AI Model Predicts Small Molecule Retention Times Across Systems

Researchers at Friedrich Schiller University Jena, in collaboration with the Helmholtz Zentrum München and the Technical University of Munich, have introduced a novel machine-learning approach that reliably predicts the chromatographic retention times of small molecules. Published in Nature Methods, the development addresses a longstanding challenge in analytical chemistry, where accurate compound identification in complex biological and environmental samples has historically depended on highly system-specific training data. Traditional liquid chromatography separates compounds based on how long they remain in a column before eluting, a duration known as retention time. However, these times fluctuate significantly with minor variations in column hardware, solvent composition, pH, temperature, and even tubing dimensions. Previous computational models required extensive calibration using standard substances from the exact target system, making them impractical for rapid or cross-platform deployment. Led by bioinformatician Prof. Dr. Sebastian Böcker and co-developed by co-first author Fleming Kretschmer, the new methodology circumvents these limitations through a streamlined two-step architecture. The model first encodes molecular structures using a directional message-passing neural network alongside feed-forward layers that process chromatographic parameters, generating a retention order index. Rather than predicting absolute elution times directly, this index establishes a molecule relative position within a chromatographic sequence. A subsequent algorithm then maps this index to precise retention times using minimal reference points. The framework, designated as 2-step, operates effectively out of the box, delivering high accuracy across untested systems and novel compounds without prior calibration. By decoupling retention prediction from instrument-specific data, the tool significantly reduces the time, cost, and logistical barriers associated with metabolomics, drug discovery, and environmental screening. Researchers can now rapidly validate presumed chemical structures against expected elution patterns in diverse analytical workflows. The open-access software package and web application are designed for seamless integration into existing liquid chromatography and mass spectrometry suites. As laboratories increasingly rely on high-throughput compound screening, the method positions itself as a critical advancement for natural product research, pharmaceutical development, and toxicological analysis, transforming retention time prediction from a labor-intensive calibration task into a reliable, automated analytical function.