GLM-Prior: a genomic language model for transferable sequence-derived priors in gene regulatory network inference Authors: Claudia Skok Gibbs, Angelica Chen, Richard Bonneau, Kyunghyun Cho Published: 2026 ABSTRACT Gene regulatory network inference depends on high-quality prior knowledge, yet curated priors are often incomplete or unavailable across species and cell types. We present GLM-Prior, a genomic language model fine-tuned to predict transcription factor-target gene interactions from nucleotide sequence. We integrate GLM-Prior with PMF-GRN in a dual-stage pipeline that combines sequence-derived priors with single-cell expression data for prior-conditioned GRN inference. Across six cell-line contexts, GLM-Prior performance scales with positive label abundance and TF coverage, showing above-chance agreement with reference networks in well-annotated mammalian settings. Single-species, species-transfer, and multi-species training show that GLM-Prior can construct informative priors across related mammalian species. Compared with accessibility-based priors, GLM-Prior achieves the highest prior performance in four of five mammalian cell lines. These benchmarks show that prior quality largely constrains GRN inference performance, positioning GLM-Prior as a transferable workflow for sequence-derived prior construction when matched experimental assays are unavailable. Introduction Gene regulatory networks (GRNs) map the transcriptional relationships between transcription factors (TFs) and their target genes, providing a framework for understanding cellular function and gene expression control in cells1,2. Constructing useful GRN models requires integrating large-scale sequencing datasets, such as single-cell RNA-seq and ATAC-seq, to infer regulatory interactions that cannot be measured experimentally3. However, the accuracy of GRN inference depends heavily on the quality of the prior knowledge used to guide the inference algorithm4,5. In the context of GRNs, prior knowledge refers to an initial matrix of putative TF-gene regulatory interactions, which provides structural guidance to the inference algorithm by defining which TF-gene pairs are more likely to represent regulatory edges. Recent methods have demonstrated that better prior knowledge, which captures biologically meaningful regulatory signals, improves agreement with reference networks and supports more informative GRN inference, highlighting the critical role of prior knowledge in improving GRN inference quality6–9. Prior knowledge for well-studied species is typically constructed using databases of experimentally validated TF-gene interactions10,11. For less-characterized organisms, prior knowledge is inferred by combining structural genomic data with sequencing information12. The Inferelator 3.0’s Inferelator-Prior performs TF motif enrichment within open chromatin regions to define edges between proximally bound TFs and target genes6. CellOracle’s base GRN draws edges between TFs bound within promoter regions of accessible target genes8. SCENIC’s CisTarget constructs prior knowledge by ranking enriched motifs in accessible regions using genome-wide motif scores and a hidden Markov model to predict TF-gene interactions9. While these methods demonstrate the importance of prior knowledge in GRN inference, they also expose key limitations. Motif annotation remains incomplete13, chromatin accessibility data is noisy and cell-type specific14, and existing methods often fail to capture long-range regulatory interactions15. Further, these existing approaches cannot create prior knowledge matrices which generalize to other species. These limitations underscore the need for benchmarked approaches to constructing transferable prior knowledge matrices, particularly for poorly characterized organisms4,16. A promising direction is to incorporate transformer-based genomic foundation models as components of prior construction workflows. Architectures such as the Nucleotide Transformer17 have demonstrated strong performance across genomic prediction tasks, including chromatin accessibility, gene expression, and transcription factor binding18–21. These models can learn sequence-associated features across large genomic corpora, making them attractive candidates for constructing TF-gene priors when matched regulatory assays are unavailable. When trained or fine-tuned across multiple species, genomic language models may preserve predictive signal across related organisms, supporting a transfer-learning paradigm in which models trained in well-annotated species can be evaluated in less-characterized systems. In this way, genomic language models offer a potential route to expand prior knowledge construction beyond organisms and cell types with existing curated interaction data, while requiring careful benchmarking against available reference networks. In this work, we present GLM-Prior, a sequence-based workflow for constructing TF-gene prior knowledge matrices to support downstream GRN inference. GLM-Prior is built by fine-tuning the 250-million parameter Nucleotide Transformer17, pretrained on the genomes of 850 species, into a sequence classification model for predicting TF-gene regulatory labels from available interaction databases. Unlike motif-based approaches that rely on the proximity of TF motifs to accessible promoter regions, GLM-Prior uses paired nucleotide sequences consisting of TF binding motifs and gene bodies. These sequences are jointly encoded by the transformer and passed through a classification head, which outputs a score for each TF-gene pair. To construct a binary prior knowledge matrix suitable for downstream GRN inference, we apply an optimized classification threshold that converts these scores into putative positive and negative prior edges. Since most TF-gene pairs are unlabeled and the set of true negatives is incomplete, the model is trained using class-weighted loss and balanced sampling to mitigate class imbalance during training. We evaluate whether this sequence-derived prior construction strategy improves agreement with held-out reference networks across species and cell-type contexts. We investigate the generalizability of GLM-Prior by evaluating three training paradigms: (1) single-species models trained and evaluated on the same organism; (2) transfer learning, in which models trained in one species are used to predict regulatory interactions in an unseen species; and (3) multi-species training, where a single model is fine-tuned on data from multiple organisms. Across these settings, performance scales with abundance and diversity of positive class labels and TF coverage in the training data. Human and mouse cell lines with rich label sets yield higher AUPRCs than yeast, where extreme class imbalance and sparse labels limit generalization. Transfer learning between evolutionarily related species (human and mouse) maintains or modestly improves performance relative to single-species models, suggesting that GLM-Prior preserves sequence-associated predictive signal across closely related mammalian species. In contrast, transfer learning between mammals and yeast does not exceed near-chance performance, indicating that this transfer signal does not extend uniformly to more evolutionarily distant organisms under the current benchmark. Multi-species training on human, mouse, and yeast preserves accuracy comparable to the best single-species or transfer models in most cell lines, with only mild trade-offs in some human contexts, and produces a sequence-derived prior with comparable benchmark performance across evaluated contexts, without requiring a separately trained model for each target species. To evaluate the utility of GLM-Prior, we incorporate it into a dual-stage GRN inference pipeline. In the first stage, GLM-Prior generates a prior knowledge matrix of putative TF-gene interactions from nucleotide sequence. In the second stage, this matrix is used to constrain PMF-GRN7, a probabilistic matrix factorization model that infers GRNs from single-cell gene expression data. This framework allows us to isolate and compare the contributions of prior construction and expression-based inference to benchmarked GRN performance against available reference networks. Building on this, we benchmark GLM-Prior against accessibility-based priors (Inferelator-Prior6 and CellOracle’s baseGRN8), and integrate each prior with its downstream GRN inference algorithm (PMF-GRN7, the Inferelator 3.06, and CellOracle8) across yeast, mouse, and human cell lines. GLM-Prior performs at or above chance across the six evaluated contexts, with the clearest gains in mammalian settings with stronger label coverage. In contrast, accessibility-based priors remain competitive in settings dominated by proximal regulatory logic, such as yeast, but show more variable agreement with the corresponding mammalian reference networks. When these priors are passed into downstream GRN inference, GRN inference typically yields modest gains when the prior is already strong, and larger improvements are observed only when the prior is weak or inaccurate. Critically, the identity of the best-performing GRN in each cell line almost always matches the identity of the best prior, and no GRN inference method is uniformly superior across priors or species. Together, these results indicate that prior construction, rather than the choice of GRN inference algorithm, is the primary driver of benchmarked GRN performance. Together, our findings position GLM-Prior as a benchmarked workflow for transferable, sequence-derived prior construction in GRN inference. Across the evaluated contexts, we assess how well sequence-derived priors align with available reference networks and how those priors shape downstream expression-based inference. Prior quality strongly influences benchmarked GRN performance, while downstream GRN methods primarily refine or contextualize the structure supplied by the prior. In this work, we characterize GLM-Prior’s generalization across species and cell-type settings, compare it with accessibility-based prior construction strategies, and dissect how different GRN inference algorithms interact with these priors. Results Dual-stage training pipeline overview We develop a dual-stage training pipeline to evaluate sequence-derived prior construction and prior-conditioned GRN inference by integrating GLM-Prior with probabilistic modeling of single-cell expression data (Fig. 1). The pipeline is designed to decouple the construction of prior knowledge from the expression-driven GRN inference step, allowing each component to be independently optimized and evaluated. In this framework, the first stage generates a sequence-informed prior knowledge matrix using a fine-tuned genomic language model (GLM-Prior), while the second stage uses this matrix as structural input to the probabilistic matrix factorization model (PMF-GRN7) to infer a GRN.Fig. 1Schematic of the dual-stage training pipeline.A GLM-Prior is trained on concatenated transcription factor (TF) binding motifs and gene body sequences labeled as annotated positive interactions or unlabeled negative examples. Input data is split into training and validation sets, where the training set is downsampled and placed into even-class batches using a custom DataLoader. These input sequences are embedded and encoded by the transformer architecture and passed through a classification head to predict the probability of a positive regulatory label. After training, inference is performed on all TF-gene pairs to generate a prior knowledge matrix. B Inferred prior knowledge is passed to PMF-GRN to provide a structural constraint to the probabilistic graphical model during