In this interview, industry expert Dr. Anthony Grice discusses how machine learning can predict LC-MS response factors, reduce reliance on surrogates, improve accuracy, and accelerate extract analysis and risk assessment workflows.
First, could you explain the core challenge of the calibration workflow that Lumo is designed to address?
If an unknown chemical is detected in an E&L study but its reference standard is not available (as is often the case), a surrogate must be used to estimate its concentration. The problem is that different molecules can react very differently in a mass spectrometer, and a poor choice of surrogates can cause errors by a factor of 10 or more.
Lumo takes the guesswork out of predicting how compounds will react directly from their molecular structure and properties. No need for standards!
Image credit: wedmoments.stock/shutterstock.com
Lumo uses machine learning with a multilayer perceptron (MLP) model. Can you explain how these models are trained and how routing compounds by chemical class contributes to improved prediction accuracy?
The model was trained on a database of over 300 compounds with experimentally measured response factors, selected to cover the widest possible range of chemical properties, not just those that are readily available. The structure of each compound is converted into a set of molecular descriptors, which a neural network learns and maps to response factor values.
We found early on that no single model could handle the complexity of LC-MS ionization. So we built a tiered system. The tool first identifies a compound’s functional group class and then routes it to specialized submodels trained specifically for that chemistry. This targeted approach gives you the precision you see in the real world.
Why do property predictions require more advanced machine learning approaches rather than relying on traditional surrogate or rule-based estimation methods?
Structural similarity is not a reliable predictor of ionization response. Two compounds that look the same on paper can behave completely differently in the detector. Simpler rule-based models fail to capture that complexity. This is possible because MLP neural networks simultaneously learn nonlinear relationships across many molecular features.
For groups of structurally related aromatic alcohols, the most favorable scenario for surrogate selection, our model more than doubled the probability of highly accurate predictions (within 80–120% of the true value) compared to surrogate selection based on expert judgment.
How does Lumo’s predictive performance compare to traditional surrogate-based estimation approaches when evaluated against expected accuracy criteria across a wide range of compounds?
We tested the model with 49 compounds that were not included in the original training data. 90% were predicted within 60% of the actual concentration. Contrast this with proxy selection based on expert judgment. There is a 1 in 5 chance of being underestimated by more than 40%. Across all training data, 92% of the predictions fell within the 3x window, well within the range required by semi-quantitative E&L analysis.
The system flags unreliable predictions for follow-up. What metrics are used to evaluate confidence levels and how does this feature improve test efficiency and reliability?
Lumo distinguishes between two completely different situations. For example, if a compound is known to be undetectable under the method conditions, such as a pure hydrocarbon in LC-MS/ESI mode, it is assigned a zero.
However, if a compound’s response may be unpredictable (for example, certain silicon- or phosphorus-containing structures), it is flagged rather than inferred. This flag tells the analyst not to trust the numbers here and to use a different approach. This is an important safeguard to prevent bad data from entering your risk assessment.
How does Lumo’s performance compare to traditional surrogate-based estimation in terms of expected accuracy criteria across a wide range of compounds?
For a group of structurally similar aromatic alcohols, the best-case scenario for surrogate selection, expert judgment yielded highly accurate results (within 80–120% of the true value) only 24% of the time. Lumo achieved the same level of accuracy 54% of the time. The probability of dangerous underestimation has decreased from 20% to 10%. Furthermore, unlike surrogate selection, Lumo’s performance does not depend on whether a suitable reference compound happens to be available.
From a practical perspective, what are the most important benefits that laboratories can expect from implementing Lumo (e.g., reduced standard usage, instrument time, improved results, etc.)?
Three things. First, there are fewer standards. Lumo generates RRF predictions from the structure only, so there’s no need to wait for standard procurement or make conservative assumptions while you wait. Second, faster turnaround. Semi-quantitative estimates are available as soon as compounds are identified, reducing the gap between detection and risk assessment. Third, consistency.
