AI/ML in Practice: Machine-learning Prediction of Chromatographic Retention Times for Small Molecules in Pharmaceutical Applications

In the paper you published entitled "Machine-learning Prediction of Chromatographic Retention Times for Small Molecules in Pharmaceutical Workflows",¹ what was the primary motivation for applying machine learning to retention time (RT) prediction, and how does this align with the needs of high-throughput chemical synthesis platforms in drug discovery?

The main motivation was to make purification more efficient in a high-throughput setting. Our chemistry platform produces a large number of compounds for early discovery projects, and each compound typically needs purification before it can move into downstream assays. Traditionally, we have relied on "scout runs" to determine when a compound elutes so that we can design an appropriate purification gradient.

What machine learning offers is essentially an in silico scout run: the ability to estimate retention time before touching the instrument. In practice, that can reduce cycle time, conserve limited analyte, and lower consumption of solvents, columns, and instrument capacity. That fits very well with the needs of high-throughput synthesis, where the bottleneck is often not making compounds, but characterizing and purifying them fast enough to keep projects moving.

Given your experience with chromatographic datasets, what factors do you think contributed to the improved performance of graph neural network models like AttentiveFP² and ChemProp³ compared to tree-based methods such as XGBoost⁴ for RT prediction?

I think there are two separate questions here: what molecular representation is most informative, and what model architecture is best able to use that representation. XGBoost remains a very strong baseline, especially when paired with well-engineered descriptors or fingerprints, so I would not frame this as a simple matter of older versus newer methods.

Where graph neural networks appear to have an advantage is that they learn directly from molecular structure rather than from a fixed, compressed representation. For retention time prediction, subtle differences in connectivity, functional groups, ring systems, and local chemical environments can have a meaningful impact on chromatographic behavior. Models such as ChemProp³and AttentiveFP² seem better able to capture those structure-property relationships.

That said, these models still operate on 2D molecular graphs. Since chromatographic retention is ultimately driven by physicochemical interactions, there is reason to expect that richer representations, including conformation-aware or 3D information where robustly available, could improve performance further.

The study used an internally generated dataset of 7,552 small molecules from high-throughput synthesis. What are the key challenges in ensuring data quality and consistency for RT prediction models in pharmaceutical environments?

Data quality is absolutely central in a project like this. In an industrial discovery setting, one of the main challenges is that early-stage chemistry generates many compounds very quickly, but each individual compound is often only sparsely characterized. Later in a project, the compounds that survive are described in much greater detail, but by then the chemical space has narrowed considerably. So there is often a tradeoff between dataset size and annotation quality.

One concrete issue for us was isomer resolution. For much of the early high-throughput data, we could not confidently distinguish all isomeric forms, whereas that level of structural confidence typically came later for compounds that progressed further. That introduces ambiguity into the training data, and ambiguity in the labels places a clear ceiling on model performance.

Another major challenge is consistency of metadata: chromatographic conditions, instrument configuration, sample handling, integration rules, and compound registration all need to be traceable and standardized. A large part of the initial effort went into recovering and harmonizing historical data from multiple repositories. Since then, we have established an automated data pipeline that captures new results in a model-ready format, supports continuous monitoring of model performance, and makes retraining on newly acquired data much more streamlined.

This research highlighted that molecular graph neural networks maintained predictive accuracy for new chemical series over time. How important is temporal robustness in RT prediction for real-world pharmaceutical workflows, and how might model drift impact routine analysis?

Temporal robustness is critical if a model is going to be useful in a live pharmaceutical workflow rather than only on a retrospective benchmark. Discovery projects evolve continuously toward new targets, new scaffolds, and sometimes entirely new chemistries. That means the compounds arriving today may differ substantially from the compounds the model saw during training.

If a model drifts or fails to generalize across those shifts, the operational consequence is straightforward: purification decisions become less reliable, and users will quickly lose trust in the predictions. In that sense, robustness over time is not just a statistical property, but a prerequisite for adoption.

We addressed this in the paper through time-series analysis, which is a more realistic way to evaluate deployment performance. We have also worked with similarity-based uncertainty or out-of-distribution assessment to estimate when an incoming molecule is unlike the training set. Operationally, that matters because a prediction is much more actionable when it is accompanied by an estimate of confidence. More recently, we have seen promising signs that fine-tuned foundation models may improve generalizability further, which is an exciting direction for future work.

