bioMoR: Biology-Guided Mixture-of-Recursions for Genomic Learning

bioMoR is a novel framework that applies Mixture-of-Recursions (MoR) to gene-level and pathway-level learning in genomics. It is the first to integrate structured biological knowledge into an MoR backbone, identifying three key locations: graph-based information sharing to refine token embeddings, a structural bias to guide self-attention toward biologically related tokens, and a graph-aware router that uses neighborhood information to determine recursion depth. The framework aims to improve efficiency in high-dimensional omics analysis by adaptively routing tokens, as only a subset of genes or pathways requires deep computation. The paper is available on arXiv under the identifier 2608.06727.