Why Order Matters in Chromatography

Machine-learning method predicts elution order before calculating absolute retention time

October 16, 2026
3 min

The Analytical Scientist

Overview

The study describes “2-step,” a method that predicts small-molecule elution order and calibrates that order to absolute retention times using a small set of identified compounds. Across 171 curated reversed-phase chromatography datasets, it showed lower retention-order error when chromatographic conditions were included and outperformed two existing retention-time models in five of six comparisons designed to reflect laboratory use.

Background

Retention time—the point at which a compound exits a chromatography column—can complement mass-spectral evidence and help distinguish compounds with similar structures. Retention time varies with the column, mobile phase, and other operating conditions, limiting the transferability of models trained on a different system. Existing approaches may require hundreds of measurements under the target conditions, whereas the reported method uses identified compounds as calibration anchors. The source also notes that the method is limited to reversed-phase chromatography and depends on accurate experimental metadata and reliable anchor identifications.

Data Highlights

Measure or evaluation	Reported result
Curated reversed-phase datasets	171 datasets from RepoRT
Retention-order error, condition-aware model	4.89%
Retention-order error, similar model without chromatographic-condition information	5.61%
Calibration anchors	Approximately 30 confidently identified compounds; as few as 15 could be sufficient
Comparison with two existing retention-time models	2-step performed better in five of six datasets and similarly in the sixth
Human fecal-sample application	19 known compounds used as anchors

Key Findings

The first stage uses a graph neural network to predict compound elution order from molecular structure, column information, and mobile-phase pH.
The second stage maps predicted order to absolute retention time for a specific laboratory method using a polynomial calculation and identified reference compounds.
In testing conditions separated from training conditions, the condition-aware model had a 4.89% retention-order error rate, compared with 5.61% for a similar model that did not account for chromatographic conditions.
When structurally similar compounds were separated between training and test sets, the method performed better than two existing models in five of six datasets and similarly in one; it did not require training on target-system data.
In human fecal LC–MS/MS data, predictions provided additional evidence for compound annotations, helped evaluate candidate bile acids, and supported differentiation of stereoisomers, using 19 known compounds as anchors.
Column chemistry had the largest effect on retention order; mobile-phase pH, solvent composition, additives, and temperature also affected some compounds. The source emphasizes recording column, mobile phase, gradient, temperature, and flow rate when generating or sharing retention data.

Clinical Implications

The source reports retention-time predictions as additional evidence for LC–MS/MS compound annotation, not as a replacement for other identification evidence. It also identifies accurate experimental metadata and reliable anchor identifications as requirements, and states that the method is currently limited to reversed-phase chromatography.

Conclusion

The two-step approach uses predicted elution order and a limited set of identified compounds to estimate retention times across reversed-phase systems. Its reported performance and fecal-sample application are bounded by the study’s stated requirements for metadata and anchor reliability.