A machine-learning method could make it easier to predict when small molecules will pass through a liquid chromatography system, according to a study published in Nature Methods.
Retention time – the point at which a compound emerges from a chromatography column – can help laboratories identify molecules during liquid chromatography-mass spectrometry. It provides information that complements mass spectra and may help distinguish between compounds with similar structures.
However, retention time depends on the column, mobile phase, and other operating conditions. A model trained using one system may therefore perform poorly on another. Existing approaches often require hundreds of compounds measured under the new conditions before they can provide reliable predictions.
Researchers developed a method called “2-step” to address this limitation. Rather than predicting retention time directly, the model first estimates the order in which compounds will leave the column. It then converts that predicted order into an absolute retention time.
The first stage uses a graph neural network, a machine-learning system designed to analyze molecular structures. The model considers each compound’s chemical structure together with information about the chromatographic column and mobile-phase pH.
The second stage calibrates the predicted order to a specific laboratory method. It uses a polynomial calculation and approximately 30 compounds that have already been confidently identified in the target dataset. These reference compounds act as anchors connecting the predicted order to measured retention times. The researchers found that as few as 15 anchors could be sufficient.
The model was trained and tested using 171 curated reversed-phase chromatography datasets from RepoRT, a public repository of retention measurements and experimental conditions. To assess its ability to work with new systems, the researchers separated the testing conditions from those used for training.
The condition-aware model had a retention-order error rate of 4.89 percent, compared with 5.61 percent for a similar model that did not account for chromatographic conditions.
The team also compared the method with two existing retention-time models. When structurally similar compounds were separated between the training and test sets – a design intended to better reflect real laboratory use – the two-step method performed better in five of six datasets and produced similar results in the sixth. Unlike the other models, it did not require training on data from the target system.
The researchers applied the approach to liquid chromatography-tandem mass spectrometry data from human fecal samples, using 19 known compounds as anchors. Retention-time predictions provided additional evidence for compound annotations, helped evaluate candidate bile acids and supported the differentiation of stereoisomers – molecules with the same formula and connectivity but different three-dimensional arrangements.
Column chemistry had the largest effect on retention order. Mobile-phase pH, solvent composition, additives, and temperature also affected some compounds. The findings reinforce the importance of recording the column, mobile phase, gradient, temperature, and flow rate when generating or sharing retention data.
The method is currently limited to reversed-phase chromatography and depends on accurate experimental metadata and reliable anchor identifications.
Newsletters
Receive the latest analytical science news, personalities, education, and career development – weekly to your inbox.
