DPD-Cancer: Graph-Attention AI Predicts Anti-Cancer Activity Across NCI-60 Panel
DPD-Cancer is a deep learning framework utilizing graph attention to forecast the anti-cancer efficacy of small molecules within the NCI-60 panel. Following a rigorous chemistry-aware data-partitioning approach during training, it recorded an AUROC of 0.87 and an AUPRC of 0.73 on the hold-out test set. Regression models for 73 individual cell lines produced a median Pearson's R of 0.64 and a median RMSE of 0.67 for predicting pGI50 values. When benchmarked against pdCSM-Cancer, MLASM, and ACLPred, DPD-Cancer consistently achieved superior MCC scores. An occlusion-based attribution analysis verified that the model successfully identifies crucial molecular substructures influencing its predictions.
Key facts
- DPD-Cancer is a graph-attention deep learning framework for predicting small-molecule anti-cancer activity across the NCI-60 panel.
- Trained under a strict chemistry-aware data-partitioning scheme.
- Achieved AUROC of 0.87 (95% CI [0.86, 0.88]) and AUPRC of 0.73 (95% CI [0.70, 0.76]) on hold-out test set.
- Per-cell-line regression models for 73 cell lines produced median Pearson's R of 0.64 and median RMSE of 0.67 for pGI50-value prediction.
- Benchmarks against pdCSM-Cancer, MLASM, and ACLPred yielded consistently higher MCC scores.
- Occlusion-based attribution analysis confirmed model identifies key molecular substructures.
Entities
—