inference. PMF-GRN infers transcription factor activity and a gene regulatory network (GRN) for the input cell-line dataset. The resulting GRN is then evaluated using area under the precision-recall curve (AUPRC) and uncertainty calibration with an independent reference network. Schematic of the dual-stage training pipeline. A GLM-Prior is trained on concatenated transcription factor (TF) binding motifs and gene body sequences labeled as annotated positive interactions or unlabeled negative examples. Input data is split into training and validation sets, where the training set is downsampled and placed into even-class batches using a custom DataLoader. These input sequences are embedded and encoded by the transformer architecture and passed through a classification head to predict the probability of a positive regulatory label. After training, inference is performed on all TF-gene pairs to generate a prior knowledge matrix. B Inferred prior knowledge is passed to PMF-GRN to provide a structural constraint to the probabilistic graphical model during inference. PMF-GRN infers transcription factor activity and a gene regulatory network (GRN) for the input cell-line dataset. The resulting GRN is then evaluated using area under the precision-recall curve (AUPRC) and uncertainty calibration with an independent reference network. Stage 1: Constructing sequence-informed prior knowledge with GLM-Prior In the first stage of the pipeline, we construct a prior knowledge edge matrix representing putative positive regulatory (1) and unlabeled or negative (0) entries between TFs and target genes (Fig. 1A). To generate these predictions, we develop GLM-Prior, a transformer-based sequence classification model built by fine-tuning the 250-million parameter Nucleotide Transformer17 model, which was pretrained on the genomes of 850 different species. The primary goal of GLM-Prior is to use sequence information to score TF-gene pairs and construct an informative prior knowledge matrix for downstream GRN inference. GLM-Prior takes as input paired nucleotide sequences consisting of TF binding motifs and gene body regions, along with labels representing annotated positive interactions and unlabeled negative examples. These sequences are concatenated and encoded into a shared representation using the transformer’s embedding and encoding layers. This representation is passed through a classification layer that outputs the probability of a positive regulatory label. During training, the model evaluates a range of classification thresholds, selecting the threshold that maximizes the F1 score on the validation set. This optimal threshold is used to binarize the predicted probabilities into positive and negative prior entries for the final prior knowledge matrix. Since GRN datasets are often highly imbalanced, with far fewer positive interactions compared to negative interactions, we apply several strategies to mitigate this imbalance during training. Specifically, we apply a downsampling rate to the negative class and use a class-weighted cross-entropy loss, tuning the negative class weight through hyperparameter search. Additionally, to further ensure stable training, we implement a custom DataLoader to create training batches with equal numbers of positive and negative examples, ensuring balanced gradient updates. After training, GLM-Prior performs inference across all provided TF-gene pairs, generating a dense probability prior knowledge matrix. The predicted probabilities are then binarized using the optimal F1 score-derived classification threshold, generating a binary prior knowledge matrix of putative regulatory edges for a given organism. Although we can use the probabilities directly, we binarize the prior knowledge matrix to align with existing methods and enable more direct comparison. We then either use this sequence-based prior against an indendent reference network or pass it into the second stage of the pipeline, enabling modular information transfer from sequence-derived prior construction to expression-based GRN inference. Stage 2: Refining GRN inference from single-cell data with PMF-GRN In the second stage of the pipeline, the prior knowledge matrix predicted by GLM-Prior serves as a structured input to our previously developed model, PMF-GRN7 (Fig. 1B). PMF-GRN leverages this sequence-informed prior knowledge to enhance the inference of GRNs from single-cell gene expression data. It does so by using probabilistic matrix factorization, which decomposes the observed gene expression into latent factors representing transcription factor activity (TFA) and TF-target gene regulatory relationships. Concretely, PMF-GRN models the expression matrix W using the likelihood function p(W∣U, V = A ⊙ B, σobs, d), where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathbb{E}}[W]\approx d\odot U{V}^{\top }$$\end{document}E[W]≈d⊙UV⊤, with per-cell sequencing depth d and observation noise σobs. Here, U is a nonnegative cells-by-TFs matrix of TF activities, and V is a genes-by-TFs matrix, which is further factorized as V = A ⊙ B. A ∈ (0, 1) is defined as the degree of existence of a TF-target gene interaction, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$B\in {\mathbb{R}}$$\end{document}B∈R is its signed effect. Each latent variable, U, A, B, d, and σobs, is assumed independent a priori. In this formulation, the GLM-Prior predicted prior knowledge matrix provides the hyperparameters for the logistic-normal prior on A, anchoring factors to specific TFs and mitigating column-label identifiability inherent to matrix factorization. Using variational inference, PMF-GRN estimates the posterior distribution of all latent variables. The posterior mean of A defines the inferred GRN, while the posterior variance provides an edge-level uncertainty score. These uncertainty estimates enable confidence-weighted interpretation of TF-gene edges and are especially valuable when reference regulatory annotations are incomplete or unavailable. The combination of GLM-Prior and PMF-GRN is motivated by their complementary roles in the workflow. GLM-Prior uses paired motif and gene-body sequence information to construct a sequence-derived scaffold of putative regulatory edges. This scaffold can be generated as a general prior matrix or tailored to specific cell lines or conditions when appropriate training data are available. However, as GLM-Prior has not seen any experimental data, its predictions require refinement using context-specific observations. PMF-GRN provides this refinement by integrating single-cell expression data within a principled generative framework, updating edge probabilities based on observed regulatory patterns while maintaining consistency with the sequence-derived prior. Moreover, like most GRN inference algorithms, PMF-GRN relies on prior knowledge to obtain a meaningful posterior GRN; without an informative prior, factor identifiability issues limit interpretability. GLM-Prior helps mitigate this limitation by providing TF-gene structural information, while PMF-GRN tailors the prior to the data, quantifies uncertainty, and produces context-specific networks. Together, this integration yields GRNs conditioned on both sequence-derived priors and single-cell expression data. GLM-Prior’s generalization scales with data composition across species and cell types We assess GLM-Prior by benchmarking across six cell lines spanning yeast, human, and mouse (Fig. 2). For each organism, we create gene-TF sequence pairs by extracting nucleotide sequences for genes from GTF annotations and TF binding motifs from CisBP22. We then assign regulatory labels to these pairs using available interaction databases: YEASTRACT11 for yeast, and STRING23–26 and TRRUST27,28 for human and mouse. These databases provide curated positive regulatory associations, but do not define an exhaustive set of true negative TF-gene pairs. We therefore treat gene-TF pairs without an annotated interaction as unlabeled negatives for training, recognizing that some sampled negatives may correspond to regulatory interactions that are absent from current databases or active only in specific cellular contexts.Fig. 2GLM-Prior model performance across yeast (orange/yellow), human embryonic stem cells (hESC) (red/pink), HepG2 hepatocellular carcinoma cells (HepG2) (blue), mouse embryonic stem cells (mESC) (green), mouse dendritic cells (mDC) (brown), and mouse hematopoietic stem cells (mHSC) (purple) cell lines.A One-epoch hyperparameter sweep over class weights (CW) and negative-class downsampling rates (DSR), evaluated by F1 score. B Validation metrics during 10-epoch training using the best hyperparameter configuration, trained across three random seeds (42, 43, and 44). Metrics include positive and negative class precision, positive and negative class recall, the optimal classification threshold, the area under the receiver operating characteristic curve (ROC-AUC), and the best F1 score. Bars represent mean validation performance, error bars indicate ±1 standard deviation, and overlaid points indicate individual seed values (n = 3). C Test area under the precision recall curve (AUPRC) on held-out reference networks across three random seeds. Large points represent mean AUPRC, error bars indicate ±1 standard deviation, and smaller points indicate individual seeds (n = 3). Baseline controls are shown as single reference points: light gray represents inference on shuffled sequences, dark gray represents shuffled predictions, and black represents chance performance based on the positive-label rate. D Train and test split composition, shown as the number of genes and transcription factors (TFs). E Class composition, shown as the number of positive and negative labels in train and test sets. Source data are provided as a Source Data file. GLM-Prior model performance across yeast (orange/yellow), human embryonic stem cells (hESC) (red/pink), HepG2 hepatocellular carcinoma cells (HepG2) (blue), mouse embryonic stem cells (mESC) (green), mouse dendritic cells (mDC) (brown), and mouse hematopoietic stem cells (mHSC) (purple) cell lines. A One-epoch hyperparameter sweep over class weights (CW) and negative-class downsampling rates (DSR), evaluated by F1 score. B Validation metrics during 10-epoch training using the best hyperparameter configuration, trained across three random seeds (42, 43, and 44). Metrics include positive and negative class precision, positive and negative class recall, the optimal classification threshold, the area under the receiver operating characteristic curve (ROC-AUC), and the best F1 score. Bars represent mean validation performance, error bars indicate ±1 standard deviation, and overlaid points indicate individual seed values (n = 3). C Test area under the precision recall curve (AUPRC) on held-out reference networks across three random seeds. Large points represent mean AUPRC, error bars indicate ±1 standard deviation, and smaller points indicate individual seeds (n = 3). Baseline controls are shown as single reference points: light gray represents inference on shuffled sequences, dark gray represents shuffled predictions, and black represents chance performance based on the positive-label rate. D Train and test split composition, shown as the number of genes and transcription factors (TFs). E Class composition, shown as the number of positive and negative labels in train and test sets. Source data are provided as a Source Data file. To assess out-of-distribution performance in this supervision setting, we evaluate on held-out test sets constructed from independent reference networks, including the literature-curated reference from29 for yeast, and cell-line-specific BEELINE-derived ChIP-seq