for RNA-seq read mapping and quantification because raw sequencing files can be large and alignment steps are computationally intensive. ◦An HPC system is not required for all users or all steps. Small to moderate RNA-seq datasets can be processed on a local workstation if sufficient memory, CPU capacity, and storage are available. An HPC system is not required for all users or all steps. Small to moderate RNA-seq datasets can be processed on a local workstation if sufficient memory, CPU capacity, and storage are available. • Minimum local workstation configuration ◦CPU: 8-core processor or higher◦Memory: At least 16 GB RAM◦Storage: At least 100 GB available disk space◦Operating system: Linux or macOS Minimum local workstation configuration ◦CPU: 8-core processor or higher◦Memory: At least 16 GB RAM◦Storage: At least 100 GB available disk space◦Operating system: Linux or macOS ◦CPU: 8-core processor or higher CPU: 8-core processor or higher ◦Memory: At least 16 GB RAM Memory: At least 16 GB RAM ◦Storage: At least 100 GB available disk space Storage: At least 100 GB available disk space ◦Operating system: Linux or macOS Operating system: Linux or macOS • Recommended configuration for larger datasets ◦Memory: 32 GB RAM or more◦Storage: More than 100 GB available disk space, depending on the number and size of FASTQ files◦For datasets containing more than 20 samples or a large number of expressed genes, additional memory may be needed for WGCNA and GENIE3 analyses. Recommended configuration for larger datasets ◦Memory: 32 GB RAM or more◦Storage: More than 100 GB available disk space, depending on the number and size of FASTQ files◦For datasets containing more than 20 samples or a large number of expressed genes, additional memory may be needed for WGCNA and GENIE3 analyses. ◦Memory: 32 GB RAM or more Memory: 32 GB RAM or more ◦Storage: More than 100 GB available disk space, depending on the number and size of FASTQ files Storage: More than 100 GB available disk space, depending on the number and size of FASTQ files ◦For datasets containing more than 20 samples or a large number of expressed genes, additional memory may be needed for WGCNA and GENIE3 analyses. For datasets containing more than 20 samples or a large number of expressed genes, additional memory may be needed for WGCNA and GENIE3 analyses. • Memory-reduction options ◦Users working with large transcriptome datasets may reduce memory demands by filtering lowly expressed genes before network construction.◦Users may also remove genes with low expression variance, use WGCNA blockwise module detection, or run memory-intensive steps on an HPC node. Memory-reduction options ◦Users working with large transcriptome datasets may reduce memory demands by filtering lowly expressed genes before network construction.◦Users may also remove genes with low expression variance, use WGCNA blockwise module detection, or run memory-intensive steps on an HPC node. ◦Users working with large transcriptome datasets may reduce memory demands by filtering lowly expressed genes before network construction. Users working with large transcriptome datasets may reduce memory demands by filtering lowly expressed genes before network construction. ◦Users may also remove genes with low expression variance, use WGCNA blockwise module detection, or run memory-intensive steps on an HPC node. Users may also remove genes with low expression variance, use WGCNA blockwise module detection, or run memory-intensive steps on an HPC node. • Expected computational burden ◦RNA-seq read alignment and quantification are usually the most computationally intensive steps.◦Downstream analyses, including DESeq2 differential expression analysis, WGCNA-based coexpression network construction, module detection, connectivity analysis, and GENIE3-based gene regulatory network inference, can generally be performed on a modern workstation for moderate-sized datasets. Expected computational burden ◦RNA-seq read alignment and quantification are usually the most computationally intensive steps.◦Downstream analyses, including DESeq2 differential expression analysis, WGCNA-based coexpression network construction, module detection, connectivity analysis, and GENIE3-based gene regulatory network inference, can generally be performed on a modern workstation for moderate-sized datasets. ◦RNA-seq read alignment and quantification are usually the most computationally intensive steps. RNA-seq read alignment and quantification are usually the most computationally intensive steps. ◦Downstream analyses, including DESeq2 differential expression analysis, WGCNA-based coexpression network construction, module detection, connectivity analysis, and GENIE3-based gene regulatory network inference, can generally be performed on a modern workstation for moderate-sized datasets. Downstream analyses, including DESeq2 differential expression analysis, WGCNA-based coexpression network construction, module detection, connectivity analysis, and GENIE3-based gene regulatory network inference, can generally be performed on a modern workstation for moderate-sized datasets. Step-by-step method details Processing RNA-seq data Timing: 1+ day, depending on genome size, read depth, and compute resources Timing: 1+ day, depending on genome size, read depth, and compute resources This section describes RNA-seq processing from raw FASTQ files through alignment and gene-level quantification for downstream differential expression analysis. The major outputs are quality-controlled FASTQ files, sorted BAM files, gene-level count tables, and DESeq2-based differentially expressed gene (DEG) lists. These steps correspond to the RNA-seq processing stage of the NEEDLE workflow shown in Figure 1.1.Verify raw FASTQ integrity and read counts.a.Count the number of lines in each FASTQ file. Bashwc -l ${INPUT_FASTQ} > ${OUTPUT_TXT}b.Divide the line count by 4 to obtain the number of reads.Note: Each read corresponds to four lines in FASTQ format.2.Perform initial quality control of raw reads using FastQC. Bashfastqc -o ${OUTPUT_DIR} -f fastq ${INPUT_FASTQ}Note: For paired-end libraries, run FastQC separately for both mates, R1 and R2.3.Trim adapters and low-quality bases using Cutadapt.a.Use paired-end trimming for paired-end RNA-seq libraries.b.Use adapter sequences appropriate for the library preparation kit and sequencing platform. Bashcutadapt \ -a AGATCGGAAGAGCACACGTCTGAACTCCAGTCAC \ -A AGATCGGAAGAGCGTCGTGTAGGGAAAGAGTGTAGATCTCGGTGGTCGCCGTATCATT \ -q 20 \ -m 30 \ -o ${OUTPUT_R1} \ -p ${OUTPUT_R2} \ ${INPUT_R1} ${INPUT_R2}CRITICAL: Confirm adapter content from FastQC reports before trimming.4.Re-check read counts after trimming. Bashwc -l ${CLEAN_FASTQ} > ${OUTPUT_TXT}Note: A reduction in read number is expected after adapter and quality trimming.5.Compress trimmed FASTQ files, if desired. Bashgzip ${CLEAN_FASTQ}Optional: Skip compression if the downstream aligner or analysis environment does not support gzipped FASTQ files.6.Perform quality control of trimmed reads using FastQC.a.Confirm that adapter contamination has been removed.b.Confirm that per-base sequence quality has improved. Bashfastqc -o ${OUTPUT_DIR} -f fastq ${CLEAN_FASTQ}7.Build a Bowtie2 index for the sorghum reference genome. Bashbowtie2-build -f ${SORGHUM_FASTA} ${INDEX_PREFIX}CRITICAL: Ensure the reference FASTA and gene annotation file, such as GFF or GTF, are from the same genome build/version.8.Map reads to the sorghum genome using TopHat2.a.Use paired-end alignment for paired-end RNA-seq libraries.b.Set --library-type to match the library strandedness.c.Set -i and -I based on the expected intron-size range of the reference genome. Bashtophat2 \ -o ${OUTPUT_DIR} \ -p $1 \ -I 30000 \ -i 10 \ -G ${SORGHUM_GFF} \ --library-type fr-firststrand \ ${INDEX_PREFIX} \ ${CLEAN_R1} ${CLEAN_R2}CRITICAL: Incorrect strandedness will reduce mapping accuracy and distort gene-level read counts.Note: The -i parameter defines the minimum intron length allowed during splice-junction discovery, whereas -I defines the maximum intron length. In this sorghum example, -i 10 and -I 30000 were used to accommodate the expected intron-size range in the reference annotation. Users applying NEEDLE to other species should adjust these values based on the intron-size distribution of the corresponding genome annotation. Setting -I too low may prevent reads spanning long introns from aligning correctly, whereas setting it too high may increase runtime and the number of candidate splice junctions.9.Sort the alignment BAM file by genomic coordinates. Bashsamtools sort -o ${SORTED_BAM} ${INPUT_BAM}10.Index the sorted BAM file. Bashsamtools index ${SORTED_BAM}11.Estimate transcript abundance by calculating FPKM values using Cufflinks, if needed. Bashcufflinks \ -G ${SORGHUM_GFF} \ -b ${SORGHUM_FASTA} \ --library-type fr-firststrand \ -I 30000 \ -q \ -o ${OUTPUT_DIR} \ ${SORTED_BAM}Optional: This step is not required for DESeq2-based differential expression analysis, which uses raw counts. Include this step only if FPKM values or transcript assembly outputs are needed for downstream analyses.12.Generate gene-level read counts using HTSeq-count. Bashhtseq-count \ --format=bam \ --order=pos \ --stranded=reverse \ --minaqual=10 \ --mode=union \ --type=gene \ --idattr=ID \ ${SORTED_BAM} ${SORGHUM_GFF} > ${COUNT_TXT}CRITICAL: The --stranded option must match the library type and aligner output. If strandedness is uncertain, evaluate different settings, such as yes, no, and reverse, on a subset of samples and confirm expected expression patterns for known genes.13.Perform differential gene expression analysis using DESeq2.Note: For clarity and to meet STAR Protocols formatting limits, repetitive pairwise DESeq2 operations are programmatically generated using stage labels inferred from sample identifiers. This analysis is equivalent to running DESeq2 separately for each pairwise comparison but is implemented as a single fitted model with extracted contrasts.a.Open R and load required packages. RsuppressPackageStartupMessages({ library(DESeq2) library(dplyr) library(RColorBrewer) library(gplots) })b.Locate HTSeq-count output files. Rfiles <- list.files(pattern = “htseq\\.txt$”, full.names = TRUE)stopifnot(length(files) > 0)c.Read count files and remove HTSeq summary rows. Rread_htseq <- function(f) { x <- read.table( f, header = FALSE, sep = “\t”, colClasses = c(“character”, “numeric”), stringsAsFactors = FALSE ) # Keep gene rows only and remove HTSeq summary lines x <- x[!grepl(“ˆ__”, x[, 1]), , drop = FALSE] return(x)}dat_list <- lapply(files, read_htseq)d.Ensure consistent gene order across files and build a count matrix. Rgenes <- dat_list[[1]][, 1]stopifnot( all(vapply(dat_list, function(x) identical(x[, 1], genes), logical(1))))count_mat <- do.call(cbind, lapply(dat_list, function(x) x[, 2]))rownames(count_mat) <- genese.Infer sample names from file names. Rsample_ids <- sub(“\\.htseq\\.txt$”, ““, basename(files))colnames(count_mat) <- sample_idscount_mat[1:5, 1:5]f.Define developmental stages from sample identifiers and construct sample metadata. R# Expected sample format: A1, A2, A3., K1, K2, K3stage <- sub(“[0-9]+$”, ““, sample_ids)coldata <- data.frame( sample = sample_ids, stage = factor(stage), row.names = sample_ids)coldataNote: The example assumes that sample names are formatted as a stage letter followed by a replicate number, such as A1, A2, A3, K1, K2, K3. Users should modify this step if their sample naming convention differs.g.Filter low-count genes across all samples. Rkeep <- rowSums(count_mat) > 50count_mat_filt <- count_mat[keep, , drop = FALSE]dim(count_mat_filt)h.Fit the DESeq2 model across all stages. Rdds <- DESeqDataSetFromMatrix( countData = count_mat_filt, colData = coldata, design = ∼ stage)dds <- DESeq(dds)ddsi.Generate an example MA plot using one contrast, such as stage B versus stage A. Rexample_contrast <- c(“stage”, “B”, “A”)res_ex <- results(dds, contrast = example_contrast)plotMA(res_ex, alpha = 0.05, ylim = c(-5, 5))j.Generate a sample distance heatmap using variance-stabilized counts. Rvsd <- vst(dds, blind = FALSE)d <- dist(t(assay(vsd)))mat <- as.matrix(d)hmcol <- colorRampPalette(brewer.pal(9, “Greens”))(100)heatmap.2( mat, trace = “none”, col = rev(hmcol), density.info = “none”, margins = c(8, 8), main = “Sample distances (VST)”)k.Generate all pairwise stage contrasts. Rstages <- levels(coldata$stage)pairs <- combn(stages, 2, simplify = FALSE)length(pairs)l.Create an output directory for DESeq2 pairwise results. Routdir <- “DESeq2_pairwise_results”dir.create(outdir, showWarnings = FALSE)m.DEG tables for each pairwise comparison using an adjusted p-value cutoff of 0.05. Ralpha <- 0.05for (p in pairs) { s1 <- p[1] s2 <- p[2] res <- results(dds, contrast = c(“stage”, s2, s1)) %>% as.data.frame() res <- res[!is.na(res$padj), , drop = FALSE] sig <- res[res$padj < alpha, , drop = FALSE] out_file <- file.path( outdir, paste0(“deseq2_gene_list_padj_”, alpha, “_”, s1, “_vs_”, s2, “.tsv”) ) write.table( sig, file = out_file, sep = “\t”, quote = FALSE, row.names = TRUE )}list.files(outdir)[1:5]Note: Users may modify the adjusted p-value cutoff, add a log2 fold-change threshold, or define biologically specific contrasts depending on the experimental design. 1.Verify raw FASTQ integrity and read counts.a.Count the number of lines in each FASTQ file. Bashwc -l ${INPUT_FASTQ} > ${OUTPUT_TXT}b.Divide the line count by 4 to obtain the number of reads.Note: Each read corresponds to four lines in FASTQ format. Verify raw FASTQ integrity and read counts.a.Count the number of lines in each FASTQ file. Bashwc -l ${INPUT_FASTQ} > ${OUTPUT_TXT}b.Divide the line count by 4 to obtain the number of reads.Note: Each read corresponds to four lines in FASTQ format. a.Count the number of lines in each FASTQ file. Bashwc -l ${INPUT_FASTQ} > ${OUTPUT_TXT} Count the number of lines in each FASTQ file. Bashwc -l ${INPUT_FASTQ} > ${OUTPUT_TXT} Bash wc -l ${INPUT_FASTQ} > ${OUTPUT_TXT} b.Divide the line count by 4 to obtain the number of reads.Note: Each read corresponds to four lines in FASTQ format. Divide the line count by 4 to obtain the number of reads.Note: Each read corresponds to four lines in FASTQ format. Note: Each read corresponds to four lines in FASTQ format. 2.Perform initial quality control of raw reads using FastQC. Perform initial quality control of raw reads using FastQC. Bash fastqc -o ${OUTPUT_DIR} -f fastq ${INPUT_FASTQ} Note: For paired-end libraries, run FastQC separately for both mates, R1 and R2. 3.Trim adapters and low-quality bases using Cutadapt.a.Use paired-end trimming for paired-end RNA-seq libraries.b.Use adapter sequences appropriate for the library preparation kit and sequencing platform. Trim adapters and low-quality bases using Cutadapt.a.Use paired-end trimming for paired-end RNA-seq libraries.b.Use adapter sequences appropriate for the library preparation kit and sequencing platform. a.Use paired-end trimming for paired-end RNA-seq libraries. Use paired-end trimming for paired-end RNA-seq libraries. b.Use adapter sequences appropriate for the library preparation kit and sequencing platform. Use adapter sequences appropriate for the library preparation kit and sequencing platform. Bash cutadapt \ -a AGATCGGAAGAGCACACGTCTGAACTCCAGTCAC \ -A AGATCGGAAGAGCGTCGTGTAGGGAAAGAGTGTAGATCTCGGTGGTCGCCGTATCATT \ -q 20 \ -m 30 \ -o ${OUTPUT_R1} \ -p ${OUTPUT_R2} \ ${INPUT_R1} ${INPUT_R2} CRITICAL: Confirm adapter content from FastQC reports before trimming. 4.Re-check read counts after trimming. Re-check read counts after trimming. Bash wc -l ${CLEAN_FASTQ} > ${OUTPUT_TXT} Note: A reduction in read number is expected after adapter and quality trimming. 5.Compress trimmed FASTQ files, if desired. Compress trimmed FASTQ files, if desired. Bash gzip ${CLEAN_FASTQ} Optional: Skip compression if the downstream aligner or analysis environment does not support gzipped FASTQ files. 