DeepEvo

The identification of adaptively driven genes underlying human evolutionary traits remains a key challenge in evolutionary genomics. Here we present DeepEvo, an interpretable Siamese neural network that predicts cross-species expression differences from orthologous sequences while pinpointing evolutionary regulatory variants. DeepEvo outperforms existing methods in cross-species modeling, with validation through documented annotations and single-base perturbation assays (MPRA/Perturb-seq). We discover that unlike population-level variations that predominantly disrupt existing regulatory motifs, evolutionary regulatory variants frequently enhance existing motifs over longer timescales. These variants are enriched in disease-associated regions, indicating their shared functions underlying human evolution and disease. Furthermore, analysis of their combinatorial cis-regulation revealed an ‘additive effect’ model. Within this framework, genes can maintain global stability through compensatory changes, while simultaneously achieving precise, cell-type-specific expression changes. This insight led to a ‘concerted drive’ strategy, prioritizing four adaptively driven genes that function across multiple systems, based on coordinated pushes from multiple evolutionary regulatory elements. As validation, we focused on PRKD2, upregulated in humans through concordant cis-regulatory changes. PRKD2-depleted rhesus macaques exhibited multi-system alterations—including reduced neuronal activity, decreased dendritic complexity, altered functional connectivity of brain, elevated insulin levels and lymphocyte counts—recapitulating key human-rhesus phenotypic differences. Our framework deciphers the cis-regulatory grammar of human transcriptome evolution and provides an effective strategy for identifying adaptively driven genes, generalizable to other cross-species comparisons.