reference networks30 for human and mouse. Before training each cell-line-specific model, we ensure strict independence between training and test sets at the level of gene-TF sequence pairs and their labels, eliminating sequence and label leakage. This design ensures that performance reflects the model’s ability to generalize to unseen sequences and correctly predict their interaction labels. To evaluate model performance within each cell line, we implement a multi-stage benchmarking procedure (Fig. 2A–E). In Fig. 2A, we perform a one-epoch hyperparameter sweep over combinations of class weights in the loss function and downsampling rates for the negative class. Negative-class downsampling is necessary to make training computationally tractable given the combinatorial number of possible gene-TF pairs; however, this sampling strategy may exclude unlabeled pairs that correspond to unannotated regulatory interactions, preventing the model from learning from those examples during training. We identify the configuration that yields the highest F1-score and use these parameters for full model training. In Fig. 2B, we train the model for ten epochs using this selected configuration and monitor performance on a held-out validation set. To assess sensitivity to model initialization, we repeat final training across three random seeds and report seed-level variability. Following training, we perform inference separately for each seed-specific model across all gene-TF pairs, including pairs unseen during training, and calculate AUPRC for each resulting prior network against the corresponding held-out reference network labels (Fig. 2C). In addition to the positive-rate chance baseline, we report two negative controls: (i) inference on test sets with gene and TF sequences shuffled to disrupt sequence-label correspondence while preserving class composition, and (ii) shuffled prediction scores on the original test set, which removes ranking signal and should approach the positive-rate baseline. To interpret what drives the resulting performance, we examine the number of genes and TFs represented in the training and test sets for each cell line (Fig. 2D), providing context for possible domain shifts between splits. Finally, we quantify the class composition of positive and negative examples in the training and test sets (Fig. 2E), allowing us to relate generalization behavior to the underlying data balance. Across the six evaluated cell lines, GLM-Prior shows context-dependent agreement with held-out reference networks, with performance varying alongside the size and composition of the training and test data (Fig. 2D, E). In yeast, GLM-Prior achieves near-chance AUPRC across all three random seeds (0.02 for each seed) compared to 0.01 for the baseline controls, despite strong within-training validation metrics (Fig. 2B). This limited out-of-distribution agreement likely reflects the extreme class imbalance in the training data, where only 660 positive interactions are available among 724,179 negatives. Although the held-out reference contains 986 positive labels, the small number of positive training examples and high validation performance suggest that the model may overfit regulatory patterns present in the training data without transferring strongly to held-out gene-TF combinations. In human embryonic stem cells (hESC), GLM-Prior shows a modest above-baseline signal, with AUPRC values ranging from 0.19 to 0.20 across three seeds, compared with 0.15 for chance and shuffled-sequence controls and 0.16 for shuffled predictions. This improvement is small but consistent across random initializations. hESC includes a larger pool of positive training examples (5305) than yeast, and the held-out reference network contains many labeled positives. However, TF coverage drops substantially from 504 TFs in training to 79 TFs in testing, indicating that performance is evaluated over a narrower set of TFs than those available during training. We therefore interpret hESC performance as modest evidence of agreement with the held-out reference network, rather than as strong evidence of broad mechanistic generalization. In HepG2, GLM-Prior achieves higher AUPRC values than in hESC, ranging from 0.29 to 0.30 across seeds, compared with 0.25 for chance and shuffled predictions and 0.24 for shuffled sequences. This performance is supported by a larger number of positive training labels (8087) and a well-represented held-out reference set (56,567 positives). Although the held-out set contains fewer TFs than the training set (53 versus 527), GLM-Prior consistently exceeds the negative controls. These results suggest that the model captures sequence-associated signal that improves ranking of labeled interactions in this cell line, while remaining limited to agreement with the available reference annotations. In mouse embryonic stem cells (mESC), GLM-Prior achieves AUPRC values ranging from 0.22 to 0.24 across seeds, compared with 0.20 for chance and shuffled predictions and 0.19 for shuffled sequences. This represents a modest but reproducible improvement over the controls. The training set includes 3129 positives among 476,260 negatives, while the held-out reference set is substantially more balanced (65,933 positives and 271,015 negatives) and covers 5711 genes unseen during training. These results indicate reproducible above-baseline agreement with the held-out reference network across unseen genes, with the modest margin over baseline suggesting that performance is measurable but limited in magnitude. In mouse dendritic cells (mDC), GLM-Prior remains at chance across all three seeds, with an AUPRC of 0.09 for GLM-Prior and all baseline controls. This indicates that the trained sequence model does not provide detectable ranking signal beyond the positive-label rate for this held-out reference. Although the mDC training set contains a large number of positive labels (11, 551), TF coverage drops sharply from 462 TFs in training to 29 TFs in testing. This difference may contribute to poor generalization, but the result primarily indicates that GLM-Prior does not improve agreement with the available mDC reference network under this evaluation design. Finally, mouse hematopoietic stem cells (mHSC) show the strongest above-baseline performance, with GLM-Prior AUPRC values ranging from 0.44 to 0.49 across seeds, compared with 0.37 for chance, shuffled sequences, and shuffled predictions. The observed improvement is larger than in the other mammalian contexts, although seed-to-seed variation is also more apparent. The held-out reference set is large (182,005 positives and 305,939 negatives) and includes 6777 genes and 72 TFs, providing a broad evaluation set. These results indicate that GLM-Prior improves the ranking of labeled interactions in the held-out reference network, producing the strongest above-baseline performance among the evaluated cell lines. Together, these results show that GLM-Prior’s performance depends strongly on the composition and coverage of the available training labels and held-out reference networks. Across seeds, the clearest improvements over negative controls occur in HepG2 and mHSC, while hESC and mESC show modest but reproducible gains, yeast remains close to chance, and mDC shows no detectable improvement. We therefore interpret above-chance performance as evidence that GLM-Prior can learn sequence-associated signals that improve agreement with held-out reference annotations in some contexts. To evaluate whether this signal could instead be attributed to simpler properties of the input data, we performed diagnostic analyses of gene body composition, sequence similarity, gene length, and database annotation features (see Supplementary Information). These diagnostics address distinct potential sources of proxy learning: gene body composition and length test whether predictions are dominated by broad sequence-scale properties, sequence similarity tests whether performance is explained by sequence relatedness among genes or gene-family structure, and annotation-based analyses test whether predictions are driven by uneven representation of genes or TFs in the interaction databases. Across these analyses, GLM-Prior’s predictions are not driven solely by broad compositional differences among gene bodies, sequence relatedness among genes, gene length, or database annotation density. This strengthens the interpretation of the main results by showing that the above-baseline agreement with held-out reference networks cannot be reduced to these simpler properties of the input data. GLM-Prior supports mammalian transfer learning and multi-species prior construction Following our six cell line evaluation of GLM-Prior, we next ask whether a sequence-derived model trained in one species can be used to construct prior knowledge in a species not seen during training (Fig. 3). This is a central motivation for GLM-Prior: existing prior construction strategies often require species- or cell-line-specific experimental data, limiting their use in systems where regulatory annotations, chromatin accessibility profiles, or matched expression datasets are unavailable. A transferable sequence-based model could provide an initial putative TF-gene prior in such settings, which can then be evaluated or refined when reference networks or expression data become available.Fig. 3Transfer learning and multi-species training with GLM-Prior.A Schematic and area under the precision recall curve (AUPRC) results of the species-transfer learning setup, showing three configurations: (i) a model trained jointly on human and mouse data evaluated on yeast, (ii) a mouse-only model evaluated on human cell lines, and (iii) a human-only model evaluated on mouse cell lines. B Schematic and AUPRC results of the multi-species model trained sequentially on human, mouse and yeast data, evaluated across six cell lines. C Comparison of training composition, shown as the number of genes and transcription factors (TFs), for single-species, species-transfer, and multi-species models. D Class composition, shown as the number of positive and negative labels, in the corresponding training sets. E Normalized AUPRC across single-species, species-transfer, and multi-species models for each cell line. Panels A and B contain the following biological illustrations. Human outline: Illustration from NIAID NIH BioArt Source (bioart.niaid.nih.gov/bioart/232). HepG2 cells: Illustration from NIAID NIH BioArt Source (bioart.niaid.nih.gov/bioart/171). mDC cells: Illustration from NIAID NIH BioArt Source (bioart.niaid.nih.gov/bioart/114). mHSC cells: Illustration from NIAID NIH BioArt Source (bioart.niaid.nih.gov/bioart/200). Source data are provided as a Source Data file. Transfer learning and multi-species training with GLM-Prior. A Schematic and area under the precision recall curve (AUPRC) results of the species-transfer learning setup, showing three configurations: (i) a model trained jointly on human and mouse data evaluated on yeast, (ii) a mouse-only model evaluated on human cell lines, and (iii) a human-only model evaluated on mouse cell lines. B Schematic and AUPRC results of the multi-species model trained sequentially on human, mouse and yeast data, evaluated across six cell lines. C Comparison of training composition, shown as the number of genes and transcription factors (TFs), for single-species, species-transfer, and multi-species models. D Class composition, shown as the number of positive and negative labels, in the corresponding training sets. E Normalized AUPRC across single-species, species-transfer, and multi-species models for each cell line. Panels A and B contain the following biological illustrations. Human outline: