Key points are not available for this paper at this time.
Background/Objectives: Due to the utility of knowing the pathway involvement of metabolites detected in biological experiments, knowledgebases such as the Kyoto Encyclopedia of Genes and Genomes (KEGG), Reactome, and MetaCyc have annotated compound entries to specific pathways defined by the knowledgebase. However, these compound-pathway annotations are largely incomplete and are costly to obtain experimentally or curate from published scientific literature. This metabolite-pathway annotation incompleteness problem is amenable to machine learning (ML)-based solutions. But to date, no machine learning model has been trained on all three knowledgebases to maximize its performance and robustness. This may be due to inconsistencies in chemical structure representation that can confuse a model and greatly reduce generalizability. Methods: We constructed a new training dataset with roughly 50,000,000 entries using compound-pathway annotations derived from KEGG, Reactome, and MetaCyc. We trained and tested a multitask classification, graph convolutional neural network-like model that classifies compound involvement with 8056 pathways that have unique pathway representations, based on annotated compound chemical structures represented with chemical substructure features. While the initial dataset contained inconsistencies in chemical structure representations across knowledgebases, we alleviated this issue by standardizing chemical structure representation using InChI (IUPAC International Chemical Identifier) canonicalization. We compared the performance of the non-standardized versus the standardized dataset and quantified their generalizability by comparing training set compounds to their knowledgebase cross-references. Results: While the non-standardized dataset scored a mean Matthews correlation coefficient (MCC) of 0.8725 ± 0.0064, the standardized dataset scored an MCC of 0.9036 ± 0.0033. When comparing model generalizability, the non-standardized chemical structure representations had a huge 0.2687 drop in mean MCC, while the standardized chemical structure representations only had a 0.0384 drop in mean MCC. Conclusions: We constructed the largest ML-ready dataset for predicting compound-pathway involvement to date. Next, we constructed, trained, and evaluated the highest performing ML model capable of predicting the highest number of pathway annotations to date. We discovered that standardizing chemical structure representation is an essential step when predicting novel chemical structures.
Huckvale et al. (Tue,) studied this question.