Brieflow: an integrated computational pipeline for high-throughput analysis of optical pooled screening data Authors: Matteo Di Bernardo, Roshan S Kern, Ana Karla Cepeda Diaz, Alexa Mallar, Samuel J Choi, Andrew Nutter-Upham, Sebastian Lourido, Paul C Blainey, Iain Cheeseman Published: 2026 ABSTRACT AbstractOptical pooled screening (OPS) has emerged as a powerful technique for functional genomics, enabling researchers to link genetic perturbations with complex cellular morphological phenotypes at scale. However, OPS data analysis presents challenges due to massive datasets, complex multi-modal integration requirements, and the absence of standardized frameworks. Here, we present Brieflow, a computational pipeline for end-to-end analysis of fixed-cell optical pooled screening data. We demonstrate Brieflow’s capabilities through reanalysis of a CRISPR-Cas9 screen encompassing 5072 fitness-conferring genes, processing more than 70 million cells with multiple phenotypic markers. To accelerate biological interpretation, we additionally present MozzareLLM, a framework leveraging large language models to identify biological processes within phenotypic clusters and prioritize gene candidates for experimental validation. Our combined analysis recovers coherent biological modules missed by existing analytical approaches, including five core mitochondrial sub-programs absent from the original study. The modular design and open-source implementation of Brieflow facilitates the integration of new analytical components while ensuring computational reproducibility and improved performance for the use of high-content phenotypic screening in biological discovery.Subject terms: Cellular imaging, Software, Functional clustering, Data processing, High-throughput screening Introduction Functional genomics has been transformed by high-throughput screening technologies that enable the systematic interrogation of gene function. Optical Pooled Screening (OPS) has emerged as a powerful technique for elucidating gene-phenotype relationships by combining pooled genetic perturbations with high-resolution cellular imaging1. In a typical OPS workflow, cells are transduced with a pooled library of genetic perturbations (e.g., CRISPR-Cas9), each identified by a unique barcode sequence (Fig. 1a, panel I). Following selection and fixation, cells undergo high-content imaging to capture phenotypic features, followed by in situ sequencing to identify barcode sequences (Fig. 1a, panel II). Brieflow: an integrated computational framework for optical pooled screening analysis. a Schematic of the optical pooled screening workflow. Cells expressing inducible Cas9 are transduced with an sgRNA library, perturbed, and subjected to both phenotypic imaging and in situ sgRNA sequencing, followed by computational analysis using Brieflow. b The Brieflow pipeline architecture comprising seven integrated modules: Preprocess converts raw microscopy files to standardized formats; Sequencing-by-Synthesis identifies guide barcodes; Phenotype extracts cellular features; Merge integrates genotype and phenotype data using triangulation-based spatial registration; Classify optionally partitions cells into relevant subpopulations; Aggregate condenses single-cell data to perturbation-level statistics; and Cluster identifies functional relationships between perturbations. c Implementation architecture. Left: the Brieflow codebase with modular structure, where a high-level Snakefile imports rules and targets from each module, rules execute individual Python scripts, and scripts call functions defined in the library. Right: the complementary Brieflow-analysis template, which separates user-facing configuration from pipeline logic to simplify parameter tuning, where a configuration notebook allows users to test and set parameters that are automatically written into the configuration file, which is then employed by execution scripts that call the Snakefile with appropriate module-specific parameters. Created in BioRender. Di Bernardo, M. (2026) https://BioRender.com/ycfpo55. Unlike traditional pooled screening approaches that rely on simple readouts such as cell viability, OPS captures rich, multidimensional phenotypic data while maintaining advantages including increased throughput and reduced reagent consumption. Furthermore, OPS integrates seamlessly into the cell biologist’s toolkit, effectively scaling standard imaging-based analyses to enable systematic investigation of stains, antibodies, and fluorescent reporters at high throughput. OPS has been successfully deployed for diverse applications, including the discovery of antiviral response pathway components2, the assignment of poorly characterized genes to known cellular pathways via morphological similarity3, and the establishment of previously unknown connections between genes and complex imaging phenotypes4. The scale and complexity of OPS data present formidable computational obstacles. These terabyte-size datasets often contain information from multi-channel fluorescence images of millions of individual cells, each characterized by thousands of phenotypic parameters. The integration of in situ sequencing and phenotypic measurements necessitates accurate matching of barcode identities with corresponding phenotypic features across different imaging modalities with distinct optical characteristics and resolution limitations. Normalization across experimental batches and conditions is essential to minimize technical variation and enable reliable biological comparisons. Although powerful computational methods have been developed that leverage processed OPS data for tasks such as perturbation prediction and phenotypic modeling5–7, the field lacks both a standardized analysis pipeline accessible to experimental biologists and a consistent output format for training computational models. The processing pipelines necessary to generate these inputs—from barcode demultiplexing and cell-perturbation assignment through phenotype feature extraction—have remained fragmented, lab-specific, and often require manual intervention at multiple stages, slowing adoption by wet-lab researchers and limiting the standardized datasets available for downstream modeling efforts8. Finally, the immense size of raw image datasets makes conventional data sharing impractical and incomplete documentation of analysis parameters undermines reproducibility. To address these challenges, we present Brieflow, a computational pipeline designed specifically for end-to-end analysis of fixed-cell OPS datasets. Brieflow provides a standardized framework for processing and analyzing OPS data, offering a critical pathway to incorporate complex cellular morphologies into emerging in silico cell models. To demonstrate the improvements offered by our approach, we benchmarked Brieflow through a comprehensive reanalysis of a large-scale CRISPR-Cas9 screen comprising over 5000 fitness-conferring genes in HeLa cells9, illustrating its efficacy in processing and extracting biological insights from terabyte-scale OPS datasets. To enable the advanced interpretation of these datasets, we introduce MozzareLLM, a specialized large language model (LLM) framework for biological interpretation that automatically analyzes phenotypic clusters for their functional processes and identifies promising gene candidates for experimental validation. By enabling reproducible, consistent analysis across diverse experimental contexts with improved quality and precision, this pipeline represents an essential step toward integrating morphological representations of cellular function into comprehensive computational models of cell biology. Results Brieflow: an integrated framework for OPS data analysis The analysis of OPS datasets is a multi-step process that requires specialized computational tools to extract, integrate, and interpret the multidimensional phenotypic data. To enable the analysis of these data, we developed a computational pipeline that we term Brieflow (Fig. 1a, panel III). Brieflow provides a unified framework that processes data from raw images to generate biological insights. Brieflow employs a modular architecture comprising seven integrated components (Fig. 1b): Preprocess, Sequencing-by-Synthesis, Phenotype, Merge, Classify, Aggregate, and Cluster. As described in detail below, the Preprocess module converts raw images to analysis-ready image tiles; Sequencing-by-Synthesis identifies in situ barcodes of genetic perturbations; Phenotype quantifies cellular features from high-content imaging; Merge aligns perturbations and phenotypes using triangle hashing; Classify partitions cells into relevant subpopulations; Aggregate collapses cellular data to the perturbation level; and Cluster identifies phenotypic relationships between perturbations. Brieflow’s modular architecture enables flexible workflows through independent components with well-defined inputs and outputs, allowing execution of either the complete pipeline or individual modules as needed. The implementation leverages Snakemake-based workflow management, which automatically determines processing steps based on file dependencies for incremental updates (Supplementary Fig. 1), distributes independent tasks across available computational resources, and captures complete analytical provenance in a version-controlled format for reproducibility10. The architecture comprises library code, Python scripts, Snakemake rules, and targets according to a standardized project development template (Fig. 1c, panel I). The Brieflow-analysis template repository provides a hierarchical, YAML-based configuration system that separates experimental parameters from processing logic, with configuration Jupyter notebooks that guide users through parameter selection for each module and execution scripts optimized for both local and high-performance computing environments (Fig. 1c, panel II). As a case study, we use Brieflow to reanalyze the OPS data from Funk et al. 2022 with the Cas9-based targeting of ~5000 fitness-conferring genes in human HeLa cells, therein designated as the “Vesuvius” dataset, with specific findings described for each step below. Preprocess: conversion of raw microscopy files to standardized tiled images The analysis of complex OPS datasets first requires standardization of diverse microscopy data formats into a unified, analysis-ready structure. The Preprocess module addresses initial challenges of OPS data analysis by providing tools for data conversion, metadata extraction, and the generation of appropriate illumination correction functions. The module supports multiple input formats through a configurable interface: users specify their file type (ND2 or TIFF) and associated metadata structure, enabling compatibility with diverse microscopy systems, including Nikon (ND2), Phenix (TIFF), and Squid (TIFF) platforms. For TIFF files, users provide accompanying metadata files. The module also supports multiple-round phenotyping via iterative staining, with each round treated as an independent phenotyping cycle that is processed separately before downstream integration. We note that the current implementation assumes one-to-one tile correspondence within phenotyping rounds and sequencing-by-synthesis cycles, which is satisfied by most standard microscope configurations. Illumination correction is applied to normalize signal intensity across the field of view (Supplementary Fig. 2)11. In our reanalysis of the “Vesuvius” screen, the Preprocess Module processed 2049 ND2 input files, generating 168,165 SBS tiles (across 11 sequencing rounds) and 58,926 phenotype tiles. The Preprocess module establishes a flexible foundation for downstream analysis by transforming heterogeneous microscopy data into standardized, efficiently-structured image tiles. Sequencing-by-synthesis: accurate identification of genetic perturbations Linking cellular phenotypes to their causative genetic perturbations requires the identification of barcodes within cells from in situ sequencing images. The Sequencing-by-synthesis module (Fig. 2a) processes multi-cycle sequencing images through alignment, spot detection, base calling, cell segmentation, and barcode-to-cell mapping, generating detailed quality metrics including read mapping rates and cell assignment statistics at each stage (Fig. 2b). Sequencing-by-synthesis and phenotype modules for analysis of optical pooled screening data. a Workflow of the Sequencing-by-Synthesis module. In situ sequencing images undergo multi-cycle alignment, spot detection, base calling, cell segmentation, and barcode-to-cell mapping, generating a comprehensive output table with cellular barcode assignments and per-base and per-barcode mapping statistics. Scale bar, 50 μm. b Quality assessment metrics for plate 1. Top: well-level heatmap showing the fraction of segmented cells in each well whose detected barcode uniquely maps to a single known barcode in the sgRNA library. Wells with reduced mapping efficiency indicate areas with low in situ sequencing signal quality (red circle). Bottom: distribution of cells per genetic perturbation (x-axis: number of cells detected for each sgRNA target gene; y-axis: number of genes at that detection frequency), indicating library coverage uniformity. c Comparison of spot detection methods on a representative tile. Spotiflow detected more spots per tile than the standard method with a spot intensity threshold of 400 arbitrary fluorescence units (13,334 vs. 9486) but resulted in comparable gene mapping efficiency (the percentage of segmented cells for which a detected barcode was successfully matched to exactly one known sgRNA target gene; 66.9% vs. 66.5%). Data are from n = 8 tiles (one randomly selected per plate). Box plot elements: center line, median; box limits, first and third quartiles; whiskers, 1.5× interquartile range. Individual data points are overlaid. Statistical significance was assessed using two-sided Mann-Whitney U tests for all pairwise comparisons shown with significance brackets in this and subsequent figures: *p < 0.05, **p < 0.01, ***p < 0.001; ns, not significant (p ≥ 0.05). d Phenotype module workflow. Multi-channel fluorescence images undergo illumination correction, channel alignment, segmentation, and feature extraction, quantifying morphological and functional characteristics across nuclear, cytoplasmic, and whole-cell compartments. e Benchmarking of segmentation approaches on a representative tile. Three methods were compared: Cellpose (v3), Cellpose4 (v4), and StarDist. Cellpose (v3) achieved superior segmentation of HeLa cells by qualitative inspection. Segmentation benchmarking was performed on n = 8 tiles (one randomly selected per plate) with consistent results across tiles. Scale bar, 25 μm. f Quality assessment of phenotypic measurements for plate 1. Top: minimum phalloidin intensity within segmented cytoplasmic masks across wells; low minimum values indicate that segmentation boundaries may extend into unstained background, serving as