Illustration from NIAID NIH BioArt Source (bioart.niaid.nih.gov/bioart/232). HepG2 cells: Illustration from NIAID NIH BioArt Source (bioart.niaid.nih.gov/bioart/171). mDC cells: Illustration from NIAID NIH BioArt Source (bioart.niaid.nih.gov/bioart/114). mHSC cells: Illustration from NIAID NIH BioArt Source (bioart.niaid.nih.gov/bioart/200). Source data are provided as a Source Data file. Building on the previous section, which established that generalization depends strongly on data composition, we test whether GLM-Prior preserves predictive signal across species with different degrees of evolutionary relatedness. We perform two complementary experiments: A) transfer learning, where models trained in one species are applied to another, and B) multi-species training, where a single model is trained sequentially on human, mouse, and yeast data. In all cases, held-out evaluation sets are fully disjoint from training data, and overlapping gene and TF identifiers are removed to prevent cross-species leakage. In Fig. 3A, we provide a schematic for the transfer learning setup. Three transfer learning configurations are evaluated: (i) a model trained jointly on human and mouse data, evaluated on yeast; (ii) a mouse-only model evaluated on human cell lines (hESC and HepG2), and (iii) a human-only model evaluated on mouse cell lines (mESC, mDC, and mHSC). Across all six cell lines, we find that transfer learning maintains or modestly improves performance relative to single-species models. In yeast, performance remains near chance (AUPRC = 0.02 vs 0.01), indicating that transfer from mammalian training data does not provide a useful prior for this more distantly related organism under the current evaluation. In contrast, transfer between human and mouse is more encouraging, where the mouse-trained model maintains performance in hESC (AUPRC = 0.19) and improves performance in HepG2 (AUPRC = 0.31), while human-trained models applied to mouse cell lines achieve performance comparable to the corresponding single-species models. These results suggest that GLM-Prior can construct informative sequence-derived priors across closely related mammalian species, while also showing that transfer does not extend uniformly across more evolutionarily distant systems. In Fig. 3B, we present the multi-species model trained sequentially on human, mouse, and yeast data and evaluated on all six cell lines. The multi-species model achieves comparable AUPRCs relative to both single-species and species-transfer settings (yeast = 0.02, hESC = 0.19, HepG2 = 0.29, mESC = 0.23, mDC = 0.09, and mHSC = 0.49). To interpret these trends, Fig. 3C, D summarize the training compositions and normalized AUPRC (Fig. 3E) across single-, transfer, and multi-species setups. Across each cell line, transfer learning consistently expands the number of genes, TFs, and positive labels seen during training. In some settings, such as with HepG2, this broader coverage allows for a modest normalized AUPRC increase (0.08 vs 0.07 single-species). In others, transfer learning maintains single-species performance. In contrast, mHSC shows a slight decrease in performance (0.14 species-transfer vs 0.19 single-species) despite larger training breadth. This drop suggests that the regulatory logic active in mouse HSCs is relatively cell-type- and species-specific and therefore not well represented in the context-agnostic human interaction data used for species-transfer training, limiting the model’s ability to fully recover mHSC-specific edges. Notably, the multi-species model maintains similar performance to the single-species models, with only slight performance drops in hESC and HepG2. This result suggests that multi-species training can produce a broadly applicable sequence-derived prior without requiring a separate model for each target context. Such a model may be useful when the target species or cell type is poorly characterized, although its utility remains dependent on the evolutionary and regulatory similarity between the training data and the target system. Together, these results show that GLM-Prior provides a practical route for constructing prior knowledge in settings where species- or cell-line-specific regulatory data are limited. Transfer between human and mouse maintains or modestly improves agreement with held-out reference networks, supporting the use of GLM-Prior for mammalian transfer-learning scenarios. However, transfer to yeast remains near chance, indicating that this capability is context-dependent and does not establish universal transfer across distant species. Multi-species training further suggests that a single sequence-derived model can provide stable priors across the evaluated contexts, while highlighting the need for additional reference networks to assess transferability across a broader range of organisms and cell types. Integrating GLM-Prior into PMF-GRN supports prior-conditioned GRN refinement To evaluate how sequence-derived prior knowledge interacts with expression-based GRN inference, we next used each cell line-specific GLM-Prior matrix as input to PMF-GRN. Whereas Stage 1 (Section 1) scores TF-gene pairs from sequence to construct a prior matrix, Stage 2 (Section 1) treats these predictions as structural prior knowledge and refines the network using single-cell expression data. For each cell line, we paired GLM-Prior with matched single-cell datasets, including GSE12516231 and GSE14482032 for yeast, and the corresponding expression datasets provided in BEELINE30 for the human and mouse cell lines. This experiment assesses whether expression-based inference improves agreement with held-out reference networks, and how much PMF-GRN modifies the sequence-derived prior during inference. To evaluate how sequence-derived prior knowledge interacts with expression-based GRN inference, we next used each cell line-specific GLM-Prior matrix as input to PMF-GRN. Whereas Stage 1 (Section 1) scores TF-gene pairs from sequence to construct a prior matrix, Stage 2 (Section 1) treats these predictions as structural prior knowledge and refines the network using single-cell expression data. For each cell line, we paired GLM-Prior with matched single-cell datasets, including GSE12516231 and GSE14482032 for yeast, and the corresponding expression datasets provided in BEELINE30 for the human and mouse cell lines. This experiment assesses whether expression-based inference improves agreement with held-out reference networks, and how much PMF-GRN modifies the sequence-derived prior during inference. Across all datasets, integrating GLM-Prior into PMF-GRN results in equal or improved AUPRC relative to the prior alone (Fig. 4A). The magnitude of improvement varies by species and cell line, where performance remains unchanged in yeast (0.02) and hESC (0.19), while gains are observed in HepG2 (0.30 prior → 0.38 GRN), mESC (0.23 prior → 0.27 GRN), and mHSC (0.49 prior → 0.52 GRN). These changes occur despite PMF-GRN modifying only a small fraction of the GLM-Prior edge matrix (Fig. 4B), with added edges ranging from 0.03% in yeast to 0.54% in mHSC, and edge removals remaining minimal across all contexts. The modifications are biased toward edge additions rather than deletions, suggesting that PMF-GRN primarily preserves the GLM-Prior scaffold while adding a limited number of expression-supported edges.Fig. 4Integration of GLM-Prior with PMF-GRN for downstream gene regulatory network (GRN) inference.A Area under the precision recall curve (AUPRC) comparison of GLM-Prior alone (yeast: yellow, human embryonic stem cells: pink, HepG2 hepatocellular carcinoma cells: light blue, mouse embryonic stem cells: light green, mouse dendritic cells: light brown, mouse hematopoietic stem cells: light purple) versus integrated with PMF-GRN (yeast: orange, human embryonic stem cells: red, HepG2 hepatocellular carcinoma cells: dark blue, mouse embryonic stem cells: dark green, mouse dendritic cells: dark brown, mouse hematopoietic stem cells: dark purple) across six cell line contexts. B Edge-flip analysis quantifying changes made to the prior knowledge matrix during PMF-GRN inference. Blue bars indicate prior edges added and red bars indicate prior edges removed during inference. Source data are provided as a Source Data file. Integration of GLM-Prior with PMF-GRN for downstream gene regulatory network (GRN) inference. A Area under the precision recall curve (AUPRC) comparison of GLM-Prior alone (yeast: yellow, human embryonic stem cells: pink, HepG2 hepatocellular carcinoma cells: light blue, mouse embryonic stem cells: light green, mouse dendritic cells: light brown, mouse hematopoietic stem cells: light purple) versus integrated with PMF-GRN (yeast: orange, human embryonic stem cells: red, HepG2 hepatocellular carcinoma cells: dark blue, mouse embryonic stem cells: dark green, mouse dendritic cells: dark brown, mouse hematopoietic stem cells: dark purple) across six cell line contexts. B Edge-flip analysis quantifying changes made to the prior knowledge matrix during PMF-GRN inference. Blue bars indicate prior edges added and red bars indicate prior edges removed during inference. Source data are provided as a Source Data file. The extent of structural change does not strictly track with performance improvement. For example, hESC and mESC exhibit tens of thousands of added or removed edges, yet show minimal differences in AUPRC. This mismatch suggests that many of the changes introduced by PMF-GRN occur outside the held-out evaluation portion of the network, which covers only a subset of TF-gene interactions. Conversely, in yeast, where GLM-Prior is close to chance (AUPRC = 0.02), and the available training interactions are sparse, PMF-GRN does not improve agreement with the held-out reference network despite access to two large single-cell expression datasets. This result suggests that expression-driven inference alone cannot compensate when the input prior has limited alignment with the available reference annotations. These findings indicate that expression-based inference provides prior-dependent refinement rather than a wholesale reconstruction of the network. When the sequence-derived prior demonstrates above-baseline agreement with the held-out reference network, PMF-GRN generally preserves most prior edges and introduces limited expression-supported modifications. In this dual-stage framework, the quality of the GLM-Prior matrix strongly influences the achievable reference-network agreement, while PMF-GRN contributes selective improvements where transcriptomic variation provides additional support. When the prior is near chance, as in yeast, expression-based GRN inference has limited capacity to recover agreement with the available reference network. Comparing prior construction strategies across yeast, mouse, and human cell lines Next, we assess how GLM-Prior compares to widely used accessibility-based prior construction approaches, evaluating performance relative to CellOracle and Inferelator priors across six cell line contexts. Unlike GLM-Prior, which uses sequence-level information to score TF-gene pairs, both CellOracle and Inferelator derive putative interactions from chromatin accessibility and motif enrichment. Importantly, these accessibility-based priors are constructed using ATAC-seq measured in the matched cell line, whereas GLM-Prior is trained on context-agnostic interaction resources (STRING and TRRUST) and evaluated without explicitly conditioning on cell-line-specific chromatin state. This comparison therefore evaluates whether sequence-derived priors can match or exceed accessibility-based priors in agreement with available reference networks across diverse biological settings. Across