Biology-Guided AI Boosts Genomic Analysis Efficiency and Accuracy

This research presents bioMoR, an innovative AI framework designed to enhance the analysis of high-dimensional omics data, such as genomics. Unlike traditional transformer models that process all genes equally, bioMoR leverages a Mixture-of-Recursions (MoR) approach, which adaptively allocates computational resources. The core innovation lies in integrating structured biological knowledge at three key points: refining token embeddings with graph-based information, guiding self-attention towards biologically related tokens, and using a graph-aware router to determine the computational depth for each token. The framework's effectiveness stems from its ability to understand and utilize token interactions based on biological context, allowing it to construct more accurate embeddings and prioritize deeper learning for relevant tokens. This results in substantial performance improvements across various omics benchmarks, achieving higher accuracy and F1 scores compared to biology-agnostic MoR baselines. Furthermore, bioMoR demonstrates remarkable efficiency, requiring significantly fewer parameters and computational operations than non-recursive transformers, while also offering biological interpretability through identified marker genes and pathways.