a quality metric for cytoplasmic segmentation accuracy. Bottom: cell count distribution, demonstrating correlation between staining anomalies and decreased segmentation efficiency (circled). Source data are provided as a Source Data file. The module offers two distinct approaches for identifying sequencing spots: a standard method based on classical image processing signal detection and a deep learning-based method using Spotiflow12. In benchmarking analysis across eight randomly selected tiles (one per plate), the standard method detected substantially more initial peaks than Spotiflow (48,133 vs. 13,344 peaks on average per tile), though after quality filtering, the standard method yielded 9486 high-quality reads compared to 13,344 per tile for Spotiflow (Fig. 2c). Critically, read mapping fractions were higher with the standard method (71.8% vs. 57.6%), although the proportion of cells successfully assigned to a single gene target remained comparable between approaches (66.5% vs. 66.9%). Processing with Spotiflow required substantially more time, averaging 40.3 s per tile compared to just 0.87 s with the standard method (Supplementary Fig. 3a). Given the superior mapping rates and dramatically faster processing speed and lower memory usage (Supplementary Fig. 3b), we employed the standard method for our reanalysis of the Vesuvius screen. Following spot detection, base calling corrects for channel cross-talk using configurable normalization strategies. For our analysis, we selected median-based normalization on the basis of its superior correction of A-C crosstalk (Supplementary Fig. 3c). Reads are mapped to a reference barcode library with support for error-tolerant matching and combinatorial barcode schemes used in PerturbView and Zombie in situ sequencing approaches13. A grid search utility evaluates combinations of detection parameters on representative tiles, reporting spot count, mapping rate, and cell assignment rate for each combination, enabling users to identify optimal settings before full-scale processing. In our reanalysis of the Vesuvius dataset, we identified a total of 74,406,684 cells with 79.2% (58,980,065) containing a sgRNA barcode, and 57.0% (42,371,543) successfully assigned to a unique gene target. To validate Brieflow’s generalizability, we applied the SBS module to an independent dataset generated using three known barcode constructs sequenced across three cycles using T7-based chemistry. Brieflow accurately mapped 78.2% of the segmented cells uniquely to a single construct (Supplementary Fig. 3d). PerturbView-style in situ sequencing produces single high-intensity nuclear peaks13 rather than distributed cytoplasmic signal; both characteristics are accommodated through Brieflow’s configurable parameters for peak detection prioritization and compartment-specific read assignment. Together, the Sequencing-by-synthesis module transforms complex multi-cycle imaging data into high-confidence genetic perturbation assignments for each cell. Phenotype: comprehensive cellular feature extraction Capturing the phenotypic impact of genetic perturbations requires the extraction of comprehensive multi-dimensional cellular features from high-content imaging data. To achieve this, the Phenotype module (Fig. 2d) implements a systematic workflow beginning with specimen-specific illumination correction to normalize signal across the field of view, followed by alignment of fluorescence channels to correct for chromatic aberrations, cellular segmentation to identify nuclear and cellular boundaries, and finally feature extraction to quantify morphological and functional characteristics across multiple cellular compartments. Channel alignment is performed using phase cross-correlation for sub-pixel precision, with integrated notebook visualizations enabling manual inspection of registration accuracy. For high-magnification or multi-round imaging contexts where automated alignment may be insufficient, the module supports a two-stage approach combining user-specified coarse offsets with automated sub-pixel refinement. Custom channel-specific offsets can additionally be specified to correct for systematic chromatic aberrations. For cellular segmentation, Brieflow supports multiple deep learning-based approaches: Cellpose314 (cyto3 model), Cellpose415 (CPSAM model, integrating the Segment Anything foundation model), and StarDist16. Segmentation masks are applied across all channels for feature extraction, with no requirement to segment each channel individually. Notebook visualizations based on visual inspection and the proportion of segmented nuclei corresponding to segmented cells enable users to test segmentation approaches on representative tiles before committing to full-scale processing. In comparative benchmarking on one random imaging tile per plate (8 tiles total), Cellpose (cyto3) achieved high retention rates (93.0% nuclei, 93.7% cell retention) with moderate processing time (186 s per tile). Cellpose4 (CPSAM) achieved comparable segmentation but with expectedly longer processing times (1161 s per tile). However, CPSAM offers strong performance for non-traditional morphologies such as elongated RPE1 cells, making it a valuable alternative for challenging cell types. StarDist was most computationally efficient (110 s per tile) but achieved lower retention rates (16.7% nuclei, 35.1% cell retention), with many detected objects lost during downstream reconciliation (Fig. 2e, and Supplementary Fig. 4a–d). Based on these results, we selected Cellpose (cyto3) for the re-analysis of the Vesuvius screen. For feature extraction, Brieflow supports two configurable approaches: a CellProfiler-inspired17 pure Python emulator developed to address integration challenges between Java-based CellProfiler and Brieflow’s Snakemake workflow and a package (cp_measure) developed by the CellProfiler team18. The emulator extracts four feature categories — intensity, texture, shape, and correlation — across nucleus, cell, and cytoplasm compartments; a comprehensive feature dictionary is provided in the Brieflow repository. Benchmarking against cp_measure revealed strong agreement for interpretable intensity and shape features, although features sensitive to local variations — particularly radial distribution features — showed divergence between implementations (Supplementary Fig. 4g, h). These discrepancies likely reflect differences in implementation details for complex multi-step computations rather than fundamental algorithmic differences. Furthermore, the cp_measure implementation required an average processing time of close to 5 hours per tile, which is impractical for high-content screens. In our reanalysis, the Phenotype module segmented 71,629,095 cells with 1651 morphological measurements extracted per cell (Fig. 2f). Together, the Phenotype module converts raw fluorescence images into rich quantitative profiles that capture the morphological consequences of genetic manipulations, informing the functional contribution of each corresponding gene product. Merge: integration of genetic and phenotypic data Connecting a genetic perturbation with its phenotypic effect necessitates spatial registration between different imaging magnifications captured during the OPS workflow. The Merge module integrates data from the Sequencing-by-synthesis and Phenotype modules, creating a unified dataset that links genetic perturbations with their corresponding phenotypic effects (Fig. 3a). Because OPS workflows typically image barcodes and phenotypes at different magnifications and potentially even on different microscopes, direct pixel-to-pixel correspondence cannot be assumed. Instead, the merge process leverages the spatial pattern of cells themselves as fiducial markers: cells occupy consistent physical positions regardless of imaging magnification, so the geometric relationships between neighboring cells can be used to identify corresponding regions across modalities. The implementation uses Delaunay triangulation to create hash-based descriptors of local cell arrangements that can be matched across imaging modalities despite differences in scale and orientation. Merge, aggregate, and cluster modules for integrating genetic and phenotypic data. a Merge workflow integrating phenotypic features with genetic perturbation data. Phenotypic measurements (left) and barcode mapping statistics (right) are combined through triangle hashing (a geometric matching algorithm that identifies congruent cell-position triplets between coordinate systems to compute spatial transformations) for spatial registration, cell merging, and deduplication to resolve multiple matches, producing a unified dataset (bottom) linking genetic perturbations with phenotypic signatures. b Quality assessment of Merge performance for plate 3. Top: heatmaps showing the fraction of SBS cells matching to phenotype cells across all wells. Bottom left: representative snapshot of the merge process for a single tile, showing the overlay of phenotype cell positions (blue) and SBS cell positions (green) plotted in their respective per-tile coordinate systems, with gray lines connecting matched cell pairs across the two coordinate frames. Close spatial proximity of connected pairs confirms successful registration. Bottom right: spatial distribution of cells with minimum marker intensity of zero (red), plotted in phenotype tile coordinates. These cells, which predominantly cluster at tile boundaries due to vignetting and edge-of-field optical artifacts, are deprioritized during deduplication to favor higher-quality measurements from tile centers. c Aggregate workflow for transforming single-cell data into perturbation-level embeddings. The module takes high-dimensional single-cell feature matrices (top), splits cells into subpopulations of interest, removes low-quality measurements and outliers, and standardizes feature distributions using non-targeting controls (cells expressing sgRNAs with no known genomic target, representing unperturbed baseline variation) as the reference distribution, generating perturbation-level statistics (bottom). d Batch effect correction through the Aggregate module. Each violin represents the distribution of measurements across cells in a single well (x-axis: individual wells). Left column: four representative raw features (mean cell DAPI, tubulin, γH2AX, and actin intensity; y-axis: original feature value). Right column: four randomly selected principal components (PCs 654, 114, 25, and 759; y-axis: aligned principal component value), shown to illustrate that batch-correlated variation extends beyond individual features into the PCA embedding space. Before Typical Variation Normalization (TVN; top), systematic well-to-well offsets are visible in both raw features and principal components; after TVN (bottom), batch effects are substantially reduced. Each violin represents one well; n = 46 wells across 8 plates. Violin elements: shape, kernel density estimate; inner lines, median and 25th/75th percentiles. Data are clipped to the 1st–99th percentile per feature per well before plotting. e Cluster module workflow. High-dimensional perturbation embeddings undergo PHATE dimensionality reduction to generate a two-dimensional representation preserving both local and global structure, followed by functional interpretation using reference database annotations, pairwise gene relationship comparisons, and group-level pathway enrichment analysis. f Cluster evaluation showing precision-recall performance across Leiden resolution parameters using the CORUM benchmark. Higher resolutions yield more granular clusters with higher precision but lower recall. Precision is calculated using an adjusted metric that only counts false positives when both genes are annotated members of different CORUM complexes; gene pairs where one or both genes lack CORUM annotation are excluded from the false-positive count, avoiding penalization of potentially real but undocumented interactions. The red circle highlights resolutions 11–15, representing an optimal balance between precision and recall. n = 20 Leiden resolution parameters were evaluated. Source data are provided as a Source Data file. The module supports two registration strategies: a tile-by-tile approach that matches cell patterns between corresponding tiles, and a well-level stitching approach for screens with low cell density that fail to produce a robust cell pattern at the tile level. The tile approach constructs a nine-edge hash directly from each tile, while the stitch method first stitches together the tiles into complete well mosaics using their recorded stage coordinates from microscope metadata. Once registration is performed at the full-well level, the same triangulation-based matching strategy is applied to the complete cell population, as aggregating cells across all tiles provides sufficient points for robust geometric matching in sparsely populated settings. Although the tile approach rapidly generates high-quality matches in most settings — as in this reanalysis — we applied the stitch approach to a challenging, elongated human retinal pigment epithelial-1 cell dataset imaged at higher magnification. The triangle hashing approach collapsed in this setting with a median density of 19 cells per phenotyping tile, recovering only 30 match pairs. When using the stitch approach, we merged 65,480 cells out of 138,007 phenotype cells. In our reanalysis, spatial registration between Sequencing-by-synthesis (74,406,684 cells) and Phenotype (71,629,095 cells) successfully matched and deduplicated 54,683,953 cell pairs, achieving a 76.4% recovery rate for phenotype cells and 70.4% for SBS cells. Of these matched cells, 58% (31,654,068) were successfully assigned to a single gene target, representing the effective mapping rate after spatial registration (Supplementary Fig. 5a). The approximately 25% cell loss during merge primarily reflects intentional filtering of cells at tile boundaries, where overlapping fields of view result in duplicate cell detections. This deduplication step is essential for preventing the same physical cell from contributing multiple observations to downstream analyzes, and the retention rates are consistent with the expected overlap between adjacent tiles in the imaging configuration. For a representative area, the triangle hash approach achieves a 97% match rate in non-overlapping tile regions compared to 53% in overlapping regions, directly confirming that cell loss is concentrated at tile boundaries where duplicate detections are expected rather than reflecting registration failures (Supplementary Fig. 5b, c). Comprehensive visualizations enable rapid assessment of spatial distribution and quality metrics across the plate (Fig. 3b). Thus, the Merge module creates a unified dataset that directly links genetic perturbations to their corresponding phenotypic signatures at single-cell resolution. Classify: cell state classification Analyzing cellular phenotypes across distinct cell cycle