all six cell lines, GLM-Prior performs at or above chance, exceeding chance performance in five of the six contexts(Fig. 5A). In comparison, Inferelator-Prior performs above chance in two out of six cell lines, and CellOracle’s Prior achieves above-chance performance in three out of six, indicating that prior performance varies substantially across methods and contexts. Yeast represents the main exception, where CellOracle produces the highest-performing prior (AUPRC 0.05 versus 0.03 for Inferelator-Prior and 0.02 for GLM-Prior). This result is consistent with the structure of accessibility-based priors, which are well suited to promoter-proximal, motif-supported regulatory assignments in yeast. In this setting, GLM-Prior is further limited by the small number of positive regulatory labels (660) available for model training (as seen in Fig. 2E).Fig. 5Comparison of prior construction strategies across species.A Area under the precision-recall curve (AUPRC) of GLM-Prior (light blue), Inferelator-Prior (light green), and CellOracle Prior (light pink) across the six cell lines (yeast, human embryonic stem cells, HepG2 hepatocellular carcinoma cells, mouse embryonic stem cells, mouse dendritic cells, mouse hematopoietic stem cells), with chance performance (gray) provided for each context. B Normalized AUPRC values with the best-performing prior (green box) highlighted in each cell line. Source data are provided as a Source Data file. Comparison of prior construction strategies across species. A Area under the precision-recall curve (AUPRC) of GLM-Prior (light blue), Inferelator-Prior (light green), and CellOracle Prior (light pink) across the six cell lines (yeast, human embryonic stem cells, HepG2 hepatocellular carcinoma cells, mouse embryonic stem cells, mouse dendritic cells, mouse hematopoietic stem cells), with chance performance (gray) provided for each context. B Normalized AUPRC values with the best-performing prior (green box) highlighted in each cell line. Source data are provided as a Source Data file. A broader comparison across species using normalized AUPRC values further illustrates these patterns (Fig. 5B), where normalized AUPRC is defined as (AUPRC − chance)/(1 − chance). While GLM-Prior provides the most consistent gains overall, CellOracle’s prior outperforms all other approaches in HepG2 (0.36 versus 0.30 for GLM-Prior and 0.24 for Inferelator-Prior). This comparatively strong performance in a mammalian cell line may reflect alignment between the construction of CellOracle’s prior and the BEELINE reference network used for evaluation. BEELINE assembles reference networks using ChIP-seq-supported TF-gene edges curated from ENCODE10, ChIP-ATLAS33, and ESCAPE34 databases, and HepG2 is among the most densely profiled ENCODE cell lines. As a result, the HepG2 reference likely contains a large and internally consistent set of TF occupancy-supported edges that are more recoverable by accessibility-conditioned, proximity-based linking, which is the type of evidence CellOracle prioritizes when constructing its prior from motifs in accessible regions. By contrast, Inferelator-Prior follows a related accessibility- and motif-based strategy but imposes stronger sparsity, which may remove a substantial number of true TF-gene pairs represented in the HepG2 reference network and thereby reduce its alignment with that reference. In contrast, GLM-Prior is the top-performing method in the four remaining cell lines: hESC, mESC, mDC, and mHSC. In hESC, all methods exceed chance, with GLM-Prior achieving the highest AUPRC (0.19 versus 0.17 for Inferelator-Prior and 0.16 for CellOracle’s prior). In mESC and mHSC, the advantage of GLM-Prior becomes more pronounced, as accessibility-based priors fall below chance while GLM-Prior remains predictive. These results suggest that sequence-derived priors can provide stronger agreement with held-out reference networks than accessibility-based priors in several mammalian contexts, despite not using matched cell-line-specific chromatin accessibility data. These comparisons show that prior construction strategies have context-dependent performance. Accessibility-driven approaches such as CellOracle can perform well when the reference network aligns with motif- and accessibility-based edge construction, as observed in yeast and HepG2. GLM-Prior, by contrast, provides a sequencing technology-independent alternative that achieves the strongest agreement with held-out reference networks in four of five mammalian contexts. The complementary performance profiles across species and cell lines indicate that no single prior construction approach is universally optimal, but that sequence-derived priors can provide a scalable strategy for prior construction when matched chromatin accessibility data are unavailable or when transfer across related mammalian contexts is desired. Paired prior-GRN strategies reveal prior-dependent gains from expression-based inference Following our comparative analysis of prior construction approaches, we next asked how each complete prior-GRN pipeline performs across the six cell lines (Fig. 6). GLM-Prior and PMF-GRN form our dual-stage pipeline, combining a finetuned sequence language model for prior construction with a probabilistic matrix factorization model for GRN inference from single-cell expression. CellOracle constructs a prior ("CellOracle baseGRN") using motif enrichment within accessible promoter regions and pairs it with a Bayesian ridge regression model for GRN inference. Similarly, Inferelator-Prior uses chromatin accessibility and TF motif enrichment to build a prior edge matrix that is then refined using regularized regression during GRN inference. By comparing these three paired pipelines, GLM-Prior + PMF-GRN, Inferelator-Prior + Inferelator, and CellOracle prior + CellOracle, we ask two related questions: (i) how well each method converts its own prior into a GRN that agrees with the corresponding reference network, and (ii) to what extent expression-based GRN inference can compensate for weak priors versus refining already strong ones.Fig. 6Paired prior-gene regulatory network (GRN) strategies reveal prior-dependent gains from expression-based GRN inference across six cell lines.A Area under the precision-recall curve (AUPRC) of each prior (GLM-Prior (light blue), Inferelator-Prior (light green), and CellOracle-Prior (light pink)) and its corresponding gene regulatory network (GRN) (PMF-GRN (dark blue), Inferelator (dark green), CellOracle (dark pink)) across six cell lines, with gray dashed lines indicating the chance level in each species. B Change in AUPRC from prior to GRN (ΔAUPRC) for each paired method, quantifying the benchmark gain from expression-based inference. C Normalized AUPRC performance for priors and GRNs in each cell line. Green boxes highlight the best prior and blue boxes highlight the best GRN per cell line. D Edge flips between prior and GRN for each method, showing how GRN inference methods modify their priors (blue indicates edge additions, while red indicates edge removal). Source data are provided as a Source Data file. Paired prior-gene regulatory network (GRN) strategies reveal prior-dependent gains from expression-based GRN inference across six cell lines. A Area under the precision-recall curve (AUPRC) of each prior (GLM-Prior (light blue), Inferelator-Prior (light green), and CellOracle-Prior (light pink)) and its corresponding gene regulatory network (GRN) (PMF-GRN (dark blue), Inferelator (dark green), CellOracle (dark pink)) across six cell lines, with gray dashed lines indicating the chance level in each species. B Change in AUPRC from prior to GRN (ΔAUPRC) for each paired method, quantifying the benchmark gain from expression-based inference. C Normalized AUPRC performance for priors and GRNs in each cell line. Green boxes highlight the best prior and blue boxes highlight the best GRN per cell line. D Edge flips between prior and GRN for each method, showing how GRN inference methods modify their priors (blue indicates edge additions, while red indicates edge removal). Source data are provided as a Source Data file. To interpret these patterns, we consider the absolute AUPRCs in Fig. 6A together with the corresponding ΔAUPRC values in Fig. 6B. In yeast and mDC, all priors are at or only slightly above chance, and GRN inference yields minimal gains (ΔAUPRC ≤ 0.02), indicating that expression-based models show limited ability to improve reference-network agreement when the starting prior is near-random. In hESC and mHSC, GLM-Prior already achieves above-chance performance, and PMF-GRN adds only small improvements, whereas Inferelator and CellOracle modestly refine their weaker priors. These contexts illustrate that when priors are already above baseline, GRN inference primarily fine-tunes edge weights rather than substantially reshaping the network. By contrast, HepG2 and mESC are the settings where GRN inference provides the largest benefits. Here, priors span a range from near- or slightly below-chance (Inferelator-Prior in HepG2; Inferelator-Prior and CellOracle prior in mESC) to moderately above-chance (GLM-Prior and CellOracle prior in HepG2; GLM-Prior in mESC), leaving partially specified structure that expression data can refine. In HepG2, GLM-Prior + PMF-GRN and Inferelator-Prior + Inferelator show sizeable increases in AUPRC, while CellOracle begins from the strongest prior and improves only modestly. In mESC, regression-based inference on top of weak or near-chance priors yields the single largest observed ΔAUPRC for CellOracle and substantial gains for Inferelator, whereas PMF-GRN makes only a small but consistent improvement over an already above-chance GLM-Prior. Taken together, these six cell lines reveal a coherent pattern: the magnitude of GRN-inference gains is strongly prior-dependent, with the largest AUPRC gains arising when priors are of intermediate quality, limited performance improvements when priors are very weak, and only modest refinements when priors are already strong. We next normalize AUPRC by the chance level in each species (as in Section 1) to directly compare performance above (or below) chance across methods and cell lines (Fig. 6C). In this view, GLM-Prior provides the best prior in four of six cell lines (hESC, mESC, mDC, mHSC), while CellOracle prior yields the strongest priors in yeast and HepG2. We observe that in four mammalian cell lines (HepG2, mESC, mDC, and mHSC), Inferelator-Prior rarely exceeds random-chance performance and typically remains close to or below chance. Similarly, in all three mouse cell lines, CellOracle’s prior does not construct a prior above chance. These patterns suggest that proximity- and accessibility-based priors do not consistently align with the held-out mammalian reference networks used here, particularly in the mouse cell lines. In addition, we find that the best GRNs largely track the best priors, with PMF-GRN producing the top-performing GRN in three of six cell lines, CellOracle in two, and Inferelator in one (Fig. 6C). Notably, there are instances where GRN inference can partially compensate for a weaker prior. In mESC, CellOracle’s prior is below chance (normalized AUPRC = − 0.04), yet its regression-based GRN achieves the highest normalized performance across all methods (0.21), outperforming GLM-Prior + PMF-GRN despite starting from a lower-performing prior. This behavior is clarified by examining how each GRN model modulates its prior in terms of edge flips (Fig. 6D). PMF-GRN primarily improves performance by adding edges to the prior, expanding the inferred network while leaving most prior edges unchanged. Inferelator both adds and removes edges, suggesting