6.Perform quality control of trimmed reads using FastQC.a.Confirm that adapter contamination has been removed.b.Confirm that per-base sequence quality has improved. Perform quality control of trimmed reads using FastQC.a.Confirm that adapter contamination has been removed.b.Confirm that per-base sequence quality has improved. a.Confirm that adapter contamination has been removed. Confirm that adapter contamination has been removed. b.Confirm that per-base sequence quality has improved. Confirm that per-base sequence quality has improved. Bash fastqc -o ${OUTPUT_DIR} -f fastq ${CLEAN_FASTQ} 7.Build a Bowtie2 index for the sorghum reference genome. Build a Bowtie2 index for the sorghum reference genome. Bash bowtie2-build -f ${SORGHUM_FASTA} ${INDEX_PREFIX} CRITICAL: Ensure the reference FASTA and gene annotation file, such as GFF or GTF, are from the same genome build/version. 8.Map reads to the sorghum genome using TopHat2.a.Use paired-end alignment for paired-end RNA-seq libraries.b.Set --library-type to match the library strandedness.c.Set -i and -I based on the expected intron-size range of the reference genome. Map reads to the sorghum genome using TopHat2.a.Use paired-end alignment for paired-end RNA-seq libraries.b.Set --library-type to match the library strandedness.c.Set -i and -I based on the expected intron-size range of the reference genome. a.Use paired-end alignment for paired-end RNA-seq libraries. Use paired-end alignment for paired-end RNA-seq libraries. b.Set --library-type to match the library strandedness. Set --library-type to match the library strandedness. c.Set -i and -I based on the expected intron-size range of the reference genome. Set -i and -I based on the expected intron-size range of the reference genome. Bash tophat2 \ -o ${OUTPUT_DIR} \ -p $1 \ -I 30000 \ -i 10 \ -G ${SORGHUM_GFF} \ --library-type fr-firststrand \ ${INDEX_PREFIX} \ ${CLEAN_R1} ${CLEAN_R2} CRITICAL: Incorrect strandedness will reduce mapping accuracy and distort gene-level read counts. Note: The -i parameter defines the minimum intron length allowed during splice-junction discovery, whereas -I defines the maximum intron length. In this sorghum example, -i 10 and -I 30000 were used to accommodate the expected intron-size range in the reference annotation. Users applying NEEDLE to other species should adjust these values based on the intron-size distribution of the corresponding genome annotation. Setting -I too low may prevent reads spanning long introns from aligning correctly, whereas setting it too high may increase runtime and the number of candidate splice junctions. 9.Sort the alignment BAM file by genomic coordinates. Sort the alignment BAM file by genomic coordinates. Bash samtools sort -o ${SORTED_BAM} ${INPUT_BAM} 10.Index the sorted BAM file. Index the sorted BAM file. Bash samtools index ${SORTED_BAM} 11.Estimate transcript abundance by calculating FPKM values using Cufflinks, if needed. Estimate transcript abundance by calculating FPKM values using Cufflinks, if needed. Bash cufflinks \ -G ${SORGHUM_GFF} \ -b ${SORGHUM_FASTA} \ --library-type fr-firststrand \ -I 30000 \ -q \ -o ${OUTPUT_DIR} \ ${SORTED_BAM} Optional: This step is not required for DESeq2-based differential expression analysis, which uses raw counts. Include this step only if FPKM values or transcript assembly outputs are needed for downstream analyses. 12.Generate gene-level read counts using HTSeq-count. Generate gene-level read counts using HTSeq-count. Bash htseq-count \ --format=bam \ --order=pos \ --stranded=reverse \ --minaqual=10 \ --mode=union \ --type=gene \ --idattr=ID \ ${SORTED_BAM} ${SORGHUM_GFF} > ${COUNT_TXT} CRITICAL: The --stranded option must match the library type and aligner output. If strandedness is uncertain, evaluate different settings, such as yes, no, and reverse, on a subset of samples and confirm expected expression patterns for known genes. 13.Perform differential gene expression analysis using DESeq2.Note: For clarity and to meet STAR Protocols formatting limits, repetitive pairwise DESeq2 operations are programmatically generated using stage labels inferred from sample identifiers. This analysis is equivalent to running DESeq2 separately for each pairwise comparison but is implemented as a single fitted model with extracted contrasts.a.Open R and load required packages. RsuppressPackageStartupMessages({ library(DESeq2) library(dplyr) library(RColorBrewer) library(gplots) })b.Locate HTSeq-count output files. Rfiles <- list.files(pattern = “htseq\\.txt$”, full.names = TRUE)stopifnot(length(files) > 0)c.Read count files and remove HTSeq summary rows. Rread_htseq <- function(f) { x <- read.table( f, header = FALSE, sep = “\t”, colClasses = c(“character”, “numeric”), stringsAsFactors = FALSE ) # Keep gene rows only and remove HTSeq summary lines x <- x[!grepl(“ˆ__”, x[, 1]), , drop = FALSE] return(x)}dat_list <- lapply(files, read_htseq)d.Ensure consistent gene order across files and build a count matrix. Rgenes <- dat_list[[1]][, 1]stopifnot( all(vapply(dat_list, function(x) identical(x[, 1], genes), logical(1))))count_mat <- do.call(cbind, lapply(dat_list, function(x) x[, 2]))rownames(count_mat) <- genese.Infer sample names from file names. Rsample_ids <- sub(“\\.htseq\\.txt$”, ““, basename(files))colnames(count_mat) <- sample_idscount_mat[1:5, 1:5]f.Define developmental stages from sample identifiers and construct sample metadata. R# Expected sample format: A1, A2, A3., K1, K2, K3stage <- sub(“[0-9]+$”, ““, sample_ids)coldata <- data.frame( sample = sample_ids, stage = factor(stage), row.names = sample_ids)coldataNote: The example assumes that sample names are formatted as a stage letter followed by a replicate number, such as A1, A2, A3, K1, K2, K3. Users should modify this step if their sample naming convention differs.g.Filter low-count genes across all samples. Rkeep <- rowSums(count_mat) > 50count_mat_filt <- count_mat[keep, , drop = FALSE]dim(count_mat_filt)h.Fit the DESeq2 model across all stages. Rdds <- DESeqDataSetFromMatrix( countData = count_mat_filt, colData = coldata, design = ∼ stage)dds <- DESeq(dds)ddsi.Generate an example MA plot using one contrast, such as stage B versus stage A. Rexample_contrast <- c(“stage”, “B”, “A”)res_ex <- results(dds, contrast = example_contrast)plotMA(res_ex, alpha = 0.05, ylim = c(-5, 5))j.Generate a sample distance heatmap using variance-stabilized counts. Rvsd <- vst(dds, blind = FALSE)d <- dist(t(assay(vsd)))mat <- as.matrix(d)hmcol <- colorRampPalette(brewer.pal(9, “Greens”))(100)heatmap.2( mat, trace = “none”, col = rev(hmcol), density.info = “none”, margins = c(8, 8), main = “Sample distances (VST)”)k.Generate all pairwise stage contrasts. Rstages <- levels(coldata$stage)pairs <- combn(stages, 2, simplify = FALSE)length(pairs)l.Create an output directory for DESeq2 pairwise results. Routdir <- “DESeq2_pairwise_results”dir.create(outdir, showWarnings = FALSE)m.DEG tables for each pairwise comparison using an adjusted p-value cutoff of 0.05. Ralpha <- 0.05for (p in pairs) { s1 <- p[1] s2 <- p[2] res <- results(dds, contrast = c(“stage”, s2, s1)) %>% as.data.frame() res <- res[!is.na(res$padj), , drop = FALSE] sig <- res[res$padj < alpha, , drop = FALSE] out_file <- file.path( outdir, paste0(“deseq2_gene_list_padj_”, alpha, “_”, s1, “_vs_”, s2, “.tsv”) ) write.table( sig, file = out_file, sep = “\t”, quote = FALSE, row.names = TRUE )}list.files(outdir)[1:5]Note: Users may modify the adjusted p-value cutoff, add a log2 fold-change threshold, or define biologically specific contrasts depending on the experimental design. Perform differential gene expression analysis using DESeq2.Note: For clarity and to meet STAR Protocols formatting limits, repetitive pairwise DESeq2 operations are programmatically generated using stage labels inferred from sample identifiers. This analysis is equivalent to running DESeq2 separately for each pairwise comparison but is implemented as a single fitted model with extracted contrasts.a.Open R and load required packages. RsuppressPackageStartupMessages({ library(DESeq2) library(dplyr) library(RColorBrewer) library(gplots) })b.Locate HTSeq-count output files. Rfiles <- list.files(pattern = “htseq\\.txt$”, full.names = TRUE)stopifnot(length(files) > 0)c.Read count files and remove HTSeq summary rows. Rread_htseq <- function(f) { x <- read.table( f, header = FALSE, sep = “\t”, colClasses = c(“character”, “numeric”), stringsAsFactors = FALSE ) # Keep gene rows only and remove HTSeq summary lines x <- x[!grepl(“ˆ__”, x[, 1]), , drop = FALSE] return(x)}dat_list <- lapply(files, read_htseq)d.Ensure consistent gene order across files and build a count matrix. Rgenes <- dat_list[[1]][, 1]stopifnot( all(vapply(dat_list, function(x) identical(x[, 1], genes), logical(1))))count_mat <- do.call(cbind, lapply(dat_list, function(x) x[, 2]))rownames(count_mat) <- genese.Infer sample names from file names. Rsample_ids <- sub(“\\.htseq\\.txt$”, ““, basename(files))colnames(count_mat) <- sample_idscount_mat[1:5, 1:5]f.Define developmental stages from sample identifiers and construct sample metadata. R# Expected sample format: A1, A2, A3., K1, K2, K3stage <- sub(“[0-9]+$”, ““, sample_ids)coldata <- data.frame( sample = sample_ids, stage = factor(stage), row.names = sample_ids)coldataNote: The example assumes that sample names are formatted as a stage letter followed by a replicate number, such as A1, A2, A3, K1, K2, K3. Users should modify this step if their sample naming convention differs.g.Filter low-count genes across all samples. Rkeep <- rowSums(count_mat) > 50count_mat_filt <- count_mat[keep, , drop = FALSE]dim(count_mat_filt)h.Fit the DESeq2 model across all stages. Rdds <- DESeqDataSetFromMatrix( countData = count_mat_filt, colData = coldata, design = ∼ stage)dds <- DESeq(dds)ddsi.Generate an example MA plot using one contrast, such as stage B versus stage A. Rexample_contrast <- c(“stage”, “B”, “A”)res_ex <- results(dds, contrast = example_contrast)plotMA(res_ex, alpha = 0.05, ylim = c(-5, 5))j.Generate a sample distance heatmap using variance-stabilized counts. Rvsd <- vst(dds, blind = FALSE)d <- dist(t(assay(vsd)))mat <- as.matrix(d)hmcol <- colorRampPalette(brewer.pal(9, “Greens”))(100)heatmap.2( mat, trace = “none”, col = rev(hmcol), density.info = “none”, margins = c(8, 8), main = “Sample distances (VST)”)k.Generate all pairwise stage contrasts. Rstages <- levels(coldata$stage)pairs <- combn(stages, 2, simplify = FALSE)length(pairs)l.Create an output directory for DESeq2 pairwise results. Routdir <- “DESeq2_pairwise_results”dir.create(outdir, showWarnings = FALSE)m.DEG tables for each pairwise comparison using an adjusted p-value cutoff of 0.05. Ralpha <- 0.05for (p in pairs) { s1 <- p[1] s2 <- p[2] res <- results(dds, contrast = c(“stage”, s2, s1)) %>% as.data.frame() res <- res[!is.na(res$padj), , drop = FALSE] sig <- res[res$padj < alpha, , drop = FALSE] out_file <- file.path( outdir, paste0(“deseq2_gene_list_padj_”, alpha, “_”, s1, “_vs_”, s2, “.tsv”) ) write.table( sig, file = out_file, sep = “\t”, quote = FALSE, row.names = TRUE )}list.files(outdir)[1:5]Note: Users may modify the adjusted p-value cutoff, add a log2 fold-change threshold, or define biologically specific contrasts depending on the experimental design. Note: For clarity and to meet STAR Protocols formatting limits, repetitive pairwise DESeq2 operations are programmatically generated using stage labels inferred from sample identifiers. This analysis is equivalent to running DESeq2 separately for each pairwise comparison but is implemented as a single fitted model with extracted contrasts. a.Open R and load required packages. RsuppressPackageStartupMessages({ library(DESeq2) library(dplyr) library(RColorBrewer) library(gplots) }) Open R and load required packages. RsuppressPackageStartupMessages({ library(DESeq2) library(dplyr) library(RColorBrewer) library(gplots) }) R suppressPackageStartupMessages({ library(DESeq2) library(dplyr) library(RColorBrewer) library(gplots) }) b.Locate HTSeq-count output files. Rfiles <- list.files(pattern = “htseq\\.txt$”, full.names = TRUE)stopifnot(length(files) > 0) Locate HTSeq-count output files. Rfiles <- list.files(pattern = “htseq\\.txt$”, full.names = TRUE)stopifnot(length(files) > 0) R files <- list.files(pattern = “htseq\\.txt$”, full.names = TRUE) stopifnot(length(files) > 0) c.Read count files and remove HTSeq summary rows. Rread_htseq <- function(f) { x <- read.table( f, header = FALSE, sep = “\t”, colClasses = c(“character”, “numeric”), stringsAsFactors = FALSE ) # Keep gene rows only and remove HTSeq summary lines x <- x[!grepl(“ˆ__”, x[, 1]), , drop = FALSE] return(x)}dat_list <- lapply(files, read_htseq) Read count files and remove HTSeq summary rows. Rread_htseq <- function(f) { x <- read.table( f, header = FALSE, sep = “\t”, colClasses = c(“character”, “numeric”), stringsAsFactors = FALSE ) # Keep gene rows only and remove HTSeq summary lines x <- x[!grepl(“ˆ__”, x[, 1]), , drop = FALSE] return(x)}dat_list <- lapply(files, read_htseq) R read_htseq <- function(f) { x <- read.table( f, header = FALSE, sep = “\t”, colClasses = c(“character”, “numeric”), stringsAsFactors = FALSE ) # Keep gene rows only and remove HTSeq summary lines x <- x[!grepl(“ˆ__”, x[, 1]), , drop = FALSE] return(x) } dat_list <- lapply(files, read_htseq) d.Ensure consistent gene order across files and build a count matrix. Rgenes <- dat_list[[1]][, 1]stopifnot( all(vapply(dat_list, function(x) identical(x[, 1], genes), logical(1))))count_mat <- do.call(cbind, lapply(dat_list, function(x) x[, 2]))rownames(count_mat) <- genes Ensure consistent gene order across files and build a count matrix. Rgenes <- dat_list[[1]][, 1]stopifnot( all(vapply(dat_list, function(x) identical(x[, 1], genes), logical(1))))count_mat <- do.call(cbind, lapply(dat_list, function(x) x[, 2]))rownames(count_mat) <- genes R genes <- dat_list[[1]][, 1] stopifnot( all(vapply(dat_list, function(x) identical(x[, 1], genes), logical(1))) ) count_mat <- do.call(cbind, lapply(dat_list, function(x) x[, 2])) rownames(count_mat) <- genes e.Infer sample names from file names. Rsample_ids <- sub(“\\.htseq\\.txt$”, ““, basename(files))colnames(count_mat) <- sample_idscount_mat[1:5, 1:5] Infer sample names from file names. Rsample_ids <- sub(“\\.htseq\\.txt$”, ““, basename(files))colnames(count_mat) <- sample_idscount_mat[1:5, 1:5] R sample_ids <- sub(“\\.htseq\\.txt$”, ““, basename(files)) colnames(count_mat) <- sample_ids count_mat[1:5, 1:5] f.Define developmental stages from sample identifiers and construct sample metadata. R# Expected sample format: A1, A2, A3., K1, K2, K3stage <- sub(“[0-9]+$”, ““, sample_ids)coldata <- data.frame( sample = sample_ids, stage = factor(stage), row.names = sample_ids)coldataNote: The example assumes that sample names are formatted as a stage letter followed by a replicate number, such as A1, A2, A3, K1, K2, K3. Users should modify this step if their sample naming convention differs. Define developmental stages from sample identifiers and construct sample metadata. R# Expected sample format: A1, A2, A3., K1, K2, K3stage <- sub(“[0-9]+$”, ““, sample_ids)coldata <- data.frame( sample = sample_ids, stage = factor(stage), row.names = sample_ids)coldataNote: The example assumes that sample names are formatted as a stage letter followed by a replicate number, such as A1, A2, A3, K1, K2, K3. Users should modify this step if their sample naming convention differs. R # Expected sample format: A1, A2, A3., K1, K2, K3 stage <- sub(“[0-9]+$”, ““, sample_ids) coldata <- data.frame( sample = sample_ids, stage = factor(stage), row.names = sample_ids ) coldata Note: The example assumes that sample names are formatted as a stage letter followed by a replicate number, such as A1, A2, A3, K1, K2, K3. Users should modify this step if their sample naming convention differs. g.Filter low-count genes across all samples. Rkeep <- rowSums(count_mat) > 50count_mat_filt <- count_mat[keep, , drop = FALSE]dim(count_mat_filt) Filter low-count genes across all samples. Rkeep <- rowSums(count_mat) > 50count_mat_filt <- count_mat[keep, , drop = FALSE]dim(count_mat_filt) R keep <- rowSums(count_mat) > 50 count_mat_filt <- count_mat[keep, , drop = FALSE] dim(count_mat_filt) h.Fit the DESeq2 model across all stages. Rdds <- DESeqDataSetFromMatrix( countData = count_mat_filt, colData = coldata, design = ∼ stage)dds <- DESeq(dds)dds Fit the DESeq2 model across all stages. Rdds <- DESeqDataSetFromMatrix( countData = count_mat_filt, colData = coldata, design = ∼ stage)dds <- DESeq(dds)dds R dds <- DESeqDataSetFromMatrix( countData = count_mat_filt, colData = coldata, design = ∼ stage ) dds <- DESeq(dds) dds i.Generate an example MA plot using one contrast, such as stage B versus stage A. Rexample_contrast <- c(“stage”, “B”, “A”)res_ex <- results(dds, contrast = example_contrast)plotMA(res_ex, alpha = 0.05, ylim = c(-5, 5)) Generate an example MA plot using one contrast, such as stage B versus stage A. Rexample_contrast <- c(“stage”, “B”, “A”)res_ex <- results(dds, contrast = example_contrast)plotMA(res_ex, alpha = 0.05, ylim = c(-5, 5)) R example_contrast <- c(“stage”, “B”, “A”) res_ex <- results(dds, contrast = example_contrast) plotMA(res_ex, alpha = 0.05, ylim = c(-5, 5)) j.Generate a sample distance heatmap using variance-stabilized counts. Rvsd <- vst(dds, blind = FALSE)d <- dist(t(assay(vsd)))mat <- as.matrix(d)hmcol <- colorRampPalette(brewer.pal(9, “Greens”))(100)heatmap.2( mat, trace = “none”, col = rev(hmcol), density.info = “none”, margins = c(8, 8), main = “Sample distances (VST)”) Generate a sample distance heatmap using variance-stabilized counts. Rvsd <- vst(dds, blind = FALSE)d <- dist(t(assay(vsd)))mat <- as.matrix(d)hmcol <- colorRampPalette(brewer.pal(9, “Greens”))(100)heatmap.2( mat, trace = “none”, col = rev(hmcol), density.info = “none”, margins = c(8, 8), main = “Sample distances (VST)”) R vsd <- vst(dds, blind = FALSE) d <- dist(t(assay(vsd))) mat <- as.matrix(d) hmcol <- colorRampPalette(brewer.pal(9, “Greens”))(100) heatmap.2( mat, trace = “none”, col = rev(hmcol), density.info = “none”, margins = c(8, 8), main = “Sample distances (VST)” ) k.Generate all pairwise stage contrasts. Rstages <- levels(coldata$stage)pairs <- combn(stages, 2, simplify = FALSE)length(pairs) Generate all pairwise stage contrasts. Rstages <- levels(coldata$stage)pairs <- combn(stages, 2, simplify = FALSE)length(pairs) R stages <- levels(coldata$stage) pairs <- combn(stages, 2, simplify = FALSE) length(pairs) l.Create an output directory for DESeq2 pairwise results. Routdir <- “DESeq2_pairwise_results”dir.create(outdir, showWarnings = FALSE) Create an output directory for DESeq2 pairwise results. Routdir <- “DESeq2_pairwise_results”dir.create(outdir, showWarnings = FALSE) R outdir <- “DESeq2_pairwise_results” dir.create(outdir, showWarnings = FALSE) m.DEG tables for each pairwise comparison using an adjusted p-value cutoff of 0.05. Ralpha <- 0.05for (p in pairs) { s1 <- p[1] s2 <- p[2] res <- results(dds, contrast = c(“stage”, s2, s1)) %>% as.data.frame() res <- res[!is.na(res$padj), , drop = FALSE] sig <- res[res$padj < alpha, , drop = FALSE] out_file <- file.path( outdir, paste0(“deseq2_gene_list_padj_”, alpha, “_”, s1, “_vs_”, s2, “.tsv”) ) write.table( sig, file = out_file, sep = “\t”, quote = FALSE, row.names = TRUE )}list.files(outdir)[1:5]Note: Users may modify the adjusted p-value cutoff, add a log2 fold-change threshold, or define biologically specific contrasts depending on the experimental design. DEG tables for each pairwise comparison using an adjusted p-value cutoff of 0.05. Ralpha <- 0.05for (p in pairs) { s1 <- p[1] s2 <- p[2] res <- results(dds, contrast = c(“stage”, s2, s1)) %>% as.data.frame() res <- res[!is.na(res$padj), , drop = FALSE] sig <- res[res$padj < alpha, , drop = FALSE] out_file <- file.path( outdir, paste0(“deseq2_gene_list_padj_”, alpha, “_”, s1, “_vs_”, s2, “.tsv”) ) write.table( sig, file = out_file, sep = “\t”, quote = FALSE, row.names = TRUE )}list.files(outdir)[1:5]Note: Users may modify the adjusted p-value cutoff, add a log2 fold-change threshold, or define biologically specific contrasts depending on the experimental design. R alpha <- 0.05 for (p in pairs) { s1 <- p[1] s2 <- p[2] res <- results(dds, contrast = c(“stage”, s2, s1)) %>% as.data.frame() res <- res[!is.na(res$padj), , drop = FALSE] sig <- res[res$padj < alpha, , drop = FALSE] out_file <- file.path( outdir, paste0(“deseq2_gene_list_padj_”, alpha, “_”, s1, “_vs_”, s2, “.tsv”) ) write.table( sig, file = out_file, sep = “\t”, quote = FALSE, row.names = TRUE ) } list.files(outdir)[1:5] Note: Users may modify the adjusted p-value cutoff, add a log2 fold-change threshold, or define biologically specific contrasts depending on the experimental design. Coexpression network construction using WGCNA Timing: 2–6 h, depending on the number of genes and samples Timing: 2–6 h, depending on the number of genes and samples This section describes construction of a signed gene coexpression network using WGCNA13 from a log2-transformed expression matrix, such as log2(FPKM+1), variance-stabilized counts, or another appropriately normalized expression matrix. The major outputs are module assignments for genes, module eigengenes summarizing expression patterns of each module, and Cytoscape-ready edge and node tables for selected modules. Representative input and output structures are shown in Figure 2.Figure 2WGCNA module detection and eigengene structure(A) Expected format of the input expression matrix for WGCNA. Rows represent genes and columns represent samples. Values correspond to normalized and transformed expression levels, shown here as log2(FPKM + 1). In this example, samples A–K represent developmental stages used for network construction.(B) Gene dendrogram generated by average-linkage hierarchical clustering based on topological overlap. Initial modules were identified using the Dynamic Tree Cut algorithm and then merged based on module eigengene similarity. Color bars indicate module assignments before and after module merging.(C) Relationships among module eigengenes. The upper panel shows hierarchical clustering of module eigengenes based on eigengene dissimilarity, defined as 1 − Pearson correlation.Modules clustering at lower heights have more similar eigengene expression patterns. The lower panel shows the eigengene adjacency heatmap, where color intensity represents eigengene adjacency among modules. This analysis evaluates higher-order relationships among modules and supports module merging decisions. WGCNA module detection and eigengene structure (A) Expected format of the input expression matrix for WGCNA. Rows represent genes and columns represent samples. Values correspond to normalized and transformed expression levels, shown here as log2(FPKM + 1). In this example, samples A–K represent developmental stages used for network construction. (B) Gene dendrogram generated by average-linkage hierarchical clustering based on topological overlap. Initial modules were identified using the Dynamic Tree Cut algorithm and then merged based on module eigengene similarity. Color bars indicate module assignments before and after module merging. (C) Relationships among module eigengenes. The upper panel shows hierarchical clustering of module eigengenes based on eigengene dissimilarity, defined as 1 − Pearson correlation. Modules clustering at lower heights have more similar eigengene expression patterns. The lower panel shows the eigengene adjacency heatmap, where color intensity represents eigengene adjacency among modules. This analysis evaluates higher-order relationships among modules and supports module merging decisions. In NEEDLE, WGCNA and GENIE3 are used sequentially but they address different analytical questions.3 WGCNA identifies groups of genes with correlated expression patterns across developmental stages, time points, tissues, genotypes, or treatment conditions. These coexpression modules help define biologically coherent gene sets associated with a gene, pathway, or phenotype of interest.3 However, WGCNA does not infer regulatory direction or distinguish transcriptional regulators from downstream target genes. GENIE3 is then applied within selected WGCNA modules to infer putative TF–target relationships using candidate TFs as regulators. Thus, WGCNA narrows the search space to relevant coexpression modules, whereas GENIE3 prioritizes candidate upstream regulators and ranks predicted regulatory edges within those modules.14.Prepare the WGCNA input expression matrix.a.Use an expression matrix with genes as rows and samples as columns.b.Place gene identifiers in the first column.c.Use normalized and transformed expression values, such as log2(FPKM + 1), variance-stabilized counts, or rlog-transformed counts.Note: The example input file used in this protocol is input_log2.FPKM.1.average.txt.CRITICAL: The input expression matrix should contain sufficient biological variation for correlation-based network modeling. Based on our testing, six or more biologically meaningful samples are generally recommended, and larger sample numbers are preferable when available. This should be considered a practical guideline rather than a strict requirement, as the optimal sample number depends on dataset complexity, variation among samples, and the biological question.15.Open R and load required packages. Roptions(stringsAsFactors = FALSE)suppressPackageStartupMessages({ library(WGCNA) library(flashClust) library(RColorBrewer) library(gplots) library(pheatmap)})# Recommended by WGCNA for speed, if supportedenableWGCNAThreads()Note:enableWGCNAThreads() may be ignored depending on the operating system or R build. The analysis will still run.16.Load the expression matrix and convert it to WGCNA format.a.Load the input expression matrix into R.b.Set gene identifiers as row names.c.Remove the gene identifier column.d.Convert expression values to a numeric matrix.e.Transpose the matrix so that samples are in rows and genes are in column.f.Inspect the dimensions of the WGCNA input matrix. R# Read expression matrix: genes x samplesdf <- read.table( “input_log2.FPKM.1.average.txt”, header = TRUE, sep = “\t”, check.names = FALSE)# The first column contains gene identifiersrownames(df) <- df[, 1]# Remove gene ID columnexpr_mat <- df[, -1]# Convert expression values to numeric matrixexpr_mat <- data.matrix(expr_mat)# WGCNA expects samples in rows and genes in columnsdatExpr <- as.data.frame(t(expr_mat))# Inspect input dimensionsdim(datExpr)datExpr[1:3, 1:3]Note: WGCNA expects samples in rows and genes in columns. Therefore, the input matrix must be transposed after loading (Figure 2A).CRITICAL: Confirm that values are numeric and properly normalized or transformed. Do not use raw count values directly for WGCNA.17.Check for missing values and remove problematic samples or genes.a.Identify genes or samples with excessive missing values.b.Remove problematic genes or samples before network construction.c.Confirm the final dimensions of the filtered expression matrix. Rgsg <- goodSamplesGenes(datExpr, verbose = 3)if (!gsg$allOK) { if (sum(!gsg$goodGenes) > 0) { message( “Removing genes: “, paste(colnames(datExpr)[!gsg$goodGenes], collapse = “, “) ) } if (sum(!gsg$goodSamples) > 0) { message( “Removing samples: ”, paste(rownames(datExpr)[!gsg$goodSamples], collapse = “, “) ) } datExpr <- datExpr[gsg$goodSamples, gsg$goodGenes]}dim(datExpr)18.Estimate an appropriate soft-thresholding power for a signed network.a.Test candidate soft-thresholding powers.b.Evaluate scale-free topology fit and mean connectivity.c.Select a power that supports approximate scale-free topology while retaining reasonable network connectivity.d.Record the selected soft-thresholding power for reproducibility. Rpowers <- c(1:10, seq(12, 40, by = 2))sft <- pickSoftThreshold( datExpr, powerVector = powers, networkType = “signed”, verbose = 5)par(mfrow = c(1, 2))cex1 <- 0.9# Scale-free topology fitplot( sft$fitIndices[, 1], -sign(sft$fitIndices[, 3]) ∗ sft$fitIndices[, 2], xlab = “Soft Threshold (power)”, ylab = “Scale Free Topology Model Fit (signed Rˆ2)”, type = “n”, main = “Scale independence”)text( sft$fitIndices[, 1], -sign(sft$fitIndices[, 3]) ∗ sft$fitIndices[, 2], labels = powers, cex = cex1)abline(h = 0.90, lty = 2)# Mean connectivityplot( sft$fitIndices[, 1], sft$fitIndices[, 5], xlab = “Soft Threshold (power)”, ylab = “Mean connectivity”, type = “n”, main = “Mean connectivity”)text( sft$fitIndices[, 1], sft$fitIndices[, 5], labels = powers, cex = cex1)CRITICAL: Choose a soft-thresholding power that achieves a high scale-free topology fit, commonly R2 ≥ 0.8–0.9, while retaining reasonable connectivity.19.Compute adjacency and topological overlap matrices.a.Set the soft-thresholding power selected in Step 18.b.Construct the adjacency matrix.c.Convert the adjacency matrix to a topological overlap matrix (TOM).d.Calculate TOM-based dissimilarity for hierarchical clustering. R# Set based on Step 17softPower <- 16adjacency <- adjacency( datExpr, type = “signed”, power = softPower)TOM <- TOMsimilarity(adjacency)dissTOM <- 1 - TOMNote: TOM computation may be memory intensive for large gene sets. For large datasets, users may prefilter genes by expression or variance before WGCNA.Optional: For large expression matrices, for example >15,000 genes, use blockwiseModules() to reduce memory usage.20.Cluster genes and identify initial coexpression modules.a.Cluster genes using TOM-based dissimilarity.b.Identify initial modules from the hierarchical clustering tree.c.Convert numeric module labels to color labels.d.Visualize the gene dendrogram and initial module assignments. RgeneTree <- flashClust( as.dist(dissTOM), method = “average”)plot( geneTree, xlab = ““, sub = ““, main = “Gene clustering on TOM-based dissimilarity”, labels = FALSE, hang = 0.04)minModuleSize <- 30dynamicMods <- cutreeDynamic( dendro = geneTree, distM = dissTOM, deepSplit = 2, pamRespectsDendro = FALSE, minClusterSize = minModuleSize)dynamicColors <- labels2colors(dynamicMods)table(dynamicColors)plotDendroAndColors( geneTree, dynamicColors, “Dynamic Tree Cut”, dendroLabels = FALSE, hang = 0.03, addGuide = TRUE, guideHang = 0.05)Note: This step identifies initial gene coexpression modules from the hierarchical clustering tree (Figure 2B).CRITICAL:minModuleSize and deepSplit affect module granularity. If modules are overly fragmented, increase minModuleSize or reduce deepSplit. If too few modules are detected, decrease minModuleSize or increase deepSplit.21.Merge modules with highly correlated eigengenes.a.Calculate module eigengenes.b.Cluster module eigengenes based on eigengene dissimilarity.c.Merge modules with highly correlated eigengenes.d.Visualize module assignments before and after merging.e.Save module eigengenes. RMEList <- moduleEigengenes(datExpr, colors = dynamicColors)MEs <- MEList$eigengenesMEDiss <- 1 - cor(MEs)METree <- hclust(as.dist(MEDiss), method = “average”)plot( METree, main = “Clustering of module eigengenes”, xlab = “”, sub = “”)MEDissThres <- 0.25abline(h = MEDissThres, col = “red”, lty = 2)merge <- mergeCloseModules( datExpr, dynamicColors, cutHeight = MEDissThres, verbose = 3)moduleColors <- merge$colorsmergedMEs <- merge$newMEsplotDendroAndColors( geneTree, cbind(dynamicColors, moduleColors), c(“Dynamic Tree Cut”, “Merged modules”), dendroLabels = FALSE, hang = 0.03, addGuide = TRUE, guideHang = 0.05)# Save module eigengeneswrite.table( mergedMEs, file = “WGCNA_module_eigengenes.tsv”, sep = “\t”, quote = FALSE, col.names = NA)Note: Modules with highly correlated eigengenes are merged to reduce redundancy and improve biological interpretability (Figure 2B). A merge threshold of 0.25 is commonly used, corresponding to module eigengene correlation ≥0.75. Users may adjust this value if modules remain redundant or become overly merged.22.Visualize eigengene dendrogram and adjacency heatmap.a.Plot the eigengene dendrogram.b.Plot the eigengene adjacency heatmap.c.Use these visualizations to evaluate higher-order relationships among modules. Rpdf(“WGCNA_plotEigengeneNetworks.pdf”, width = 7, height = 6)plotEigengeneNetworks( mergedMEs, “”, marDendro = c(0, 4, 1, 2), marHeatmap = c(3, 4, 1, 2))dev.off()Note: This visualization summarizes relationships among module eigengenes and helps evaluate higher-order relationships among modules (Figure 2C).23.Export gene lists for each module.a.Assign genes to final merged modules.b.Export a complete gene-to-module assignment table.c.Export one gene list for each module. Rprobes <- colnames(datExpr)# Save complete module assignment tablemodule_assignment <- data.frame( gene_id = probes, module = moduleColors)write.table( module_assignment, file = “WGCNA_gene_module_assignments.tsv”, sep = “\t”, quote = FALSE, row.names = FALSE)# Export one gene list per modulefor (col in sort(unique(moduleColors))) { genes_in_module <- probes[moduleColors == col] out <- data.frame( gene_id = genes_in_module ) write.table( out, file = paste0(“WGCNA_module_”, col, “_genes.tsv”), sep = “\t”, quote = FALSE, row.names = FALSE )}Note: Module-specific gene lists can be used for enrichment analysis, candidate gene prioritization, and downstream GENIE3-based GRN construction.24.Export TOM-based network edges and nodes for selected modules.a.Select one or more modules for network export.b.Extract the TOM values corresponding to genes in the selected modules.c.Apply a TOM threshold to define edges.d.Export edge and node files for Cytoscape visualization. R# Modify this vector to export modules of interestmodules_to_export <- c(“yellowgreen”)inModule <- moduleColors %in% modules_to_exportmodProbes <- probes[inModule]modTOM <- TOM[inModule, inModule]dimnames(modTOM) <- list(modProbes, modProbes)exportNetworkToCytoscape( modTOM, edgeFile = paste0( “Cytoscape_edges_”, paste(modules_to_export, collapse = “_”), “.txt” ), nodeFile = paste0( “Cytoscape_nodes_”, paste(modules_to_export, collapse = “_”), “.txt” ), weighted = TRUE, threshold = 0.2, nodeNames = modProbes, nodeAttr = moduleColors[inModule])Note: The exported edge and node files can be imported into Cytoscape for visualization and network refinement.CRITICAL: The TOM threshold strongly affects network density. If Cytoscape files are too large, increase the threshold, for example to 0.3–0.4, or export fewer modules. If the resulting network is too sparse, reduce the threshold. 