The model applies the same logic every time, eliminating inter-analyst variation in surrogate selection. The result is increased protection and faster delivery.
Lumo is described as being consistent with ISO 10993-18’s risk-based approach. How is toxicity assessment supported within this framework?
ISO 10993-18 requires chemical characterization to be performed within a risk management framework and to explicitly account for analytical uncertainty in analytical evaluation thresholds. Good RRF predictions mean low uncertainty in the analysis. This means that AET calculations are more accurate and require less overly conservative uncertainty factors.
The flagging system also creates a transparent audit trail. The meaning of Lumo is simple. Improved RRF predictions better characterize the uncertainties of real-world methods, supporting more defensible and accurate UFs, and ultimately supporting more reliable and less underestimated AETs.
What safeguards are in place to prevent the production of chemically implausible results, and how do these quality checks contribute to the overall robustness of the system?
There are multiple layers. Invalid structures are caught at the input stage. The substructure routing system ensures that compounds are evaluated only by models trained on chemically relevant data. Forced zero logic prevents the model from reporting chemically impossible signals. Additionally, the flagging system ensures that no unreliable estimates ever reach the analyst.
These are not just quality inspections. These make the tool reliable enough to be used in a regulatory context.
Have you observed any measurable results from early adopters of Lumo, especially in terms of accelerating non-target screening and improving decision-making timelines?
The published work is a proof of concept, so we are at the beginning of the adoption curve. But what we can say is that achieving a 90% success rate with off-sample compounds without standards translates directly into fewer subsequent analyzes and faster decision-making.
One of the biggest bottlenecks in non-targeted screening has always been the gap between “I found something” and “I know what to do about it.” Lumo significantly closes that gap.
How does Lumo integrate with existing laboratory workflows and software platforms? Is it designed as a standalone tool or as part of a broader analytical ecosystem?
The input is a SMILES string and the output is a numeric prediction. A simple, universal format that connects to any data pipeline. The underlying software (RDKit, scikit-learn, Python) is completely open source and platform independent. In practice, Lumo is included as an automated post-identification step.
Once a compound is tentatively identified from a spectral match, an RRF prediction is immediately made. It is designed to integrate without requiring workflow redesign.
Looking to the future, how do you envision predictive tools like Lumo evolving to further support extract screening and regulatory compliance in medical and toxicology research?
The near-term roadmap includes more advanced molecular characterization, including but not limited to Morgan fingerprints, graph neural networks, 3D conformational features, and more. Incorporating these should improve the accuracy of complex or special structures.
In the long term, as confidence in ML-based RRF predictions increases through prospective validation, we expect regulators to increasingly accept model-derived values as a defensible basis for risk assessment. This vision is a seamless pipeline from raw analytical data to regulatory submissions, where uncertainties are quantified rather than assumed, speeding decision-making without compromising rigor.
Where can our readers find more information?
Deng, Y. Others. (2026). Neural network prediction of extractables and leachables response factors for 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 the Jordi Institute, an RQM+ company. He trained as a polymer chemist and received his Ph.D. A graduate of the University of Warwick (UK), he has spent over 12 years leading complex research projects across a wide range of analytical techniques. We provide our customers with practical and rational solutions in both problem-solving and regulatory contexts.
About the Jordi Institute
Jordi Labs provides the highest quality contract analytical services and polymer HPLC columns to the world’s leading consumer product, polymer, pharmaceutical, and medical device manufacturers. Our team of PhD analytical chemists specializes in chemical identification. One of our core competencies is extractables and leachables testing.
We are also a world leader in:
- Method development/validation
- Particulate and residue analysis
- Comparison of good and bad points
- polymer analysis
- Polymer breakage
We also assist companies from Fortune 500 companies to innovative startups with method development, preparative HPLC, training seminars, depositions, and consulting. As a family company, we take pride in manufacturing all of our products and providing analytical services. Our goal is to help our customers overcome their analytical challenges by providing superior products and the personal assistance of our staff of highly trained PhD chemists.
