Interpreting Machine Learning Pipelines Produced by Evolutionary AutoML for Biochemical Property Prediction

Machine learning (ML) has been playing a crucial role in drug discovery, mainly through quantitative structure-activity relationship models that relate molecular structures to properties, such as absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties. However, traditional ML approaches often lack customisation to a particular biochemical task and fail to generalise to new biochemical spaces, resulting in reduced predictive performance. Automated machine learning (AutoML) has emerged to address these limitations by automatically selecting the suitable ML pipelines for a given input dataset. Despite its potential, AutoML is underutilised in cheminformatics, and its decisions often lack interpretability, reducing user trust - especially among non-experts. Accordingly, this paper proposes an evolutionary AutoML method for biochemical property prediction that outputs an interpretable model for understanding the evolved ML pipelines. It combines grammar-based genetic programming with Bayesian networks to guide search and enhance the searched pipelines' interpretability. The evaluation on 12 benchmark ADMET datasets showed that the proposed AutoML method obtained similar or better results than three existing methods. Additionally, the interpretable Bayesian network identified, among the ML pipelines' components generated by the AutoML method (i.e. components like biochemical feature extraction methods, preprocessing techniques and ML algorithms), which components affect the ML pipelines' predictive performance.