14.Prepare the WGCNA input expression matrix.a.Use an expression matrix with genes as rows and samples as columns.b.Place gene identifiers in the first column.c.Use normalized and transformed expression values, such as log2(FPKM + 1), variance-stabilized counts, or rlog-transformed counts. Prepare the WGCNA input expression matrix.a.Use an expression matrix with genes as rows and samples as columns.b.Place gene identifiers in the first column.c.Use normalized and transformed expression values, such as log2(FPKM + 1), variance-stabilized counts, or rlog-transformed counts. a.Use an expression matrix with genes as rows and samples as columns. Use an expression matrix with genes as rows and samples as columns. b.Place gene identifiers in the first column. Place gene identifiers in the first column. c.Use normalized and transformed expression values, such as log2(FPKM + 1), variance-stabilized counts, or rlog-transformed counts. Use normalized and transformed expression values, such as log2(FPKM + 1), variance-stabilized counts, or rlog-transformed counts. Note: The example input file used in this protocol is input_log2.FPKM.1.average.txt. CRITICAL: The input expression matrix should contain sufficient biological variation for correlation-based network modeling. Based on our testing, six or more biologically meaningful samples are generally recommended, and larger sample numbers are preferable when available. This should be considered a practical guideline rather than a strict requirement, as the optimal sample number depends on dataset complexity, variation among samples, and the biological question. 15.Open R and load required packages. Open R and load required packages. R options(stringsAsFactors = FALSE) suppressPackageStartupMessages({ library(WGCNA) library(flashClust) library(RColorBrewer) library(gplots) library(pheatmap) }) # Recommended by WGCNA for speed, if supported enableWGCNAThreads() Note:enableWGCNAThreads() may be ignored depending on the operating system or R build. The analysis will still run. 16.Load the expression matrix and convert it to WGCNA format.a.Load the input expression matrix into R.b.Set gene identifiers as row names.c.Remove the gene identifier column.d.Convert expression values to a numeric matrix.e.Transpose the matrix so that samples are in rows and genes are in column.f.Inspect the dimensions of the WGCNA input matrix. Load the expression matrix and convert it to WGCNA format.a.Load the input expression matrix into R.b.Set gene identifiers as row names.c.Remove the gene identifier column.d.Convert expression values to a numeric matrix.e.Transpose the matrix so that samples are in rows and genes are in column.f.Inspect the dimensions of the WGCNA input matrix. a.Load the input expression matrix into R. Load the input expression matrix into R. b.Set gene identifiers as row names. Set gene identifiers as row names. c.Remove the gene identifier column. Remove the gene identifier column. d.Convert expression values to a numeric matrix. Convert expression values to a numeric matrix. e.Transpose the matrix so that samples are in rows and genes are in column. Transpose the matrix so that samples are in rows and genes are in column. f.Inspect the dimensions of the WGCNA input matrix. Inspect the dimensions of the WGCNA input matrix. R # Read expression matrix: genes x samples df <- read.table( “input_log2.FPKM.1.average.txt”, header = TRUE, sep = “\t”, check.names = FALSE ) # The first column contains gene identifiers rownames(df) <- df[, 1] # Remove gene ID column expr_mat <- df[, -1] # Convert expression values to numeric matrix expr_mat <- data.matrix(expr_mat) # WGCNA expects samples in rows and genes in columns datExpr <- as.data.frame(t(expr_mat)) # Inspect input dimensions dim(datExpr) datExpr[1:3, 1:3] Note: WGCNA expects samples in rows and genes in columns. Therefore, the input matrix must be transposed after loading (Figure 2A). CRITICAL: Confirm that values are numeric and properly normalized or transformed. Do not use raw count values directly for WGCNA. 17.Check for missing values and remove problematic samples or genes.a.Identify genes or samples with excessive missing values.b.Remove problematic genes or samples before network construction.c.Confirm the final dimensions of the filtered expression matrix. Check for missing values and remove problematic samples or genes.a.Identify genes or samples with excessive missing values.b.Remove problematic genes or samples before network construction.c.Confirm the final dimensions of the filtered expression matrix. a.Identify genes or samples with excessive missing values. Identify genes or samples with excessive missing values. b.Remove problematic genes or samples before network construction. Remove problematic genes or samples before network construction. c.Confirm the final dimensions of the filtered expression matrix. Confirm the final dimensions of the filtered expression matrix. R gsg <- goodSamplesGenes(datExpr, verbose = 3) if (!gsg$allOK) { if (sum(!gsg$goodGenes) > 0) { message( “Removing genes: “, paste(colnames(datExpr)[!gsg$goodGenes], collapse = “, “) ) } if (sum(!gsg$goodSamples) > 0) { message( “Removing samples: ”, paste(rownames(datExpr)[!gsg$goodSamples], collapse = “, “) ) } datExpr <- datExpr[gsg$goodSamples, gsg$goodGenes] } dim(datExpr) 18.Estimate an appropriate soft-thresholding power for a signed network.a.Test candidate soft-thresholding powers.b.Evaluate scale-free topology fit and mean connectivity.c.Select a power that supports approximate scale-free topology while retaining reasonable network connectivity.d.Record the selected soft-thresholding power for reproducibility. Estimate an appropriate soft-thresholding power for a signed network.a.Test candidate soft-thresholding powers.b.Evaluate scale-free topology fit and mean connectivity.c.Select a power that supports approximate scale-free topology while retaining reasonable network connectivity.d.Record the selected soft-thresholding power for reproducibility. a.Test candidate soft-thresholding powers. Test candidate soft-thresholding powers. b.Evaluate scale-free topology fit and mean connectivity. Evaluate scale-free topology fit and mean connectivity. c.Select a power that supports approximate scale-free topology while retaining reasonable network connectivity. Select a power that supports approximate scale-free topology while retaining reasonable network connectivity. d.Record the selected soft-thresholding power for reproducibility. Record the selected soft-thresholding power for reproducibility. R powers <- c(1:10, seq(12, 40, by = 2)) sft <- pickSoftThreshold( datExpr, powerVector = powers, networkType = “signed”, verbose = 5 ) par(mfrow = c(1, 2)) cex1 <- 0.9 # Scale-free topology fit plot( sft$fitIndices[, 1], -sign(sft$fitIndices[, 3]) ∗ sft$fitIndices[, 2], xlab = “Soft Threshold (power)”, ylab = “Scale Free Topology Model Fit (signed Rˆ2)”, type = “n”, main = “Scale independence” ) text( sft$fitIndices[, 1], -sign(sft$fitIndices[, 3]) ∗ sft$fitIndices[, 2], labels = powers, cex = cex1 ) abline(h = 0.90, lty = 2) # Mean connectivity plot( sft$fitIndices[, 1], sft$fitIndices[, 5], xlab = “Soft Threshold (power)”, ylab = “Mean connectivity”, type = “n”, main = “Mean connectivity” ) text( sft$fitIndices[, 1], sft$fitIndices[, 5], labels = powers, cex = cex1 ) CRITICAL: Choose a soft-thresholding power that achieves a high scale-free topology fit, commonly R2 ≥ 0.8–0.9, while retaining reasonable connectivity. 19.Compute adjacency and topological overlap matrices.a.Set the soft-thresholding power selected in Step 18.b.Construct the adjacency matrix.c.Convert the adjacency matrix to a topological overlap matrix (TOM).d.Calculate TOM-based dissimilarity for hierarchical clustering. Compute adjacency and topological overlap matrices.a.Set the soft-thresholding power selected in Step 18.b.Construct the adjacency matrix.c.Convert the adjacency matrix to a topological overlap matrix (TOM).d.Calculate TOM-based dissimilarity for hierarchical clustering. a.Set the soft-thresholding power selected in Step 18. Set the soft-thresholding power selected in Step 18. b.Construct the adjacency matrix. Construct the adjacency matrix. c.Convert the adjacency matrix to a topological overlap matrix (TOM). Convert the adjacency matrix to a topological overlap matrix (TOM). d.Calculate TOM-based dissimilarity for hierarchical clustering. Calculate TOM-based dissimilarity for hierarchical clustering. R # Set based on Step 17 softPower <- 16 adjacency <- adjacency( datExpr, type = “signed”, power = softPower ) TOM <- TOMsimilarity(adjacency) dissTOM <- 1 - TOM Note: TOM computation may be memory intensive for large gene sets. For large datasets, users may prefilter genes by expression or variance before WGCNA. Optional: For large expression matrices, for example >15,000 genes, use blockwiseModules() to reduce memory usage. 20.Cluster genes and identify initial coexpression modules.a.Cluster genes using TOM-based dissimilarity.b.Identify initial modules from the hierarchical clustering tree.c.Convert numeric module labels to color labels.d.Visualize the gene dendrogram and initial module assignments. Cluster genes and identify initial coexpression modules.a.Cluster genes using TOM-based dissimilarity.b.Identify initial modules from the hierarchical clustering tree.c.Convert numeric module labels to color labels.d.Visualize the gene dendrogram and initial module assignments. a.Cluster genes using TOM-based dissimilarity. Cluster genes using TOM-based dissimilarity. b.Identify initial modules from the hierarchical clustering tree. Identify initial modules from the hierarchical clustering tree. c.Convert numeric module labels to color labels. Convert numeric module labels to color labels. d.Visualize the gene dendrogram and initial module assignments. Visualize the gene dendrogram and initial module assignments. R geneTree <- flashClust( as.dist(dissTOM), method = “average” ) plot( geneTree, xlab = ““, sub = ““, main = “Gene clustering on TOM-based dissimilarity”, labels = FALSE, hang = 0.04 ) minModuleSize <- 30 dynamicMods <- cutreeDynamic( dendro = geneTree, distM = dissTOM, deepSplit = 2, pamRespectsDendro = FALSE, minClusterSize = minModuleSize ) dynamicColors <- labels2colors(dynamicMods) table(dynamicColors) plotDendroAndColors( geneTree, dynamicColors, “Dynamic Tree Cut”, dendroLabels = FALSE, hang = 0.03, addGuide = TRUE, guideHang = 0.05 ) Note: This step identifies initial gene coexpression modules from the hierarchical clustering tree (Figure 2B). CRITICAL:minModuleSize and deepSplit affect module granularity. If modules are overly fragmented, increase minModuleSize or reduce deepSplit. If too few modules are detected, decrease minModuleSize or increase deepSplit. 21.Merge modules with highly correlated eigengenes.a.Calculate module eigengenes.b.Cluster module eigengenes based on eigengene dissimilarity.c.Merge modules with highly correlated eigengenes.d.Visualize module assignments before and after merging.e.Save module eigengenes. Merge modules with highly correlated eigengenes.a.Calculate module eigengenes.b.Cluster module eigengenes based on eigengene dissimilarity.c.Merge modules with highly correlated eigengenes.d.Visualize module assignments before and after merging.e.Save module eigengenes. a.Calculate module eigengenes. Calculate module eigengenes. b.Cluster module eigengenes based on eigengene dissimilarity. Cluster module eigengenes based on eigengene dissimilarity. c.Merge modules with highly correlated eigengenes. Merge modules with highly correlated eigengenes. d.Visualize module assignments before and after merging. Visualize module assignments before and after merging. e.Save module eigengenes. Save module eigengenes. R MEList <- moduleEigengenes(datExpr, colors = dynamicColors) MEs <- MEList$eigengenes MEDiss <- 1 - cor(MEs) METree <- hclust(as.dist(MEDiss), method = “average”) plot( METree, main = “Clustering of module eigengenes”, xlab = “”, sub = “” ) MEDissThres <- 0.25 abline(h = MEDissThres, col = “red”, lty = 2) merge <- mergeCloseModules( datExpr, dynamicColors, cutHeight = MEDissThres, verbose = 3 ) moduleColors <- merge$colors mergedMEs <- merge$newMEs plotDendroAndColors( geneTree, cbind(dynamicColors, moduleColors), c(“Dynamic Tree Cut”, “Merged modules”), dendroLabels = FALSE, hang = 0.03, addGuide = TRUE, guideHang = 0.05 ) # Save module eigengenes write.table( mergedMEs, file = “WGCNA_module_eigengenes.tsv”, sep = “\t”, quote = FALSE, col.names = NA ) Note: Modules with highly correlated eigengenes are merged to reduce redundancy and improve biological interpretability (Figure 2B). A merge threshold of 0.25 is commonly used, corresponding to module eigengene correlation ≥0.75. Users may adjust this value if modules remain redundant or become overly merged. 22.Visualize eigengene dendrogram and adjacency heatmap.a.Plot the eigengene dendrogram.b.Plot the eigengene adjacency heatmap.c.Use these visualizations to evaluate higher-order relationships among modules. Visualize eigengene dendrogram and adjacency heatmap.a.Plot the eigengene dendrogram.b.Plot the eigengene adjacency heatmap.c.Use these visualizations to evaluate higher-order relationships among modules. a.Plot the eigengene dendrogram. Plot the eigengene dendrogram. b.Plot the eigengene adjacency heatmap. Plot the eigengene adjacency heatmap. c.Use these visualizations to evaluate higher-order relationships among modules. Use these visualizations to evaluate higher-order relationships among modules. R pdf(“WGCNA_plotEigengeneNetworks.pdf”, width = 7, height = 6) plotEigengeneNetworks( mergedMEs, “”, marDendro = c(0, 4, 1, 2), marHeatmap = c(3, 4, 1, 2) ) dev.off() Note: This visualization summarizes relationships among module eigengenes and helps evaluate higher-order relationships among modules (Figure 2C). 23.Export gene lists for each module.a.Assign genes to final merged modules.b.Export a complete gene-to-module assignment table.c.Export one gene list for each module. Export gene lists for each module.a.Assign genes to final merged modules.b.Export a complete gene-to-module assignment table.c.Export one gene list for each module. a.Assign genes to final merged modules. Assign genes to final merged modules. b.Export a complete gene-to-module assignment table. Export a complete gene-to-module assignment table. c.Export one gene list for each module. Export one gene list for each module. R probes <- colnames(datExpr) # Save complete module assignment table module_assignment <- data.frame( gene_id = probes, module = moduleColors ) write.table( module_assignment, file = “WGCNA_gene_module_assignments.tsv”, sep = “\t”, quote = FALSE, row.names = FALSE ) # Export one gene list per module for (col in sort(unique(moduleColors))) { genes_in_module <- probes[moduleColors == col] out <- data.frame( gene_id = genes_in_module ) write.table( out, file = paste0(“WGCNA_module_”, col, “_genes.tsv”), sep = “\t”, quote = FALSE, row.names = FALSE ) } Note: Module-specific gene lists can be used for enrichment analysis, candidate gene prioritization, and downstream GENIE3-based GRN construction. 