Intuitive Explainable Artificial Intelligence for Molecular Design

Abstract

With the onset of artificial intelligence (AI), machine learning (ML) approaches have become increasingly embedded into the drug discovery pipeline with the aim of accelerating research and reducing cost. Exemplary applications include quantitative structure-property relationship modeling and <em>de novo</em> design of candidate molecules. However, the practical impact of these ML models has been limited by the opaque nature of their predictions, termed the black-box problem. The proliferation of large-scale public chemical datasets combined with advances in high-performance computing has driven a shift from traditional ML approaches toward more complex deep learning (DL) methods. While powerful, these models further exacerbate the inability to rationalize model outcomes. Therefore, the field of explainable AI (XAI) has attracted substantial interest to improve the transparency of model decisions. Given the wide range of available XAI approaches and lack of established practices, these methods are often applied without careful consideration for their suitability to the task at hand. This represents a problem in drug discovery where interpretable explanations are critical to guide experimental follow-up. To address this issue, this dissertation proposes the adaptation of established XAI approaches to operate on the level of chemical structure and facilitate interpretable ML for molecular design. The first study demonstrates the use of Shapley additive explanations (SHAP) to analyze feature sets driving the correct prediction of compound promiscuity. This is followed by a benchmarking study comparing several methods for the calculation of feature importance values, revealing inconsistencies across approaches that complicate the interpretation of feature-importance explanations. To support chemically meaningful interpretability, SHAP values are then combined with counterfactual reasoning, yielding the SHAP-CF methodology. Thereafter, a novel counterfactual method is introduced that generates candidate counterfactuals through the iterative recombination of molecular cores and substituents. This method is evaluated on a challenging multi-class kinase inhibitor prediction task, demonstrating its capacity to generate large numbers of counterfactuals, enabling in-depth analysis. Next, a closely related XAI concept, contrastive explanations, is adapted for molecular design by introducing rational chemical modifications to test compounds through the exchange of scaffolds or substituents with structural analogues, termed MolCE. This approach is shown to provide chemically relevant contrastive explanations that enable direct causal insights into the structural factors that drive model predictions. Finally, MolAnchor is introduced, a domain-adapted variant of the rule-based Anchors method. By conforming the principles underlying the Anchors framework to operate on retrosynthetically meaningful molecular fragments, MolAnchor produces interpretable fragment-level decision rules. A follow-up study showcases how the if-then formulation of the fragment rules is particularly complementary to causal reasoning. Taken together, the studies presented hereindemonstrate the utility of XAI methods that have been augmented with domain-specific information to provide chemically intuitive explanations for predictive models applied to molecular design.