A Rigorous Statistical Benchmark of the Tsetlin Machine: A Novel Rule-Based Method Exceeding SOTA for QSAR Prioritisation

Quantitative Structure Activity Relationship (QSAR) remains an effective tool for early-stage chemical modelling and virtual screening in drug design. The advancements in this field are led by two core paradigms, 1) descriptor engineering, where complex fixed-length vectors of compounds are generated and conventional ML methods are applied to those representations and 2) raw chemical inputs (e.g., SMILES, 2D-graph) are provided to complex deep learning neural network models which construct their own internal representations of molecules to quantify their biological activity. Conventional ML uses simple, computationally efficient algorithms and achieves state of the art accuracy on dataset sizes typical of practical QSAR applications. NN-based ML may not be as accurate for these dataset sizes but is inherently flexible and can learn iteratively. Here we present the Tsetlin Machine as a new potential QSAR method - which combines the accuracy and easy-use of existing rule-based QSAR ML methods (e.g., RF and XGBoost) with the continuous learning aspect and flexibility of NN algorithms. The Tsetlin Machine uses teams of finite-state automata which capture frequent patterns as propositional logic (clauses) in a reinforcement learning based game. We compare the TM's performance to existing rule-based QSAR methods using a benchmarking pipeline developed by the Polaris group. It uses repeated measures ANOVA and Tukey’s HSD to measure the difference in mean performance distributions. Across all tested targets, compound groups and descriptor types, the TM consistently achieved high ROC AUC and PPV scores - making it well-suited for QSAR-based compound prioritization, especially in early stages of virtual screening pipelines. SOTA results were achieved with the TM for MOR (ROC-AUC = 0.87, PRC-AUC = 0.77) and CYPA4 (ROC-AUC = 0.92, PRC-AUC = 0.63), demonstrating comparable and in certain cases superior performance of TM over the established SOTA models. Additionally, this work contains practical guidelines for applying TMs to QSAR tasks including descriptor selection, model parameters and number of training epochs that were used to produce the reported SOTA results.