Using attribution to decode binding mechanism in neural network models for chemistry.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Deep neural networks (DNNs) excel at predicting molecular binding to protein targets.
• Identifying specific molecular fragments (pharmacophores) driving binding is crucial for drug discovery.
• Interpreting DNNs to reveal these binding mechanisms remains a significant challenge.

Purpose of the Study:

• To investigate the reliability of DNNs in revealing causal binding mechanisms.
• To assess whether DNNs learn genuine binding interactions or spurious correlations.
• To develop methods for interrogating DNNs and ensuring their interpretability.

Main Methods:

• Utilized a recently described attribution method to interrogate DNNs.
• Employed carefully constructed synthetic datasets with known binding features.
• Developed adversarial examples to test model robustness and identify non-spurious correlations.

Main Results:

• DNNs achieving high accuracy can still learn spurious correlations unrelated to actual binding.
• Adversarial examples successfully fooled models, demonstrating their unreliability.
• The learned mechanisms were found to be nonrobust, indicating potential flaws.

Conclusions:

• Current DNNs may be unreliable for accurately revealing protein-ligand binding mechanisms.
• A proposed test can identify if a DNN has learned a valid mechanism.
• Model simplification, regularization, or training data augmentation may be necessary for reliable interpretation.