nabbo/tankbind-modal

TankBind Modal

Modal-first inference wrapper for TankBind, using A100 GPU workers and checkpoints mirrored on Hugging Face for faster container builds.

Model weights: nabbo/tankbind-modal

Credits

This repository is based on TankBind from Lu, Wei; Wu, Qifeng; Zhang, Jixian; Rao, Jiahua; Li, Chengtao; and Zheng, Shuangjia:

TankBind: Trigonometry-Aware Neural NetworKs for Drug-Protein Binding Structure Prediction, NeurIPS 2022.

Original project: TankBind

What Changed

• Refactored the repo for Modal inference instead of local notebook workflows.
• Moved TankBind checkpoints to Hugging Face: nabbo/tankbind-modal.
• Updated Modal images to pull self_dock.pt from Hugging Face during image build.
• Updated panel inference to install p2rank in the Modal image instead of vendoring the local tool directory.
• Kept input/ and output/ as the user-facing data folders.

Inputs

Put proteins and ligands under:

input/structures/   # protein structures, PDB format
input/ligands/      # ligand structures, SDF format

TankBind expects structure files, not PDF or SVG files. Use:

• Protein: .pdb
• Ligand: .sdf

For panel inference, prepare a CSV like input/panel_prediction_inputs.csv:

id,protein_path,ligand_path
pik3ca_wt_alpelisib,input/structures/pik3ca_wt.pdb,input/ligands/alpelisib.sdf

Paths must be relative to the repository root.

Local CLI Setup

You only need a lightweight local environment for the Modal CLI. The model runs remotely on Modal.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
modal setup

If you use Modal profiles:

export MODAL_PROFILE=csblabigib

Single Pair Inference

modal run scripts/modal_tankbind.py \
  --protein input/structures/pik3ca_wt.pdb \
  --ligand input/ligands/alpelisib.sdf

Or:

./scripts/run_modal_tankbind.sh \
  --protein input/structures/pik3ca_wt.pdb \
  --ligand input/ligands/alpelisib.sdf

Output is printed as JSON:

{
  "affinity_pred": 7.12,
  "device": "cuda"
}

Panel Inference

Run many protein-ligand pairs from a CSV:

modal run scripts/modal_run_prepared_panel.py \
  --input-csv input/panel_prediction_inputs.csv \
  --run-id panel_a100 \
  --out-dir output/modal_panel

The panel runner:

1. Stages input files into a Modal volume.
2. Runs p2rank pocket detection for each unique protein.
3. Scores each protein-ligand pair on A100 workers.
4. Writes summaries to output/modal_panel/.

Expected outputs:

output/modal_panel/<run_id>_pair_affinity_summary.csv
output/modal_panel/<run_id>_run_summary.json

Checkpoints

The Modal build downloads checkpoints from:

https://huggingface.co/nabbo/tankbind-modal

Uploaded files:

self_dock.pt
re_dock.pt

Only self_dock.pt is used by the current Modal inference scripts.

Citation

@article{lu2022tankbind,
  title={Tankbind: Trigonometry-aware neural networks for drug-protein binding structure prediction},
  author={Lu, Wei and Wu, Qifeng and Zhang, Jixian and Rao, Jiahua and Li, Chengtao and Zheng, Shuangjia},
  journal={Advances in Neural Information Processing Systems},
  year={2022}
}