Journal: bioRxiv

Article Title: MetaReact: A Reaction-Aware Transformer for End-to-End Prediction of Drug Metabolism

doi: 10.64898/2026.03.14.711529

Figure Lengend Snippet: (A) MetaReact model architecture. The encoder uses self-attention to generate context-aware representations of the input sequence. The decoder generates output tokens autoregressively, using both self-attention and cross-attention to integrateinformation from the encoder. (B) Pretraining on general chemical reactions and fine-tuning on metabolic reactions, both represented using the reaction-aware ReactSeq format that encodes atomic and bond transformations. (C) Three task-specific input formats guide the model to predict metabolites from substrates ( enzyme-agnostic ), infer both enzyme and metabolite ( enzyme-completion ), or predict metabolites given an enzyme ( enzyme-conditioned ).

Article Snippet: We curated a large-scale collection of enzymatic metabolic reactions from multiple public sources, including DrugBank , MetXBioD , SMPDB , HumanCyc , Recon3 , and the Accelrys Metabolite Database (AMD; now BIOVI) .

Techniques: Sequencing

Journal: bioRxiv

Article Title: MetaReact: A Reaction-Aware Transformer for End-to-End Prediction of Drug Metabolism

doi: 10.64898/2026.03.14.711529

Figure Lengend Snippet: (A) Number of correctly identified metabolites for each enzyme family in the MetaTrans dataset, based on the Top-13 predictions per reaction. Bar segments represent different enzyme families. Dataset denotes the total number of metabolites for each enzyme family in the dataset. (B) Accuracy comparison using three reaction-coverage metrics within the Top-13 predictions. (C) Comparison of precision and recall for metabolite identification at Top-5, Top-10, and Top-20 predictions. BioTransformer is excluded here because its output size is fixed at 13 candidates. (D) Accuracy for predicting major metabolites on the L-data dataset at Top-1, Top-3, and Top-5 cutoffs.

Article Snippet: We curated a large-scale collection of enzymatic metabolic reactions from multiple public sources, including DrugBank , MetXBioD , SMPDB , HumanCyc , Recon3 , and the Accelrys Metabolite Database (AMD; now BIOVI) .

Techniques: Comparison

Journal: bioRxiv

Article Title: MetaReact: A Reaction-Aware Transformer for End-to-End Prediction of Drug Metabolism

doi: 10.64898/2026.03.14.711529

Figure Lengend Snippet: (A, B) Performance of MetaReact vs. random guessing on the internal test set, for (A) enzyme families and (B) CYP subtypes. Dot size reflects the sample size (number of reactions) for each enzyme family or subtype. ( C-E ) Comparison of MetaReact and BioTransformer3.0 on the external D-data benchmark. Panels show prediction performance at the (C) enzyme family, (D) enzyme subtype, (E) and enzyme subtype-metabolite pair levels, with precision, recall, and F1-score reported across Top-k thresholds. Accuracy is also evaluated under three coverage criteria: at least one, at least half, and all matched.

Article Snippet: We curated a large-scale collection of enzymatic metabolic reactions from multiple public sources, including DrugBank , MetXBioD , SMPDB , HumanCyc , Recon3 , and the Accelrys Metabolite Database (AMD; now BIOVI) .

Techniques: Comparison

Journal: bioRxiv

Article Title: MetaReact: A Reaction-Aware Transformer for End-to-End Prediction of Drug Metabolism

doi: 10.64898/2026.03.14.711529

Figure Lengend Snippet: (A) Effect of enzyme context on metabolite prediction in the internal test set. Performance was evaluated by precision, recall, and F1-score with and without enzyme subtype information. (B) Metabolite-prediction performance of MetaReact and the rule-based tool CyProduct under matched CYP450 isoform conditions. Accuracy, precision, recall, and F1-score are compared at multiple Top-k cutoffs. (C-D) SOM-prediction performance of MetaReact and baseline models at two levels: (C) CYP450 family level, showing Top-1, Top-2, and Top-3 accuracies, and (D) isoform level, showing Top-1 accuracy.

Article Snippet: We curated a large-scale collection of enzymatic metabolic reactions from multiple public sources, including DrugBank , MetXBioD , SMPDB , HumanCyc , Recon3 , and the Accelrys Metabolite Database (AMD; now BIOVI) .

Techniques:

Journal: bioRxiv

Article Title: MetaReact: A Reaction-Aware Transformer for End-to-End Prediction of Drug Metabolism

doi: 10.64898/2026.03.14.711529

Figure Lengend Snippet: (A) Falnidamol and SGX523, showing correct AOX1 ranking and metabolite prediction. (B) Representative site-directed modifications for CXCR4 antagonist TIQ-15 (I) and ZAK inhibitor ZAK-14 (II). MLM (%) refers to the percentage of compound remaining after 10 minutes of incubation with mouse liver microsomes. HLM Cl int (mL/min/kg) represent the intrinsic clearance by human liver microsomes. The light-blue shaded fragment denotes the experimentally identified metabolic site region, within which metabolism is known to occur but without atom-level resolution. The arrow and highlighted atom(s) mark the specific metabolic site predicted by MetaReact.

Article Snippet: We curated a large-scale collection of enzymatic metabolic reactions from multiple public sources, including DrugBank , MetXBioD , SMPDB , HumanCyc , Recon3 , and the Accelrys Metabolite Database (AMD; now BIOVI) .

Techniques: Incubation

Journal: Foods

Article Title: Metabolic and Microbial Community Profiles of Century-Old Pu-Erh Tea: An Integrative Metabolomic and Microbiomic Analysis

doi: 10.3390/foods15050916

Figure Lengend Snippet: Statistical analysis of differentially expressed metabolites (DEMs). ( A ) Bar chart of DEM statistics. ( B ) Volcano plot of DEMs for S vs. Y. ( C ) Volcano plot of DEMs for S vs. Q. ( D ) Volcano plot of DEMs for Y vs. Q. (The x -axis represents the log2-transformed fold change in the quantitative value of a metabolite between the two samples; the y -axis represents the −log10-transformed p -value. Each point in the plot corresponds to one metabolite. The greater the absolute value on the x -axis, the larger the difference in expression level of the metabolite between the two samples. The larger the value on the y -axis, the more significant the differential expression, and thus the more reliable the selected differentially expressed metabolite. The size of each point reflects the magnitude of the VIP value; red points indicate metabolites that are significantly up-regulated, blue points indicate those that are significantly down-regulated, and grey points represent metabolites that do not meet the criteria for differential expression).

Article Snippet: Metabolite identification was performed by searching and matching against spectral databases, including HMDB [ ], MassBank [ ], LipidMaps [ ], mzCloud [ ], KEGG [ ], and a self-built metabolite standard database from PANOMICS.

Techniques: Transformation Assay, Expressing, Quantitative Proteomics