Interpretable deep learning architectures for improving drug response prediction performance: myth or reality?

Area of Science:

• Biomedical informatics
• Computational biology
• Machine learning in drug discovery

Background:

• Interpretable deep learning (DL) models are crucial for drug response prediction (DRP) in the biomedical field.
• Recently, DL models integrating biological signaling pathways have emerged for DRP, aiming to enhance interpretability.
• The impact of pathway integration on DRP accuracy and generalizability remains an open question.

Purpose of the Study:

• To systematically evaluate state-of-the-art interpretable DL models for DRP.
• To compare models that explicitly versus implicitly incorporate pathway information.
• To assess model performance and generalizability on independent datasets.

Main Methods:

• Comprehensive assessment of four interpretable DL models using three pathway collections.
• Evaluation of prediction accuracy on unseen samples and generalizability to an independent dataset.
• Comparison with black-box models (multilayer perceptron) and baseline (random forests).

Main Results:

• Models explicitly incorporating pathway information performed worse than those with implicit integration.
• Black-box multilayer perceptron models generally achieved the best performance.
• Random forests baseline performance was comparable to interpretable models; random pathways yielded similar results.
• Model performance significantly decreased when applied to an independent dataset.

Conclusions:

• Explicit pathway integration in DL models does not necessarily improve drug response prediction accuracy.
• Systematic evaluation with appropriate baselines is essential for assessing new DL models.
• Black-box models and simpler methods may offer competitive or superior performance in DRP.
• Generalizability to independent datasets remains a significant challenge for current interpretable DL models.