GUTchetp: integrated prediction of gut microbial biotransformation profiles using ensemble monolingual and multilingual neural machine translation and enzyme class consistency-reaction similarity.

The analysis of the metabolic fate of chemical compounds in the human gut remains a fundamental challenge, as comprehensive experimental characterization across the vast chemical space is infeasible. Moreover, despite progress in computational biology, a comprehensive tool for predicting microbe-dependent metabolism of diverse chemicals is still lacking. To address these challenges, we introduce the GUTchetp framework, which comprises two components: one predicts gut biotransformation products and was tested on a benchmark of six compound categories, whereas the other identifies the associated microbial enzymes and species and was evaluated on a benchmark of 95 metabolism events. The first component employs an ensemble of monolingual and multilingual neural machine translation models, integrating transfer learning, mixed fine-tuning, multiple molecular representations, and adaptive data augmentation. The ensemble model outperformed previous approaches, achieving a Top-20 BLEU score and MaxFrag accuracy of 66.59% and 67.19%, respectively. The second component applies a re-ranking rule that combines the prediction consistency between two Enzyme Commission (EC) number classifiers with chemical reaction similarity, increasing accuracy by 34.38 percentage points over existing tools on the hidden dataset. Both components showed statistically significant improvements compared with existing tools, enabling the accurate prediction of 68.42% of known gut microbial biotransformation profiles, which is highly promising for anticipating gut microbial biotransformation outcomes. GUTchetp will thus pave the way for predicting the capacity of personalized gut microbiomes to metabolize intentional xenobiotics, such as pharmaceuticals and nutrients, and unintentional ones, such as environmental chemicals.