24.Export TOM-based network edges and nodes for selected modules.a.Select one or more modules for network export.b.Extract the TOM values corresponding to genes in the selected modules.c.Apply a TOM threshold to define edges.d.Export edge and node files for Cytoscape visualization. Export TOM-based network edges and nodes for selected modules.a.Select one or more modules for network export.b.Extract the TOM values corresponding to genes in the selected modules.c.Apply a TOM threshold to define edges.d.Export edge and node files for Cytoscape visualization. a.Select one or more modules for network export. Select one or more modules for network export. b.Extract the TOM values corresponding to genes in the selected modules. Extract the TOM values corresponding to genes in the selected modules. c.Apply a TOM threshold to define edges. Apply a TOM threshold to define edges. d.Export edge and node files for Cytoscape visualization. Export edge and node files for Cytoscape visualization. R # Modify this vector to export modules of interest modules_to_export <- c(“yellowgreen”) inModule <- moduleColors %in% modules_to_export modProbes <- probes[inModule] modTOM <- TOM[inModule, inModule] dimnames(modTOM) <- list(modProbes, modProbes) exportNetworkToCytoscape( modTOM, edgeFile = paste0( “Cytoscape_edges_”, paste(modules_to_export, collapse = “_”), “.txt” ), nodeFile = paste0( “Cytoscape_nodes_”, paste(modules_to_export, collapse = “_”), “.txt” ), weighted = TRUE, threshold = 0.2, nodeNames = modProbes, nodeAttr = moduleColors[inModule] ) Note: The exported edge and node files can be imported into Cytoscape for visualization and network refinement. CRITICAL: The TOM threshold strongly affects network density. If Cytoscape files are too large, increase the threshold, for example to 0.3–0.4, or export fewer modules. If the resulting network is too sparse, reduce the threshold. GRN construction using GENIE3 Timing: 1–2 h, depending on the number of genes in the selected module and available CPU resources Timing: 1–2 h, depending on the number of genes in the selected module and available CPU resources This section describes construction of a GRN for a WGCNA-derived coexpression module using GENIE3, a tree-based regulatory inference method. GENIE3 predicts putative regulatory relationships by using candidate TFs as potential regulators and expression profiles of module genes as potential targets. The output is a ranked list of predicted TF–target edges that can be used to prioritize candidate regulators for downstream analysis and validation. Unlike WGCNA, which groups genes into modules based on correlated expression patterns, GENIE3 infers putative directional relationships from candidate TFs to predicted target genes within a selected module. Thus, WGCNA first defines a biologically relevant coexpression module, and GENIE3 is then used to prioritize candidate upstream regulators within that module.25.Prepare GENIE3 input files.a.Prepare a module-specific expression matrix formatted as genes × samples, with gene identifiers in the first column.b.Use normalized and transformed expression values, such as log2(FPKM + 1), variance-stabilized counts, or an equivalent normalized expression matrix.c.Prepare a single-column file containing candidate TF gene identifiers present in the selected module.Note: We recommend obtaining species-specific TF lists from PlantTFDB15 or another curated TF annotation resource. Users should download the TF annotations corresponding to the same genome version used for RNA-seq processing to ensure gene identifier consistency.CRITICAL: GENIE3 requires that TF identifiers match the gene IDs in the expression matrix exactly. If different genome versions, transcript identifiers, or naming conventions are mixed, GENIE3 may fail to recognize candidate regulators.26.Launch R and load required packages. Rlibrary(GENIE3)Note: GENIE3 uses tree-based inference and may produce slightly different results across runs unless a random seed is specified. Set a seed before running GENIE3 to improve reproducibility.27.Load the module expression table and convert it to a numeric expression matrix.a.Load the module-specific expression table into R.b.Set gene identifiers as row names.c.Remove non-expression columns.d.Convert the expression table to a numeric matrix.e.Inspect the expression matrix. R# Read module-specific expression matrixmod_df <- read.table( “saddlebrown.c_sorghum_log2fpkm1.txt”, header = TRUE, sep = “\t”, check.names = FALSE)# The first column contains gene identifiersgene_ids <- mod_df[, 1]# Convert expression values to a numeric matrixexprMatr <- data.matrix(mod_df[, 2:ncol(mod_df)])# Assign gene identifiers as row namesrownames(exprMatr) <- gene_ids# Optional: assign generic sample namescolnames(exprMatr) <- paste(“Sample”, 1:ncol(exprMatr), sep = ““)# Inspect the expression matrixdim(exprMatr)head(exprMatr[, 1:5])Note: GENIE3 expects an expression matrix with genes in rows and samples in columns. Gene identifiers must be stored as row names.CRITICAL: Confirm that all expression values are numeric. Non-numeric columns, missing values, or duplicated gene identifiers may cause GENIE3 to fail or produce unreliable results.28.Load the candidate TF list and retain TFs present in the module expression matrix.a.Load the candidate TF list.b.Extract unique candidate TF identifiers.c.Intersect the candidate TF list with row names of the expression matrix.d.Confirm that at least one candidate regulator is retained. R# Read candidate TF listtf_df <- read.table( “TFs_included_saddlebrown-c.txt”, header = TRUE, sep = “\t”, check.names = FALSE)# Extract TF identifiersregulators <- unique(tf_df$tracking_id)# Retain only TFs present in the expression matrixregulators <- intersect(regulators, rownames(exprMatr))# Inspect candidate regulatorslength(regulators)head(regulators)CRITICAL: If length (regulators) is close to 0, the TF identifiers likely do not match the gene IDs in the expression matrix. Check for differences in genome version, gene model version, transcript versus gene identifiers, or formatting differences such as prefixes and suffixes.Note: Restricting regulators to TFs present in the selected WGCNA module allows GENIE3 to infer regulatory relationships among module-associated TFs and target genes. Alternatively, users may provide a broader list of expressed TFs if they wish to infer potential regulators outside the module.29.Infer regulatory links using GENIE3.a.Set a random seed to improve reproducibility.b.Run GENIE3 using the module expression matrix and candidate TF list.c.Inspect the weight matrix. R# Set seed for reproducibilityset.seed(123)# Run GENIE3wm <- GENIE3( exprMatr, regulators = regulators, K = “sqrt”)# Inspect the weight matrixdim(wm)wm[1:5, 1:5]Note: GENIE3 returns a weight matrix in which each value represents the inferred importance of a regulatory relationship between a candidate regulator and a target gene. Larger weights indicate stronger predicted regulatory relationships.Note: Runtime increases with the number of genes and regulators included in the analysis. For large modules, consider running GENIE3 on an HPC system or reducing the input to biologically relevant genes.Note: In this protocol, K = “sqrt” was used, which is commonly applied for regulatory network inference with modest sample sizes.30.Convert the GENIE3 weight matrix to a ranked TF-target edge list and export the results.a.Transform the GENIE3 weight matrix into a ranked TF–target edge list.b.Apply a weight threshold or select the top-ranked interactions.c.Export the edge list for downstream visualization.d.Import the ranked edge list into Cytoscape for network visualization and refinement, if desired.Figure 3GENIE3-enabled GRN construction(A) Representative ranked TF–target edge list generated by GENIE3. Each row corresponds to a predicted regulatory interaction. The regulatoryGene column denotes candidate TFs, the targetGene column denotes predicted target genes, and the weight column indicates the GENIE3 importance score reflecting the relative strength of the inferred interaction.(B) Example GRN visualization generated in Cytoscape. Triangles represent TFs and circles represent non-TF target genes. Directed edges indicate putative regulatory relationships from candidate TFs to predicted target genes. Network topology highlights differences in regulatory connectivity among candidate TFs.(C) Distribution of GENIE3 edge weights for the top-ranked TFs within the selected WGCNA module.Ridgeline plots display the density distribution of importance scores for TF–target interactions. The x-axis represents GENIE3 weight scores, and the y-axis lists the top-ranked TFs.Differences in distribution shape illustrate variability in predicted regulatory connectivity patterns among TFs, such as broad moderate-weight interactions versus fewer high-weight interactions. This analysis supports threshold selection and prioritization of candidate hub regulators for downstream validation. R# Convert GENIE3 weight matrix to ranked edge listedge_list <- getLinkList( wm, threshold = 0.15)# Inspect edge listdim(edge_list)head(edge_list)# Export edge list for Cytoscape or downstream analysiswrite.table( edge_list, file = “GENIE3_edges_saddlebrown_c_threshold_0.15.tsv”, quote = FALSE, sep = “\t”, row.names = FALSE)Note: The resulting edge list contains three columns: regulatoryGene, targetGene, and weight. The regulatoryGene column indicates the candidate TF, the targetGene column indicates the predicted target gene, and the weight column indicates the GENIE3 importance score for the inferred TF–target relationship (Figures 3A and 3B).CRITICAL: The optimal weight threshold is dataset- and module-dependent. If the inferred network is overly sparse, decrease the threshold, for example to 0.05–0.10. If the network is excessively dense, increase the threshold, for example to 0.20–0.30, or export only the top-ranked edges.Optional: Instead of applying a fixed weight threshold, users can export a defined number of top-ranked interactions. This approach can be useful when comparing networks across modules or species. R# Export top-ranked links instead of using a fixed thresholdtop_links <- 5000edge_list_top <- getLinkList( wm, reportMax = top_links)write.table( edge_list_top, file = “GENIE3_edges_saddlebrown_c_top5000.tsv”, quote = FALSE, sep = “\t”, row.names = FALSE)Note: After generating the ranked edge list, examine the distribution of GENIE3 weight scores for highly connected TFs. Visualization of weight distributions, such as ridgeline plots, can help distinguish TFs with broad moderate-weight interactions from TFs with fewer high-confidence interactions (Figure 3C). This inspection can guide threshold selection and prioritization of candidate hub regulators for downstream validation. 25.Prepare GENIE3 input files.a.Prepare a module-specific expression matrix formatted as genes × samples, with gene identifiers in the first column.b.Use normalized and transformed expression values, such as log2(FPKM + 1), variance-stabilized counts, or an equivalent normalized expression matrix.c.Prepare a single-column file containing candidate TF gene identifiers present in the selected module. Prepare GENIE3 input files.a.Prepare a module-specific expression matrix formatted as genes × samples, with gene identifiers in the first column.b.Use normalized and transformed expression values, such as log2(FPKM + 1), variance-stabilized counts, or an equivalent normalized expression matrix.c.Prepare a single-column file containing candidate TF gene identifiers present in the selected module. a.Prepare a module-specific expression matrix formatted as genes × samples, with gene identifiers in the first column. Prepare a module-specific expression matrix formatted as genes × samples, with gene identifiers in the first column. b.Use normalized and transformed expression values, such as log2(FPKM + 1), variance-stabilized counts, or an equivalent normalized expression matrix. Use normalized and transformed expression values, such as log2(FPKM + 1), variance-stabilized counts, or an equivalent normalized expression matrix. c.Prepare a single-column file containing candidate TF gene identifiers present in the selected module. Prepare a single-column file containing candidate TF gene identifiers present in the selected module. Note: We recommend obtaining species-specific TF lists from PlantTFDB15 or another curated TF annotation resource. Users should download the TF annotations corresponding to the same genome version used for RNA-seq processing to ensure gene identifier consistency. CRITICAL: GENIE3 requires that TF identifiers match the gene IDs in the expression matrix exactly. If different genome versions, transcript identifiers, or naming conventions are mixed, GENIE3 may fail to recognize candidate regulators. 26.Launch R and load required packages. Launch R and load required packages. R library(GENIE3) Note: GENIE3 uses tree-based inference and may produce slightly different results across runs unless a random seed is specified. Set a seed before running GENIE3 to improve reproducibility. 27.Load the module expression table and convert it to a numeric expression matrix.a.Load the module-specific expression table into R.b.Set gene identifiers as row names.c.Remove non-expression columns.d.Convert the expression table to a numeric matrix.e.Inspect the expression matrix. Load the module expression table and convert it to a numeric expression matrix.a.Load the module-specific expression table into R.b.Set gene identifiers as row names.c.Remove non-expression columns.d.Convert the expression table to a numeric matrix.e.Inspect the expression matrix. a.Load the module-specific expression table into R. Load the module-specific expression table into R. b.Set gene identifiers as row names. Set gene identifiers as row names. c.Remove non-expression columns. Remove non-expression columns. d.Convert the expression table to a numeric matrix. Convert the expression table to a numeric matrix. e.Inspect the expression matrix. Inspect the expression matrix. R # Read module-specific expression matrix mod_df <- read.table( “saddlebrown.c_sorghum_log2fpkm1.txt”, header = TRUE, sep = “\t”, check.names = FALSE ) # The first column contains gene identifiers gene_ids <- mod_df[, 1] # Convert expression values to a numeric matrix exprMatr <- data.matrix(mod_df[, 2:ncol(mod_df)]) # Assign gene identifiers as row names rownames(exprMatr) <- gene_ids # Optional: assign generic sample names colnames(exprMatr) <- paste(“Sample”, 1:ncol(exprMatr), sep = ““) # Inspect the expression matrix dim(exprMatr) head(exprMatr[, 1:5]) Note: GENIE3 expects an expression matrix with genes in rows and samples in columns. Gene identifiers must be stored as row names. CRITICAL: Confirm that all expression values are numeric. Non-numeric columns, missing values, or duplicated gene identifiers may cause GENIE3 to fail or produce unreliable results. 28.Load the candidate TF list and retain TFs present in the module expression matrix.a.Load the candidate TF list.b.Extract unique candidate TF identifiers.c.Intersect the candidate TF list with row names of the expression matrix.d.Confirm that at least one candidate regulator is retained. Load the candidate TF list and retain TFs present in the module expression matrix.a.Load the candidate TF list.b.Extract unique candidate TF identifiers.c.Intersect the candidate TF list with row names of the expression matrix.d.Confirm that at least one candidate regulator is retained. a.Load the candidate TF list. Load the candidate TF list. b.Extract unique candidate TF identifiers. Extract unique candidate TF identifiers. c.Intersect the candidate TF list with row names of the expression matrix. Intersect the candidate TF list with row names of the expression matrix. d.Confirm that at least one candidate regulator is retained. Confirm that at least one candidate regulator is retained. R # Read candidate TF list tf_df <- read.table( “TFs_included_saddlebrown-c.txt”, header = TRUE, sep = “\t”, check.names = FALSE ) # Extract TF identifiers regulators <- unique(tf_df$tracking_id) # Retain only TFs present in the expression matrix regulators <- intersect(regulators, rownames(exprMatr)) # Inspect candidate regulators length(regulators) head(regulators) CRITICAL: If length (regulators) is close to 0, the TF identifiers likely do not match the gene IDs in the expression matrix. Check for differences in genome version, gene model version, transcript versus gene identifiers, or formatting differences such as prefixes and suffixes. Note: Restricting regulators to TFs present in the selected WGCNA module allows GENIE3 to infer regulatory relationships among module-associated TFs and target genes. Alternatively, users may provide a broader list of expressed TFs if they wish to infer potential regulators outside the module. 