Qizhi Pei, Zhimeng Zhou, Yi Duan, Yiyang Zhao, Wei Li, Han Guo, Liang He, Chengping Li, Chang-Yu Hsieh, Conghui He, Rui Yan, Lijun Wu
5 min
Biological research is currently fragmented across different data modalities—sequences, 3D structures, and natural language—and different entity types—molecules and proteins. Existing foundation models typically address these in isolation, either by focusing on a single entity type or by using adapter-based architectures that cannot natively generate the modalities they read. BioMatrix asks whether a single, unified decoder-only model can integrate all these modalities for both molecules and proteins under a shared next-token prediction objective.
The researchers developed BioMatrix by mapping diverse biological data into a shared discrete token space. This includes molecular sequences (SMILES and SELFIES), protein sequences, and 3D structural data for both. To handle 3D structures, the model uses specialized tokenizers: a modified MolStrucTok for molecules and a GCP-VQVAE-based tokenizer for proteins. These structural codes are interleaved with standard text and sequence tokens, allowing the model to treat tasks like protein folding, molecule generation, and text-based property prediction as unified conditional generation problems. The model was built on the Qwen3 backbone (1.7B and 4B parameters) and pretrained on a massive 304.4-billion-token corpus of scientific text and biomolecular data.
BioMatrix demonstrates that a single, natively multimodal generalist model can outperform or match specialized models across a wide range of biological tasks. By avoiding external encoders and projection adapters, the model maintains a seamless flow between modalities. Empirical results show that the model excels particularly in cross-modal and cross-entity tasks, such as generating molecules from text descriptions, predicting protein structures from sequences, and calculating molecule-protein binding affinities. The unified architecture allows for flexible, instruction-tuned interaction, enabling the model to perform 80 distinct tasks across 6 categories with high accuracy.
This work represents a significant step toward a "generalist" biological AI. By reconciling the disparate ways we represent molecules and proteins, BioMatrix allows researchers to query biological systems using natural language while simultaneously leveraging the precision of 3D structural data. This unified approach simplifies the development of AI-driven drug discovery and protein engineering pipelines, as it removes the need for complex, task-specific model architectures.
We present BioMatrix, the first multimodal foundation model that natively integrates sequences, structures, and natural language for both molecules and proteins within a single decoder-only architecture. Existing biological foundation models pursue native multimodality and broad entity coverage separately: those that fuse multiple modalities under a shared objective remain confined to a single entity type, while those spanning multiple entity types either omit explicit structural modeling or rely on adapter-based designs in which the model cannot natively generate the very modalities it can read. BioMatrix closes this gap by mapping molecular sequences (supporting both SMILES and SELFIES notations), molecular structures, protein sequences, protein structures, and natural language into a shared discrete token space through a unified tokenization scheme, so that all modalities are consumed and produced uniformly under a single next-token prediction objective -- without external encoders, projection adapters, or modality-specific output heads. Built upon the Qwen3 language model (1.7B and 4B), BioMatrix is continually pretrained on 304.4 billion tokens spanning general and domain-specific text, sequence and structure views of molecules and proteins, and cross-modal corpora that interleave biomolecular entities with scientific text and link distinct entities through molecule-protein and protein-protein interaction data. After tuning on a comprehensive suite of downstream applications covering 80 tasks across 6 categories -- encompassing single-entity and multi-entity understanding and generation tasks across and within modalities -- BioMatrix achieves state-of-the-art or competitive performance on 77 out of 80 tasks, demonstrating that a single, natively multimodal generalist model can effectively match or surpass specialized approaches across a wide range of biological tasks.
Sam: So each data type gets its own specialist encoder, but the output of all those encoders feeds into one shared language the model can reason across.
Alex: That's the key innovation. It's what allows the model to handle 80 different biological tasks—from predicting a molecule's properties to generating a protein's three-dimensional structure—all within one architecture, rather than needing a separate system for each.
Sam: And how does it actually perform? Does being a generalist cost it anything compared to the specialized tools?
Alex: According to the paper, largely no. The paper suggests it matches or outperforms specialized tools on 77 out of 80 tasks tested. The authors argue this happens because training across all data types together allows the model to learn the underlying patterns that are common to biological data broadly—patterns that a narrowly trained specialist might miss.
Sam: That's a meaningful result. Though I'd imagine there are still cases where that generalist approach hits a wall?
Alex: There is one clear limitation the authors acknowledge. Because the system maps continuous, smooth three-dimensional shapes into a fixed set of codes, it introduces what's called quantization error. Here's a simple way to think about it: imagine trying to draw a smooth curve using only square tiles. You can get close, but the edges will always be slightly blocky. That's the kind of precision loss that happens when you convert real-world geometry into a finite digital vocabulary.
Sam: So for applications that need atomic-level exactness—like detailed physics simulations—that slight loss of detail could matter.
Alex: It could. The authors note, though, that for many practical applications, a short refinement step after the model generates a structure is enough to recover that precision. The model provides the broad design, and the refinement handles the fine detail.
Sam: So it's a genuine trade-off. You gain the ability to link text, sequences, and structures in one system, and you give up a small amount of geometric exactness—which can often be recovered anyway.
Alex: That's a fair summary. And the authors point to DNA and RNA as the natural next frontier—extending this shared vocabulary to cover even more of the biological world.
Sam: So the bottleneck wasn't the data itself. It was how we were encoding it. By giving everything a common language, the model can reason across boundaries that used to require entirely separate tools.
Alex: That's the most significant takeaway from this work. Whether a generalist architecture like this becomes the standard approach in biological AI remains to be seen, but this paper makes a considered case that unified, multimodal systems are worth taking seriously. Thanks for listening to ResearchPod.