ChemProp achieved an average error of 38.70 seconds on the public METLIN SMRT dataset. From a chromatographer's perspective, how acceptable is this level of prediction error, and in what contexts) would it be most useful?

The acceptability of that error depends very much on the use case. For high-precision analytical tasks, an average error of 38.70 seconds may be too large to support direct decision-making. But for triage, screening, or early method development, it can still be highly useful if it narrows the expected elution window enough to reduce experimental search space.

In our own workflow, the practical threshold is stricter: beyond roughly 15 seconds, predictions begin to lose value for downstream purification decisions. So from that perspective, 38.70 seconds would not be sufficient for the most demanding operational use cases. At the same time, the METLIN SMRT result remains informative because it provides an external benchmark against the published literature, even though the chromatographic conditions in that dataset differ substantially from our own setup.

What are the main barriers to integrating machine-learning-based RT prediction tools into routine chromatographic method development and validation pipelines in pharmaceutical labs, and how might these be overcome?

The biggest barrier is usually not the machine-learning model itself, but the surrounding data and workflow infrastructure. To deploy RT prediction routinely, you need data in a model-ready format, rich metadata, standardized experimental records, and a clear link between predictions and the actual decisions chromatographers are making. In many organizations, the raw information exists, but it is fragmented across instruments, file systems, registration platforms, and legacy repositories.

Another barrier is trust. Even a strong model will struggle to gain adoption unless users understand when it works, when it is likely to fail, and how prediction uncertainty should be interpreted. That is why deployment needs to include monitoring, retraining, and some form of confidence or out-of-domain assessment rather than only a point prediction.

We addressed this by building an automated data pipeline that captures new data in a reusable format and supports rapid retraining and evaluation. In our case, that level of infrastructure only became realistic once we had established proof of concept and could justify scaling the effort.

Anything else you would like to add?

One broader point is that retention time prediction should not be viewed only as a modeling exercise. Its real value comes from how well it integrates into an operational workflow: reducing assay turnaround time, increasing purification throughput, and helping scientists make decisions earlier with less material and fewer experiments.

I also think the field is moving from asking whether machine learning can predict retention time to asking how robustly, how transparently, and under what uncertainty it can do so in production. That shift is important, because in pharmaceutical settings the most useful models are not necessarily the ones with the best retrospective benchmark, but the ones that remain reliable as chemistry evolves over time.

References

1. Vik, D.; Pii, D.; Mudaliar, C.; Nørregaard-Madsen, M.; Kontijevskis, A. Performance and Robustness of Small Molecule Retention Time Prediction with Molecular Graph Neural Networks in Industrial Drug Discovery Campaigns. Sci. Rep. 2024, 14, 8733. https://doi.org/10.1038/s41598-024-59620-4.

2. Xiong, Z.; Wang, D.; Liu, X.; Zhong, F.; Wan, X.; Li, X.; Li, Z.; Luo, X.; Chen, K.; Jiang, H.; Zheng, M. Pushing the Boundaries of Molecular Representation for Drug Discovery with the Graph Attention Mechanism. J. Med. Chem. 2020, 63 (16), 8749–8760. https://doi.org/10.1021/acs.jmedchem.9b00959.

3. Graff, D. E.; Morgan, N. K.; Burns, J. W.; Doner, A. C.; Li, B.; Li, S.-C.; Manu, J.; Menon, A.; Pang, H.-W.; Wu, H.; Zalte, A. S.; Zheng, J. W.; Coley, C. W.; Green, W. H.; Greenman, K. P. Chemprop v2: An Efficient, Modular Machine Learning Package for Chemical Property Prediction. J. Chem. Inf. Model. 2026, 66 (1), 28–33. https://doi.org/10.1021/acs.jcim.5c02332.

4. Chen, T.; Guestrin, C. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '16); ACM: New York, NY, 2016; pp 785–794. https://doi.org/10.1145/2939672.2939785.

Biography

Daniel Vik is a Senior Data Scientist at Amgen Research Copenhagen, Denmark where he develops production machine learning systems for drug discovery, incl. GNN-based prediction. He holds a PhD in quantitative molecular biology from the University of Copenhagen and previously worked in R&D/AI at LEO Pharma.