Hierarchical Encoding of Regulatory Mechanisms and Expression Syntax by a foundational genomic sequence-to-function model

Deciphering the comprehensive regulatory rules encoded in the genomic sequence remains a central challenge in functional genomics, requiring a paradigm shift from descriptive annotations to a mechanistic understanding of the entire genome. Here, we introduce HERMES (Hierarchical Encoding of Regulatory Mechanisms and Expression Syntax), a framework that progressively defines a fundamental sequence vocabulary for the complex regulatory genome and parses gene expression syntax into transparent biological insights. Specifically, we trained a foundational sequence-to-function model on a massive compendium of 137,127 functional genomics profiles spanning diverse biochemical marks and cellular conditions, including DNA methylation, transcription factor binding, polymerase binding, histone marks, chromatin accessibility and RNA expression across various tissues, cell lines and cell types. By harnessing the high-fidelity sequence representations of HERMES, we established a context-specific regulatory vocabulary comprising 40 distinct sequence classes and fine-grained subclusters spanning the entire genome. We interrogated the sequence determinants of diverse promoters and tissue-specific enhancers, and further identified a housekeeping genes-specific promoter sub-class, driven by ETS, YY1 and CCAAT motifs. To further decode transcriptional regulation, HERMES was leveraged to predict cell-type-specific gene expression and the enhancer perturbation effects. Extensive evaluations confirmed that the cis-regulatory elements and their interactions were captured to predict gene expression underlying different cellular conditions, and simultaneously revealed divergent enhancer dependencies between housekeeping genes (HKGs) and highly variable genes (HVGs). Ultimately, we distilled the complex sequence-to-function model into a biologically interpretable rulebook of Enhancer-Promoter interaction grammar. Synthesizing all the insights, we propose a unified compatibility model, where HKGs utilize a strong-promoter architecture for high-output expression, while HVGs depend on a context-dependent syntax driven by compatible promoters and enhancers. In summary, HERMES bridges the gap between predictive modeling and biological mechanism, transforming sequence representations into a comprehensive functional encyclopedia and a quantitative grammar of the complex regulatory genome.