29.Infer regulatory links using GENIE3.a.Set a random seed to improve reproducibility.b.Run GENIE3 using the module expression matrix and candidate TF list.c.Inspect the weight matrix. Infer regulatory links using GENIE3.a.Set a random seed to improve reproducibility.b.Run GENIE3 using the module expression matrix and candidate TF list.c.Inspect the weight matrix. a.Set a random seed to improve reproducibility. Set a random seed to improve reproducibility. b.Run GENIE3 using the module expression matrix and candidate TF list. Run GENIE3 using the module expression matrix and candidate TF list. c.Inspect the weight matrix. Inspect the weight matrix. R # Set seed for reproducibility set.seed(123) # Run GENIE3 wm <- GENIE3( exprMatr, regulators = regulators, K = “sqrt” ) # Inspect the weight matrix dim(wm) wm[1:5, 1:5] Note: GENIE3 returns a weight matrix in which each value represents the inferred importance of a regulatory relationship between a candidate regulator and a target gene. Larger weights indicate stronger predicted regulatory relationships. Note: Runtime increases with the number of genes and regulators included in the analysis. For large modules, consider running GENIE3 on an HPC system or reducing the input to biologically relevant genes. Note: In this protocol, K = “sqrt” was used, which is commonly applied for regulatory network inference with modest sample sizes. 30.Convert the GENIE3 weight matrix to a ranked TF-target edge list and export the results.a.Transform the GENIE3 weight matrix into a ranked TF–target edge list.b.Apply a weight threshold or select the top-ranked interactions.c.Export the edge list for downstream visualization.d.Import the ranked edge list into Cytoscape for network visualization and refinement, if desired.Figure 3GENIE3-enabled GRN construction(A) Representative ranked TF–target edge list generated by GENIE3. Each row corresponds to a predicted regulatory interaction. The regulatoryGene column denotes candidate TFs, the targetGene column denotes predicted target genes, and the weight column indicates the GENIE3 importance score reflecting the relative strength of the inferred interaction.(B) Example GRN visualization generated in Cytoscape. Triangles represent TFs and circles represent non-TF target genes. Directed edges indicate putative regulatory relationships from candidate TFs to predicted target genes. Network topology highlights differences in regulatory connectivity among candidate TFs.(C) Distribution of GENIE3 edge weights for the top-ranked TFs within the selected WGCNA module.Ridgeline plots display the density distribution of importance scores for TF–target interactions. The x-axis represents GENIE3 weight scores, and the y-axis lists the top-ranked TFs.Differences in distribution shape illustrate variability in predicted regulatory connectivity patterns among TFs, such as broad moderate-weight interactions versus fewer high-weight interactions. This analysis supports threshold selection and prioritization of candidate hub regulators for downstream validation. Convert the GENIE3 weight matrix to a ranked TF-target edge list and export the results.a.Transform the GENIE3 weight matrix into a ranked TF–target edge list.b.Apply a weight threshold or select the top-ranked interactions.c.Export the edge list for downstream visualization.d.Import the ranked edge list into Cytoscape for network visualization and refinement, if desired.Figure 3GENIE3-enabled GRN construction(A) Representative ranked TF–target edge list generated by GENIE3. Each row corresponds to a predicted regulatory interaction. The regulatoryGene column denotes candidate TFs, the targetGene column denotes predicted target genes, and the weight column indicates the GENIE3 importance score reflecting the relative strength of the inferred interaction.(B) Example GRN visualization generated in Cytoscape. Triangles represent TFs and circles represent non-TF target genes. Directed edges indicate putative regulatory relationships from candidate TFs to predicted target genes. Network topology highlights differences in regulatory connectivity among candidate TFs.(C) Distribution of GENIE3 edge weights for the top-ranked TFs within the selected WGCNA module.Ridgeline plots display the density distribution of importance scores for TF–target interactions. The x-axis represents GENIE3 weight scores, and the y-axis lists the top-ranked TFs.Differences in distribution shape illustrate variability in predicted regulatory connectivity patterns among TFs, such as broad moderate-weight interactions versus fewer high-weight interactions. This analysis supports threshold selection and prioritization of candidate hub regulators for downstream validation. a.Transform the GENIE3 weight matrix into a ranked TF–target edge list. Transform the GENIE3 weight matrix into a ranked TF–target edge list. b.Apply a weight threshold or select the top-ranked interactions. Apply a weight threshold or select the top-ranked interactions. c.Export the edge list for downstream visualization. Export the edge list for downstream visualization. d.Import the ranked edge list into Cytoscape for network visualization and refinement, if desired.Figure 3GENIE3-enabled GRN construction(A) Representative ranked TF–target edge list generated by GENIE3. Each row corresponds to a predicted regulatory interaction. The regulatoryGene column denotes candidate TFs, the targetGene column denotes predicted target genes, and the weight column indicates the GENIE3 importance score reflecting the relative strength of the inferred interaction.(B) Example GRN visualization generated in Cytoscape. Triangles represent TFs and circles represent non-TF target genes. Directed edges indicate putative regulatory relationships from candidate TFs to predicted target genes. Network topology highlights differences in regulatory connectivity among candidate TFs.(C) Distribution of GENIE3 edge weights for the top-ranked TFs within the selected WGCNA module.Ridgeline plots display the density distribution of importance scores for TF–target interactions. The x-axis represents GENIE3 weight scores, and the y-axis lists the top-ranked TFs.Differences in distribution shape illustrate variability in predicted regulatory connectivity patterns among TFs, such as broad moderate-weight interactions versus fewer high-weight interactions. This analysis supports threshold selection and prioritization of candidate hub regulators for downstream validation. Import the ranked edge list into Cytoscape for network visualization and refinement, if desired.Figure 3GENIE3-enabled GRN construction(A) Representative ranked TF–target edge list generated by GENIE3. Each row corresponds to a predicted regulatory interaction. The regulatoryGene column denotes candidate TFs, the targetGene column denotes predicted target genes, and the weight column indicates the GENIE3 importance score reflecting the relative strength of the inferred interaction.(B) Example GRN visualization generated in Cytoscape. Triangles represent TFs and circles represent non-TF target genes. Directed edges indicate putative regulatory relationships from candidate TFs to predicted target genes. Network topology highlights differences in regulatory connectivity among candidate TFs.(C) Distribution of GENIE3 edge weights for the top-ranked TFs within the selected WGCNA module.Ridgeline plots display the density distribution of importance scores for TF–target interactions. The x-axis represents GENIE3 weight scores, and the y-axis lists the top-ranked TFs.Differences in distribution shape illustrate variability in predicted regulatory connectivity patterns among TFs, such as broad moderate-weight interactions versus fewer high-weight interactions. This analysis supports threshold selection and prioritization of candidate hub regulators for downstream validation. GENIE3-enabled GRN construction (A) Representative ranked TF–target edge list generated by GENIE3. Each row corresponds to a predicted regulatory interaction. The regulatoryGene column denotes candidate TFs, the targetGene column denotes predicted target genes, and the weight column indicates the GENIE3 importance score reflecting the relative strength of the inferred interaction. (B) Example GRN visualization generated in Cytoscape. Triangles represent TFs and circles represent non-TF target genes. Directed edges indicate putative regulatory relationships from candidate TFs to predicted target genes. Network topology highlights differences in regulatory connectivity among candidate TFs. (C) Distribution of GENIE3 edge weights for the top-ranked TFs within the selected WGCNA module. Ridgeline plots display the density distribution of importance scores for TF–target interactions. The x-axis represents GENIE3 weight scores, and the y-axis lists the top-ranked TFs. Differences in distribution shape illustrate variability in predicted regulatory connectivity patterns among TFs, such as broad moderate-weight interactions versus fewer high-weight interactions. This analysis supports threshold selection and prioritization of candidate hub regulators for downstream validation. R # Convert GENIE3 weight matrix to ranked edge list edge_list <- getLinkList( wm, threshold = 0.15 ) # Inspect edge list dim(edge_list) head(edge_list) # Export edge list for Cytoscape or downstream analysis write.table( edge_list, file = “GENIE3_edges_saddlebrown_c_threshold_0.15.tsv”, quote = FALSE, sep = “\t”, row.names = FALSE ) Note: The resulting edge list contains three columns: regulatoryGene, targetGene, and weight. The regulatoryGene column indicates the candidate TF, the targetGene column indicates the predicted target gene, and the weight column indicates the GENIE3 importance score for the inferred TF–target relationship (Figures 3A and 3B). CRITICAL: The optimal weight threshold is dataset- and module-dependent. If the inferred network is overly sparse, decrease the threshold, for example to 0.05–0.10. If the network is excessively dense, increase the threshold, for example to 0.20–0.30, or export only the top-ranked edges. Optional: Instead of applying a fixed weight threshold, users can export a defined number of top-ranked interactions. This approach can be useful when comparing networks across modules or species. R # Export top-ranked links instead of using a fixed threshold top_links <- 5000 edge_list_top <- getLinkList( wm, reportMax = top_links ) write.table( edge_list_top, file = “GENIE3_edges_saddlebrown_c_top5000.tsv”, quote = FALSE, sep = “\t”, row.names = FALSE ) Note: After generating the ranked edge list, examine the distribution of GENIE3 weight scores for highly connected TFs. Visualization of weight distributions, such as ridgeline plots, can help distinguish TFs with broad moderate-weight interactions from TFs with fewer high-confidence interactions (Figure 3C). This inspection can guide threshold selection and prioritization of candidate hub regulators for downstream validation. Promoter scanning and conservation analysis Timing: 1–2 h, depending on the number of genes analyzed Timing: 1–2 h, depending on the number of genes analyzed This section describes the identification of conserved non-coding sequences (CNSs) within promoters of a target gene and its homologs in related grass species using mVISTA.10 In this protocol, the promoter of SbCSLF6 and its homologs is used as an example. CNSs that overlap predicted cis-regulatory elements (CREs) or candidate TF-binding motifs provide supportive in silico evidence for TF–target relationships inferred from GENIE3-based GRN modeling. In this protocol, mVISTA is used as an online web platform, not as locally installed software. Users should upload promoter FASTA sequences through the mVISTA submission page, select LAGAN alignment, specify the sorghum promoter as the reference sequence, and adjust the conservation parameters as described below. No local mVISTA installation or command-line execution is required.CRITICAL: Promoter conservation analysis does not prove direct binding or regulatory activity. It provides complementary in silico evidence to prioritize predicted regulatory interactions for downstream experimental validation.31.Identify homologs of the target gene in related grass species.a.Search for homologs of the sorghum target gene in related monocot species using Phytozome.b.Use SbCSLF6; Sobic.007G050600 as the example sorghum target gene.c.Retrieve homologous genes from related grass species, including Zea mays RefGen_v4, Setaria italica v2.2, Panicum virgatum v5.1, Hordeum vulgare r1, and Brachypodium distachyon v3.2.Note: The example Phytozome gene page for SbCSLF6 is https://phytozome-next.jgi.doe.gov/report/gene/Sbicolor_v3_1_1/Sobic.007G050600.Note: When multiple isoforms or homolog candidates are present, select the transcript with the strongest sequence similarity, such as the lowest E-value and highest query coverage. If candidates have comparable similarity, prioritize the longest coding sequence or the gene model most consistent with synteny or orthology information when available.Optional: For more stringent homolog identification, perform a manual BLASTP search and filter candidates by E-value, query coverage, and percentage identity.32.Retrieve promoter sequences.a.Download upstream promoter sequences using Phytozome Biomart or another genome browser/sequence retrieval tool.b.Extract promoter sequences upstream of the transcription start site (TSS) for each homolog.c.Confirm that promoter sequences are extracted from the correct strand.d.Confirm that selected promoter regions do not overlap upstream coding sequences.Note: The example Phytozome BioMart page is https://phytozome-next.jgi.doe.gov/biomart/martview/c38ea25dab0e78338af09d12c561ac40.Note: Promoter length should be adjusted based on biological context and genome size and gene density. For many transcriptionally inducible genes, 1 kb upstream of the TSS may capture core CREs. In species with larger intergenic regions, such as maize, functional regulatory elements may extend farther upstream, for example 5–10 kb.CRITICAL: For genes located on the negative strand, the upstream promoter region is downstream in genomic coordinate space.33.Perform promoter alignment using mVISTA.a.Open the mVISTA online submission page.b.Upload promoter FASTA sequences.c.Select LAGAN alignment.d.Specify the sorghum promoter sequence, such as the SbCSLF6 promoter, as the reference sequence.e.Adjust conservation parameters under “Adjust Conservation Parameters” using the following values: Min_Y = 20, Min_Id = 70, and Min_Length = 30.f.Export or save the resulting VISTA conservation plot.Note: These settings display conserved alignment windows of at least 30 bp with a minimum sequence identity of 70%. Conserved regions are highlighted in the VISTA plot.Note: Conserved non-coding regions that overlap predicted TF-binding motifs or CREs strengthen the biological plausibility of regulatory interactions inferred from GENIE3. However, these results should be interpreted as prioritization evidence rather than direct proof of TF binding.Optional: Users may adjust the conservation parameters depending on evolutionary distance among species. If no conserved regions are detected, consider lowering the minimum identity threshold or increasing promoter length. If too many conserved regions are detected, consider increasing the minimum identity or minimum length threshold. CRITICAL: Promoter conservation analysis does not prove direct binding or regulatory activity. It provides complementary in silico evidence to prioritize predicted regulatory interactions for downstream experimental validation. 