QizhiPei / BioMatrix-4B-Base

BioMatrix-4B-Base

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 Base model , obtained via multimodal continual pretraining of Qwen3-4B-Base on 304.4 billion tokens spanning text, molecular and protein 1D/3D data, and cross-modal corpora. This base checkpoint is intended for further fine-tuning on downstream tasks. For an instruction-tuned model ready for inference, see BioMatrix-4B-SFT .

• 📄 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 maps all biological 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 Stage : Multimodal Continual Pretraining only (not instruction-tuned)
• Training Tokens : 304.4B
• 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

Embedding Initialization

New vocabulary entries are initialized via a description-based scheme : each new token is grounded in the pretrained Qwen3 embedding space by averaging the embeddings of the subword tokens of a short natural-language description (e.g., <A_W> → "Tryptophan"), plus a small isotropic Gaussian perturbation to break symmetry. This provides a more stable starting point than random initialization.

Pretraining Corpus (304.4B tokens)

Training Configuration

• Framework : LLaMA-Factory
• Hardware : 64 NVIDIA H100 GPUs
• Global Batch Size : 1,024
• Maximum Sequence Length : 8,192 tokens
• Optimizer : AdamW
• Peak Learning Rate : 2.0 × 10⁻⁴ (cosine schedule)
• Warmup Steps : 2,000
• Total Steps : ~36.4K (1 epoch over the full 304.4B-token corpus)

Intended Use

This Base model is not instruction-tuned . It is suitable for:

• Further fine-tuning on custom biological tasks
• Continued pretraining on domain-specific corpora
• Research on representation learning across biomolecular modalities
• Embedding extraction for downstream classification/regression tasks

For ready-to-use instruction-following capabilities (e.g., molecule captioning, protein design, property prediction), please use the SFT variant .

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "QizhiPei/BioMatrix-4B-Base"
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: Continue a SMILES sequence
prompt = "<|mol_smi_start|>CC(=O)"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=False))

Modality Wrapping

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

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

Limitations

• This model is not instruction-tuned and is unlikely to follow natural-language instructions out-of-the-box. Use the SFT variant for instruction-following.
• 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.

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 He, Liang and Hsieh, Chang-Yu and He, Conghui and Yan, Rui and Wu, Lijun},
  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.