BioMatrix-4B-SFT

BioMatrix is a multimodal biological foundation model that natively integrates 1D sequences, 3D structures, and natural language for both molecules and proteins within a single decoder-only architecture.

This is the 4B-parameter SFT (Supervised Fine-Tuned) variant, instruction-tuned across 80 downstream biological tasks spanning 6 categories.

• 📄 Paper: BioMatrix: Towards a Comprehensive Biological Foundation Model Spanning the Modality Matrix of Sequences, Structures, and Language
• 💻 Code: https://github.com/QizhiPei/BioMatrix
• 🤗 Model & Data Collection: https://huggingface.co/collections/QizhiPei/biomatrix

Model Overview

BioMatrix closes the gap between native multimodality and broad entity coverage in biological foundation models. Unlike adapter-based approaches that bolt external encoders onto a language model, or prior native-tokenization models confined to a single entity type, BioMatrix maps all modalities into a shared discrete token space via a unified tokenization scheme:

• Molecular 1D sequences (both SMILES and SELFIES notations)
• Molecular 3D structures (via MolStrucTok with branch-decoupled decoder)
• Protein 1D sequences (residue-level tokens)
• Protein 3D structures (via GCP-VQVAE backbone tokenizer)
• Natural language (inherited from Qwen3 tokenizer)

All modalities are consumed and produced uniformly under a single next-token prediction objective—without external encoders, projection adapters, or modality-specific output heads.

Model Details

• Base Architecture: Qwen3-4B-Base
• Parameters: 4B
• Training Stages:
  • Continual Pretraining on 304.4B tokens (general/scientific text, molecular & protein 1D/3D data, cross-modal interleaved corpora)
  • Instruction Tuning on a comprehensive suite of 80 downstream tasks across 6 categories
• Context Length: 8,192 tokens
• Tokenizer: Extended Qwen3 vocabulary with:
  • 11,294 joint molecular 3D tokens (composed from SELFIES atom × MolStrucTok codes)
  • 4,096 protein 3D tokens (GCP-VQVAE codebook)
  • 26 protein 1D tokens (amino acids + non-standard/unknown)
  • SELFIES atom tokens and modality-specific control tokens

Pretraining Corpus (304.4B tokens)

Performance Highlights

BioMatrix achieves state-of-the-art or competitive performance on 77 out of 80 tasks. Selected highlights for the 4B-SFT variant:

Molecular Tasks

• Unconditional 1D Generation (GuacaMol): 0.998 validity, 1.000 uniqueness, 0.986 novelty
• Name Conversion (I2S EM): 92.83% (vs. SciReasoner-8B: 84.40%)
• Text-Based Molecule Generation (EM): 65.07% (vs. SciReasoner-8B: 48.00%)
• MoleculeQA Total Accuracy: 73.78% (vs. prior best MolCA-1.3B: 64.79%)
• Property-Conditioned 3D Generation: ~3-4× error reduction on QM9 electronic-structure targets

Protein Tasks

• Fold Type Prediction (Family level): 87.25% accuracy
• Annotation Prediction (UniProtSeq Keywords F1): 91.26%
• Inverse Folding AAR: 75.50% (vs. DPLM-2-3B: 61.67%)
• Sequence–Structure Co-generation: scTM = 0.965, scRMSD = 2.80
• Unconditional Backbone Generation: scTM = 0.963 (joint frontier with RFDiffusion)

Interaction Tasks

• BindingDB Affinity (RMSE): 1.030 (new SOTA, surpasses prior literature SOTA of 1.340)
• PDBBindv2020 3D Affinity: RMSE = 1.260, Pearson = 0.737, MAE = 0.972

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "QizhiPei/BioMatrix-4B-SFT"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)

# Example: Molecule captioning with SELFIES input
instruction = "I need a brief explanation of the molecule denoted in this SELFIES notation. <|mol_sfi_start|>[Te]<|mol_sfi_end|>"

messages = [
    {"role": "user", "content": instruction}
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=False)
print(response)

Modality Wrapping

When constructing prompts, biomolecular content must be wrapped with the corresponding control tokens:

Natural language text is left unwrapped and serves as the default carrier modality.

Supported Tasks

BioMatrix-4B-SFT was instruction-tuned across the following task categories:

Molecule (1D): unconditional generation, name conversion, property prediction, captioning, text-based generation, forward/retrosynthesis, editing, optimization, customized generation, question answering

Molecule (3D): unconditional generation, property-conditioned generation

Protein (1D): sequence understanding, annotation prediction, knowledge mining, text-based design, unconditional generation

Protein (3D): structure understanding, folding, inverse folding, sequence-structure co-generation, unconditional backbone generation

Interaction: molecule-protein binding affinity (1D & 3D), protein-protein interaction

Note on task-group variants: As detailed in the paper, the released SFT model is trained on the union of all sub-task corpora with mild oversampling for small-data tasks. For best performance on specific benchmarks, please refer to the paper's task-group-specific variants.

SMILES vs. SELFIES

BioMatrix supports both notations as parallel 1D molecular representations. Empirically:

• SELFIES excels on tasks requiring validity-by-construction (unconditional generation, property optimization)
• SMILES excels on tasks requiring surface-level structural anchoring (customized generation with atom/bond/functional-group constraints, forward synthesis, retrosynthesis)

See Section 9.2 of the paper for detailed analysis.

Limitations

• Molecular and protein 3D structures are tokenized in disjoint geometric reference frames, so the model cannot natively represent biomolecular complexes (e.g., docking poses).
• Heavy domain specialization may erode some general-purpose language capabilities of the underlying Qwen3 backbone.
• Coverage is limited to small molecules and proteins; nucleic acids, carbohydrates, and lipids are not currently supported.
• Fine-grained 3D geometry (e.g., bond lengths) shows residual quantization error from finite codebooks; a lightweight post-hoc force-field refinement (e.g., MMFF) closes most of this gap.

Citation

If you find BioMatrix useful, please cite:

@article{pei2026biomatrix,
  title={BioMatrix: Towards a Comprehensive Biological Foundation Model Spanning the Modality Matrix of Sequences, Structures, and Language},
  author={Pei, Qizhi and Zhou, Zhimeng and Duan, Yi and Zhao, Yiyang and Li, Wei and Guo, Han and He, Liang and Li, Chengping and Hsieh, Chang-Yu and He, Conghui and Yan, Rui and Wu, Lijun},
  journal={arXiv preprint arXiv:2606.22138},
  year={2026}
}

License

This model is released under the Apache 2.0 license. The base model (Qwen3-4B-Base) is subject to its own license terms.