lamm-mit TDC – Binding Effect Prediction
Predict binding-related effects for small molecules using pre-trained models from Therapeutics Data Commons (TDC) on Hugging Face. Uses SMILES as input and returns classification or scores.
Overview
- Blood–brain barrier (BBB): Will the compound cross the BBB? (binary)
- hERG blockade: Cardiotoxicity risk – does it block hERG? (binary)
- CYP3A4 inhibition: Metabolism – does it inhibit CYP3A4? (binary)
Models: AttentiveFP (graph), CNN, or Morgan fingerprints. Same task, different architectures.
Prerequisites
Install TDC and DeepPurpose (optional; needed for prediction). See ScienceClaw requirements.txt or:
pip install PyTDC DeepPurpose
pip install 'dgl' 'torch'
Usage
Run with the conda environment tdc (PyTDC/DGL are installed there). Use: conda run -n tdc python ... or activate the env first.
Predict with one model (SMILES required)
conda run -n tdc python {baseDir}/scripts/tdc_predict.py --smiles "CC(=O)OC1=CC=CC=C1C(=O)O" --model BBB_Martins-AttentiveFP
Predict hERG blockade (cardiotoxicity)
conda run -n tdc python {baseDir}/scripts/tdc_predict.py --smiles "CN1C=NC2=C1C(=O)N(C(=O)N2C)C" --model herg_karim-AttentiveFP
List available models
conda run -n tdc python {baseDir}/scripts/tdc_predict.py --list-models
Parameters
| Parameter | Description | Default |
|---|---|---|
--smiles |
Single SMILES string | - |
--smiles-file |
File with one SMILES per line | - |
--model |
TDC model name (see --list-models) | BBB_Martins-AttentiveFP |
--list-models |
Print available models and exit | - |
--format |
Output: summary, json | summary |
Notes
- First run downloads the model from Hugging Face (cached in
~/.scienceclaw/tdc_models). - Input must be valid SMILES; get SMILES from PubChem or ChEMBL if you have a name or ID.
- References: TDC, Hugging Face tdc.