Retention time has always been one of the quieter costs of building a chromatographic method, measured in scouting runs, patience and instrument hours that rarely make it into a published result. At HTC-19 in Leuven, Sam Whitmarsh of CatSci offered a view of one attempt to lighten that burden: an ensemble of machine learning models that weighs the analyte's molecular character, the column's measured properties and the gradient's shape to anticipate where a compound will elute. Trained on open interlaboratory data from RepoRT alongside in-house records, the approach reaches toward the edge of w
Ensemble ML Predicts Reversed-Phase LC Retention Near Experimental Reproducibility Limits
Close to the limits of experimental reproducibility
Mark
Why does retention prediction matter so much?
Mimi
Because it's where method development burns time. Scouting runs, trial and error, long hours on the instrument. A good prediction could cut some of that.
Luke
Could. The source says predictions come close to experimental reproducibility limits, but gives no error numbers.
Mark
What goes into the model?
Mimi
Molecular descriptors, column characterisation and gradient information, combined as an ensemble.
Mark
And the data?
Mimi
RepoRT, an open repository, plus internal datasets. Making retention comparable across labs and instruments was the curation challenge he was asked about.
Luke
Worth noting we only have the questions, not his answers. How he solved comparability isn't in this text.
Mark
Where might it fail?
Mimi
Novel chemotypes, unusual stationary phases, ionisable compounds, gradient extrapolation. All cases far from past data.
Luke
Those are areas he was asked about, not confirmed failures. Whether the model actually struggles there isn't stated.
Mark
What should a lab take from it?
Mimi
Building in-house takes enough data, good metadata and computational expertise.
The Pulse
- Retention time prediction promises to cut the repeated scouting runs that consume pharmaceutical method development instrument time.
- An ensemble model fed by molecular descriptors, column characterisation and gradient data is landing predictions near the limits of experimental reproducibility.
- Curating interlaboratory data from RepoRT into a comparable training set has been the first hurdle, since identical methods behave differently across instruments.
- Novel chemotypes, unusual stationary phases, ionisable compounds and gradient extrapolation remain the points where the model is most likely to stumble.
- Labs eyeing in-house versions must secure enough data, quality metadata and computational expertise before the tool becomes practically useful.
Retention time has always been one of the quieter costs of building a chromatographic method, measured in scouting runs, patience and instrument hours that rarely make it into a published result. At HTC-19 in Leuven, Sam Whitmarsh of CatSci offered a view of one attempt to lighten that burden: an ensemble of machine learning models that weighs the analyte's molecular character, the column's measured properties and the gradient's shape to anticipate where a compound will elute. Trained on open interlaboratory data from RepoRT alongside in-house records, the approach reaches toward the edge of what repeated experiments can agree on with themselves, though its honest boundaries lie in unfamiliar chemistry, unusual phases, shifting charge states and gradients beyond what it has seen. For laboratories considering their own versions, the lesson is less about the algorithm than about the discipline of data, metadata and skill it quietly requires.
At HTC-19 in Leuven, Belgium, Sam Whitmarsh of CatSci sat down for a short video interview about a stubborn problem in pharmaceutical labs: predicting how long a compound will stay on a reversed-phase LC column. Anyone who has developed a method knows the grind. Retention time usually demands repeated scouting runs, trial and error and long hours on the instrument.
His presentation describes an ensemble machine learning model that draws on three inputs: molecular descriptors for the analyte, characterisation data for the column and details of the gradient. According to the source, its predictions approach the limits of experimental reproducibility, meaning the gap between predicted and measured retention nearly matches the variation seen when the same experiment is simply repeated. The exact error figures are left to the presentation itself.
The training data came from RepoRT, an open chromatographic repository, and internal datasets. Making those records comparable across laboratories and instruments proved a central challenge, since the same method can behave differently from one system to the next. Whitmarsh also named where the model is weakest: novel chemotypes, unusual stationary phases, ionisable compounds and gradient extrapolation. Each asks a model to predict beyond the precedent it was built on.
For laboratories considering their own tools, the message is that a published model is a starting point. Useful in-house versions need sufficient data, well-documented metadata and the computational skill to build and maintain them. The piece is the second part of a series, and the open questions are how much discipline the approach demands and how it performs on the hard cases.
At HTC-19 in Leuven, Belgium, Sam Whitmarsh of CatSci in the UK sat down for a short video interview about a problem that eats instrument time in pharmaceutical labs: predicting how long a compound will be retained on a reversed-phase LC column.
The difficulty is familiar to anyone who develops methods. Retention time is one of the hardest parts of reversed-phase method development, and getting a handle on it usually means repeated scouting runs, trial and error, and long hours on the instrument. Whitmarsh's presentation, titled "Predicting Reversed-Phase LC Retention Using Ensemble Machine Learning and Open Chromatographic Data," describes an attempt to shorten that process by predicting retention computationally.
The model is an ensemble. It draws on three kinds of input: molecular descriptors for the analyte, characterisation data for the column, and information about the gradient. According to the source, its predictions come close to the limits of experimental reproducibility, meaning the gap between predicted and measured retention approaches the variation you would see simply from repeating the experiment. The source does not give the specific error figures, so how close "close" is remains for the presentation itself.
In the interview, Whitmarsh walks through how the model was built, what drives its accuracy, and where it could be improved for real pharmaceutical workflows. The training data comes from two places: RepoRT, an open repository of chromatographic data, and internal datasets. Combining them raises the first of the questions he was asked: what were the biggest challenges in curating that data, and how did he make retention times comparable across different laboratories and instruments? Interlaboratory comparability is a long-standing problem, since the same method can behave differently from one system to the next.
The second question concerns limits. He was asked where the model is weakest in pharmaceutical development, with four areas named: novel chemotypes, unusual stationary phases, ionisable compounds and gradient extrapolation. Each is a place where a model trained on existing data may be asked to predict something it has little precedent for. A new chemical scaffold, an uncommon column chemistry, a compound whose charge state shifts with conditions, or a gradient outside the range of the training data can each stress a model built from past measurements.
The third question is practical and aimed at other laboratories: what minimum dataset, metadata quality and computational expertise would a lab need to build a useful in-house model? The framing suggests the published model is not the end point, and that companies may want versions tuned to their own compounds and columns. The source summarises the answer only by pointing to the three requirements: enough data, well-documented metadata, and the computational skill to build and maintain the model.
The piece is the second part of a series on the work. For labs weighing whether to build their own tools, the thing to watch is how much data and metadata discipline the approach demands, and how the model performs on the difficult cases Whitmarsh was asked about.
Notable Quotes
The ensemble model's predictions come close to the limits of experimental reproducibility (paraphrase of the source's summary).— Chromatography Online, on Sam Whitmarsh's presentation
Want the full story?
Read the original at Chromatography Online ↗