Mahmood Raisul Awal (Kyungpook National University)

Sunghwan Kim (Kyungpook National University)

Mahmood Raisul Awal (Kyungpook National University)

Sunghwan Kim (Kyungpook National University)

Retention time (RT) prediction provides an orthogonal information in liquid chromatography–mass spectrometry (LC–MS) based proteomics. However, accurate RT prediction of peptides is crucial for enhancing identification confidence, in-silico library generation for data-independent acquisition (DIA), designing efficient targeted experiments, and improving the interpretation of complex proteomic datasets. In this study, we developed a transformer machine learning (ML) framework for predicting the RTs of unmodified and modified peptides. The model performance was assessed using standard regression metrics, including the coefficient of determination, mean absolute error, and root mean square error followed by statistical analysis. The evaluation metrics showed that datasets with low MAE values ​​(≤ 0.5 min​) are associated with minimal ΔRT, confirming the effective and reliable prediction results from large training sets of unmodified sequence patterns. Datasets with high MAE values ​​(> 0.5 min) were associated with modified peptides, identified as larger ΔRT deviations due to having modifications and side-chains. Further, we predicted the RTs of lysine-acetylated modified peptides to observe the robustness of developed transformer model. Within 95% confidence interval, the predicted RTs of lys-acetylated modified peptides showed good agreement with the experimental RT values, validated by the statistical analysis. Overall, the transformer learning approach can contribute to enhance the peptide identification through RT prediction in proteomics workflow.