Sebastian Böcker discusses how predicting retention order, not time, makes liquid chromatography retention time prediction transferable across laboratories.
Retention times are useful for identifying small molecules in LC–MS, but they are notoriously hard to predict across laboratories. Even small changes in column, mobile phase, pH or instrument plumbing can shift them. Most existing models must be trained on data from the same system they will later predict for.
In this episode of AI/ML Now, Sebastien Böcker discusses the group’s recent Nature Methods paper: “Times Are Changing but Order Matters: Transferable Prediction of Small-Molecule Liquid Chromatography Retention Times.1 This research introduces “2-step,” a machine learning approach that splits the problem in two. A graph neural network first predicts a retention order index from a compound’s structure and the column and mobile-phase conditions. A small set of anchor compounds then converts that index into retention times on the user’s own system. No model retraining is needed, and the method outperforms approaches that were trained on the target system, according to Böcker.
In this episode Sebastian Böcker answers the following questions:
• Your recent Nature Methods paper, “Times Are Changing but Order Matters: Transferable Prediction of Small-molecule Liquid Chromatography Retention Times,” tackles a long-standing problem. What motivated this research, and why has transferable retention time prediction been so difficult to achieve?
• How many anchor compounds does a laboratory need in practice, and can library hits replace authentic standards?
• What should users do if their column lacks hydrophobic subtraction model (HSM) or Tanaka parameters?
Watch the first part of the interview here:
Watch the second episode here:
1. Kretschmer, F.; Harrieder, E.-M.; Witting, M.; Böcker, S. Times Are Changing but Order Matters: Transferable Prediction of Small-Molecule Liquid Chromatography Retention Times. Nat. Methods2026. DOI: 10.1038/s41592-026-03243-2.
