Evaluating molecular representations in machine learning models for drug response prediction and interpretability.

Area of Science:

• Computational chemistry
• Chemoinformatics
• Machine learning in drug discovery

Background:

• Machine learning (ML) is crucial for modern drug discovery.
• Traditional ML methods rely on precomputed molecular descriptors or fingerprints.
• End-to-end deep learning (DL) offers an alternative by learning representations directly from molecular data.

Purpose of the Study:

• To compare the suitability of various compound representation methods for drug sensitivity prediction in cancer cell lines.
• To evaluate the performance of end-to-end DL models against traditional fingerprint-based approaches.
• To assess the impact of ensemble methods and feature attribution on predictive performance and explainability.

Main Methods:

• Benchmarking twelve different compound representation methods.
• Utilizing the DeepMol chemoinformatics package for analysis.
• Testing on five diverse compound screening datasets for cancer cell line drug sensitivity prediction.

Main Results:

• End-to-end DL models achieved predictive performance comparable to, and sometimes exceeding, models using molecular fingerprints.
• This advantage was observed even with limited training data.
• Ensemble methods combining multiple representations improved overall performance.

Conclusions:

• End-to-end DL methods are highly effective for drug sensitivity prediction, offering a competitive alternative to traditional feature engineering.
• Ensembling representations and employing post hoc feature attribution can enhance model performance and interpretability.
• The findings support the adoption of DL for more efficient and explainable drug discovery pipelines.