a balance between expansion and pruning. In contrast, CellOracle’s regression-based design only allows removal of edges: it can shrink edge weights to zero but cannot introduce edges that were absent from the accessibility-derived prior. The large performance gains observed for CellOracle in settings such as mESC therefore arise from pruning a dense prior rather than adding new edges during GRN inference. This pruning behavior can make a low-performing but dense prior align more strongly with the held-out reference network after GRN inference, yet also highlights an important limitation: when single-cell expression is used only to filter a fixed network, not to propose new edges, it cannot expand the regulatory scaffold beyond what is already encoded in the prior. Together, these results show that while GRN inference can substantially improve performance in some settings, particularly when priors are neither near-random nor already highly aligned with the reference network, the overall ranking of methods is largely dictated by prior quality. Expression-based GRN inference models mostly refine and re-weight the structure supplied by the prior, rather than overturning it, underscoring the central role of prior construction in determining benchmarked GRN performance. Cross-method comparison disentangles prior and GRN inference contributions to performance To disentangle the contribution of prior construction and GRN inference to reference-network agreement, we next performed a full cross-comparison in which each GRN inference method is applied to each prior across all six cell lines (Fig. 7). Specifically, we combine GLM-Prior, Inferelator-Prior, and CellOracle baseGRN with PMF-GRN, Inferelator, and CellOracle, yielding nine prior-GRN combinations per cell line. In Fig. 7A, we show the resulting AUPRC values, along with species-specific chance levels (gray dashed line), and the baseline performance of the prior (colored dashed line) used for GRN inference. In Fig. 7B, we report the same results normalized relative to chance, where values above zero indicate performance above random. We focus our interpretation on the normalized AUPRC values, which allow for easier comparison across species and methods, and refer to the raw AUPRCs in Fig. 7A where relevant.Fig. 7Cross-method comparison disentangles prior and gene regulatory network (GRN) inference contributions.A Area under the precision-recall curve (AUPRC) for all nine combinations of prior and GRN inference methods across six cell lines (PMF-GRN: dark blue, Inferelator: dark green, CellOracle: dark pink). Gray dashed lines indicate chance performance, while light blue, light green, and light pink dashed lines represent baseline performance of GLM-Prior, Inferelator-Prior (Inf-Prior), and CellOracle-Prior (CO-Prior), respectively, across each cell line. B Normalized AUPRC of all nine combinations of prior and GRN for direct comparison across methods and cell lines. Blue boxes indicate the highest GRN performance per cell line. Source data are provided as a Source Data file. Cross-method comparison disentangles prior and gene regulatory network (GRN) inference contributions. A Area under the precision-recall curve (AUPRC) for all nine combinations of prior and GRN inference methods across six cell lines (PMF-GRN: dark blue, Inferelator: dark green, CellOracle: dark pink). Gray dashed lines indicate chance performance, while light blue, light green, and light pink dashed lines represent baseline performance of GLM-Prior, Inferelator-Prior (Inf-Prior), and CellOracle-Prior (CO-Prior), respectively, across each cell line. B Normalized AUPRC of all nine combinations of prior and GRN for direct comparison across methods and cell lines. Blue boxes indicate the highest GRN performance per cell line. Source data are provided as a Source Data file. Across species, the dominant pattern is that the highest-performing inferred GRN for a given cell line is largely determined by the underlying prior, not by the choice of GRN inference method. In yeast, for example, any GRN method paired with CellOracle’s prior attains the same normalized AUPRC (0.05), whereas all combinations using GLM-Prior or Inferelator-Prior remain at or below 0.03. Once CellOracle’s prior is selected, the choice among PMF-GRN, Inferelator, or CellOracle has no quantifiable impact, indicating that the prior itself, rather than the choice of downstream inference algorithm, is the main constraint on benchmark performance. Similar patterns appear in the mammalian cell lines. In hESC, GLM-Prior and Inferelator-Prior yield the highest-scoring priors (raw AUPRC: 0.19 and 0.17, respectively), and nearly all GRNs inferred from these priors achieve competitive reference-network agreement. In HepG2, CellOracle’s prior is best (raw AUPRC = 0.36), and the two highest-performing GRNs are obtained by pairing this prior with PMF-GRN or CellOracle, demonstrating again that performance is largely determined by the prior rather than the choice of inference algorithm. An analogous effect is observed in mHSC, where GLM-Prior combined with PMF-GRN yields the highest-performing combination (normalized AUPRC = 0.24), while all GRNs built on Inferelator-Prior or CellOracle’s prior plateau at 0.08 or lower. These results support a consistent conclusion: the quality and structure of the prior largely constrain achievable benchmark performance, while GRN inference methods primarily modulate performance within a range set by the prior. The full cross-comparison design also allows us to ask whether any GRN inference method is inherently superior across priors and species. Across PMF-GRN, Inferelator, and CellOracle, we find no evidence that any single GRN inference method is uniformly superior. Each approach wins within different regimes defined by the combination of species and prior. PMF-GRN delivers the highest normalized performance in HepG2 when paired with CellOracle’s prior (0.20) and in mHSC when paired with GLM-Prior (0.24). PMF-GRN also performs well in yeast when paired with CellOracle’s prior, and in mESC when paired with GLM-Prior or Inferelator-Prior. Inferelator is best in hESC when paired with Inferelator-Prior (0.06) and in mDC when paired with CellOracle’s prior (0.02), and remains competitive in HepG2 across multiple priors. CellOracle is uniquely optimal in mESC when paired with CellOracle’s prior (0.21), substantially outperforming all other prior-GRN combinations in that species. The ranking of each GRN inference method thus changes with both prior and cell line, consistent with the interpretation that GRN methods act as context-dependent refiners of prior information using expression data, rather than serving as uniformly superior drivers of performance across settings. The cross-method comparison further reveals characteristic interactions between each GRN inference method and the priors. PMF-GRN appears highly prior sensitive, where switching priors can produce large swings in performance within a given species. In mHSC, for instance, normalized performance for PMF-GRN ranges from −0.05 with CellOracle-Prior to 0.24 with GLM-Prior. In mESC, performance ranges from − 0.04 with CellOracle Prior to 0.09 with GLM-Prior. This pattern indicates that PMF-GRN is most effective when initialized with priors that already show strong reference-network agreement, but has limited ability to improve priors that are near or below chance. The Inferelator, by contrast, is comparatively robust to prior choice. Within each species, the variation in normalized AUPRC across priors is typically modest (often within the range of 0.02 to 0.04). In mHSC, all three priors yield identical normalized performance (0.08) with the Inferelator. This suggests that the Inferelator relies more heavily on expression-based regularized regression, making it less sensitive to exact prior structure, while also limiting its ability to reach the highest benchmark performance in most cell lines. CellOracle exhibits strong method-prior compatibility in specific settings. For example, in mESC, CellOracle’s prior + CellOracle performs well (normalized AUPRC: 0.21), significantly outperforming all other combinations, indicating that the regression-based pruning used by CellOracle is particularly well matched to dense priors that require substantial edge removal to improve agreement with the reference network. As each GRN inference method is applied to every prior, we can also ask whether cross-pairing ever allows a high-performing GRN inference method to elevate a lower-performing prior above combinations that use a higher-performing prior with a different inference method. Such reversals are rare. In mHSC, for example, no method applied to the Inferelator-Prior or CellOracle prior achieves performance comparable to GLM-Prior + PMF-GRN. In HepG2, there is a modest example of beneficial cross-pairing, where CellOracle prior + PMF-GRN achieves a normalized AUPRC of 0.20, outperforming the original CellOracle prior + CellOracle pairing (normalized AUPRC: 0.17). The performance gains here are likely due to PMF-GRN adding expression-supported edges to the CellOracle prior. In mDC, CellOracle prior + Inferelator (normalized AUPRC: 0.02) is better than CellOracle prior + CellOracle (normalized AUPRC: 0) or CellOracle prior + PMF-GRN (normalized AUPRC: 0.02), although all values remain near chance. These exceptions are incremental, not complete reversals of the prior-driven ranking, and reinforce the view that GRN inference can fine-tune benchmark performance but typically does not elevate a low-performing prior to the level of the highest-performing priors. Overall, the cross-method comparison in Fig. 7 supports three main conclusions. First, prior quality remains the primary determinant of benchmarked GRN performance, even when each GRN method is given access to each prior. Second, there is no universally better GRN inference algorithm: PMF-GRN, the Inferelator, and CellOracle perform best in different prior and species regimes, reflecting differences in how they use expression data to refine or re-weight the prior. Third, GRN inference methods modulate, rather than overturn, the structure imposed by the prior. Changing the inference algorithm can yield meaningful gains in some contexts, particularly when paired with an appropriate prior, but rarely elevates a weak prior to the level of a strong one. These findings reinforce the conclusion that prior construction is the main driver of reference-network agreement in this benchmark, with expression-based GRN inference primarily serving a prior-dependent refining role. Discussion Constructing useful GRN models remains a central challenge in genomics, particularly in systems where regulatory annotations and matched experimental assays are incomplete or unavailable. In this work, we present GLM-Prior as a sequence-derived workflow for constructing TF-gene prior knowledge matrices and evaluate how these priors interact with downstream expression-based GRN inference. By integrating GLM-Prior with PMF-GRN, a probabilistic matrix factorization framework for single-cell expression data, we decouple prior construction from GRN inference and systematically assess how each component contributes to benchmarked GRN performance across yeast, mouse, and human cell-line contexts. Our results first demonstrate that GLM-Prior performance is tightly linked to the composition of the available training data. Across six evaluated contexts, agreement with held-out reference networks scales with positive-label abundance, TF coverage, and the relationship between training and evaluation labels. Human and mouse cell lines with richer regulatory annotations support stronger