BioMatrix: Towards a Comprehensive Biological Foundation Model Spanning the Modality Matrix of Sequences, Structures, and Language

BioMatrix is introduced as the first multimodal foundation model that integrates molecular sequences, structures, and natural language within a single decoder-only architecture. It utilizes a unified tokenization scheme to map various biological modalities into a shared discrete token space, allowing for uniform processing under a next-token prediction objective. Built on the Qwen3 model with sizes of 1.7B and 4B parameters, it has been pretrained on 304.4 billion tokens and demonstrates state-of-the-art performance across 77 out of 80 tasks in biological applications, highlighting its potential as a versatile tool for practitioners in the field.