Abstract
Computational models predicting the sites of metabolism (SOM) of small organic molecules have become invaluable tools for studying and optimizing the metabolic properties of xenobiotics. However, the performance of SOM predictors has shown signs of plateauing in recent years, primarily due to the limited availability of training data. While vast amounts of biotransformation data in the form of substrate-metabolite pairs exist, their potential for SOM prediction remains largely untapped due to the absence of annotations. Annotating SOMs requires expert knowledge and is highly time-consuming. To address this challenge, we introduce AutoSOM, the first open-source tool that automatically extracts SOMs by mapping structural differences using transformation rules. AutoSOM is both fast and highly accurate, achieving over 90% labeling accuracy on a diverse validation set of 5,000+ reactions within minutes. Moreover, its annotation process is fully transparent and interpretable, which we hope will facilitate its adoption in high-stakes downstream applications such as drug discovery campaigns and regulatory assessments. Beyond accelerating annotation, AutoSOM enables standardized and consistent SOM labeling across institutions without requiring direct data sharing. This capability lays the foundation for federated learning approaches in metabolism prediction, fostering collaborative model improvement while preserving data confidentiality.
Supplementary materials
Title
Supporting information file 1
Description
PDF document detailing the software used in this work, the composition of the data set, the data preprocessing steps, and a discussion of representative examples of annotated substrate-metabolite pairs
Actions
Title
Supporting information file 2
Description
CSV file containing the complete list of evaluated MetaTrans substrate-metabolite pairs
Actions