Deciphering Important Metabolic Pathways through Reaction Pair Prediction with a Siamese Transformer Network

Abstract

Important pathway identification is essential for unraveling biological mechanisms in functional metabolomics. However, current pathway enrichment is often biased toward well-characterized pathways due to low annotation rates in untargeted metabolomics and incomplete pathway coverage. It leads to potential misinterpretation of metabolomics data. Herein, Siamese transformer reaction pair (STRP) prediction, an approach for important pathway exploration in metabolomics, is introduced. STRP leverages a weight-sharing Siamese network and a multihead attention Transformer encoder to predict metabolic reaction pairs, utilizing molecular fingerprints derived from either known metabolites or tandem mass spectra of unannotated metabolic features. Pathway labels are then deduced for metabolic features from known pathway metabolites within the reaction pairs. STRP can achieve crossvalidation metrics of 98.10%/98.13% accuracy, 97.98%/98.01% precision, 97.94%/97.97% recall, 97.96%/97.99% F1 score, and 99.56%/99.57% area under the receiver operating characteristic curve of spectral pairs in ESI⁺/ESI⁻ modes. It is applied to metabolomics datasets from prostate cancer and diabetic retinopathy. STRP successfully identifies and interprets important metabolic pathways, demonstrating its robust utility for important pathway identification. Besides, STRP-based molecular network showcases potential application in metabolome annotation. This approach reveals a significant advancement in leveraging high-resolution mass spectrometry-based metabolomics data, with the potential to transform understanding of complex biological processes.