31.Identify homologs of the target gene in related grass species.a.Search for homologs of the sorghum target gene in related monocot species using Phytozome.b.Use SbCSLF6; Sobic.007G050600 as the example sorghum target gene.c.Retrieve homologous genes from related grass species, including Zea mays RefGen_v4, Setaria italica v2.2, Panicum virgatum v5.1, Hordeum vulgare r1, and Brachypodium distachyon v3.2. Identify homologs of the target gene in related grass species.a.Search for homologs of the sorghum target gene in related monocot species using Phytozome.b.Use SbCSLF6; Sobic.007G050600 as the example sorghum target gene.c.Retrieve homologous genes from related grass species, including Zea mays RefGen_v4, Setaria italica v2.2, Panicum virgatum v5.1, Hordeum vulgare r1, and Brachypodium distachyon v3.2. a.Search for homologs of the sorghum target gene in related monocot species using Phytozome. Search for homologs of the sorghum target gene in related monocot species using Phytozome. b.Use SbCSLF6; Sobic.007G050600 as the example sorghum target gene. Use SbCSLF6; Sobic.007G050600 as the example sorghum target gene. c.Retrieve homologous genes from related grass species, including Zea mays RefGen_v4, Setaria italica v2.2, Panicum virgatum v5.1, Hordeum vulgare r1, and Brachypodium distachyon v3.2. Retrieve homologous genes from related grass species, including Zea mays RefGen_v4, Setaria italica v2.2, Panicum virgatum v5.1, Hordeum vulgare r1, and Brachypodium distachyon v3.2. Note: The example Phytozome gene page for SbCSLF6 is https://phytozome-next.jgi.doe.gov/report/gene/Sbicolor_v3_1_1/Sobic.007G050600. Note: When multiple isoforms or homolog candidates are present, select the transcript with the strongest sequence similarity, such as the lowest E-value and highest query coverage. If candidates have comparable similarity, prioritize the longest coding sequence or the gene model most consistent with synteny or orthology information when available. Optional: For more stringent homolog identification, perform a manual BLASTP search and filter candidates by E-value, query coverage, and percentage identity. 32.Retrieve promoter sequences.a.Download upstream promoter sequences using Phytozome Biomart or another genome browser/sequence retrieval tool.b.Extract promoter sequences upstream of the transcription start site (TSS) for each homolog.c.Confirm that promoter sequences are extracted from the correct strand.d.Confirm that selected promoter regions do not overlap upstream coding sequences. Retrieve promoter sequences.a.Download upstream promoter sequences using Phytozome Biomart or another genome browser/sequence retrieval tool.b.Extract promoter sequences upstream of the transcription start site (TSS) for each homolog.c.Confirm that promoter sequences are extracted from the correct strand.d.Confirm that selected promoter regions do not overlap upstream coding sequences. a.Download upstream promoter sequences using Phytozome Biomart or another genome browser/sequence retrieval tool. Download upstream promoter sequences using Phytozome Biomart or another genome browser/sequence retrieval tool. b.Extract promoter sequences upstream of the transcription start site (TSS) for each homolog. Extract promoter sequences upstream of the transcription start site (TSS) for each homolog. c.Confirm that promoter sequences are extracted from the correct strand. Confirm that promoter sequences are extracted from the correct strand. d.Confirm that selected promoter regions do not overlap upstream coding sequences. Confirm that selected promoter regions do not overlap upstream coding sequences. Note: The example Phytozome BioMart page is https://phytozome-next.jgi.doe.gov/biomart/martview/c38ea25dab0e78338af09d12c561ac40. Note: Promoter length should be adjusted based on biological context and genome size and gene density. For many transcriptionally inducible genes, 1 kb upstream of the TSS may capture core CREs. In species with larger intergenic regions, such as maize, functional regulatory elements may extend farther upstream, for example 5–10 kb. CRITICAL: For genes located on the negative strand, the upstream promoter region is downstream in genomic coordinate space. 33.Perform promoter alignment using mVISTA.a.Open the mVISTA online submission page.b.Upload promoter FASTA sequences.c.Select LAGAN alignment.d.Specify the sorghum promoter sequence, such as the SbCSLF6 promoter, as the reference sequence.e.Adjust conservation parameters under “Adjust Conservation Parameters” using the following values: Min_Y = 20, Min_Id = 70, and Min_Length = 30.f.Export or save the resulting VISTA conservation plot. Perform promoter alignment using mVISTA.a.Open the mVISTA online submission page.b.Upload promoter FASTA sequences.c.Select LAGAN alignment.d.Specify the sorghum promoter sequence, such as the SbCSLF6 promoter, as the reference sequence.e.Adjust conservation parameters under “Adjust Conservation Parameters” using the following values: Min_Y = 20, Min_Id = 70, and Min_Length = 30.f.Export or save the resulting VISTA conservation plot. a.Open the mVISTA online submission page. Open the mVISTA online submission page. b.Upload promoter FASTA sequences. Upload promoter FASTA sequences. c.Select LAGAN alignment. Select LAGAN alignment. d.Specify the sorghum promoter sequence, such as the SbCSLF6 promoter, as the reference sequence. Specify the sorghum promoter sequence, such as the SbCSLF6 promoter, as the reference sequence. e.Adjust conservation parameters under “Adjust Conservation Parameters” using the following values: Min_Y = 20, Min_Id = 70, and Min_Length = 30. Adjust conservation parameters under “Adjust Conservation Parameters” using the following values: Min_Y = 20, Min_Id = 70, and Min_Length = 30. f.Export or save the resulting VISTA conservation plot. Export or save the resulting VISTA conservation plot. Note: These settings display conserved alignment windows of at least 30 bp with a minimum sequence identity of 70%. Conserved regions are highlighted in the VISTA plot. Note: Conserved non-coding regions that overlap predicted TF-binding motifs or CREs strengthen the biological plausibility of regulatory interactions inferred from GENIE3. However, these results should be interpreted as prioritization evidence rather than direct proof of TF binding. Optional: Users may adjust the conservation parameters depending on evolutionary distance among species. If no conserved regions are detected, consider lowering the minimum identity threshold or increasing promoter length. If too many conserved regions are detected, consider increasing the minimum identity or minimum length threshold. Expected outcomes Application of this protocol generates a network-guided set of candidate transcriptional regulators associated with genes or pathways of interest using transcriptome data alone. The workflow begins with RNA-seq data processing and proceeds through WGCNA-based coexpression network construction, GENIE3-based GRN inference, and promoter conservation analysis (Figure 1). Starting from raw RNA-seq data, users should obtain quality-controlled trimmed FASTQ files, sorted alignment BAM files, and gene-level count matrices suitable for downstream analysis. For the sorghum stem RNA-seq dataset used in this protocol, mapping rates are typically expected to be above 85%, although this value may vary depending on genome quality, read depth, library type, tissue, and species. Differential expression analysis using DESeq2 will produce contrast-specific gene lists with log2 fold changes and adjusted p-values. The number of DEGs will depend on the biological contrast, sample number, sequencing depth, and statistical thresholds. WGCNA-based coexpression network analysis will yield discrete gene modules, each assigned a unique color label (Figure 2). The number of modules depends on sample size, biological variation, and parameter choices, including soft-thresholding power, minimum module size, and module merging threshold. Each module is represented by a module eigengene summarizing its expression pattern across samples. These eigengenes can be compared across developmental stages, time points, tissues, or treatments to identify biologically relevant modules. Module-specific gene lists and eigengene tables are the key outputs of this stage. GENIE3-based GRN inference will generate a ranked list of predicted TF–target interactions within selected WGCNA modules (Figure 3). The output includes weighted edges representing inferred regulatory importance. Depending on the selected weight threshold or number of top-ranked edges retained, the resulting network may contain hundreds to thousands of candidate interactions. When integrated with WGCNA-derived module membership and connectivity metrics, this output enables prioritization of candidate upstream regulators associated with genes or pathways of interest. Promoter conservation analysis using mVISTA is expected to identify conserved CNSs among homologous promoters across related grass species. Conserved regions that overlap predicted CREs or candidate TF-binding motifs provide supportive in silico evidence for predicted regulatory interactions. These results should be interpreted as prioritization evidence rather than proof of direct TF binding or regulatory activity. Collectively, this protocol enables users to produce quality-controlled RNA-seq count matrices, differential expression gene lists, coexpression modules with eigengenes, ranked TF–target edge lists, and conserved promoter regions supporting predicted regulatory relationships. These outputs provide a reproducible, transcriptome-driven strategy for prioritizing regulatory genes in crop and non-model plant species where multi-omics datasets are limited. Limitations Although NEEDLE provides a structured workflow for transcriptome-driven gene regulatory inference, several limitations should be considered when interpreting results. First, the accuracy of coexpression and regulatory network modeling depends strongly on the quality, size, and biological relevance of the input transcriptome dataset. Datasets with too few biologically meaningful samples may reduce the robustness of correlation-based module detection and increase the likelihood of spurious associations. Based on our testing, six or more biologically meaningful samples are generally recommended as a practical starting point, but this should not be interpreted as a strict cutoff. The optimal sample number depends on dataset complexity, biological variation, and the specific question being addressed. Likewise, datasets lacking sufficient biological variation across time points, developmental stages, tissues, genotypes, or treatment may fail to produce well-defined modules or meaningful regulatory hierarchies. Second, coexpression does not imply causation.3,13 WGCNA identifies genes with correlated expression patterns, and GENIE3 infers putative regulatory relationships based on statistical dependencies, but neither approach directly demonstrates physical TF–DNA binding or regulatory activity in vivo. Network predictions are therefore hypothesis-generating and require experimental validation. In addition, parameter choices, including soft-thresholding power in WGCNA, module merging thresholds, and weight cutoffs in GENIE3, can influence network topology and regulator ranking. Although commonly used parameters are provided in this protocol, users may need to optimize these settings according to dataset characteristics. Third, transcriptome-only approaches cannot capture regulatory mechanisms mediated through chromatin accessibility, DNA methylation, protein–protein interactions, post-translational modifications, translational control, or protein stability. As a result, true regulators with low, transient, or constitutive expression may be missed, and inferred networks may not fully represent the underlying biological complexity. Similarly, promoter conservation analysis provides indirect evidence of regulatory potential but does not confirm direct TF binding or functional activity of CREs. Finally, technical and annotation-related factors, such as batch effects, sequencing-depth variability, inaccurate genome annotations, inconsistent gene identifiers, or mismatched genome versions, can affect network construction and regulatory inference. Careful preprocessing, metadata verification, and consistent use of genome and annotation versions are critical to minimize these issues. Taken together, NEEDLE should be viewed as a systematic, transcriptome-based prioritization pipeline that guides downstream functional validation rather than as definitive proof of regulatory relationships. Troubleshooting NEEDLE is designed as a transcriptome-driven prioritization strategy. If unexpected results occur at multiple stages, such as unstable modules or inconsistent regulator rankings, re-evaluate upstream preprocessing steps, sample metadata accuracy, and genome annotation consistency before adjusting downstream network parameters. Below are common issues that may arise during execution of the NEEDLE workflow, along with recommended solutions. Each problem is linked to the corresponding step in the protocol. Problem 1 Low mapping rate, for example <60%, after alignment. Related to RNA-seq processing, Step 8. Potential causes •Incorrect genome version or mismatched GFF/GTF annotation•Incorrect library-type specification•Poor-quality or contaminated FASTQ files•Incorrect pairing of R1 and R2 reads •Incorrect genome version or mismatched GFF/GTF annotation Incorrect genome version or mismatched GFF/GTF annotation •Incorrect library-type specification Incorrect library-type specification •Poor-quality or contaminated FASTQ files Poor-quality or contaminated FASTQ files •Incorrect pairing of R1 and R2 reads Incorrect pairing of R1 and R2 reads Potential solution •Confirm that the reference FASTA and annotation files are from the same genome build.•Verify --library-type parameter, such as fr-firststrand versus fr-secondstrand.•Inspect FastQC reports for adapter contamination or low-quality reads and adjust trimming parameters if needed.•Confirm correct pairing of R1 and R2 reads, especially if paired-end reads were split from an interleaved FASTQ file. •Confirm that the reference FASTA and annotation files are from the same genome build. Confirm that the reference FASTA and annotation files are from the same genome build. •Verify --library-type parameter, such as fr-firststrand versus fr-secondstrand. Verify --library-type parameter, such as fr-firststrand versus fr-secondstrand. •Inspect FastQC reports for adapter contamination or low-quality reads and adjust trimming parameters if needed. Inspect FastQC reports for adapter contamination or low-quality reads and adjust trimming parameters if needed. •Confirm correct pairing of R1 and R2 reads, especially if paired-end reads were split from an interleaved FASTQ file. Confirm correct pairing of R1 and R2 reads, especially if paired-end reads were split from an interleaved FASTQ file. Problem 2 DESeq2 returns very few or no differentially expressed genes. Related to differential gene expression analysis, Step 13. Potential causes •Insufficient biological variation among samples•Incorrect sample metadata or experimental design matrix•Overly stringent filtering or statistical thresholds•Normalized values used as input instead of raw counts •Insufficient biological variation among samples Insufficient biological variation among samples •Incorrect sample metadata or experimental design matrix Incorrect sample metadata or experimental design matrix •Overly stringent filtering or statistical thresholds Overly stringent filtering or statistical thresholds •Normalized values used as input instead of raw counts Normalized values used as input instead of raw counts Potential solution •Confirm that metadata correctly reflects experimental groups, stages, or contrasts.