README.md

metadata

license: mit
tags:
  - biology
  - protein-design
  - ligand-binding
  - antibody-design

InstructPro2

Natural Language and Biological Grounding Guided De Novo Protein Design

Paper · Code · Dataset

InstructPro2 designs proteins from natural-language function descriptions. A text encoder reads the instruction, frozen ligand (ChemBERTa) and antigen (ESM-2) encoders supply biological context through gated cross-attention adapters, and a 3.2B-parameter decoder initialized from ProGen2-large-BFD90 generates the sequence.

Checkpoints

File	Parameters	Task
instructpro2_stage1.pt	2.87B	design from text; also the starting point for finetuning
instructpro2_stage2_ligand.pt	3.22B	ligand-binding design, conditioned on SMILES
instructpro2_stage2_antibody.pt	3.22B	HCDR1-3 design, conditioned on an antigen
progen2/progen2-large-BFD90/	2.7B	base decoder weights

progen2/progen2-large-BFD90/ is required by all three models, including at inference — the decoder is built from it before the checkpoint loads.

Usage

Clone the code and follow its README to create the environment, then:

export WORK_ROOT=/path/to/workspace

huggingface-cli download charlesnovak/InstructPro2 --local-dir $WORK_ROOT/model
huggingface-cli download charlesnovak/InstructPro2Bench --repo-type dataset \
    --local-dir $WORK_ROOT/data

# design from a description
python3 scripts/prepare_inputs.py --task ligand \
    --input examples/ligand_designs.csv \
    --out $WORK_ROOT/my_designs/test_examples.jsonl

MODEL=ligand DATA_PATH=$WORK_ROOT/my_designs SUBSET=test_examples bash infer.sh

infer.sh selects the architecture mode, dataset reader and generation flags that match each checkpoint, so prefer it over calling the underlying fairseq entry point directly.

The whole repository is 24 GB. To take only one task's weights:

huggingface-cli download charlesnovak/InstructPro2 \
    --include "instructpro2_stage2_ligand.pt" "progen2/*" \
    --local-dir $WORK_ROOT/model

Evaluation

Stage 1 designs are folded with ESMFold and scored against the ground-truth cluster structures published with the dataset. eval_stage1.sh reproduces that end to end.

The ligand and antibody results in the paper are computed from AlphaFold3 complexes — ligand designs folded with their SMILES, antibody designs with the light chain and antigen. That code is available on request.

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}
}

License

MIT. The base ProGen2 weights are BSD-3-Clause.