DeepETD: a novel deep-learning based model for endogenous metabolite target discovery

Metabolites are involved in almost all fundamental biological processes. The identification of their protein targets is crucial for elucidating non-canonical signaling roles and evaluating their therapeutic potential. In recent years, due to the continuous development of chemical proteomics, atypical phenotypic functions of classic metabolites have been continuously discovered. However, the discovery of endogenous metabolites for their disease-related functions is still progressing slowly. To accelerate the identification of disease-related targets for endogenous metabolites, we propose a new hypothesis: endogenous metabolites and their molecular targets are expected to have similar disease associations. Following this hypothesis, here we report the development of a novel deep-learning model called DeepETD, which integrates bioinformatics data and introduces an attention mechanism to predict functional targets of specific metabolite phenotypes. Using this model, we constructed a publicly accessible database named EMTDD containing potential targets for 3,382 common human endogenous metabolites. Overall, this study presents a new computational method and resource for endogenous metabolite target discovery as an important supplement to experimental methods such as chemical proteomics.