states or other biological conditions requires the ability to systematically partition cells into relevant subpopulations. The Classify module provides an integrated framework for training custom machine learning classifiers to categorize cells based on their morphological features, enabling cell-state-specific downstream analyzes. The classification workflow is designed for rapid, interactive classifier development, even for lowly represented cell populations. An interactive labeling interface rapidly displays cellular images across all wells (Supplementary Fig. 6a). For rare cell states—such as mitotic cells, which constitute only ~4% of asynchronous cultures—feature-based gating prioritizes cells with relevant characteristics for labeling, enabling the creation of balanced training datasets without manually screening thousands of majority-class cells. The module then trains and evaluates multiple classifier architectures in parallel, automatically identifying the best-performing model. For the Vesuvius dataset, we manually annotated 800 cells (276 mitotic, 524 interphase) in less than 3 h (Supplementary Fig. 6b). A high-performing model with a low number of features (XGBoost with top 50 features) achieved 96.3% accuracy on held-out training data, with strong performance for both mitotic and interphase cells (Supplementary Fig. 6c). Per-class confidence thresholds are then set interactively (Supplementary Fig. 6d), and classified cells are automatically partitioned into separate analysis tracks for cell-state-specific downstream profiling. The Classify module thus enables biologically informed partitioning of heterogeneous cell populations, ensuring that downstream analyzes capture cell-state-specific phenotypic variation rather than conflating distinct biological programs. Aggregate: from single-cell measurements to perturbation-level profiles OPS approaches obtain phenotype information from thousands of individual cells for each perturbation. Although this generates rich single-cell data that is useful for some applications, aggregating these measurements into perturbation-level profiles provides statistically meaningful biological insights regarding gene function. The Aggregate module transforms single-cell perturbation data into robust gene-level embeddings suitable for downstream analysis (Fig. 3c). This module allows the user to split their dataset into different subpopulations using any classifier as appropriate, before employing a standard cell feature processing pipeline with steps to filter data, align to reduce variation across experimental batches, and aggregate data to the perturbation level19. An optional perturbation scoring step assesses the strength of each cell’s perturbation phenotype, enabling filtering of cells with weak or noisy phenotypic effects before aggregation20. Batch effects arising from plate and well position are addressed through Typical Variation Normalization (TVN), which uses non-targeting control samples to estimate and remove technical variation while preserving biological signal21. Separately, the original interpretable features (such as nuclear size and stain intensities) are aggregated to the perturbation level to provide quantitative metrics that directly inform our understanding of which specific cellular processes drive the observed functional gene clusters. These features can be optionally bootstrapped over non-targeting control distributions to assess statistical significance, identifying which specific morphological measurements are most strongly perturbed for each gene relative to baseline variation. This enables users to move beyond aggregate profile similarity and pinpoint the individual cellular properties that drive a gene’s phenotypic signature. A critical design choice in the aggregation pipeline is the order of dimensionality reduction and aggregation. Brieflow implements single-cell Principal Component Analysis (PCA), computing principal components across all individual cells before aggregating per gene. This approach captures dominant modes of cell-level phenotypic variation, so genes that push cells into similar phenotypic states have similar aggregated profiles even if their molecular mechanisms differ — emphasizing phenotypic convergence that naturally groups functionally related genes. This contrasts with an alternative approach where features are first aggregated per gene before applying PCA, which instead emphasizes morphological specificity and excels at resolving fine-grained sub-pathway distinctions. Brieflow’s phenotypic convergence architecture is optimized for systematic functional discovery — identifying which genes participate in similar biological processes and placing poorly characterized genes into pathway contexts. For the final aggregation step, cellular measurements are collapsed to the perturbation level using median aggregation, which provides robustness to outliers common in single-cell data; alternative methods, including mean and trimmed mean, are also supported. In the reanalysis, TVN batch correction was validated by analysis-of-variance (ANOVA) F-test for batch-associated variation, with p-values increasing from highly significant levels (10−3 for interphase) to non-significant levels (0.52), confirming successful removal of technical variation (Fig. 3d). This produced perturbation-level embeddings for 5299 perturbations in both interphase (30,271,668 cells; median 5732 cells per perturbation) and mitotic (885,424 cells; median 163 cells per perturbation) populations. In sum, the Aggregate module transforms single-cell data into two complementary perturbation-level datasets: standardized principal component embeddings that serve as the foundation for clustering analysis, and aggregated interpretable feature values that enable quantitative assessment of specific cellular phenotypes for downstream biological interpretation. Cluster: identification of functional relationships The Cluster module reveals functional relationships between genes by grouping genetic perturbations with similar phenotypic profiles. This final analytical stage transforms the high-dimensional embeddings from the Aggregate module into interpretable clusters that illuminate biological pathways and functional connections (Fig. 3e). The clustering pipeline employs Potential of Heat-diffusion for Affinity-based Transition Embedding (PHATE)22 for dimensionality reduction, preserving both local and global structure in a two-dimensional representation. The resulting graph weights serve as input to the Leiden community detection algorithm23, which identifies gene communities at a configurable resolution parameter. To objectively evaluate clustering quality, we implemented a benchmarking framework that tests gene clusters against established biological databases, including STRING protein-protein interactions24, CORUM protein complexes25, and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways26. Clustering is performed across a range of resolutions, and automated evaluation functions enable users to select an optimal resolution by balancing cluster granularity against enrichment for known biological relationships (Fig. 3f). Beyond discrete cluster assignments, we quantified phenotypic divergence for each perturbation using the PHATE diffusion potential space, computing the average Euclidean distance from each gene’s potential vector to all non-targeting control vectors. These “potential distance” scores provide a continuous measure of phenotypic effect strength that complements cluster membership. Thus, the Cluster module translates high-dimensional phenotypic profiles into biologically meaningful gene groups that illuminate pathway relationships and functional connections. For our final analysis, we selected k = 12 for interphase clustering (producing 227 clusters) and k = 5 for mitotic clustering (producing 222 clusters). At these resolutions, interphase clusters showed 33.0% CORUM enrichment (75 clusters), 18.1% KEGG enrichment (41 clusters), and a STRING F1 score of 0.098 (precision=0.064, recall=0.203); mitotic clusters showed 5.4% CORUM enrichment (12 clusters), 3.2% KEGG enrichment (7 clusters), and a STRING F1 score of 0.065 (precision=0.043, recall=0.128). Compared to the Funk et al. clustering at their optimal resolutions (k = 10 interphase, k = 9 mitotic; both producing 222 clusters), Brieflow’s interphase clustering showed a higher STRING F1 (0.098 vs. 0.067), CORUM enrichment (33.0% vs. 23.9%), and KEGG enrichment (18.1% vs. 13.5%). Mitotic clustering showed higher STRING F1 (0.065 vs. 0.044) but lower CORUM (5.4% vs. 7.2%) and KEGG enrichment (3.2% vs. 5.4%) (Supplementary Fig. 7a). The improved biological resolution relative to the original Funk et al. analysis likely reflects cumulative contributions across multiple pipeline stages: enhanced segmentation and feature extraction capture subtler morphological distinctions, and single-cell PCA preserves phenotypic variation that gene-level averaging would obscure. Interactive visualization and exploration dashboard Finally, Brieflow incorporates an interactive visualization dashboard for quality control assessment and exploratory analysis (Fig. 4a). The dashboard comprises five displays — Cluster Analysis, Pipeline Stats, Quality Control, Screen Overview, and Analysis Overview — enabling users to explore gene clusters and LLM annotations, review pipeline statistics, assess data integrity, inspect experimental metadata and perturbation libraries, and examine configuration and dependencies. The dashboard additionally integrates with MozzareLLM, a LLM framework for biological interpretation described below. All finalized outputs of the reanalysis can be accessed at our interactive web portal. Interactive visualization dashboard and MozzareLLM for biological interpretation. a Sample of the interactive visualization dashboard. The dashboard comprises five displays: Cluster Analysis for visualization of gene relationships and MozzareLLM output (displayed), Pipeline Stats for processing statistics, Quality Control for data integrity assessment, Screen Overview for experimental metadata, and Analysis Overview for configuration and dependency information. b MozzareLLM workflow for automated biological interpretation of phenotypic clusters. Cluster data (left) is processed through a structured prompt engineering approach using multiple large language model providers (Anthropic Claude, Google Gemini, and OpenAI GPT; center), each run independently, producing pathway annotations with confidence levels and functional classification of constituent genes (right). Each LLM assigns a confidence level (high, medium, or low) based on the proportion of genes in the cluster that correspond to the identified dominant biological pathway. MozzareLLM identifies the dominant biological process for each cluster and categorizes genes as established pathway members, characterized genes with potential novel roles, or uncharacterized genes, assigning prioritization scores based on evidence strength and discovery potential. MozzareLLM: automated biological interpretation and prioritization Although benchmarking against established biological datasets, such as CORUM and KEGG, provides confidence in cluster quality, extracting actionable biological insights from these clusters remains a significant challenge relying on time-consuming and imperfect manual evaluation. LLMs have recently emerged as powerful tools for analyzing gene sets and interpreting biological pathways27, but their application to identifying novel or uncharacterized genes with potential functional significance has been limited. To address this gap, we developed MozzareLLM, a specialized LLM framework for automating the interpretation of gene-phenotype relationships and prioritizing candidates for functional investigation (Fig. 4b). For each cluster, MozzareLLM is provided with UniProt functional annotations for all member genes, identifies the dominant biological pathway, classifies each gene as established, novel-role, or uncharacterized relative to that pathway, and assesses pathway confidence based on the proportion of genes the model categorizes as established pathway members. When applied to the reanalysis of the Vesuvius screen, we ran MozzareLLM separately on the clusters generated from both interphase and mitotic cell populations to capture cell-cycle-specific functional organization. For interphase clusters (n = 227 at k = 12), MozzareLLM identified distinct high-confidence biological processes for 73 clusters (32.2%; Supplementary Data 1), compared with 52 high-confidence clusters (23.4% of 222 clusters) from the original Funk et al. analysis at its optimal resolution (Supplementary Fig. 7b, and Supplementary Data 2). These 73 high-confidence clusters captured 1874 genes compared to Funk et al.’s 1114 genes, a 68% increase in genes assigned to interpretable biological pathways. For mitotic clusters (n = 222 at k = 5), MozzareLLM identified high-confidence processes for 10 clusters (4.5%), capturing 731 genes compared to Funk’s 16 high-confidence mitotic clusters containing 521 genes. To establish that MozzareLLM pathway confidence scores reflect genuine biological signal rather than artifacts of the annotation framework, we generated a shuffled negative control by randomly permuting gene-to-cluster assignments within the Brieflow interphase dataset while preserving cluster size distributions. Running MozzareLLM on this permuted clustering yielded 256 clusters, of which only 1 (0.4%) received a high-confidence pathway annotation; that single cluster contained 6 genes, and subsequent inspection revealed it arose by chance from a random co-assignment of annotated paralogs. The approximately 80-fold enrichment of high-confidence annotations in real versus shuffled data (32.2% vs. 0.4%) demonstrates that MozzareLLM’s confidence assessment is not trivially satisfied by arbitrary gene groupings; coherent pathway annotations arise only when cluster members share genuine functional relationships as reflected in their phenotypic profiles (Supplementary Fig. 7b). Consistent with this, the shuffled clustering showed 0% CORUM and KEGG enrichment (compared to 33.0% and 18.1% for real interphase clusters) and a STRING F1 score of 0.003 (compared to 0.098), confirming that the biological signal captured by Brieflow’s clustering is absent in randomized controls. High-confidence Brieflow clusters contained, on average, 16.6 genes with established functions in the identified pathway, 7.4 genes with established functions but potential novel roles in the identified processes, and 1.7 genes with limited functional annotation — representing candidates that co-cluster with well-characterized pathway components (Supplementary Fig. 7c). Reanalysis with Brieflow recovers mitochondrial clusters A powerful feature of OPS is its ability to report on diverse cellular properties regardless of the markers used, as core cellular processes are intimately interrelated. Despite