A hybrid machine learning model for predicting gene expression from epigenetics across fungal species

Understanding and controlling gene expression in organisms is essential for optimizing biological processes, whether in service of bioeconomic processes, human health, or environmental regulation. Epigenetic modifications play a significant role in regulating gene expression by altering chromatin structure, DNA accessibility and protein binding. While a significant amount is known about the combinatorial effects of epigenetics on gene expression, our understanding of the degree to which the orchestration of these mechanisms is conserved in gene expression regulation across species, particularly for non-model organisms, remains limited. In this study, we aim to predict gene expression levels based on epigenetic modifications in chromatin across different fungal species, to enable transferring information about well characterized species to poorly understood species. We developed a custom hybrid deep learning model, EAGLE (Evolutionary distance-Adaptable Gene expression Learned from Epigenomics), which combines convolutional layers and multi-head attention mechanisms to capture both local and global dependencies in epigenetic data. We demonstrate the cross-species performance of EAGLE across fungi, a kingdom containing both pathogens and biomanufacturing chassis and where understanding epigenetic regulation in under-characterized species would be transformative for bioeconomic, environmental, and biomedical applications. EAGLE outperformed shallow learning models and a modified transformer benchmarking model, achieving up to 80% accuracy and 89% AUROC for intra-species validation and 77% accuracy and 83% AUROC in cross-species prediction tasks. SHAP analysis revealed that EAGLE identifies important epigenetic features that drive gene expression, providing insights for experimental design and potential future epigenome engineering work. Our findings demonstrate the potential of EAGLE to generalize across fungal species, offering a versatile tool for optimizing fungal gene expression in multiple sectors. In addition, our architecture can be adapted for cross-species tasks across the tree of life where detailed molecular and genetic information can be scarce.