ABCoRT: Retention Time Prediction for Metabolite Identification via Atom-Bond Co-Learning

Area of Science:

• Computational chemistry
• Metabolomics
• Machine learning

Background:

• Metabolite identification in untargeted metabolomics relies heavily on accurate liquid chromatography retention time (RT) prediction.
• Developing robust molecular representations is key to achieving reliable RT predictions.

Purpose of the Study:

• To introduce ABCoRT (Atom-Bond Co-learning for Retention Time), a novel framework for learning molecular representations to predict metabolite retention times.
• To evaluate the performance of ABCoRT on large-scale datasets and its utility in metabolite screening.

Main Methods:

• ABCoRT transforms molecular graphs into dual hypergraphs for collaborative atomic and bond information updating.
• The model was evaluated on the Small Molecule Retention Time (SMRT) dataset (80,038 molecules) and fine-tuned on six PredRet datasets.
• Metabolite screening was performed on MetaboBASE and RIKEN_PlaSM datasets.

Main Results:

• ABCoRT achieved a mean absolute error (MAE) of 25.75 s and a mean relative error (MRE) of 3.24% on the SMRT dataset.
• Fine-tuned ABCoRT models achieved the lowest MAEs on five out of six PredRet datasets.
• The model effectively filtered 38.35% and 28.46% of candidate compounds in MetaboBASE and RIKEN_PlaSM datasets, respectively.

Conclusions:

• ABCoRT provides highly informative molecular representations for accurate retention time prediction.
• The framework demonstrates significant potential for improving metabolite identification and screening in metabolomics research.