above-baseline performance than yeast, where sparse positive labels and extreme class imbalance limit generalization despite strong within-training validation metrics. Here, hESC and mESC show modest but reproducible gains, HepG2 and mHSC show clearer improvements, and mDC remains at chance. These findings indicate that genomic language models can learn sequence-associated signals that improve agreement with available regulatory references in some contexts, but that their performance remains constrained by incomplete supervision and the coverage of current benchmark networks. Transfer-learning experiments further support GLM-Prior as a workflow for constructing priors in settings where species- or cell-line-specific regulatory data are limited. Transfer between human and mouse maintains or modestly improves agreement with held-out reference networks compared to single-species models, suggesting that GLM-Prior can preserve useful sequence-associated predictive signal across related mammalian species. In contrast, transfer from mammalian training data to yeast remains near chance, indicating that this capability is context-dependent and does not extend uniformly across evolutionarily distant organisms. Further, multi-species training produces performance comparable to single-species or transfer models across the evaluated contexts, suggesting that a single sequence-derived model can provide stable priors without requiring separate model training for every target species. These results support the practical use of GLM-Prior for mammalian transfer-learning scenarios, while emphasizing that broader claims about transferability require additional reference networks across more organisms, cell types, and regulatory regimes. Furthermore, benchmarking GLM-Prior against accessibility-based priors further shows that prior construction strategies have context-dependent strengths. Accessibility-driven approaches such as CellOracle perform well when their construction assumptions align with the reference network, as observed in yeast and HepG2. GLM-Prior, by contrast, achieves the strongest agreement with held-out reference networks in four of five mammalian contexts, despite not using matched cell-line-specific chromatin accessibility data. This comparison does not establish that sequence-derived priors recover specific classes of regulatory mechanisms, such as distal enhancer regulation or long-range TF-gene interactions. Rather, it shows that sequence-derived priors can provide competitive or stronger reference-network agreement in several mammalian settings and offer a scalable alternative when matched chromatin accessibility data are unavailable. Integrating priors with multiple GRN inference algorithms reveals a consistent pattern: the quality and structure of the prior largely constrain achievable benchmark performance, while expression-based inference provides prior-dependent refinement. When GLM-Prior is used as input to PMF-GRN, AUPRC is equal to or higher than the prior alone across the evaluated contexts, with gains varying by cell line. Edge-flip analyses show that PMF-GRN generally preserves most prior edges and introduces limited expression-supported modifications, rather than reconstructing the network independently from expression data. The fully crossed comparison among GLM-Prior, Inferelator-Prior, CellOracle’s baseGRN, PMF-GRN, Inferelator, and CellOracle further shows that no GRN inference algorithm is uniformly superior across priors or species. Instead, the highest-performing inferred GRNs usually arise from combinations that use the highest-performing prior for that context. These results support the conclusion that prior construction is a major determinant of reference-network agreement in current GRN benchmarks, with downstream inference methods refining or reweighting the structure supplied by the prior. Several limitations of our study point to important directions for future work. First, regulatory interaction databases are incomplete and noisy. Positive labels are derived from curated interaction resources, while pairs without annotated interactions are treated as unlabeled negatives for training. Some of these unlabeled negatives may correspond to real regulatory interactions that are absent from current databases or active only in specific cellular contexts. The use of negative-class downsampling in our GLM-Prior approach is necessary to make training computationally tractable given the combinatorial number of possible TF-gene pairs, but this sampling strategy may also exclude informative unlabeled pairs from training. Future work should explore sampling strategies and positive-unlabeled learning frameworks that better reflect the incompleteness of regulatory labels. Second, evaluation is limited by the availability of benchmarkable reference networks. The six cell-line contexts used here represent the settings in which matched sequence inputs, training labels, and independent reference networks were available for systematic evaluation. These references provide useful standardized benchmarks, but they are not exhaustive maps of regulatory interactions. Expanding the number and diversity of reference networks will be essential for evaluating whether sequence-derived prior construction generalizes beyond the mammalian contexts where GLM-Prior performs best. Third, the gene body input design remains an important area for further investigation. Gene body sequence provides a consistent anchor for constructing gene-level inputs, but it may also encode proxy features such as sequence composition, gene length, gene-family structure, or annotation density. To address this concern, we performed diagnostic analyses showing that GLM-Prior performance is not reducible to broad compositional differences among gene bodies, sequence relatedness among genes, gene length, or database annotation density. These analyses support the interpretation that GLM-Prior captures sequence-associated information beyond these simple proxy features, while leaving open the need for future model designs that incorporate richer regulatory context, including promoters, enhancers, chromatin state, three-dimensional genome organization, and cell-type-specific regulatory annotations. Finally, while we compare multiple GRN inference algorithms, all evaluated methods retain an edge-centric view of network inference. Future work could extend this framework by modeling regulatory modules, temporal trajectories, perturbation responses, or dynamic TF activity states. Such extensions may help clarify when expression data provides information beyond prior refinement and how sequence-derived priors can be combined with multimodal measurements in a more flexible inference framework. In summary, this work positions GLM-Prior as a benchmarked workflow for transferable, sequence-derived prior construction in GRN inference. Across the evaluated contexts, GLM-Prior provides a scalable strategy for constructing TF-gene priors when matched experimental assays are unavailable, with strongest support for transfer across related mammalian systems. More broadly, our results show that prior quality strongly shapes benchmarked GRN performance and that expression-based inference methods primarily provide prior-dependent refinement. These findings shift emphasis in GRN modeling toward the construction, evaluation, and transferability of prior knowledge, and provide a foundation for future methods that integrate sequence-derived priors with expression and other regulatory modalities. Methods The GLM-prior model We developed GLM-Prior, a genomic language model fine-tuned to score transcription factor (TF)-target gene pairs from nucleotide sequence and construct TF-gene prior knowledge matrices for downstream inference. Specifically, we fine-tuned the 250-million-parameter Nucleotide-Transformer model17, originally pre-trained on the genomes of 850 species, to predict TF-target gene interaction probabilities from nucleotide sequences. The resulting interactions are used to define a sequence-derived prior knowledge matrix, which provides structural information for downstream GRN inference. The model uses a transformer encoder architecture to process two distinct sequence types, TF motif sequences from the CisBP database22 and gene body sequences derived from genome annotations in a GTF file. We append a classification head to the transformer encoder to estimate the probability that each TF-gene pair belongs to the annotated positive class. Positive labels are derived from curated regulatory interaction resources such as STRING23–26 and TRRUST27,28, whereas TF-gene pairs without an annotated interaction are treated as unlabeled negative examples for supervised training. These unlabeled pairs should not be interpreted as experimentally verified non-interactions, since some may correspond to regulatory interactions that are absent from current databases or active only in specific cellular contexts. This design allows the model to use sequence information together with database-derived labels to construct prior knowledge matrices whose performance can be benchmarked against held-out reference networks. Each training example is constructed by concatenating the nucleotide sequence of a TF with that of its candidate target gene, separated by a special classification token (< cls >). This composite sequence is tokenized, during which a model-specific (< cls >) token is automatically prepended to the input. The tokenized sequence is passed through the pre-trained transformer, which produces contextual embeddings across the full input. The embedding corresponding to the prepended (< cls >) token is fed into the classification head composed of a dropout layer, a linear projection, a \tanhtanh activation, a second dropout, and a final linear layer that maps to two logits representing the binary interaction label. This setup allows the model to learn sequence-associated features that improve prediction of database-derived TF-gene labels. Dataset construction and preprocessing To train the model, we constructed a dataset of TF-gene pairs in which annotated positive examples were derived from curated regulatory interaction databases (YEASTRACT11, STRING23–26 and TRRUST27,28), and unlabeled negative examples were sampled from the remaining TF-gene sequence pairs without an annotated interaction. These unlabeled pairs were used as negative examples for supervised learning, but they do not represent experimentally verified non-interactions. This dataset was split into a 99% training and 1% validation set. Given the substantial imbalance between annotated positive and unlabeled negative samples, we implemented a downsampling strategy to reduce the number of negative samples in the training set. This preserved diversity among unlabeled negative samples while preventing the model from overfitting to the negative class. Specifically, the number of retained unlabeled negative samples after downsampling can be defined as, 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{sampled}=\lfloor r\cdot {N}_{-}\rfloor,$$\end{document}Nsampled=⌊r⋅N−⌋, where r ∈ (0, 1) is the downsampling rate, and N− is the total number of unlabeled negative examples in the dataset. To further address class imbalance during training, we designed a custom DataLoader that performs even class batch sampling. Each batch was constructed to contain an equal number of annotated positive and unlabeled negative samples, ensuring a balanced signal during training. Specifically, each batch is defined as, 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$B=\{({x}_{i}^{+},{x}_{j}^{-}):{x}_{i}^{+}\in {X}_{+},{x}_{j}^{-}\in {X}_{-}\},$$\end{document}B={(xi+,xj−):xi+∈X+,xj−∈X−}, where annotated positive examples \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i}^{+}$$\end{document}xi+ were sampled uniformly with replacement from X+ and unlabeled negative examples \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{j}^{-}$$\end{document}xj− were cycled through without replacement: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x}_{i}^{+} \sim {X}_{+},\quad {x}_{j}^{-} \sim {X}_{-}\quad {{{\rm{with}}}}\quad j\,{{\mathrm{mod}}}\,\,{N}_{-}.