the use of only four cellular markers — DNA, a DNA damage marker, actin, and microtubules — both the original Funk et al. analysis and our reanalysis detected coherent gene clusters for protein translation, protein degradation, vesicle trafficking, and numerous other cellular processes. A notable absence from both the initial Funk et al. analysis relative to complementary Perturb-seq analyzes28 were gene clusters corresponding to mitochondrial function. Mitochondrial biology provides a demanding test case due to the absence of imaging channels in the Vesuvius screen that directly stain mitochondria, such that any signal must arise from indirect morphological consequences of mitochondrial perturbation. MozzareLLM reanalysis of the Funk et al. clustering identified four interphase clusters containing groups of mitochondrial proteins (Fig. 5B, and Supplementary Data 3), with established-gene fractions ranging from 47.4% to 71.4% and mixed functional content throughout. However, these clusters were missing the vast majority of established mitochondrial components and contained non-mitochondrial genes, forming coherent sub-programs rather than isolated outliers. For example, Funk interphase cluster 149 (19 genes, 52.6% established) co-clusters the small GTPase KRAS and the kinase BRAF with its OXPHOS core, and Funk interphase cluster 147 (19 genes, 47.4% established) embeds an intact actin disassembly module (CAP1, CFL1, WDR1) alongside genuine OXPHOS subunits. Recovery of mitochondrial sub-programs by Brieflow. A PHATE visualization of Brieflow interphase clustering with five high-confidence mitochondrial sub-modules highlighted. Cluster labels in bold indicate sub-modules independently validated as coherent mitochondrial functional units in MitoCarta3.0 (a curated inventory of mammalian mitochondrial proteins); labels in regular weight indicate mitochondrial-related modules identified by MozzareLLM that do not correspond to predefined MitoCarta3.0 groupings. In gene callout tables, text color indicates gene classification: bold text for established pathway members, regular colored text for genes with characterized functions but potential novel roles in the indicated pathway, and italic text for uncharacterized genes. B Corresponding PHATE visualization from the Funk et al. analysis with four mitochondrial-containing clusters highlighted. Fewer mitochondrial sub-modules were resolved in the Funk et al. analysis, consistent with its coarser clustering resolution. Gene callout tables use the same text conventions as in A. C Pearson correlation matrix of PCA-based perturbation profiles for all mitochondrial-related genes, with genes organized by Brieflow cluster assignment. The upper triangle shows pairwise correlations from the Brieflow pipeline; the lower triangle shows correlations from the Funk et al. pipeline. Both triangles are computed from PCA-based profiles (single-cell feature vectors projected into PCA space, then aggregated per perturbation via median). Each triangle is independently normalized to its own 99th-percentile absolute value and clipped to [−1, 1], preventing one pipeline’s correlation magnitudes from dominating the color scale (per-triangle normalization). Gene labels are colored by Brieflow cluster membership as in A; bold labels indicate established pathway genes, italic labels indicate genes with novel or uncharacterized roles. D Pearson correlation matrix of raw feature-based perturbation profiles for the same genes, with genes organized by Funk et al. cluster assignment. The upper triangle shows pairwise correlations from the Brieflow pipeline; the lower triangle shows correlations from the Funk et al. pipeline. Both triangles are computed from raw feature-based profiles (original measured features aggregated per perturbation via median without dimensionality reduction), with the same per-triangle normalization as in C. Gene labels are colored by Funk et al. cluster membership as in B, with the same bold/italic convention. Comparison of C and D illustrates how PCA-based profiles (C) and raw feature profiles (D) capture different aspects of phenotypic similarity between mitochondrial gene perturbations. In contrast, application of MozzareLLM to Brieflow clusters highlights five distinct high-confidence mitochondrial sub-modules (Fig. 5A, and Supplementary Data 3), encompassing 119 genes of which 81 (68.1%) are MitoCarta-annotated29—nearly twice the number of mitochondrial genes recovered by the Funk et al. analysis at higher purity. Brieflow interphase cluster 22 (48 genes, 79.2% established) covers OXPHOS assembly and mitochondrial translation; cluster 69 (24 genes, 87.5% established) concentrates mitochondrial ribosome proteins and gene expression factors including TFAM; clusters 121 and 129 resolve ATP synthase assembly and mitochondrial translation machinery, respectively; and cluster 160 captures mitochondrial membrane organization (Fig. 5C). The few novel-role genes present appear as isolated individuals with plausible mitochondrial connections—such as the ISR effector DDIT3 and the mitophagy chaperone HSPB8—rather than as coherent non-mitochondrial sub-programs. Together, clusters 22 and 69 alone account for 24 distinct MRPL/MRPS mitochondrial ribosome subunits distributed across just two clusters, compared with the scattering of mitoribosomal proteins across multiple diverse clusters in the Funk et al. analysis. Most distinctively, cluster 160 captures a membrane organization sub-program — MFN2 (fusion), DNAJC11 (cristae architecture), PISD (phospholipid synthesis), VPS13D (ER–mitochondria contacts) — with none of its established genes appearing in any Funk mitochondrial cluster, representing sub-compartment specificity that the original analysis did not recover (Fig. 5C). Pairwise gene correlations computed from raw interpretable features (Fig. 5D) show relatively similar structure between the Brieflow (upper triangle) and Funk et al. (lower triangle) pipelines, likely because the large number of cells per perturbation produces stable feature medians. In contrast, correlations computed from PCA-based embeddings, which are used for clustering in both approaches (Fig. 5C), reveal more distinct differences between the two pipelines, suggesting that computing principal components at the single-cell level before aggregation captures phenotypic variation that direct feature aggregation does not. To assess whether Brieflow retains the principal biological findings of Funk et al., we tracked all 57 clusters highlighted in that study as biologically interpretable, 15 of which were discussed individually in the text as exemplary functional modules and the remaining 42 were organized into pathway groups and presented as heatmaps or annotated on PHATE plots. Using MozzareLLM, we independently assigned high-confidence pathway annotations to 50 of 57 (87.7%) of the original Funk clusters. In the Brieflow clustering, 77.2% of these curated clusters were at least partially preserved (≥50% of genes co-occurring in a single Brieflow cluster), with 26 (45.6%) well-preserved at a mean retention of 91.8% (Supplementary Data 4). Among the 15 highlighted clusters, 7 were well-preserved, and the remaining 8 showed partial preservation or fragmentation (Supplementary Fig. 8a); however, 82% of the redistributed genes from these 8 partially preserved clusters were assigned to high-confidence Brieflow modules (Supplementary Fig. 8b), indicating that gene redistribution typically reflects reassignment into more specific modules rather than loss of signal. Beyond recapitulating the findings of Funk et al., we next asked whether Brieflow resolves functional programs not captured by the original analysis. We computed bidirectional Jaccard similarity indices between all high-confidence clusters across both interphase and mitotic populations, defining low concordance as Jaccard <0.15 (Supplementary Data 5). This revealed a pronounced asymmetry: 29 of 83 high-confidence Brieflow clusters (35%) had no adequate match in the Funk clustering, compared with 14 of 68 Funk clusters (21%) that were reciprocally unmatched (Supplementary Fig. 9a, b). Notably, 297 genes in Brieflow-unique clusters were flagged as novel-role or uncharacterized, representing candidates whose functional assignments are potential targets for follow-up investigation. Discussion Although technologies such as Perturb-seq30 have benefited from extensive methodological refinement through the widespread use of gene expression data, OPS represents an emerging data type with substantial room for methodological innovation compared with more established functional genomics approaches. The unique analytical challenges inherent to OPS—including multi-modal data integration across different imaging modalities, complex image processing requirements, and the extraction of biologically meaningful features from high-dimensional cellular measurements—necessitate specialized computational approaches that continue to evolve rapidly. These challenges present significant opportunities for innovation in computational methods tailored specifically to the complexities of image-based functional genomics. Brieflow addresses key challenges in the analysis of OPS data, addressing key bottlenecks that have limited the widespread adoption of this powerful technique. Although experimental OPS platforms have become increasingly accessible, challenges in analysis represent a significant barrier to entry for biologists seeking to adopt OPS, underscoring the need for accessible, integrated computational tools that can bridge the gap between experimental accessibility and analytical complexity. Although existing OPS analysis approaches have made valuable contributions, they often focus on specific analytical challenges rather than providing integrated and validated end-to-end workflows, with some relying on specialized software environments that limit scalability and broader adoption (Table 1). By providing a unified framework for end-to-end analysis, Brieflow eliminates the need for fragmented pipelines and format conversions, substantially reducing the technical barriers to entry for researchers interested in applying OPS to their biological questions, while its modular architecture allows users to adapt individual components to their specific experimental contexts. Our reanalysis of the large-scale “Vesuvius” optical pooled screen demonstrates Brieflow’s capabilities in handling large-scale datasets to reveal enhanced biological insights while maintaining high data quality and analytical rigor. Table 1Comparison of existing optical pooled screening analysis pipelines and their capabilitiesFeaturePERISCOPE3OpticalPooledScreens9PerturbView13CellPaint-POSH7BrieflowEnd-to-End Integration✗✗✗✗✓Unified Framework✗✗✗✗✓Analysis DocumentationComprehensiveMinimalMinimalMinimalComprehensiveWorkflow ManagementManualSnakemakeSnakemakeCustomSnakemakeModular Architecture✗✗✗✗✓Community Extensibility✗✗✗✗✓Primary FocusCellProfiler-based analysisBasic phenotype/SBSTissue spatial alignmentML-based morphological profilingComplete OPS analysisData IntegrationMultiple separate pipelinesFragmented workflowSpecialized use casePre-processed data onlySingle integrated pipelineAnalytical ScopeComplete analysis, no interpretationIncomplete pipelineSpecialized applicationFeature extraction and gene networksComplete analysis and interpretationFeatures are compared across five computational frameworks for OPS data analysis. Periscope relies on CellProfiler-based architecture requiring integration of multiple separate analytical pipelines. OpticalPooledScreens provides basic phenotype and sequencing analysis components but lacks comprehensive integration and documentation for end-to-end workflows. PerturbView focuses specifically on tissue spatial alignment applications with limited broader analytical scope. CellPaint-POSH provides analysis notebooks and self-supervised deep learning models for morphological profiling of pre-processed OPS data, but does not implement raw image processing or an integrated analysis pipeline. Brieflow provides a unified framework encompassing complete analysis from raw data to biological interpretation. Extensive benchmarking of alternative pipelines was not feasible due to limited accessibility and the time-intensive nature of adapting existing frameworks to perform the comprehensive end-to-end analysis demonstrated with Brieflow. Comparison of existing optical pooled screening analysis pipelines and their capabilities Features are compared across five computational frameworks for OPS data analysis. Periscope relies on CellProfiler-based architecture requiring integration of multiple separate analytical pipelines. OpticalPooledScreens provides basic phenotype and sequencing analysis components but lacks comprehensive integration and documentation for end-to-end workflows. PerturbView focuses specifically on tissue spatial alignment applications with limited broader analytical scope. CellPaint-POSH provides analysis notebooks and self-supervised deep learning models for morphological profiling of pre-processed OPS data, but does not implement raw image processing or an integrated analysis pipeline. Brieflow provides a unified framework encompassing complete analysis from raw data to biological interpretation. Extensive benchmarking of alternative pipelines was not feasible due to limited accessibility and the time-intensive nature of adapting existing frameworks to perform the comprehensive end-to-end analysis demonstrated with Brieflow. The standardization and computational reproducibility provided by Brieflow will be particularly important as the adoption of OPS approaches continues. Through Snakemake-based workflow management and comprehensive configuration documentation, the Brieflow pipeline enables researchers to share not only their results but also the exact methodologies used to generate them, enhancing transparency and facilitating collaborative research. The open-source implementation of Brieflow, with its well-documented interfaces and extensible design, provides a foundation for community-driven improvement and customization. The modular architecture enables rapid development of new capabilities—for example, swapping in a custom segmentation method or extending feature extraction to additional cellular structures can be accomplished rapidly while maintaining the pipeline’s organization and validation framework. As with any complex analytical pipeline, certain implementation choices reflect practical constraints that users should consider. Although Brieflow provides a comprehensive framework for OPS analysis, several technical considerations should be noted when applying the pipeline to new experimental contexts. The current implementation assumes one-to-one tile correspondence within phenotyping rounds and sequencing-by-synthesis cycles, a constraint satisfied by most standard microscope configurations. Notably, despite images being acquired on two different