In this interview, industry expert Dr. Anthony Grice explains how machine learning predicts LC-MS response factors, reducing reliance on surrogates, improving accuracy, and accelerating extractables analysis and risk assessment workflows
Please can you start by explaining the core challenge in calibration workflows that Lumo is designed to address?
When you detect an unknown chemical in an E&L study but don't have a reference standard for it, which is frequently, you must estimate its concentration using a surrogate. The problem is that different molecules can respond very differently in the mass spectrometer, and a poor surrogate choice can produce errors of ten-fold or more.
Lumo eliminates that guesswork by predicting how a compound will respond directly from its molecular structure and properties. No standard required!
Image Credit: wedmoments.stock/shutterstock.com
Lumo uses machine learning in multi-layer perceptron (MLP) models. Could you talk us through how these models are trained, and explain how routing compounds by chemical class contributes to improved prediction accuracy?
The models were trained on a database of over 300 compounds with experimentally measured response factors, chosen to cover the widest possible range of chemical properties, not just what's easy to source. Each compound's structure is converted into a set of molecular descriptors, which the neural network learns to map to a response factor value.
Early on, we found that a single model couldn't handle the complexity of LC-MS ionization. So, we built a tiered system: the tool first identifies a compound's functional group class, then routes it to a specialized sub-model trained specifically on that chemistry. That targeted approach is what gives us the accuracy we see in practice.
Why is a more advanced machine-learning approach necessary for property prediction, rather than relying on traditional surrogate or rule-based estimation methods?
Structural similarity doesn't reliably predict ionization response; two compounds that look alike on paper can behave very differently in the detector. Simpler rule-based models can't capture that complexity. MLP neural networks can, because they learn nonlinear relationships across many molecular features simultaneously.
For a group of structurally related aromatic alcohols, the most favorable possible scenario for surrogate selection, our model more than doubled the probability of a highly accurate prediction (within 80-120 % of the true value) compared to expert-judgment surrogate selection.
How does Lumo’s predictive performance compare with traditional surrogate-based estimation approaches when evaluated against expected accuracy criteria across a broad range of compounds?
We tested the models on 49 compounds that were never part of the original training data. 90 % were predicted within 60 % of their true concentration. Compare that to expert-judgment surrogate selection, where there's a 1-in-5 chance of underestimating by more than 40 %. Across all training data, 92 % of predictions fell within a three-fold window, well within what semi-quantitative E&L analysis demands.
The system flags low-confidence predictions for follow-up. What indicators are used to assess confidence levels, and how does this feature improve testing efficiency and reliability?
Lumo distinguishes between two very different situations. If a compound is known to be non-detectable under the method conditions, a pure hydrocarbon in LC-MS/ESI mode, for example, it gets assigned a zero.
But, if a compound's response is likely to be unpredictable (certain silicon- or phosphorus-containing structures, for example), it gets flagged rather than estimated. That flag tells the analyst not to trust a number here, to use a different approach. It's an essential safeguard that prevents bad data from flowing into a risk assessment.
How does Lumo’s performance compare with traditional surrogate-based estimations in terms of expected accuracy criteria across a broad range of compounds?
For a group of structurally similar aromatic alcohols, the best-case scenario for surrogate selection, expert judgment achieved a highly accurate result (within 80-120 % of the true value) only 24 % of the time. Lumo achieved that same level of accuracy 54 % of the time. The probability of a risky underestimate dropped from 20 % to 10 %. Additionally, unlike surrogate selection, Lumo's performance doesn't depend on whether a suitable reference compound happens to be available.
From a practical standpoint, what are the most significant advantages – such as reduced use of standards, instrument time, or improvements to results – that laboratories can expect when incorporating Lumo?
Three things. First, fewer standards. Lumo generates RRF predictions from structure alone, so you're not waiting on standard procurement or making conservative assumptions while you wait. Second, faster turnaround. Semi-quantitative estimates are available as soon as a compound is identified, shortening the gap between detection and risk assessment. Thirdly, consistency.
The model applies the same logic every time, removing analyst-to-analyst variability in surrogate selection. The result is more defensible and delivered faster.
Lumo is described as aligning with risk-based approaches in ISO 10993-18. How does it support toxicological assessments within this framework?
