Topology-Aware Deep Learning on Higher-Order Structures for Drug Response Prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning in Oncology

Background:

• Accurate prediction of anticancer drug response is crucial for precision oncology.
• Current methods often use pairwise modeling, failing to capture complex drug-cell line interactions.
• Higher-order dependencies among drugs and cell lines are overlooked in existing approaches.

Purpose of the Study:

• To develop a novel deep learning framework, TopDr, for enhanced prediction of anticancer drug response.
• To incorporate multiscale topological structures for richer data representation.
• To provide mechanism-level interpretability for drug response predictions.

Main Methods:

• TopDr encodes drugs and cell lines as multiscale simplicial complexes, capturing interactions at 0-, 1-, and 2-simplex levels.
• The framework integrates local neighborhoods and global topological structures for enriched representations.
• Performance was evaluated on six benchmark datasets using regression and classification tasks.

Main Results:

• TopDr consistently matched or surpassed state-of-the-art baselines across multiple datasets.
• The model demonstrated robust performance in both regression and classification tasks.
• TopDr provided mechanism-level interpretability, identifying significant pathway enrichments and biologically coherent cell line modules.

Conclusions:

• Modeling multiscale higher-order topology significantly improves the accuracy and robustness of drug response prediction.
• TopDr offers valuable biological interpretability, aiding in understanding drug mechanisms and cell line responses.
• This topology-aware approach paves the way for more reliable precision oncology drug modeling.