microscopes for the SBS and phenotyping workflows in our validation dataset, the merge registration performs robustly, demonstrating that the pipeline can successfully integrate data across imaging platforms when tile layouts are consistent. Registration quality depends on the ability to identify corresponding cellular landmarks across imaging rounds. The default tile-by-tile registration performs robustly under standard screening conditions, but high-magnification screens with sparse cellular populations may not provide sufficient landmarks per tile for robust geometric matching. In these cases, Brieflow’s well-level stitching strategy aggregates cells across tiles to recover sufficient match points, as demonstrated in our retinal pigment epithelial cell dataset. Conversely, the stitching approach introduces greater computational demands during pairwise distance matrix construction for screens with large numbers of cells per well. These trade-offs are being actively addressed through ongoing improvements to Brieflow’s registration algorithms and memory handling. Our CellProfiler-inspired feature extraction, although effective for morphological profiling, reflects a broader challenge in the field: reliable methods to extract single-cell image features in a fully Pythonic format are not widely available. As the field of single-cell image featurization continues to advance with improved deep learning-based approaches and standardized feature representations, Brieflow’s modular architecture is designed to readily incorporate these improvements. Brieflow is in active development and will remain so for the foreseeable future. The pipeline benefits from contributions from an expanding user community, and improvements to individual modules—from segmentation to feature extraction to pathway analysis—will continue to enhance the quality of biological insights derived from optical pooled screens. A central priority is aligning Brieflow with next-generation file formats, particularly OME-Zarr, to improve interoperability with the broader imaging ecosystem and enable more efficient processing of increasingly large datasets. The pipeline’s modular architecture positions it as a central hub for OPS analysis, where advances in component methods can be seamlessly integrated to benefit the entire workflow. Furthermore, the standardized outputs produced through Brieflow constitute valuable resources for the development of machine learning models on functional genomics data. As perturbational datasets capture a broader spectrum of cellular phenotypes, the analysis of multiple OPS experiments processed through a common framework enables the creation of comprehensive training datasets with enhanced generalizability and predictive power. These standardized datasets will significantly advance computational cell modeling efforts, allowing researchers to build increasingly accurate in silico predictions of cellular behaviors that integrate multi-dimensional phenotypic data across diverse genetic perturbations. Finally, the addition of MozzareLLM for automated biological interpretation represents a complementary approach to extracting actionable insights from high-dimensional phenotypic data. By leveraging LLMs specifically oriented towards biological knowledge extraction, this component accelerates the transition from data to hypothesis formation, a critical step in scientific discovery of gene function. Methods System design and implementation architecture Brieflow follows a standardized Snakemake project development template with four principal components: library code containing core analytical functions, Python scripts for individual rules, Snakemake rules defining execution logic, and targets specifying pipeline outputs. This architecture separates core analytical logic from execution scripts, enabling modification of specific components without disrupting the overall pipeline. The data structure employs standardized file formats, including TIFF for image data, TSV and parquet for tabular data, and PNG for evaluation outputs. Consistent naming conventions facilitate automated data organization and rule matching, while clearly separating raw, intermediate, and analysis data. Brieflow-analysis, provided as a separate template repository, incorporates a hierarchical YAML-based configuration system organized in a tree structure that separates experimental parameters from processing logic. Users create a new repository from this template for each screen analysis, with Brieflow integrated as a git submodule to ensure consistent versioning while enabling project-specific customization. The repository includes comprehensive testing procedures to verify correct installation across computing environments. Brieflow implements multi-level parallelization across sample, plate, well, and tile levels, with GPU acceleration available for computationally intensive tasks such as deep learning-based segmentation. The framework includes specialized execution scripts that optimize job submission for different computing environments, processing plates in sequence to avoid overwhelming computational resources and consolidating multiple small tasks into fewer, larger jobs. The system includes configuration profiles for different execution environments that specify resource allocation, job submission parameters, and environment variables, and makes use of the Snakemake plugin catalog for use of Brieflow across different user-specific interfaces. The framework supports interchangeable components through specialized branches designed for different experimental contexts, enabling community contributions in specific areas such as spot calling, segmentation, or feature extraction. Standardized interfaces between modules ensure that modifications to specific components can be implemented without disrupting the overall workflow. Preprocess module The Preprocess module converts raw microscopy files to standardized TIFF format while extracting comprehensive metadata, including spatial information, optical parameters, and experimental identifiers. The module employs a uniform file reading architecture that supports multiple input formats through a configurable interface: users specify their file type (ND2 or TIFF) and associated metadata structure, and the pipeline automatically adapts its reading and parsing logic accordingly. This design enables compatibility with diverse microscopy systems, including Nikon (ND2), Phenix (TIFF), and Squid (TIFF) platforms. The module processes diverse file organizations, including well-based image storage or individual tile-based acquisitions, combined or separated imaging rounds, and single-channel or multi-channel file structures. Brieflow also supports multiple-round phenotyping via iterative staining, with each round treated as an independent phenotyping cycle that is processed and aligned separately before downstream integration. The current implementation assumes one-to-one tile correspondence within phenotyping rounds and sequencing-by-synthesis cycles, which is satisfied by most standard microscope configurations. Configuration parameters are set through interactive notebooks, where users specify their input files and file format type. For ND2 files, metadata is extracted directly from the proprietary format; for TIFF files, users provide accompanying metadata files or configure parsing rules for filename-encoded information (e.g., plate, well, tile, and channel identifiers). This automatically generates a comprehensive file mapping all combinations of plates, wells, tiles, and channels, eliminating the need for hardcoded values in downstream analysis. Illumination correction is implemented through batch-level correction field calculation, computing illumination patterns by averaging multiple fields of view across an experimental batch. Users can specify a sample fraction (ranging from 0−1) to control the proportion of images used for correction field estimation, with our implementation using the complete dataset. After smoothing with a median filter, each image undergoes correction by division with the illumination function in downstream modules. The implementation leverages multi-threading capabilities to accelerate processing of large image datasets. Sequencing by synthesis module The Sequencing-by-synthesis module processes in situ sequencing images through a sequential workflow: alignment of sequencing cycles, illumination correction optimized for sequencing signal detection, application of a Laplacian-of-Gaussian (LoG) filter to enhance spots, and computation of intensity standard deviation to identify consistent signals across cycles. Alignment supports multiple methods, including reference-based alignment using DAPI as a fiduciary marker (if imaged in each round) and signal-based alignment using the mean sequence signal. Alignment is performed using phase cross-correlation with a configurable upsampling factor for sub-pixel precision; our implementation used integer-pixel resolution, which was sufficient. To account for sub-pixel alignment errors, the module applies maximum filtering with a configurable width (set to 3 pixels in our analysis) that dilates sequencing signals slightly, ensuring robust signal capture despite minor registration imperfections. For barcode detection, two distinct approaches are implemented: a standard method based on statistical signal detection and a deep learning-based method using Spotiflow. The standard approach (which we implemented with a peak threshold of 400) first applies a LoG filter (with kernel width σ = 1 pixel) to enhance spots across sequencing channels. After applying this filter to all sequencing cycles, the standard deviation across cycles and mean across channels are computed to identify consistent signals. Local maxima are then identified within a defined neighborhood (default width of 5 pixels) using maximum and minimum filters: at each local maximum, the peak score is calculated as the difference between the maximum and minimum values within the neighborhood. Peaks within the neighborhood width of image borders are excluded to avoid edge artifacts. This approach effectively captures spots that maintain a consistent signal across multiple sequencing cycles while filtering out noise and non-specific background. The Spotiflow method provides an alternative approach, utilizing a pre-trained U-Net model to predict spot locations from a selected sequencing cycle across all four base channels. Each base channel is processed independently, and results are combined while enforcing minimum distance constraints through iterative distance-based filtering to exclude overlapping signals. For benchmarks, Spotiflow was run with a probability threshold of 0.3, using the first cycle for spot detection, and a minimum distance of 1 pixel between detected spots. Base calling implements a statistical approach that considers channel cross-talk and signal-to-noise ratios. Each called base receives a quality score reflecting confidence in the call, calculated as: where Ifirst and Isecond are the intensities of the highest and second-highest channels, respectively. The module corrects for varying base intensities using either median-based or percentile-based normalization (we selected the median-based approach, based on better correction of A-C crosstalk). For median-based correction, a transformation matrix is constructed by: (1) for each channel, identifying spots where that channel has maximum intensity; (2) computing the median intensity profile across all channels for these spots; (3) assembling these profiles into a matrix and normalizing columns to sum to 1; and (4) inverting this matrix to obtain the correction transformation. For percentile-based correction, the 95th percentile of relative intensities (intensity divided by row sum) is used instead of the median to identify high-confidence spots for each channel. Both methods support restriction of correction estimation to reads within cells, thereby excluding aberrant signals. Following base calling, reads are mapped to a reference barcode library to identify genetic perturbations. The mapping process supports both exact matching and error-tolerant matching with configurable edit distance thresholds. For error correction, two complementary approaches are implemented: (1) reference-based correction, which maps reads to the nearest barcode in the reference library within a specified Hamming distance threshold, resolving ambiguous mappings by selecting the barcode with the smallest edit distance or flagging reads with multiple equidistant matches; and (2) quality-weighted correction, which incorporates per-base quality scores to prioritize high-confidence base calls when resolving mismatches. The module also supports combinatorial barcode schemes used in PerturbView and Zombie in situ sequencing approaches, where multiple barcode segments are read and concatenated to form a complete perturbation identifier. For these multi-segment barcodes, each segment is called independently, and segments are then recombined according to user-specified rules to reconstruct the full barcode sequence before mapping to the reference library. Segmentation of cells in SBS images is performed after illumination correction, with support for Cellpose, MicroSAM, StarDist, and Watershed; we employed Cellpose with the cyto3 model (nuclei diameter 9.44 pixels, cell diameter 18.85 pixels) using base C of the last sequencing cycle due to high background signal in other channels. Segmentation masks are passed to base extraction, which integrates intensities at detected peak locations restricted to either the cellular or nuclear compartment as configured — the nuclear compartment option directly supports T7-based and Zombie in situ sequencing methods where amplified DNA barcodes are confined to the nucleus. Barcode-to-cell assignment resolves a single perturbation identity per cell from the full set of mapped reads. A critical parameter governing this step is the barcode prioritization strategy, which can be set to either peak- or count-prioritized calling. Peak-prioritized calling selects the barcode associated with the highest-intensity spot and is the appropriate mode for T7-based and Zombie in situ sequencing chemistries, where each cell produces a single bright amplified locus. Count-prioritized calling instead selects the most frequently observed barcode identity across all reads within a cell and is appropriate for mRNA-based barcoding approaches where multiple distinct transcript species from different loci are detected per cell. The two modes produce mutually exclusive output metrics — peak intensity scores in peak mode, read counts in count mode — and selecting the correct mode for the sequencing chemistry used is essential for accurate perturbation assignment. For our primary analysis, we used count-prioritized calling. The module is designed to accommodate diverse in situ sequencing chemistries through configurable parameters, including spot