Ensemble modeling with machine learning and deep learning to provide interpretable generalized rules for classifying...

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Machine learning in drug discovery

Background:

• The trade-off between predictability and interpretability is a key challenge in central nervous system quantitative structure-activity relationship (CNS-QSAR) analysis.
• Many advanced predictive models lack interpretability, hindering the derivation of simple rules for drug design.
• Existing CNS-QSAR models often fail to provide clear structural insights or mechanistic understanding.

Purpose of the Study:

• To develop a novel protocol combining machine learning (ML) and deep learning (DL) for CNS-QSAR.
• To generate a set of simple, interpretable rules with high predictive power for CNS drug activity.
• To overcome the limitations of black-box models in CNS-QSAR analysis.

Main Methods:

• A hybrid ensemble protocol integrating ML (Support Vector Machine) and DL (Graph Convolutional Network) algorithms was developed.
• A dataset of 940 market drugs (315 CNS-active, 625 CNS-inactive) was used for model training.
• Model performance was rigorously evaluated using an external dataset of 117 market drugs and fingerprint-split validation.

Main Results:

• The novel hybrid ensemble model achieved high performance, with an accuracy of 0.96 and an F1 score of 0.95.
• The protocol successfully generated simple physicochemical rules based on six sub-structural features for CNS drug prediction.
• These rules demonstrated superior classification ability, specificity, and mechanistic insights compared to classical guidelines.

Conclusions:

• The developed hybrid protocol effectively balances predictability and interpretability in CNS-QSAR.
• The generated rules offer valuable mechanistic insights beyond simple blood-brain barrier permeability predictions.
• This approach holds potential for application in predicting other drug properties.