TANKBind

Overview

TANKBind accompanies the NeurIPS 2022 paper “TANKBind: Trigonometry-Aware Neural NetworKs for Drug-Protein Binding Structure Prediction.” The repository is intended to support reproduction of the paper’s results and further research built on the model. Its documented prediction targets are protein–ligand binding structure and binding affinity, positioning it as a predictive component in computational drug-discovery workflows rather than an experimental binding assay.

The project README organizes practical work around notebooks. Its structure-prediction example uses ABL1 with Imatinib and compound6, identified by PDB: 6HD6. Other notebooks cover construction of PDBbind training and test datasets, self-docking evaluation, and high-throughput virtual screening against the WDR domain of LRRK2. At a workflow level, the model operates on protein and ligand information to produce binding predictions; the excerpt does not specify complete input schemas, preprocessing requirements or output file formats. Those details need to be checked in the notebooks before adapting the examples.

The installation guidance names a Python-based environment, PyTorch and several scientific dependencies, alongside an external p2rank download. It also notes that the CUDA toolkit choice may need adjustment for the GPU. The README directs users to a separate account-based service for a newer model experience, so that service should not be treated as identical to the repository implementation. Reported screening throughput is a source claim, not an independently established benchmark for another dataset or hardware configuration.

Key Features

• Predicts protein–ligand binding structures and binding affinity.
• Provides a structure-prediction notebook using ABL1, Imatinib and compound6 with PDB: 6HD6.
• Includes a notebook for reproducing the paper’s self-docking test-set evaluation.
• Provides a notebook for constructing PDBbind training and test datasets, plus a documented model-training invocation.
• Includes a high-throughput virtual-screening notebook for the WDR domain of LRRK2.

Use Cases

• Intended evaluation: use the ABL1 example to assess whether the documented structure-prediction workflow can be adapted to a research protein–ligand pair.
• Intended evaluation: reconstruct the documented datasets and run the self-docking notebook to investigate reproducibility of the paper’s evaluation.
• Intended evaluation: explore the LRRK2 WDR screening notebook as a starting point for computational candidate prioritization, measuring throughput and predictive usefulness on the intended workload.

How to Use

1. Read the official README to choose between structure prediction, self-docking evaluation, dataset construction and virtual screening. Keep the repository workflow separate from the linked account-based service.
2. Follow the README’s environment setup. It specifies Python 3.8 and scientific dependencies, and advises adjusting the CUDA toolkit choice for the GPU; these instructions are not evidence of compatibility with other configurations.
3. Check the documented p2rank download and inspect the selected notebook for its actual input and preprocessing requirements before preparing research data.
4. Begin with examples/prediction_example_using_PDB_6hd6.ipynb for the ABL1 example, or examples/high_throughput_virtual_screening_LRRK2_WDR.ipynb for the screening workflow. Examine the generated outputs before adapting either example.
5. For reproduction work, inspect examples/construction_PDBbind_training_and_test_dataset.ipynb.ipynb, then examples/testset_evaluation_cleaned.ipynb. Record local settings and measured results rather than assuming the README’s reported throughput transfers to your setup.