detection area, peak detection width, and intensity thresholds. To facilitate parameter optimization before full-scale processing, a grid search utility systematically evaluates combinations of detection parameters on a representative subset of tiles, reporting spot count, mapping rate, and cell assignment rate for each combination. To validate the SBS module’s generalizability, we applied it to an independent in situ sequencing dataset with three known barcode constructs sequenced via T7-based chemistry across three cycles. We used the standard spot detection method with a peak threshold of 400, a peak width of 5, peak-prioritized barcode calling, and reads assigned to the nuclear compartment, consistent with T7-based chemistry, where amplified barcodes localize to the nucleus. Phenotype module The Phenotype module extracts cellular measurements through multiple processing steps: specimen-specific illumination correction, alignment of fluorescence channels to correct for chromatic aberrations, and segmentation of nuclear and cellular boundaries. The implementation supports both single-step and multi-step alignment strategies, with options for custom offsets when specific channels require targeted adjustment. In alignment with previous methodologies, phenotype images were first maximum intensity projected to compress z-slices into a single plane. Channel alignment is performed using phase cross-correlation with a configurable upsampling factor (default of 2) for sub-pixel precision. A centered window of the image data can be used for alignment calculation to reduce noise from image edges. The computed offsets are applied via similarity transformations to shift channels into alignment. At high magnifications or when imaging multiple phenotyping rounds with iterative staining protocols, automated alignment may be insufficient due to larger inter-round shifts or differences in sample positioning. For these cases, the module supports a two-stage manual alignment approach: first, users specify coarse-grain offsets (in pixels) to bring rounds into approximate registration based on visual inspection of landmark features; second, the automated phase cross-correlation refinement is applied to achieve precise sub-pixel alignment. This hierarchical strategy enables robust alignment even when initial offsets exceed the capture range of correlation-based methods. Custom channel-specific offsets can also be specified when particular channels exhibit systematic chromatic aberrations that differ from the global alignment solution. For cell segmentation, the module implements a primary workflow to identify nuclear and cellular boundaries, followed by a derived segmentation step that extracts cytoplasmic regions by subtracting nuclear masks from cellular masks. Multiple segmentation approaches are supported, including deep learning-based methods such as Cellpose, MicroSAM, and StarDist, each optimized for different cellular morphologies and imaging conditions. Our implementation used Cellpose with the cyto3 model, with nuclei diameter of 42.68 pixels and cell diameter of 68.06 pixels, and default threshold parameters. Brieflow supports two configurable feature extraction approaches: a CellProfiler-inspired pure Python emulator developed as part of Brieflow, and the cp_measure package, a pure Python CellProfiler implementation developed by the CellProfiler team. The Brieflow emulator was developed to address integration challenges between Java-based CellProfiler and Brieflow’s Snakemake-based workflow architecture, and integrates algorithms from scikit-image, mahotas, and SciPy. The emulator extracts four categories of features. Intensity features include integrated intensity, mean, standard deviation, maximum, minimum, edge intensities (using morphological operations with connectivity of 2), mass displacement, quartiles, median, and median absolute deviation. Texture features comprise Haralick features computed via mahotas and PFTAS (Parameter-Free Threshold Adjacency Statistics). Shape features include Zernike moments, Feret diameters, Hu moments, convex hull metrics, eccentricity, and solidity. Correlation and colocalization features include Pearson correlation between channel pairs, least-squares slope, overlap coefficient, Manders’ coefficients (M1 and M2), and rank-weighted colocalization coefficients, computed using Otsu thresholding. For user-specified channels, specialized foci detection and quantification are implemented using a white tophat filter followed by LoG enhancement to enable counting and characterization of subcellular punctate structures, which we configured for the GH2AX channel (index 2). Detected foci are refined using watershed segmentation on the distance transform of the thresholded image. Feature extraction operates at three distinct cellular compartments—nucleus, cell, and cytoplasm—allowing precise characterization at each level. For spatial relationships between cells, the module calculates neighbor metrics, including counts of adjacent cells and percent boundary contact. A complete description of all extracted features and their definitions is provided in the Brieflow repository. Brieflow’s modular architecture is designed to accommodate future integration of improved feature extraction methods as they mature. Merge module The Merge module integrates data from Sequencing by synthesis and Phenotype modules through spatial registration between imaging modalities. The foundation of the merge process is accurate alignment that identifies corresponding fields of view between datasets, leveraging metadata from the Preprocess module. The implementation employs a Delaunay triangulation strategy to create hash-based descriptors for cell patterns, enabling robust matching even with partial cell correspondence. For each tile, a Delaunay triangulation is computed from cell centroid coordinates, requiring at least 4 valid cells per tile. For each triangle in the triangulation (excluding those on the outer boundary), a nine-edge hash descriptor is computed: this includes the three internal edges of the triangle plus six edges connecting to the three neighboring triangles. The edges are ordered by starting from the longest internal edge to ensure consistent orientation. Each hash consists of the 18-dimensional vector of edge displacements (9 edges × 2 coordinates), along with the triangle center computed as the mean of vertex coordinates. The nine-edge hash method extracts vector displacements for edges connected to each triangle, creating a distinctive signature that can be matched across imaging modalities despite differences in magnification and optical characteristics. The merge process uses a multi-step alignment algorithm that iteratively refines the transformation model as more matching points are identified. In the tile-by-tile approach, initial alignment begins with a small set of candidate tile-site pairs, either specified by the user or automatically determined. Our implementation used six predefined tile pairs, providing starting points for the iterative alignment. For each candidate pair, the algorithm evaluates the match by: (1) computing normalized edge vectors (dividing by magnitude); (2) finding nearest neighbors between vector sets using squared Euclidean distance; (3) filtering triangle matches below a distance threshold (default 0.3); (4) fitting a RANSAC regressor to the matched triangle centers to estimate the affine transformation; and (5) computing a match score as the proportion of transformed triangle centers within a point threshold (default 2 pixels) of their nearest neighbors. To control alignment quality, the pipeline offers configurable parameters for determinant range, score threshold, and distance threshold. Our implementation used a determinant range of [0.06, 0.065], ensuring the scaling factor remained within acceptable bounds. The matching score threshold was set to 0.1, requiring a minimum level of triangle pattern overlap to consider a match valid. For cell-level matching, a distance threshold of 2 pixels controlled the maximum distance between transformed coordinates for cells to be considered matches. The iterative alignment proceeds by: (1) filtering matches by determinant range and score threshold; (2) using RANSAC regression on matched tile-site coordinates to predict additional candidate pairs; (3) evaluating new candidates in parallel batches (default batch size of 180); and (4) repeating until no new matches are found. This approach progressively expands from initial known matches to identify all valid tile-site correspondences. With spatial registration established, individual cells between modalities are linked based on their spatial coordinates, applying the optimized transformation parameters. The matching process incorporates distance-based thresholds to filter low-confidence associations. A deduplication pipeline resolves cases where multiple cells from one modality match a single cell in the other modality, applying a hierarchical strategy that first selects the best Sequencing by synthesis match for each phenotype cell (based on whether a cell has a known barcode), then selects the best phenotype match for each remaining Sequencing by synthesis cell (based on the highest minimum channel intensity). Quality assessment evaluates alignment quality using determinant range checks to verify scaling consistency and scoring parameters to assess matching accuracy. After cell matching, match rates for both modalities are calculated, reporting the percentage of cells successfully integrated into the final dataset. For experimental conditions where cell density is low, or tiles contain insufficient cells for robust nine-edge hash construction (which requires at least 4 cells per tile for Delaunay triangulation), the module implements an alternative well-level stitching approach. In this method, individual tiles from each imaging modality are first stitched together into complete well mosaics using their recorded stage coordinates from microscope metadata. The stitching process assembles tiles into complete, well-mosaics according to their recorded stage positions, with optional blending at tile boundaries to reduce edge artifacts. Once both modalities are reconstructed as full-well images, registration is performed at the well level rather than tile level, using the same triangulation-based matching strategy but applied to the complete cell population within each well. This approach aggregates cells across all tiles, providing sufficient points for robust geometric matching even when individual tiles are sparsely populated. The well-level alignment transformation is then applied to map cell coordinates between modalities. This stitching-based strategy is particularly valuable for screens with low cell confluence, small tile sizes, or high cell loss during sample preparation, where tile-by-tile matching would otherwise fail due to inadequate cell numbers for triangulation. Classify module The Classify module implements a complete machine learning pipeline for training and applying custom cell classifiers. The implementation is organized into four main components: labeling, training, calibration, and application. The interactive labeling system presents cellular images in batches, with each batch containing a mixture of in-gate cells (prioritized by user-specified feature thresholds) and out-of-gate cells (randomly sampled for diversity). For our implementation, we configured the interface to display 25 cells per batch with 1 out-of-gate example, using DAPI and Tubulin channels for visualization. Feature gating prioritized cells with high nuclear DAPI median absolute deviation (90th percentile and above) to efficiently identify mitotic cells with condensed chromatin. The labeling interface supports incremental dataset expansion, loading previously labeled examples and filtering them from the candidate pool to avoid redundant annotations. Users can optionally enable relabeling mode to revisit and modify existing annotations. Labeled cells are continuously saved to a checkpoint file to prevent data loss during annotation sessions, with final training datasets exported as parquet files containing both morphological features and class assignments. For each training run, multiple model architectures are trained in parallel with different configurations. Our implementation trained six XGBoost-based classifiers with no feature scaling (tree-based models do not benefit from scaling) and varying numbers of selected features: no feature selection (all 1633 features), top 100 features selected by mutual information, and top 50 features. We also trained linear baselines (logistic regression with StandardScaler) and ensemble models (random forest) for comparison. Feature filtering removes metadata columns (plate, well, tile, cell identifiers, spatial coordinates, and barcode assignments) before training, retaining only morphological measurements. For our mitotic/interphase classifier, features were extracted across all four imaging channels (DAPI, Tubulin, yH2AX, Actin) for nucleus, cell, and cytoplasm compartments, yielding 1633 features after metadata filtering. The training pipeline implements stratified fivefold cross-validation to estimate generalization performance. For each model configuration, the pipeline splits data into training (80%) and test (20%) sets with stratified sampling to preserve class balance, fits the specified scaler on training data if applicable, applies optional feature selection using SelectKBest with mutual information criterion, trains the classifier on transformed training data, evaluates on held-out test data to compute accuracy, precision, recall, and F1 scores for each class, generates confusion matrices and per-class evaluation plots, and serializes the complete pipeline (scaler plus feature selector plus classifier) using dill for deployment. Feature importance analysis is performed for tree-based models, with importance scores visualized as horizontal bar plots showing the top 20 contributing features. The module optionally supports post-hoc confidence calibration to improve the reliability of predicted probabilities. Two calibration methods are implemented: isotonic regression (recommended for datasets with more than 100 examples) and sigmoid calibration (Platt scaling, suitable for small datasets). Calibration fits a monotonic transformation that maps raw classifier confidence scores to calibrated probabilities that better reflect true classification accuracy. For calibration, a separate labeled dataset (or the training dataset if no calibration set is provided) is used to fit the calibration function. The calibrated classifier is then applied to test data, and calibration quality is assessed by comparing confidence score distributions before and after calibration. In our analysis, we did not apply confidence calibration as the raw XGBoost confidence scores exhibited good agreement with classification accuracy. Trained classifiers are