Modelling transcription with explainable AI uncovers context-specific epigenetic gene regulation at promoters and gene bodies

Transcriptional regulation involves complex interactions with chromatin-associated proteins, but disentangling these mechanistically remains challenging. Here, we generate deep learning models to predict RNA Pol-II occupancy from chromatin-associated protein profiles in unperturbed conditions. We evaluate the suitability of Shapley Additive Explanations (SHAP), a widely used explainable AI (XAI) approach, to infer functional relevance and analyse regulatory mechanisms across diverse datasets. We aim to validate these insights using data from degron-based perturbation experiments. Remarkably, genes ranked by SHAP importance predict direct targets of perturbation even from unperturbed data, enabling inference without costly experimental interventions. Our analysis reveals that SHAP not only predicts differential gene expression but also captures the magnitude of transcriptional changes. We validate the cooperative roles of SET1A and ZC3H4 at promoters and uncover novel regulatory contributions of ZC3H4 at gene bodies in influencing transcription. Cross-dataset validation uncovers unexpected connections between ZC3H4, a component of the Restrictor complex, and INTS11, part of the Integrator complex, suggesting crosstalk mediated by H3K4me3 and the SET1/COMPASS complex in transcriptional regulation. These findings highlight the power of integrating predictive modelling and experimental validation to unravel complex context-dependent regulatory networks and generate novel biological hypotheses.Author summary: Genes are turned on or off through complex processes involving many proteins that interact with DNA wrapped histones and modify their structure. These changes, known as epigenetic modifications, help control how genes are expressed without altering the DNA sequence itself. In this study, we wanted to understand how different proteins influence gene activity in mouse stem cells by looking at their positions along the genome, particularly whether they act near the gene’s start site (promoter) or within the gene body. To do this, we used machine learning models and a method called SHAP, which helps explain the model’s decisions. By comparing our predictions to data from experiments where specific proteins were removed, we found that some proteins have context-specific effects, acting not only at the promoter but also along the whole gene body. Our approach highlighted both well-known and unexpected regulators of transcription and revealed that gene body signals, which are often overlooked, can play key roles. These findings show how explainable AI can help uncover new insights into how epigenetic features shape gene regulation, and offer a powerful way to generate testable hypotheses from complex genomic data.