•Inspect sample clustering or principal component analysis plots to evaluate separation among biological groups.•Adjust filtering thresholds if they are overly restrictive.•Verify that count files contain raw gene-level counts, because DESeq2 requires raw counts as input. •Confirm that metadata correctly reflects experimental groups, stages, or contrasts. Confirm that metadata correctly reflects experimental groups, stages, or contrasts. •Inspect sample clustering or principal component analysis plots to evaluate separation among biological groups. Inspect sample clustering or principal component analysis plots to evaluate separation among biological groups. •Adjust filtering thresholds if they are overly restrictive. Adjust filtering thresholds if they are overly restrictive. •Verify that count files contain raw gene-level counts, because DESeq2 requires raw counts as input. Verify that count files contain raw gene-level counts, because DESeq2 requires raw counts as input. Problem 3 WGCNA fails with a memory error or excessive runtime. Related to Network construction and TOM calculation, Step 19. Potential causes •Large gene set, for example >15,000 genes•Limited RAM availability•Full TOM calculation applied to a large expression matrix •Large gene set, for example >15,000 genes Large gene set, for example >15,000 genes •Limited RAM availability Limited RAM availability •Full TOM calculation applied to a large expression matrix Full TOM calculation applied to a large expression matrix Potential solution •Pre-filter genes by expression level or variance before WGCNA.•Use blockwiseModules() instead of full TOM calculation.•Run the analysis on an HPC node with higher memory allocation.•If biologically appropriate, reduce the input to DEGs or genes associated with the process of interest. •Pre-filter genes by expression level or variance before WGCNA. Pre-filter genes by expression level or variance before WGCNA. •Use blockwiseModules() instead of full TOM calculation. Use blockwiseModules() instead of full TOM calculation. •Run the analysis on an HPC node with higher memory allocation. Run the analysis on an HPC node with higher memory allocation. •If biologically appropriate, reduce the input to DEGs or genes associated with the process of interest. If biologically appropriate, reduce the input to DEGs or genes associated with the process of interest. Problem 4 WGCNA produces too many small modules or one very large module. Related to Dynamic tree cut and module merging, Step 20 and 21. Potential causes •Inappropriate soft-thresholding power•Suboptimal deepSplit or minModuleSize parameters•Weak scale-free topology fit•Insufficient biological variation in the input dataset •Inappropriate soft-thresholding power Inappropriate soft-thresholding power •Suboptimal deepSplit or minModuleSize parameters Suboptimal deepSplit or minModuleSize parameters •Weak scale-free topology fit Weak scale-free topology fit •Insufficient biological variation in the input dataset Insufficient biological variation in the input dataset Potential solution •Re-evaluate soft-threshold power selection, Step 18.•Increase minModuleSize to reduce module fragmentation.•Adjust deepSplit: lower values reduce splitting, whereas higher values increase module granularity.•Confirm that the dataset contains sufficient biological variation across time points, developmental stages, tissues, genotypes, or treatments. •Re-evaluate soft-threshold power selection, Step 18. Re-evaluate soft-threshold power selection, Step 18. •Increase minModuleSize to reduce module fragmentation. Increase minModuleSize to reduce module fragmentation. •Adjust deepSplit: lower values reduce splitting, whereas higher values increase module granularity. Adjust deepSplit: lower values reduce splitting, whereas higher values increase module granularity. •Confirm that the dataset contains sufficient biological variation across time points, developmental stages, tissues, genotypes, or treatments. Confirm that the dataset contains sufficient biological variation across time points, developmental stages, tissues, genotypes, or treatments. Problem 5 GENIE3 returns very few or zero regulatory edges. Related to GENIE3 input preparation, network inference, and edge extraction, Steps 25 and 28–30. Potential causes •TF identifiers do not match expression matrix gene IDs•Too few candidate regulators are present in the selected module•Weight threshold is set too high•Input expression matrix is not numeric or is incorrectly formatted •TF identifiers do not match expression matrix gene IDs TF identifiers do not match expression matrix gene IDs •Too few candidate regulators are present in the selected module Too few candidate regulators are present in the selected module •Weight threshold is set too high Weight threshold is set too high •Input expression matrix is not numeric or is incorrectly formatted Input expression matrix is not numeric or is incorrectly formatted Potential solution •Confirm exact gene ID matching between TF list and expression matrix.•Check whether TF identifiers use gene IDs, transcript IDs, or different genome-version naming conventions.•Lower the weight threshold, for example from 0.15 to 0.05–0.10.•Expand the regulator list to include all expressed TFs if appropriate.•Verify that the input expression matrix is numeric, contains genes as rows and samples as columns, and has gene IDs stored as row names. •Confirm exact gene ID matching between TF list and expression matrix. Confirm exact gene ID matching between TF list and expression matrix. •Check whether TF identifiers use gene IDs, transcript IDs, or different genome-version naming conventions. Check whether TF identifiers use gene IDs, transcript IDs, or different genome-version naming conventions. •Lower the weight threshold, for example from 0.15 to 0.05–0.10. Lower the weight threshold, for example from 0.15 to 0.05–0.10. •Expand the regulator list to include all expressed TFs if appropriate. Expand the regulator list to include all expressed TFs if appropriate. •Verify that the input expression matrix is numeric, contains genes as rows and samples as columns, and has gene IDs stored as row names. Verify that the input expression matrix is numeric, contains genes as rows and samples as columns, and has gene IDs stored as row names. Problem 6 GENIE3 produces an excessively dense network. Related to edge extraction, Step 30. Potential causes •Weight threshold is set too low•Selected module contains many genes•Regulator list is too broad for the biological question •Weight threshold is set too low Weight threshold is set too low •Selected module contains many genes Selected module contains many genes •Regulator list is too broad for the biological question Regulator list is too broad for the biological question Potential solution •Increase weight threshold, for example to 0.20–0.30.•Export the top N edges instead of using threshold-based filtering.•Restrict the regulator list to TFs with high intramodular connectivity or TFs associated with the selected module.•Export and inspect the edge-weight distribution before choosing a final threshold. •Increase weight threshold, for example to 0.20–0.30. Increase weight threshold, for example to 0.20–0.30. •Export the top N edges instead of using threshold-based filtering. Export the top N edges instead of using threshold-based filtering. •Restrict the regulator list to TFs with high intramodular connectivity or TFs associated with the selected module. Restrict the regulator list to TFs with high intramodular connectivity or TFs associated with the selected module. •Export and inspect the edge-weight distribution before choosing a final threshold. Export and inspect the edge-weight distribution before choosing a final threshold. Problem 7 Promoter sequences cannot be retrieved or appear truncated. Related to promoter extraction, Step 32. Potential causes •Incorrect gene model or genome version•Gene is located near a chromosome or scaffold boundary•Strand orientation was not considered correctly•Promoter extraction overlaps an upstream coding region •Incorrect gene model or genome version Incorrect gene model or genome version •Gene is located near a chromosome or scaffold boundary Gene is located near a chromosome or scaffold boundary •Strand orientation was not considered correctly Strand orientation was not considered correctly •Promoter extraction overlaps an upstream coding region Promoter extraction overlaps an upstream coding region Potential solution •Confirm that the genome assembly version matches the annotation version.•Check the TSS coordinates and gene orientation.•For genes on the negative strand, verify that the upstream promoter region is extracted from the correct genomic direction.•Manually retrieve upstream region using genome browser if automated extraction fails.•Confirm that the selected promoter region does not overlap neighboring coding sequences. •Confirm that the genome assembly version matches the annotation version. Confirm that the genome assembly version matches the annotation version. •Check the TSS coordinates and gene orientation. Check the TSS coordinates and gene orientation. •For genes on the negative strand, verify that the upstream promoter region is extracted from the correct genomic direction. For genes on the negative strand, verify that the upstream promoter region is extracted from the correct genomic direction. •Manually retrieve upstream region using genome browser if automated extraction fails. Manually retrieve upstream region using genome browser if automated extraction fails. •Confirm that the selected promoter region does not overlap neighboring coding sequences. Confirm that the selected promoter region does not overlap neighboring coding sequences. Problem 8 No conserved regions are detected in mVISTA. Related to promoter alignment, Step 33. Potential causes •Promoter sequences are too short•Evolutionary distance among selected species is too large•Alignment parameters are too stringent•Incorrect homologs were selected•Sequences are not oriented consistently •Promoter sequences are too short Promoter sequences are too short •Evolutionary distance among selected species is too large Evolutionary distance among selected species is too large •Alignment parameters are too stringent Alignment parameters are too stringent •Incorrect homologs were selected Incorrect homologs were selected •Sequences are not oriented consistently Sequences are not oriented consistently Potential solution •Extend promoter length, for example from 1 kb to 2–5 kb.•Lower the conservation identity threshold, for example from 70% to 60%.•Confirm homolog selection using BLASTP, orthology information, or synteny when available.•Verify that promoter sequences are aligned in the correct orientation.•Consider using more closely related species if conserved regions remain undetectable. •Extend promoter length, for example from 1 kb to 2–5 kb. Extend promoter length, for example from 1 kb to 2–5 kb. •Lower the conservation identity threshold, for example from 70% to 60%. Lower the conservation identity threshold, for example from 70% to 60%. •Confirm homolog selection using BLASTP, orthology information, or synteny when available. Confirm homolog selection using BLASTP, orthology information, or synteny when available. •Verify that promoter sequences are aligned in the correct orientation. Verify that promoter sequences are aligned in the correct orientation. •Consider using more closely related species if conserved regions remain undetectable. Consider using more closely related species if conserved regions remain undetectable. Problem 9 Inconsistent gene identifiers across steps. Related to DESeq2, WGCNA, and GENIE3 input preparation, Steps 13, 14, 16, 25, 27, and 28. Potential causes •Different genome builds are mixed across FASTA, GFF/GTF, expression matrices, and TF lists•Gene IDs and transcript IDs are used inconsistently•Identifier prefixes, suffixes, or version numbers differ across files •Different genome builds are mixed across FASTA, GFF/GTF, expression matrices, and TF lists Different genome builds are mixed across FASTA, GFF/GTF, expression matrices, and TF lists •Gene IDs and transcript IDs are used inconsistently Gene IDs and transcript IDs are used inconsistently •Identifier prefixes, suffixes, or version numbers differ across files Identifier prefixes, suffixes, or version numbers differ across files Potential solution •Standardize gene identifiers at the beginning of the workflow.•Use consistent genome FASTA, annotation, expression matrix, and TF annotation files throughout the analysis.•Convert identifiers using official mapping tables when necessary.•Before running GENIE3, confirm that candidate TF identifiers are present in the expression matrix using intersect(regulators, rownames(exprMatr)). •Standardize gene identifiers at the beginning of the workflow. Standardize gene identifiers at the beginning of the workflow. •Use consistent genome FASTA, annotation, expression matrix, and TF annotation files throughout the analysis. Use consistent genome FASTA, annotation, expression matrix, and TF annotation files throughout the analysis. •Convert identifiers using official mapping tables when necessary. Convert identifiers using official mapping tables when necessary. •Before running GENIE3, confirm that candidate TF identifiers are present in the expression matrix using intersect(regulators, rownames(exprMatr)). Before running GENIE3, confirm that candidate TF identifiers are present in the expression matrix using intersect(regulators, rownames(exprMatr)). Resource availability Lead contact Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Federica Brandizzi (fb@msu.edu). Technical contact Technical questions on executing this protocol should be directed to and will be answered by the technical contact, Dae Kwan Ko (dkko@msu.edu). Materials availability This study did not generate new materials. Data and code availability •The accession number for RNA-seq data reported in this study is available from the following link: https://genome.jgi.doe.gov/portal/SorbicEProfiling_14/SorbicEProfiling_14.info.html.•Code has been deposited in the GitHub repository (https://github.com/DaeKwan-Ko/needle) and archived at Zenodo (https://doi.org/10.5281/zenodo.14560598). The code is publicly available as of the date of publication. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. •The accession number for RNA-seq data reported in this study is available from the following link: https://genome.jgi.doe.gov/portal/SorbicEProfiling_14/SorbicEProfiling_14.info.html. The accession number for RNA-seq data reported in this study is available from the following link: https://genome.jgi.doe.gov/portal/SorbicEProfiling_14/SorbicEProfiling_14.info.html. •Code has been deposited in the GitHub repository (https://github.com/DaeKwan-Ko/needle) and archived at Zenodo (https://doi.org/10.5281/zenodo.14560598). The code is publicly available as of the date of publication. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Code has been deposited in the GitHub repository (https://github.com/DaeKwan-Ko/needle) and archived at Zenodo (https://doi.org/10.5281/zenodo.14560598). The code is publicly available as of the date of publication. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments This study was supported primarily by the 10.13039/100015814Great Lakes Bioenergy Research Center, 10.13039/100000015U.S. Department of Energy, 10.13039/100006132Office of Science, 10.13039/100006206Office of Biological and Environmental Research (DE-SC0018409), with contributing support from the 10.13039/100000002National Institutes of Health (R35GM136637); the Chemical Sciences, Geoscience and Biosciences Division, 10.13039/100006151Office of Basic Energy Sciences, Office of Science, U.S. Department of Energy (DE-FG02-91ER20021); and 10.13039/100007250MSU AgBioResearch (MICL02598). Author contributions D.K.K. and F.B. conceptualized the study. D.K.K. designed and conducted the analyses, developed and implemented the method, and wrote the manuscript draft. D.K.K. and F.B. edited the manuscript. F.B. provided supervision and funding. Declaration of interests The authors declare no competing interests.