applied to full datasets through the CellClassifier class, which loads the serialized model pipeline and applies it to feature matrices extracted from phenotype or merge data. The classifier assigns each cell a predicted class label (1-indexed integer corresponding to the class list order) and a confidence score (probability of the predicted class). Per-class confidence thresholds are implemented with two filtering modes: “exclude” mode removes cells with confidence below threshold from downstream analysis, while “reassign” mode tests low-confidence cells against alternative class thresholds and reassigns them if they exceed another class’s threshold, otherwise excluding them. This configuration enables asymmetric filtering strategies where minority classes, such as mitotic cells, are held to stricter standards while majority classes, such as interphase cells, use more permissive thresholds to maximize cell retention. The rankline UI provides an interactive tool for empirically determining appropriate confidence thresholds. Cells are displayed in a scrollable grid sorted by prediction confidence (descending order for each class), with actual images shown alongside predicted class and confidence score. Users can visually inspect high-confidence and low-confidence predictions to identify the confidence level at which classification errors become frequent, then set thresholds accordingly. The UI supports filtering by class, minimum confidence difference to focus on ambiguous cases, and direct navigation to specific confidence ranges. Classified cells are automatically partitioned by the Aggregate module based on their assigned class labels. Each class is processed independently through the aggregation pipeline, generating separate gene-level embeddings for each cell state. This enables cell-state-specific phenotypic analyzes, such as separate clustering of interphase and mitotic cell populations to identify genes with distinct perturbation effects in different cell cycle stages. For our analysis, we configured the classifier to create two subpopulations (mitotic and interphase), which were independently aggregated and clustered using the parameters optimized for each cell state. Aggregate module The Aggregate module transforms single-cell perturbation data into robust gene-level embeddings through a systematic workflow that enables analysis of specific cellular subpopulations. It accepts classified cell populations from the Classify module and can further subdivide datasets by channel combinations of interest, enabling analysis of phenotypic effects specific to particular cellular structures or markers. The filter step implements a comprehensive strategy addressing common data quality issues. Query-based filtering allows removal of specific cell subsets based on configurable criteria; our implementation filtered for cells with a single gene mapping. Perturbation filtering restricts analysis to cells with assigned genetic perturbations, while missing value filtering drops columns and rows with a configurable proportion of missing data (we used thresholds of 0.05 for columns and 0.01 for rows). The module offers optional imputation of remaining missing values using K-nearest neighbors with k = 5, implemented in batches (default 1000 rows per batch with 10000 sampled complete cases) to handle large datasets efficiently (enabled in our implementation). The Local Outlier Factor algorithm identifies outlier cells with anomalous marker patterns for removal, using mean channel intensities as input features, with a configurable contamination parameter (set to 0.01). The module optionally implements per-cell perturbation scoring to assess the strength of each cell’s perturbation phenotype, which is particularly valuable for perturbations with heterogeneous penetrance or for identifying off-target effects. For each gene, the algorithm samples an equal number of non-targeting control cells matched to the perturbed cell count, centers and scales features on controls within each plate-well batch, selects the top 200 differential features using ANOVA F-test, trains a logistic regression classifier (L2 regularization, balanced class weights, max 2000 iterations) to distinguish perturbed from control cells, generates out-of-fold probability predictions via stratified k-fold cross-validation, and computes the ROC AUC as a measure of overall perturbation strength. The resulting per-cell perturbation scores (probability of being perturbed) and gene-level AUC metrics enable fine-grained quality control. Cells with perturbation scores below a specified threshold can be filtered before aggregation, removing cells with weak or noisy phenotypic effects that would dilute gene-level profiles. Genes with insufficient cell counts (less than 100 cells by default) receive NaN scores and are retained without filtering. The perturbation scoring implementation processes genes in parallel batches to manage memory usage, with each batch independently computing scores before updating the main dataset. For our analysis, we did not apply perturbation score filtering so as to more tightly compare with the previous analysis, retaining all cells with assigned gene mappings. However, when conducting larger genome-wide screens, this becomes a tractable way to subset to perturbations that result in a real morphological deviation from control cells. Batch alignment addresses technical variation by first organizing samples by experimental batch (defined in our case by plate and well identifiers), then applying PCA to reduce dimensionality while preserving variation. For large datasets exceeding 50,000 cells, PCA is first fit on a subsample of 50,000 cells to determine the number of components, then the full dataset is transformed. Users can configure either the target variance to preserve (we used 0.99) or a specific component count. The critical TVN step leverages non-targeting control samples to standardize data through a multi-step process: (1) center and scale all data based on control sample mean and standard deviation; (2) fit PCA on control samples and transform all data to this space; (3) center and scale again based on controls (optionally per batch); (4) for each batch, compute the source covariance matrix from batch controls, then apply the transformation Xaligned=X⋅Σsource−1/2⋅Σtarget1/2, where Σtarget is the global control covariance. The matrix fractional powers are computed using scipy’s fractional_matrix_power function. A small regularization term (0.5 × identity matrix) is added to covariance matrices to ensure numerical stability. This generates globally representative principal components at the single-cell level that capture essential morphological feature relationships while removing technical variation. The final aggregation stage combines cellular measurements into perturbation-level representations through two complementary approaches. First, batch-normalized principal components are aggregated to the perturbation level to create embeddings used for clustering analysis that reveals functional relationships between genetic manipulations. Second, the original interpretable features (such as nuclear size and stain intensities) are aggregated to the perturbation level to provide quantitative metrics that directly inform understanding of which specific cellular processes drive observed functional clusters. The aggregation supports multiple statistical methods to accommodate different analytical objectives: mean aggregation provides the arithmetic average of cellular measurements, appropriate when the goal is to capture the central tendency of normally distributed features; median aggregation offers robustness to outliers and skewed distributions, which we selected for our analysis; and trimmed mean aggregation excludes extreme values before averaging, providing a balance between sensitivity and robustness. For experiments with variable cell counts per perturbation, weighted aggregation schemes can incorporate cell count as a weighting factor. To assess the stability and significance of aggregated features, the module implements bootstrapping at the feature level. For each perturbation, cells are resampled with replacement across multiple iterations, and aggregation statistics are computed for each resample. This generates empirical distributions for each aggregated feature, from which confidence intervals and standard errors can be derived. The bootstrapped distributions enable statistical comparison between perturbations, identification of features with high measurement uncertainty, and assessment of whether observed differences exceed sampling variability. Bootstrap-derived p-values can be computed by comparing each perturbation’s feature distribution against the non-targeting control distribution, with multiple testing correction applied across features and perturbations. Quality assessment tools are integrated throughout the module, generating comprehensive reports of missing value patterns and feature distributions before and after batch correction. The implementation leverages efficient data structures and lazy evaluation to handle large datasets, with support for parallel processing and memory optimization for high-performance computing environments. Cluster module The Cluster module identifies functional relationships between genes by grouping genetic perturbations with similar phenotypic profiles. This final analytical stage transforms the high-dimensional embeddings from the Aggregate module into interpretable clusters that reveal biological pathways and functional connections. At the core of the Cluster module is a dimensionality reduction and clustering pipeline built on PHATE and Leiden community detection. PHATE is run with k = 10 nearest neighbors and a configurable distance metric, producing both a 2D embedding and a diffusion potential matrix that captures the underlying manifold structure. PHATE preserves both local and global structure while reducing high-dimensional data to a visually interpretable two-dimensional representation. This approach leverages a diffusion-based metric that captures continuous phenotypic transitions better than linear methods like PCA or nonlinear methods like t-SNE. The PHATE graph weights serve as input to the Leiden algorithm. Before clustering, the weight matrix is symmetrized by averaging with its transpose: Wsymmetric=W+WT/2. This symmetric matrix is then converted to an undirected weighted graph, and the Leiden algorithm is applied using the RBConfigurationVertexPartition method with a configurable resolution parameter controlling cluster granularity. The implementation uses the leidenalg library with n_iterations = −1 for convergence. To objectively evaluate clustering quality, the module implements a comprehensive benchmarking framework that tests gene clusters against established biological databases. The evaluation employs two complementary approaches: pairwise gene relationship analysis using STRING protein-protein interactions, and group-level enrichment analysis using CORUM protein complexes and KEGG pathways. For pairwise evaluation, the system identifies true positive interactions when genes from known interacting pairs appear in the same cluster, while false positives and negatives are determined by mismatched clustering relative to known interactions. This yields standard precision and recall metrics, with an optional adjusted precision calculation that accounts for incomplete knowledge in reference databases. For group-level evaluation, the module applies Fisher’s exact test with Benjamini–Hochberg correction to assess enrichment of known gene complexes or pathways within each cluster. The system filters reference complexes to ensure robustness, requiring at least three genes present in the dataset and at least two-thirds of complex components represented. An additional filter removes larger complexes that share more than 10% of gene pairs with smaller complexes, preventing redundant enrichment calls. The pipeline performs parallel analysis against randomly shuffled controls to establish baseline performance and confirm that identified patterns represent true biological signal rather than technical artifacts. Beyond clustering, the module computes a phenotypic distance metric for each perturbation relative to non-targeting controls using the PHATE diffusion potential space. For each gene, the average Euclidean distance from its potential vector to all non-targeting control potential vectors is calculated, providing a quantitative measure of phenotypic divergence. These distances are then min-max normalized across all perturbations to yield a 0−1 scaled “potential distance” score, where higher values indicate stronger phenotypic effects. This metric serves as a continuous measure of perturbation strength that complements discrete cluster assignments. To focus analyzes on perturbations with significant phenotypic effects, the module implements filtering based on the per-gene AUC scores computed during the Aggregate module’s perturbation scoring step. Perturbations with AUC values below a user-specified threshold (e.g., 0.6 or 0.7) are excluded from downstream clustering and visualization, restricting analysis to genes that demonstrate distinguishable phenotypes relative to non-targeting controls. This filtering is particularly valuable for large screens where all cells are retained through aggregation without perturbation score-based cell filtering; by removing low-AUC perturbations before clustering, researchers can reduce noise from inactive or poorly penetrant knockouts and focus on genes most likely to yield interpretable biological insights. The Cluster module provides extensive visualization capabilities, generating PHATE embeddings that allow intuitive exploration of gene relationships in two-dimensional space. Cluster assignments are highlighted through color coding, enabling rapid identification of gene communities. When AUC filtering is applied, filtered embeddings display only perturbations exceeding the significance threshold, with non-targeting controls shown for reference. Additional visualizations include cluster size distribution plots to assess clustering granularity, enrichment pie charts showing the proportion of clusters supported by different reference databases, precision-recall curves to identify optimal clustering resolutions, and AUC distribution histograms to characterize the overall phenotypic effect landscape of the screen. These visualization tools facilitate both quality control assessment and biological interpretation of the clustering results. MozzareLLM As part of the Brieflow pipeline, we implemented MozzareLLM, a framework that leverages LLMs to analyze phenotypic clusters and prioritize gene candidates for experimental validation. MozzareLLM employs a structured prompt engineering approach organized into sequential analytical stages. For each cluster, the model is provided with the list of member genes along with their UniProt