Differences in learning characteristics between support vector machine and random forest models for compound...

Area of Science:

• Molecular machine learning
• Computational chemistry
• Cheminformatics

Background:

• Random Forest (RF) and Support Vector Machine (SVM) are key algorithms in molecular machine learning (ML) for predicting compound properties.
• Understanding the prediction mechanisms of these ML models is crucial for reliable compound property prediction.

Purpose of the Study:

• To investigate the prediction mechanisms of RF and SVM binary classification models in molecular ML.
• To apply and extend explainable artificial intelligence (XAI) techniques, specifically Shapley values, for model interpretability.

Main Methods:

• Utilized RF and SVM algorithms for large-scale activity-based compound classification.
• Employed Shapley value analysis, adapted from game theory, to interpret model predictions.
• Trained models using training sets of increasing size to observe learning characteristics.

Main Results:

• RF and SVM models with the Tanimoto kernel produced highly similar predictions in compound classification tasks.
• Shapley value analysis demonstrated systematic differences in the learning characteristics of RF and SVM.
• Chemically intuitive explanations for accurate predictions derived from RF and SVM models originated from different feature attributions.

Conclusions:

• Despite similar predictive performance, RF and SVM models exhibit distinct internal learning processes in molecular ML.
• XAI methods like Shapley values are essential for uncovering these underlying differences and providing chemically meaningful interpretations.
• This research enhances the understanding of how ML models make predictions in drug discovery and chemical informatics.