InstructPro2

Model weights for InstructPro2: Natural Language and Biological Grounding Guided De Novo Protein Design — a 3.2B protein design model that generates sequences from natural-language function descriptions, optionally conditioned on a ligand or an antigen.

Code: chovak/InstructPro2 · Data: charlesnovak/InstructPro2Bench · Paper: arXiv:2610.19592

Download

export WORK_ROOT=/path/to/workspace    # default: ~/instructpro2
huggingface-cli download charlesnovak/InstructPro2 --local-dir $WORK_ROOT/model

Checkpoints

File / Directory	Parameters	Use
instructpro2_stage1.pt	2.87B	Pretrained on InstructPro2Bench. Text-conditioned generation, and the starting point for finetuning.
instructpro2_stage2_ligand.pt	3.22B	Ligand-binding protein design, conditioned on SMILES.
instructpro2_stage2_antibody.pt	3.22B	Antibody design — HCDR1–3 infilling conditioned on an antigen.
progen2/progen2-large-BFD90/	2.7B	Base ProGen2 decoder configuration and weights (Nijkamp et al., 2023; BSD-3-Clause), used for model initialization and Stage 1 pretraining.

The Stage 2 checkpoints are larger because they carry the gated cross-attention adapters and the frozen ChemBERTa and ESM-2 encoders.

Use

Pass a checkpoint to the scripts in the code repository:

MODEL_CHECKPOINT=$WORK_ROOT/model/instructpro2_stage1.pt          bash validate_stage1.sh
MODEL_CHECKPOINT=$WORK_ROOT/model/instructpro2_stage2_ligand.pt   bash eval_ligand.sh
MODEL_CHECKPOINT=$WORK_ROOT/model/instructpro2_stage2_antibody.pt bash eval_antibody.sh

The scripts set the architecture mode and the generation settings for each checkpoint; see the code repository's README.

Citation

@article{instructpro2,
  title   = {InstructPro2: Natural Language and Biological Grounding Guided De Novo Protein Design},
  author  = {Charles Novak and Zhenqiao Song and Chuan Li and Jianwen Xie and Lei Li},
  journal = {arXiv preprint arXiv:2610.19592},
  year    = {2026}
}