functional annotations, which supply the primary biological context for reasoning about gene function and pathway membership. The model is first tasked with identifying the dominant biological pathway that explains why the genes in the cluster exhibit similar phenotypic profiles. Pathway confidence is assessed according to stringent criteria based on the proportion of genes that fit the proposed pathway: the model is instructed that high confidence requires greater than 70% of genes with strong literature support, medium confidence requires 50–70%, and low confidence requires 30–50%, with clusters below 30% designated as having no coherent pathway. The model is then tasked with classifying each gene into one of three mutually exclusive categories relative to the identified pathway. Established pathway genes are well-documented members with clear functional roles supported by multiple publications. Uncharacterized genes have minimal to no functional annotation, with limited experimental validation and few publications. Novel role genes have established functions in other pathways but may contribute to the identified pathway in a previously unrecognized way. Classifications are informed by the UniProt annotations provided in the prompt, which the model uses to assess the extent of existing functional characterization for each gene. For both uncharacterized and novel role genes, the model assigns a prioritization score from 1 to 10. Uncharacterized genes receive higher scores when they are virtually unstudied with unknown molecular function, and lower scores when partial evidence for pathway involvement exists. Novel role genes receive higher scores when a compelling rationale exists for a previously unrecognized role with minimal existing literature, and lower scores when existing data already support involvement in the pathway. To select the optimal language model for this task, we evaluated three frontier LLMs—Claude Opus 4.6 (Anthropic), GPT-5.2 (OpenAI), and Gemini 3 Pro Preview (Google)—on their ability to annotate gene clusters from functional genomics screens. Each model was tasked with identifying the dominant biological process for clusters drawn from three benchmark datasets: Optical Pooled Screens (OPS; 7 clusters)9, DepMap genetic dependencies (2 clusters)31, and proteomics co-complex data (2 clusters)32, totaling 11 ground-truth-annotated and published clusters. Claude Opus 4.6 and GPT-5.2 performed equivalently, each correctly identifying the dominant biological process for 10 of 11 clusters (91%), while Gemini 3 Pro Preview matched 9 of 11 (82%). All three models appropriately identified the dominant biological processes in the DepMap and Proteomics benchmarks. Performance differences emerged exclusively in the OPS benchmark, which contains more heterogeneous clusters spanning multiple sub-complexes. Gemini failed to identify the mTOR signaling pathway in a multi-complex cluster (cluster 37), returning “Unknown,” whereas both Claude Opus and GPT-5.2 correctly synthesized the mTOR/Golgi trafficking/Integrator sub-complex architecture. The sole cluster missed by all three models (cluster 197) had a ground-truth annotation of m6A mRNA modification—all models instead identified cell cycle and chromatin regulation as the dominant process, reflecting the greater representation of these genes within the cluster. Based on equivalent performance and platform stability, we selected Claude Opus 4.6 (temperature=0.0) for all cluster annotations in the current analysis. Our implementation supports integration with multiple LLM providers, including OpenAI, Anthropic, and Google, allowing researchers to select the most appropriate model for their analytical needs. The MozzareLLM methodology produces standardized outputs that include the dominant biological process, pathway confidence level, gene classifications with prioritization scores, and rationales explaining why particular genes merit investigation. This structured approach facilitates efficient experimental planning by highlighting the most promising candidates for functional validation, particularly those representing potential novel pathway components or genes with unexpected functions. To verify that high-confidence pathway annotations reflect genuine biological signal rather than spurious pattern recognition, we performed a negative control experiment. We shuffled cluster assignments while preserving cluster size distributions, destroying biological relationships while maintaining the overall structure of the embedding space. MozzareLLM was then applied to these scrambled clusters using identical parameters (Claude Opus 4.6, temperature=0.0). This scrambled validation provides objective evidence that the pathway interpretations require genuine functional relationships rather than arising from random gene associations or LLM hallucination. For each benchmark dataset, the screen context provided to the model was adapted to reflect the specific experimental design of that dataset. In the evaluation of the clusters generated by Brieflow, the prompt additionally included the imaging markers used and the cell population (interphase or mitotic) for each cluster. All comparisons between Funk et al. clusters and Brieflow clusters were performed using identical prompts. HeLa screen generation and analysis The “Vesuvius” HeLa screen dataset comprised a CRISPR-Cas9 screen targeting 5072 fitness-conferring genes with four sgRNAs per gene and 250 non-targeting control sgRNAs. Following library transduction into HeLa cells with integrated doxycycline-inducible Cas9, cells underwent puromycin selection for 4 days. Cas9 expression was then induced with doxycycline for 78 h, identified as the optimal timepoint for maximizing observable phenotypes while minimizing dropout of knockout cells. The dataset consisted of eight 6-well plates with 46 wells containing fixed-cell imaging data capturing four cellular markers (DNA, DNA damage response, actin, and microtubules). More comprehensive details on how the screen was conducted are available in the original publication9. Statistics and reproducibility No statistical method was used to predetermine sample size; the dataset comprises all available cells from the previously published Vesuvius screen9. No data were excluded from the analyzes beyond the quality-control filtering steps described in the Aggregate module (outlier removal via Local Outlier Factor with contamination = 0.01; missing-value filtering at thresholds of 0.05 for columns and 0.01 for rows). The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment, as this study involves computational reanalysis of previously published data, where blinding is not applicable. For benchmarking comparisons (spot detection, segmentation, feature extraction), one tile was randomly selected per plate (n = 8 tiles total) to provide representative sampling across experimental batches, and statistical significance for pairwise comparisons was assessed using two-sided Mann–Whitney U tests, as reported in the figure legends. Clustering quality was evaluated against CORUM, KEGG, and STRING reference databases using Fisher’s exact test with Benjamini–Hochberg correction for multiple comparisons. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Reporting Summary Transparent Peer Review file Source data Source Data Acknowledgements We extend our gratitude to David Feldman for the original implementation of the pipeline. We thank Luke Funk for useful information on how the screen was conducted and in-depth discussions on how it was analyzed, and both Luke Funk and Kuan Chung Su for conducting the HeLa screen. We acknowledge members of the Blainey and Cheeseman laboratories for their valuable discussions regarding pipeline development, and especially Russell Walton for his guidance and for providing the secondary ISS dataset for additional benchmarking. We are particularly grateful to members of the Whitehead Institute Information Technology Core for their assistance with pipeline acceleration and implementation on high-performance computing infrastructure. We thank Charlie Bushman of Sunbeam for his guidance on using Snakemake to build adaptable pipelines for biological data. We also thank Alán F. Munoz and members of the Broad Imaging Platform for their insights regarding CellProfiler implementation. Additionally, we acknowledge Albert Dominguez Mantes for his thoughtful discussions concerning the implementation of Spotiflow. Author contributions M.D. and R.K. conceptualized the work. M.D., R.K., A.K.C.D., A.M., S.J.C., and A.N-U. implemented the software. M.D., I.M.C., and P.C.B. drafted the manuscript. S.L., I.M.C., and P.C.B. provided supervision and resources. All co-authors reviewed and edited the final manuscript. Peer review Peer review information Nature Communications thanks Takamasa Kudo and Manuel Leonetti for their contribution to the peer review of this work. A peer review file is available. Funding This work was supported by grants from the NIH (R35GM126930 to I.M.C., R01HG009283 to P.C.B., R01AI144369 to S.L.), and the Chan Zuckerberg Initiative to I.M.C. and P.C.B. M.D. is supported in part by an NSF GRFP fellowship (000955563). This research was also supported by the Whitehead Innovation Initiative (to I.M.C. and S.L.) and the Massachusetts Life Science Center Data Science Internship Program. A.M. was supported by the MIT Undergraduate Research Opportunities Program. Data availability The raw imaging data, in TIFF format, used in this study were originally generated by Funk et al. and are deposited in BioImage Archive under accession S-BIAD394. The raw nd2 files are now deposited in BioImage Archive under accession S-BIAD3221. A stable interactive interface for exploring the results of the HeLa screen analysis, including visualizations of phenotypic clusters and gene relationships, quality control plots, details on the Brieflow run conducted, and experimental metadata, is publicly accessible [https://screens.wi.mit.edu/aconcagua], from which cluster groupings and feature matrices can be downloaded. Source data are provided with this paper. Code availability All code is publicly available. The core Brieflow pipeline is available via GitHub [https://github.com/cheeseman-lab/brieflow]. The specific implementation for the HeLa OPS dataset analysis described in this paper, including benchmarking experiments between methods, is available [https://github.com/cheeseman-lab/aconcagua-analysis], which used Brieflow v1.4.10. MozzareLLM, designed to scale to different types of gene grouping analyses beyond its integration with Brieflow, is available as a standalone tool [https://github.com/cheeseman-lab/mozzarellm]. Comprehensive documentation for Brieflow, including tutorial videos and detailed implementation instructions, can be accessed at the readthedocs page [https://brieflow.readthedocs.io]. For researchers interested in analyzing optical pooled screening data, we recommend consulting this documentation before initiating a new analysis project based on the Brieflow-analysis template. Any questions, issues, or code bugs should be reported as issues on the Brieflow GitHub [https://github.com/cheeseman-lab/brieflow/issues]. A small test image dataset for validating Brieflow installation and testing pipeline functionality is included in the Brieflow repository [https://github.com/cheeseman-lab/brieflow/tree/main/tests/small_test_analysis]. This dataset provides representative examples of OPS images for both sequencing-by-synthesis and phenotyping outputs. All benchmarking scripts conducted as part of this study are also available [https://github.com/cheeseman-lab/aconcagua-analysis/tree/main/benchmarks]. All repositories are released under the MIT License. Brieflow is archived on Zenodo [10.5281/zenodo.19488678]. The analysis implementation is archived on Zenodo [10.5281/zenodo.19488746]. MozzareLLM is archived on Zenodo [10.5281/zenodo.19488677]. Competing interests P.C.B. is a consultant to or holds equity in 10X Genomics, General Automation Lab Technologies/Isolation Bio, Next Gen Diagnostics, Cache DNA, Concerto Biosciences, Stately Bio, Ramona Optics, Bifrost Biosystems, and Amber Bio. His laboratory has received research funding from Calico Life Sciences, Merck, and Genentech for work related to genetic screening. The remaining authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-73643-7. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Reporting Summary Transparent Peer Review file Source Data Data Availability Statement The raw imaging data, in TIFF format, used in this study were originally generated by Funk et al. and are deposited in BioImage Archive under accession S-BIAD394. The raw nd2 files are now deposited in BioImage Archive under accession S-BIAD3221. A stable interactive interface for exploring the results of the HeLa screen analysis, including visualizations of phenotypic clusters and gene relationships, quality control plots, details on the Brieflow run conducted, and experimental metadata, is publicly accessible [https://screens.wi.mit.edu/aconcagua], from which cluster groupings and feature matrices can be downloaded. Source data are provided with this paper. All code is publicly available. The core Brieflow pipeline is available via GitHub [https://github.com/cheeseman-lab/brieflow]. The specific implementation for the HeLa OPS dataset analysis described in this paper, including benchmarking experiments between methods, is available [https://github.com/cheeseman-lab/aconcagua-analysis], which used Brieflow v1.4.10. MozzareLLM, designed to scale to different types of gene grouping analyses beyond its integration with Brieflow, is available as a standalone tool [https://github.com/cheeseman-lab/mozzarellm]. Comprehensive documentation for Brieflow, including tutorial videos and detailed implementation instructions, can be accessed at the readthedocs page [https://brieflow.readthedocs.io]. For researchers interested in analyzing optical pooled screening data, we recommend consulting this documentation before initiating a new analysis project based on the Brieflow-analysis template. Any questions, issues, or code bugs should be reported as issues on the Brieflow GitHub [https://github.com/cheeseman-lab/brieflow/issues]. A small test image dataset for validating Brieflow installation and testing pipeline functionality is included in the Brieflow repository [https://github.com/cheeseman-lab/brieflow/tree/main/tests/small_test_analysis]. This dataset provides representative examples of OPS images for both sequencing-by-synthesis and phenotyping outputs. All benchmarking scripts conducted as part of this study are also available [https://github.com/cheeseman-lab/aconcagua-analysis/tree/main/benchmarks]. All repositories are released under the MIT License. Brieflow is archived on Zenodo [10.5281/zenodo.19488678]. The analysis implementation is archived on Zenodo [10.5281/zenodo.19488746]. MozzareLLM is archived on Zenodo [10.5281/zenodo.19488677].