Machine-learning helps predict retention times across chromatography systems

A two-step model predicts the order in which molecules leave a chromatography column, then uses a few reference points to estimate their retention times

A machine-learning method could predict when small molecules emerge from liquid chromatography (LC) columns, even when the instrument system differs from those used to train earlier models. The approach – called 2-step – has been reported by researchers Dr. Fleming Kretschmer and Professor Sebastian Böcker of the Friedrich Schiller University of Jena, Germany.

LC separates molecules in a sample as they pass through a column. Each compound is retained for a different length of time, so its retention time can help identify it. However, those times depend on the column, solvent, acidity, temperature and other system details. Even replacing a tube can shift the measurements.

Many existing prediction models need to be calibrated with data from the same system they will later assess. The 2-step method instead first predicts a molecule’s position in the retention order. It then converts that index into an estimated time using a small number of known reference compounds. The method focuses on reversed-phase chromatography, a widely used form of the technique.

Co-first author Kretschmer said the model performed better than approaches that require extensive training on the target system. The team said the software could help identify unknown small molecules in natural-product research, environmental analysis, food chemistry and pharmaceutical research, where samples may contain many compounds.

The 2-step software is available as a package and web application, and the researchers said it could be integrated into analytical programmes used with LC and mass spectrometry (MS). Its predictions should still be interpreted alongside other chemical evidence when identifying compounds.

The model’s first output is an index rather than a time, which allows it to predict relative order before adjusting to the instrument. Reference compounds with known retention times provide the anchor points for the second step. This separation between order prediction and time calibration is intended to make the method less dependent on measurements from the exact chromatography system in use.

Small molecules are especially diverse compared with DNA and proteins, and biological or environmental samples can contain metabolites, natural compounds, toxins, degradation products and medicines. Retention time is one clue used to identify them, alongside MS and other measurements. A prediction that transfers between systems could reduce the number of standards that researchers must measure before analysing an unfamiliar sample.

Potential applications include searches for biologically active compounds in bacteria, fungi and plants, as well as food, environmental and pharmaceutical analysis. The researchers have released the 2-step software as a package and web application. They said it could be incorporated into analytical software used for LC and MS, although retention time remains one part of compound identification and should be interpreted with other evidence.

AS described, retention time varies with the solvent gradient, acidity, temperature and column properties. The Jena team’s two-step model first estimates where a compound falls in the order of elution, then translates that relative position into an instrument-specific time. It is intended to reduce the need for extensive calibration on each target system, especially when a laboratory makes small changes to its equipment.

The software is aimed at reversed-phase LC, in which a non-polar stationary phase separates compounds according to their interactions with the mobile solvent. The researchers said that more transferable predictions could help when a laboratory has a complex sample but few standards for each possible compound. They emphasised that retention time is one piece of evidence, which should be considered with mass spectra and other analytical measurements.

For further reading please visit: 10.1038/s41592-026-03243-2