Researchers have engineered a two-step machine learning method to predict small-molecule liquid chromatography (LC–MS) retention times, solving a recurring bottleneck in analytical chemistry.
The approach, detailed in Nature Methods, removes the requirement for model retraining when switching between different laboratory systems, as reported by chromatographyonline.com.
Retention times in LC–MS have long frustrated chemists, as they fluctuate wildly based on column type, mobile phase, pH, and instrument plumbing. Historically, these discrepancies forced scientists to train predictive models on data specific to the individual hardware in their lab. A slight change in equipment configuration often rendered existing models obsolete.
The new framework, developed by Sebastien Böcker’s group, bypasses the need for system-specific training. The process begins with a graph neural network that calculates a “retention order index.” This initial step relies solely on the compound’s structure and the specific conditions of the chromatography column, rather than attempting to force a prediction of the absolute time.
To bridge the gap between abstract order and physical time, the researchers employ a small set of “anchor compounds.” These known molecules serve as a calibration tool, allowing the system to map the predicted retention index onto the specific time scale of the user’s equipment. According to Böcker, this method facilitates transferable predictions without the time-consuming burden of retraining models for every new configuration.
The research team claims this two-step approach outperforms traditional models that are trained directly on a target system. By decoupling chemical structure prediction from the physical variations inherent in laboratory hardware, the model maintains accuracy even when environmental factors or equipment setups shift. This advancement offers a more reliable solution for labs struggling to maintain consistent results across diverse instrument configurations.
Related reading