DPD-Cancer: Explainable Graph-based Deep Learning for Small Molecule Anti-Cancer Activity Prediction
Summary: arXiv:2603.26114v1 Announce Type: cross
Abstract
Accurate drug response prediction is a critical bottleneck in computational biochemistry, limited by the challenge of modeling the interplay between molecular structure and cellular context. In cancer research, this challenge is particularly acute due to tumor heterogeneity and genomic variability, which hinder the identification of effective therapies. Conventional approaches often fail to capture non-linear relationships between chemical features and biological outcomes across diverse cell lines.
Introduction to DPD-Cancer
To address these challenges, researchers have introduced DPD-Cancer, a deep learning method based on a Graph Attention Transformer (GAT) framework. This innovative approach is designed specifically for small molecule anti-cancer activity classification and for the quantitative prediction of cell-line specific responses, particularly the growth inhibition concentration (pGI50).
Performance Metrics
DPD-Cancer has been benchmarked against several state-of-the-art methods including pdCSM-cancer, ACLPred, and MLASM. The results demonstrate superior performance for DPD-Cancer, achieving an Area Under ROC Curve (AUC) of up to 0.87 on strictly partitioned NCI60 data and up to 0.98 on both ACLPred and MLASM datasets.
Quantitative Predictions
For the prediction of pGI50 across 10 different cancer types and 73 cell lines, the model achieved Pearson’s correlation coefficients of up to 0.72 on independent test sets. These findings confirm the effectiveness of attention-based mechanisms in extracting meaningful molecular representations, establishing DPD-Cancer as a competitive tool for prioritizing drug candidates.
Explainability and Insights
One of the standout features of DPD-Cancer is its explainability. By leveraging the attention mechanism, the model is capable of identifying and visualizing specific molecular substructures. This feature provides actionable insights for lead optimization, allowing researchers to better understand why certain compounds exhibit anti-cancer activity.
Accessibility
DPD-Cancer is freely available as a web server, allowing researchers and practitioners easy access to its predictive capabilities. Interested users can explore the tool at the following link:
https://biosig.lab.uq.edu.au/dpd_cancer/.
Conclusion
The introduction of DPD-Cancer marks a significant advancement in the field of computational biochemistry and cancer research. By utilizing a graph-based deep learning approach and providing explainability, DPD-Cancer not only enhances the accuracy of drug response predictions but also offers valuable insights that can guide future research and drug development efforts.
Key Features of DPD-Cancer
- Graph Attention Transformer framework for enhanced molecular representation.
- High performance with AUCs up to 0.98 on benchmark datasets.
- Quantitative pGI50 predictions across multiple cancer types.
- Explainability through molecular substructure visualization.
- Freely accessible web server for users.