The Tsetlin Machine: A "Third Way" in QSAR Modeling.

Area of Science:

• Computational Chemistry and Cheminformatics
• Machine Learning in Drug Discovery
• Quantitative Structure-Activity Relationship (QSAR) Modeling

Background:

• Current Quantitative Structure-Activity Relationship (QSAR) approaches rely on descriptor engineering with conventional machine learning or deep learning on graphical inputs.
• Existing methods often lack intrinsic interpretability or require complex feature engineering.
• There is a need for QSAR models that combine accuracy, efficiency, and interpretability for virtual screening.

Purpose of the Study:

• To introduce the Tsetlin Machine (TM) as a novel QSAR methodology.
• To evaluate the performance of TM-QSAR against established QSAR methods.
• To demonstrate the interpretability features of TM-QSAR for understanding molecular bioactivity.

Main Methods:

• The Tsetlin Machine (TM), utilizing finite-state automata and reinforcement learning, was employed for QSAR modeling.
• TM-QSAR was benchmarked using ECFP4 descriptors and compared against Random Forest (RF) and XGBoost.
• Interpretability was assessed using molecule property maps (TM-MPM) and Weights × Activations × Clauses (WAC) scores.

Main Results:

• TM-QSAR coupled with ECFP4 descriptors outperformed RF and XGBoost in ROC-AUC, PRC-AUC, and PPV, showing strong interscaffold generalization.
• Particularly high classification scores were achieved for MOR (ROC-AUC = 0.87, PRC-AUC = 0.77) and CYPA4 (ROC-AUC = 0.92, PRC-AUC = 0.63).
• Interpretability methods revealed atom-wise contributions and conditional relationships, aligning with known ligand-protein interactions for MOR.

Conclusions:

• The Tsetlin Machine (TM) provides an accurate, computationally efficient, and highly interpretable QSAR methodology.
• TM-QSAR demonstrates significant potential for virtual screening applications, offering advantages over existing rule-based and deep learning methods.
• The developed interpretability tools enable deeper insights into molecular bioactivity predictions.