Calculation of exact Shapley values for support vector machines with Tanimoto kernel enables model interpretation

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Support Vector Machine (SVM) algorithms are widely used in chemistry and drug discovery but often function as black boxes.
• Interpreting SVM predictions typically involves feature weighting or model-agnostic methods like Shapley Additive Explanations (SHAP), which approximate Shapley values (SVs).
• Accurate explanation of complex models is crucial for rationalizing predictions in sensitive fields like drug discovery.

Purpose of the Study:

• To introduce a novel algorithm, SV-expressed Tanimoto similarity (SVETA), for the exact calculation of Shapley values (SVs) in SVM models.
• To assess the reliability of exact SVs compared to approximate SHAP values in a drug discovery context.
• To demonstrate the utility of SVETA in providing consistent and interpretable explanations for SVM-based compound classification.

Main Methods:

• Development and implementation of the SV-expressed Tanimoto similarity (SVETA) algorithm for exact SV calculation.
• Application of SVETA to an SVM model utilizing the Tanimoto kernel for molecular similarity assessment.
• Comparison of exact SVs derived from SVETA with SHAP values in an SVM-based compound classification task.
• Atom-based mapping of prioritized features to identify substructures responsible for predictions.

Main Results:

• The SVETA algorithm enables the exact calculation of Shapley values for SVM models with Tanimoto kernels.
• A limited correlation was observed between exact SVs and SHAP values in a drug discovery classification task, indicating potential limitations of SHAP for rationalizing these specific SVM predictions.
• Atom-based feature mapping using exact SVs identified coherent substructures, consistent with explanations from independent Random Forest models.

Conclusions:

• SVETA provides an accurate method for explaining SVM models in cheminformatics and drug discovery.
• Exact Shapley values are essential for reliable interpretation, as approximate methods like SHAP may not fully capture the model's decision-making process.
• The consistency of explanations across different modeling techniques (SVM with SVETA and Random Forest) enhances confidence in the identified molecular substructures.