Cross-domain transfer learning from peptides to metabolites using a multi-property fine-tuned LLM

Area of Science:

• Computational Chemistry
• Cheminformatics
• Metabolomics

Background:

• Accurate retention time (RT) prediction is crucial for compound identification in metabolomics and lipidomics.
• Existing methods struggle with limited experimental RT data, hindering model generalization.
• Transfer learning offers a potential solution, but its application to metabolite RT prediction needs further investigation.

Purpose of the Study:

• To develop and evaluate a transfer learning framework for improving metabolite retention time (RT) prediction.
• To leverage large peptide datasets for enhanced RT prediction in data-sparse metabolomics conditions.
• To assess the impact of multi-task learning on the robustness and generalization of RT prediction models.

Main Methods:

• Developed a transfer learning framework using ChemBERTa, pretraining on large peptide datasets.
• Employed a multi-task learning objective to jointly predict RT and molecular descriptors.
• Evaluated model performance in low-data regimes for metabolite RT prediction.

Main Results:

• The multi-task pretrained model achieved superior generalization to metabolites compared to an RT-only model (median test R² of 0.842 vs. 0.820).
• Transfer learning significantly outperformed baseline models in low-data scenarios (e.g., 3% data: R² 0.322 vs. 0.216, MAE 114.9 vs. 131.7).
• Multi-task molecular property learning, not peptide pretraining alone, drove the performance gains.

Conclusions:

• Multi-task transfer learning is an effective and scalable strategy for enhancing metabolite RT prediction, particularly with limited experimental data.
• This approach improves compound identification accuracy in metabolomics and lipidomics.
• The developed framework provides a robust method for building comprehensive RT libraries.