ISO 10993-18 requires chemical characterization to be conducted within a risk management framework, with analytical uncertainty explicitly accounted for in the Analytical Evaluation Threshold. Better RRF predictions mean lower analytical uncertainty, which means more accurate AET calculations and less need for overly conservative uncertainty factors.
The flagging system also creates a transparent audit trail. The implication for Lumo is straightforward: better RRF predictions lead to better characterization of actual method uncertainty, supporting a more defensible and accurate UF, and ultimately a more reliable yet not overly underestimated AET.
What safeguards are in place to prevent the generation of chemically implausible results, and how do these quality checks contribute to the system’s overall robustness?
There are multiple layers. Invalid structures are caught at the input stage. The substructure routing system ensures compounds are only evaluated by models trained on chemically relevant data. The forced-zero logic prevents the model from reporting a signal where none is chemically possible. And the flagging system stops unreliable estimates from reaching the analyst at all.
These aren't simply quality checks. They're what make the tool trustworthy enough to use in a regulatory context.
Have you observed measurable outcomes from early adopters of Lumo, particularly in terms of accelerating non-targeted screening or improving decision-making timelines?
The published work is a proof-of-concept, so we're at the beginning of the adoption curve. But what we can say is that a 90 % success rate on out-of-sample compounds, without any standards, translates directly into fewer follow-up analyses and faster decisions.
One of the biggest bottlenecks in non-targeted screening has always been the gap between 'we’ve found something' and 'we know what to do about it.' Lumo compresses that gap significantly.
How does Lumo integrate with existing laboratory workflows or software platforms? Is it designed as a standalone tool or as part of a broader analytical ecosystem?
The inputs are SMILES strings and the outputs are numerical predictions; simple, universal formats that connect to any data pipeline. The underlying software (RDKit, scikit-learn, Python) is entirely open-source and platform-agnostic. In practice, Lumo slots in as an automated post-identification step.
Once a compound is tentatively identified from a spectral match, the RRF prediction follows immediately. It's designed to integrate, not to require a workflow redesign around it.
Looking to the future, how do you envision predictive tools like Lumo evolving to further support extractables screening and regulatory compliance in medical and toxicological research?
The near-term roadmap includes more advanced molecular featurization, including, but not limited to, Morgan fingerprints, graph neural networks, and 3D conformational features. Incorporation of these should improve accuracy for complex or unusual structures.
Longer term, as confidence in ML-based RRF predictions grows through prospective validation, we expect regulators to increasingly accept model-derived values as a defensible basis for risk assessment. The vision is a seamless pipeline from raw analytical data to regulatory submission, where uncertainty is quantified rather than assumed, and decisions are faster without being less rigorous.
Where can readers find more information?
Deng, Y., et al. (2026). Neural Network Prediction of Response Factors for Extractables and Leachables in Pharmaceuticals and Medical Devices. PDA Journal of Pharmaceutical Science and Technology, (online) pp.pdajpst.2025-000061.1. DOI: 10.5731/pdajpst.2025-000061.1. https://journal.pda.org/content/early/2026/01/30/pdajpst.2025-000061.1.
About Dr. Anthony Grice 
Dr. Anthony Grice is a Principal Scientist at Jordi Labs, an RQM+ Company. A polymer chemist by training, with a Ph.D. from the University of Warwick (UK), he has spent over 12 years leading complex investigative projects across a broad range of analytical techniques; delivering practical, well-reasoned solutions for customers in both problem-solving and regulatory contexts.
About Jordi Labs
Jordi Labs provides the highest quality contract analytical services and polymer HPLC columns to some of the world’s leading consumer products, polymers, pharmaceutical, and medical device manufacturers. Our team of PhD analytical chemists specializes in chemical identification. One of the core competencies is Extractables & Leachables testing.
We are also worldwide leaders in;
- Method development/validation
- Particulates & residue analysis
- Good-bad comparisons
- Polymer analysis
- Polymer failure
We also help companies from Fortune 500s to innovative startups with method development, preparative HPLC, training seminars, depositions, and consulting. As a family company, we take pride in the production of all of our products and analytical service offerings. It is our goal to help our customers overcome their analytical challenges by providing excellent products and personal assistance from our highly-trained staff of PhD chemists.