MOTIVATION: Transcription factors (TFs) are key players in gene regulation and development, where they activate and repress gene expression through DNA binding. Predicting transcription factor binding sites (TFBSs) has long been an active area of research, with many deep learning methods developed to tackle this problem. These models are often trained on TF ChIP-seq data, which is generally seen as only providing positive samples. The choice of datasets and negative sampling techniques is a critical yet often overlooked aspect of this work. RESULTS: In this study, we investigate the impact of different negative sampling techniques on TFBS prediction performance. We create high-quality test datasets based on ChIP-seq and ATAC-seq data, where true negatives can be identified as positions that are accessible but not bound by the TF in question. We then train models using various negative sampling techniques, including genomic sampling, shuffling, dinucleotide shuffling, neighborhood sampling, and cell line specific sampling, simulating cases where matching ATAC-seq data is not available. Our results show that, generally, metrics calculated on training datasets give inflated performance scores. Of the tested techniques, genomic sampling of negatives based on similarity to the positives performed by far the best, although still not reaching the performance of baseline models trained on high-quality datasets. Models trained on dinucleotide shuffled negatives performed poorly, despite being a common practice in the field. Our findings highlight the importance of carefully selecting negative sampling techniques for TFBS prediction, as they can significantly impact model performance and the interpretation of results. AVAILABILITY AND IMPLEMENTATION: The code used in this study is available at https://github.com/NatanTourne/TFBS-negatives (DOI: 10.5281/zenodo.18007567).

Eukaryotic gene regulation is a combinatorial, dynamic, and quantitative process that plays a vital role in development and disease and can be modeled at a systems level in gene regulatory networks (GRNs). The wealth of multi-omics data measured on the same samples and even on the same cells has lifted the field of GRN inference to the next stage. Combinations of (single-cell) transcriptomics and chromatin accessibility allow the prediction of fine-grained regulatory programs that go beyond mere correlation of transcription factor and target gene expression, with enhancer GRNs (eGRNs) modeling molecular interactions between transcription factors, regulatory elements, and target genes. In this review, we highlight the key components for successful (e)GRN inference from (sc)RNA-seq and (sc)ATAC-seq data exemplified by state-of-the-art methods as well as open challenges and future developments. Moreover, we address preprocessing strategies, metacell generation and computational omics pairing, transcription factor binding site detection, and linear and three-dimensional approaches to identify chromatin interactions as well as dynamic and causal eGRN inference. We believe that the integration of transcriptomics together with epigenomics data at a single-cell level is the new standard for mechanistic network inference, and that it can be further advanced with integrating additional omics layers and spatiotemporal data, as well as with shifting the focus towards more quantitative and causal modeling strategies.

BACKGROUND: Representing the complex interplay between different types of biomolecules across different omics layers in multi-omics networks bears great potential to gain a deep mechanistic understanding of gene regulation and disease. However, multi-omics networks easily grow into giant hairball structures that hamper biological interpretation. Module detection methods can decompose these networks into smaller interpretable modules. However, these methods are not adapted to deal with multi-omics data nor consider topological features. When deriving very large modules or ignoring the broader network context, interpretability remains limited. To address these issues, we developed a SUbgraph BAsed mulTi-OMIcs Clustering framework (SUBATOMIC), which infers small and interpretable modules with a specific topology while keeping track of connections to other modules and regulators. RESULTS: SUBATOMIC groups specific molecular interactions in composite network subgraphs of two and three nodes and clusters them into topological modules. These are functionally annotated, visualized and overlaid with expression profiles to go from static to dynamic modules. To preserve the larger network context, SUBATOMIC investigates statistically the connections in between modules as well as between modules and regulators such as miRNAs and transcription factors. We applied SUBATOMIC to analyze a composite Homo sapiens network containing transcription factor-target gene, miRNA-target gene, protein-protein, homologous and co-functional interactions from different databases. We derived and annotated 5586 modules with diverse topological, functional and regulatory properties. We created novel functional hypotheses for unannotated genes. Furthermore, we integrated modules with condition specific expression data to study the influence of hypoxia in three cancer cell lines. We developed two prioritization strategies to identify the most relevant modules in specific biological contexts: one considering GO term enrichments and one calculating an activity score reflecting the degree of differential expression. Both strategies yielded modules specifically reacting to low oxygen levels. CONCLUSIONS: We developed the SUBATOMIC framework that generates interpretable modules from integrated multi-omics networks and applied it to hypoxia in cancer. SUBATOMIC can infer and contextualize modules, explore condition or disease specific modules, identify regulators and functionally related modules, and derive novel gene functions for uncharacterized genes. The software is available at https://github.com/CBIGR/SUBATOMIC .