$$\end{document}xi+~X+,xj−~X−withjmodN−. This strategy ensures that each unlabeled negative example was seen exactly once during training, while annotated positive examples were reused as needed to maintain class balance. Even-class batching was critical for stabilizing training dynamics and improving the model’s sensitivity to annotated positive interactions while limiting bias toward the much larger unlabeled negative class. Model architecture and classification The model architecture builds on the pre-trained Nucleotide Transformer by retaining its embedding and transformer encoding layers. A classification head is applied to the final hidden state corresponding to the prepended (< cls >) token to compute logits for a binary classification task. This classification head outputs a two-dimensional logit vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$z\in {{\mathbb{R}}}^{2}$$\end{document}z∈R2 representing unnormalized scores for the positive and negative classes, 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\bf{z}}}}={{{\bf{W}}}}\cdot {{{\bf{h}}}}+{{{\bf{b}}}},$$\end{document}z=W⋅h+b, where h is the transformed hidden state of the (< cls >) token following the intermediate projection and non-linearity, and W, b are the weights and bias of the final linear layer. The components of z are denoted z+ and z−, corresponding to the logits for the positive and negative classes, respectively. Class probabilities are computed using a softmax function, 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{P}}}}(y=1| {x}_{{{{\rm{TF}}}}},{x}_{{{{\rm{gene}}}}})=\frac{\exp ({z}_{+})}{\exp ({z}_{+})+\exp ({z}_{-})},$$\end{document}P(y=1∣xTF,xgene)=exp(z+)exp(z+)+exp(z−), Training used a class-weighted cross-entropy loss function to account for residual class imbalance after downsampling. The loss assigns a fixed weight of 1.0 to positive examples, while the negative class weight w− is tuned through hyperparameter search to optimize the balance between precision and recall. The loss is computed as, 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{L}}}}=-{w}_{+}y\log p-{w}_{-}(1-y)\log (1-p),\quad \,{\mbox{where}}\,\quad p=\frac{\exp ({z}_{+})}{\exp ({z}_{+})+\exp ({z}_{-})}.$$\end{document}L=−w+ylogp−w−(1−y)log(1−p),wherep=exp(z+)exp(z+)+exp(z−). Here, y ∈ {0, 1} is the true label, z+ and z− are the predicted logits for the positive and negative classes, respectively. w− is the down-weighted negative class weight, varied during hyperparameter search, and the positive class weight w+ is fixed at 1.0. This approach maintains sensitivity to positive predictions while mitigating bias toward the negative class. Training procedure We trained the model on 4 H100 GPUs using PyTorch’s Distributed Data Parallel (DDP) framework35 to enable efficient multi-GPU scaling. The training process was distributed across GPUs to accelerate computation and ensure consistent gradient updates. We used a per-device batch size of 32 and set the gradient accumulation steps to 32, resulting in an effective batch size of 4096. The model was optimized using Adam with a learning rate of 10−5. Training spanned 10 epochs, using optimal hyperparameters selected through a sweep over the negative class weight (w−) and downsampling rate (see Supplementary Information for more details). We selected the configuration that achieved the highest F1 score on the validation set for final training. After training, the model was used to score TF-gene sequence pairs without labels and construct a prior knowledge matrix for downstream GRN inference. This matrix represents sequence-derived prior evidence for putative TF-gene edges and can be further refined using cell-type, cell-line, or condition-specific expression data during downstream GRN inference. Performance and evaluation We evaluated model performance using standard binary classification metrics, with a focus on metrics that remain robust under class imbalance. Specifically, we report precision, recall, F1 score, area under the receiver operating characteristic curve (AUC-ROC), area under the precision-recall curve (AUPRC), and Matthews correlation coefficient (MCC). Precision and recall were computed separately for the positive and negative classes to assess the model’s ability to minimize false positives and false negatives, respectively. Let TP, FP, FN, and TN denote the true positives, false positives, false negatives, and true negatives. Then, 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{Precision}}}}=\frac{{{{\rm{TP}}}}}{{{{\rm{TP}}}}+{{{\rm{FP}}}}},\quad {{{\rm{Recall}}}}=\frac{{{{\rm{TP}}}}}{{{{\rm{TP}}}}+{{{\rm{FN}}}}}.$$\end{document}Precision=TPTP+FP,Recall=TPTP+FN. The F1 score, which represents the harmonic mean of precision and recall, was used as the primary metric for model selection and hyperparameter optimization. It can be computed as, 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{F}}}}}_{{{{\rm{1}}}}}=2\cdot \frac{{{{\rm{Precision}}}}\cdot {{{\rm{Recall}}}}}{{{{\rm{Precision}}}}+{{{\rm{Recall}}}}}.$$\end{document}F1=2⋅Precision⋅RecallPrecision+Recall. To account for class imbalance and provide a threshold-independent measure of performance, we also computed AUC-ROC and AUPRC. the ROC curve plots true positive rate (TPR) against false positive rate (FPR), defined as: 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\,{{{\rm{True}}}}\;{{{\rm{Positive}}}}\;{{{\rm{Rate}}}}\;{{{\rm{(TPR)}}}}\,=\frac{{{{\rm{TP}}}}}{{{{\rm{TP}}}}+{{{\rm{FN}}}}},\quad \,{{{\rm{False}}}}\;{{{\rm{Positive}}}}\;{{{\rm{Rate}}}}\;{{{\rm{(FPR)}}}}\,=\frac{{{{\rm{FP}}}}}{{{{\rm{FP}}}}+{{{\rm{TN}}}}}.$$\end{document}TruePositiveRate(TPR)=TPTP+FN,FalsePositiveRate(FPR)=FPFP+TN. While AUC-ROC captures the model’s general discriminative ability, AUPRC is more informative in imbalanced settings, as it directly reflects the trade-off between precision and recall. We used AUPRC to benchmark model predictions against reference networks of TF-gene interactions. To determine the optimal classification threshold, we performed a grid search over the predicted positive class probabilities. The threshold t* that maximized the F1 score on the validation set was selected for final inference, 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${t}^{*}={{{\rm{argmax}}}}\,{{{{\rm{F}}}}}_{{{{\rm{1}}}}}(t).$$\end{document}t*=argmaxF1(t). Finally, we report the Matthews correlation coefficient (MCC), a balanced measure of classification quality that incorporates all four confusion matrix components, 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\rm{MCC}}}}=\frac{{{{\rm{TP}}}}\cdot {{{\rm{TN}}}}-{{{\rm{FP}}}}\cdot {{{\rm{FN}}}}}{\sqrt{({{{\rm{TP}}}}+{{{\rm{FP}}}})({{{\rm{TP}}}}+{{{\rm{FN}}}})({{{\rm{TN}}}}+{{{\rm{FP}}}})({{{\rm{TN}}}}+{{{\rm{FN}}}})}}.$$\end{document}MCC=TP⋅TN−FP⋅FN(TP+FP)(TP+FN)(TN+FP)(TN+FN). MCC ranges from − 1 to + 1, where + 1 indicates perfect predictions, 0 indicates random predictions, and − 1 indicates incorrect predictions. MCC remains informative even when classes are highly imbalanced, making it a useful complement to F1 and AUPRC. GRN inference with PMF-GRN We performed GRN inference using our previously published method, PMF-GRN (Probabilistic Matrix Factorization for Gene Regulatory Network inference)7, to estimate putative regulatory relationships between TFs and target genes from single-cell expression data. PMF-GRN decomposes an observed gene expression matrix into latent factors that represent TF activity and regulatory interactions between TFs and their target genes. These latent factors provide an inferred, prior-conditioned representation of GRN structure, which is not directly measured from gene expression data alone. Further details regarding the PMF-GRN model and the inference strategy used to obtain GRNs can be found in ref. 6. Using PMF-GRN, we perform inference independently on each single-cell dataset to obtain dataset-specific prior-conditioned GRNs. These inferred networks are then combined post-inference using a simple averaging strategy to produce a consensus GRN, 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{GRN}}}}}_{{{{\rm{Consensus}}}}}=\frac{1}{N}{\sum}_{i=1}^{N}{{{{\rm{GRN}}}}}_{i},$$\end{document}GRNConsensus=1N∑i=1NGRNi, where N is the number of datasets and GRNi is the inferred network for dataset i. This consensus GRN summarizes regulatory edges that are consistently assigned high scores across datasets, while preserving dataset-specific networks for context-specific analyses. To evaluate the relationship between posterior uncertainty and predictive accuracy, we rank all TF-gene interaction pairs by their posterior variance, constructing 10 cumulative bins corresponding to the variances in increments of 10%. For each bin k, we use the posterior point estimates of the interactions within the bin and compute the AUPRC using a set of TF-gene interactions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{G}}}}$$\end{document}G from a reference network. All interactions in each bin are included in the evaluation, regardless of whether they appear in the reference network. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{B}}}}}_{k}$$\end{document}Bk be the set of posterior point estimates in the k% of variances. Then the AUPRC for bin k is given by: 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\rm{AUPRC}}}}}_{k}={{{\rm{AUPRC}}}}({{{{\mathcal{B}}}}}_{k};{{{\mathcal{G}}}}),$$\end{document}AUPRCk=AUPRC(Bk;G), where the AUPRC is computed by comparing the predicted scores in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{{\mathcal{B}}}}}_{k}$$\end{document}Bk to labels derived from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{{\mathcal{G}}}}$$\end{document}G. The cumulative binning strategy ensures that each bin contains a sufficient number of interactions for stable AUPRC estimation. Statistics and reproducibility No statistical method was used to predetermine sample size. Dataset sizes were determined by the publicly available benchmark datasets and regulatory interaction databases used in this study. Data were excluded only according to predefined criteria described in the manuscript. For all cell lines, genes and regulatory interactions overlapping the independent reference network were removed prior to model training to prevent information leakage. During model training, unlabeled negative examples were downsampled according to the predefined sampling strategy to mitigate class imbalance. No additional data were excluded after inspection of the results. Training and validation splits were generated programmatically using fixed random seeds. Model training and evaluation were repeated using three independent random seeds (42, 43, and 44), and summary statistics are reported as mean ± one standard deviation unless otherwise indicated. Statistical analyses were performed using SciPy. Comparisons of train and test sequence-property distributions used two-sided Kolmogorov-Smirnov and two-sided Mann-Whitney U tests, and exact p values are reported. No adjustment for multiple comparisons was applied. The investigators were not blinded during model training, evaluation, or outcome assessment. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Source Data