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

Researchers introduce BioMatrix, a multimodal foundation model that integrates molecular sequences, structures, protein data, and natural language within a single decoder-only architecture. The model achieves state-of-the-art performance on 77 of 80 downstream tasks, demonstrating that a unified generalist AI can match or exceed specialized biological tools across diverse applications.

BioMatrix represents a significant advancement in biological AI by solving a critical architectural challenge: previous foundation models either integrated multiple data modalities for a single entity type or covered multiple biological entities without native structural understanding. This new approach unifies sequences (SMILES/SELFIES), molecular structures, protein sequences, protein structures, and scientific text into a shared token space, enabling seamless multimodal processing without adapters or specialized output heads.

The model builds on Qwen3 (1.7B and 4B parameters) and underwent continual pretraining on 304.4 billion tokens encompassing general text, domain-specific literature, and cross-modal corpora linking molecules to proteins and their interactions. This training strategy mirrors broader trends in foundation models where scale and diverse data substantially improve generalization.

For the biotech and computational biology sectors, BioMatrix's competitive performance across 80 tasks—spanning drug discovery, protein function prediction, and structure-guided generation—suggests unified models can replace expensive specialized pipelines. This has implications for research acceleration and democratization of biological AI tools, potentially reducing development costs for small biotech teams. The model's ability to jointly process molecules and proteins addresses real scientific workflows where understanding protein-drug interactions requires reasoning across entity types.

Looking ahead, the critical metric is how this architecture scales beyond 4B parameters and whether the unified tokenization approach maintains performance gains at larger scales. Integration with wet-lab validation pipelines and commercial drug discovery platforms will determine real-world adoption. The work also opens questions about fine-tuning efficiency and whether generalist models reduce hallucinations compared to specialized predecessors in high-stakes applications like drug safety assessment.

• →BioMatrix integrates sequences, structures, and language for molecules and proteins in one native multimodal model without adapters or specialized heads
• →Achieves state-of-the-art or competitive results on 77 of 80 biological tasks after training on 304.4 billion tokens
• →Unified tokenization scheme enables all modalities to be consumed and produced under single next-token prediction, unlike prior adapter-based approaches
• →Model built on Qwen3 (1.7B-4B parameters) with continual pretraining on general, domain-specific, and cross-modal biomedical corpora
• →Success suggests single generalist models can match or exceed specialized biological tools across drug discovery, protein prediction, and structure generation