Combining Motifs, CRE Activity, And Gene Expression Data Using ML Greatly Improves the Accuracy of Tissue-Specific TF Network Maps

Transcription factor (TF) network maps link TFs to their direct, functional gene targets whose transcription they regulate by binding cis-regulatory elements (CREs). Existing methods to reconstruct these networks typically rely either on TF motifs in CREs or gene expression data alone, limiting their accuracy. Motif data alone often fail to identify actual TF binding sites, while expression data cannot distinguish direct from indirect regulatory relationships. Additionally, accurate TF networks must be tissue-specific due to varied TF activities and expression patterns across tissues.

We introduce METANets (Motif Expression TF Association Networks), a novel supervised ensemble learning approach integrating TF motifs, TF binding locations, CRE activity, and gene expression data. Using XGBoost models, we predict TF binding in CREs based on TF motif and gene expression features, derived from linear (LASSO) and non-linear (BART) regression models trained on tissue-specific and aggregated RNA-seq data. This approach was applied to 36 human tissues from GTEx.

METANets significantly outperform existing motif-only and expression-only approaches, capturing more direct, functional TF targets. Evaluations against ChIP-seq binding data and gene ontology enrichment demonstrate METANets’ superiority in identifying functional targets directly bound by TFs. Furthermore, tissue specificity assessed through tissue-specific expression quantitative trait loci (eQTLs) confirms METANets effectively capture tissue-specific regulation, performing comparably with other networks.

Our approach markedly improves TF network reconstruction by combining complementary data types, enhancing the accuracy and utility of tissue-specific transcriptional regulatory maps. METANets provide robust resources for researchers investigating TF-mediated regulation within human tissues.