AI/ML In Practice: Predicting the Measurable Chemical Space for Nontargeted LC–ESI–HRMS Analysis Workflows

In the paper “Measurable Feature Prediction for Estimating Chemical Space Coverage in LC–ESI–HRMS Nontargeted Analysis”,1 what was the primary motivation for developing this framework, and how does it address current limitations in defining experimentally accessible regions of chemical space in LC–ESI–HRMS workflows?

The primary motivation behind the framework was to address a longstanding (and often underestimated) challenge in non-targeted analysis (NTA): although liquid chromatography electrospray ionization high-resolution mass spectrometry (LC–ESI–HRMS) workflows are routinely described as “comprehensive”, the actual portion of chemical space that is experimentally accessible under a given method remains poorly defined. In practice, chemical space measurability depends on a combination of chromatographic retention, ionization efficiency, and instrument-specific response. Together, these factors inherently limit the measurable regions of the sample chemical space and the identification of new/unexpected compounds.

Current approaches to defining a method’s measurable chemical space are largely empirical and heavily dependent on available native analytical standards or, in some cases, retrospective observations from previous studies.²

This creates two important limitations. First, experimentally characterized compounds represent only a very small fraction (about 10%) of the environmentally relevant chemical space. Second, method performance evaluated after experimental acquisition, is often overestimated beyond the selectivity domain.

The framework was designed to move beyond this trial-and-error paradigm by introducing a bottom-up, structure-driven approach capable of predicting whether compounds are likely to be experimentally measurable before acquisition takes place. By integrating molecular fingerprints with predictive models for retention index and ionization efficiency, the framework estimates the likelihood that a compound falls within the detectability window of a specific LC–ESI–HRMS method.

An important aspect of the work is that measurability is treated as a continuum rather than a binary detectable/non-detectable classification. This allows chemical space to be mapped according to regions of higher or lower analytical accessibility, providing a more realistic representation of method behavior. The framework therefore helps bridge the gap between theoretical chemical diversity and experimentally observable space, offering a scalable strategy for evaluating analytical coverage across millions of candidate structures.

More broadly, the study highlights that measurable chemical space is not a fixed property of a mass spectrometer alone, but rather an emergent property of the entire analytical workflow. Through this, the framework provides a foundation for comparing methods, identifying blind spots, and ultimately improving confidence in exposomic and environmental NTA studies.

How does the integration of molecular fingerprints with quantitative structure–property relationships, such as retention index and ionization efficiency, improve the predictive power of measurability compared to traditional empirical or rule-based approaches?

Traditional approaches for estimating measurability in LC–ESI–HRMS often rely on simplified physicochemical descriptors such as logP, molecular weight, and pKa. While useful, these descriptors only capture a limited portion of the structural complexity governing measurability by chromatographic retention and ionization processes.

In contrast, the framework combines molecular fingerprints with quantitative structure–property relationship (QSPR) modeling to encode much richer structural information. Molecular fingerprints capture detailed substructural patterns, functional group distributions, and connectivity relationships that are directly linked to analyte behavior in LC–ESI–HRMS systems. This enables the model to map structural similarities between known (e.g., analytical standards) and sample-related compounds in a well resolved chemical space, even when physicochemical descriptors substantially differ.

The integration of fingerprints with predicted retention index (RI) and ionization efficiency (IE) is particularly important because these two properties define complementary aspects of measurability. Retention determines whether a compound can be chromatographically separated and detected within the acquisition window, whereas ionization efficiency governs the likelihood that sufficient signal will be generated in the electrospray source (relevant also to the fragmentation process and structural elucidaiton). Modeling these properties together therefore provides a more holistic description of analytical accessibility.

Once the model is trained on the chemical space of interest (chemical repositories such as CompTox).³

What considerations guided the selection and optimization of the k-nearest neighbor regression model for predicting retention index and ionization efficiency, and how sensitive are the predictions to the choice of distance metrics?

The selection of k-nearest neighbor (k-NN) regression was motivated by both interpretability and flexibility. In the context of chemical similarity modeling, k-NN offers a relatively intuitive framework: measurable features are predicted based on the behavior of structurally similar neighbors in the fingerprint space. This aligns well with the underlying assumption that structurally related molecules tend to exhibit comparable chromatographic and ionization properties. Distance metric here is especially important because it defines how similarity neighbors are assigned in high-dimensional fingerprint space, representing the applicability domain of the model itself.

An additional advantage is that k-NN does not impose a predefined functional relationship between molecular descriptors and analytical response. This is particularly relevant for LC–ESI–HRMS applications, where retention and ionization processes are often highly nonlinear and influenced by multiple structural features and experimental parameters.

Model optimization focused on balancing predictive accuracy with generalizability across diverse chemical domains. Parameters such as the number of neighbors (k), fingerprint representation by distance metric, were systematically evaluated through cross-validation procedures. The goal was not only to minimize prediction error, but also to preserve local chemical similarity relationships that remain chemically meaningful.

How does the framework ensure that the sampled chemical space, for example, from CompTox, is representative of real-world exposomic mixtures, and what biases might arise from database-dependent coverage?

Representing real-world exposomic chemical space is one of the central challenges in non-targeted analysis (NTA) research because no database can fully capture the diversity of chemicals present in environmental or biological systems. Additionally, leveraging large-scale repositories is computationally demanding and often approximations are required.

Databases like CompTox provide a practical approximation of chemically plausible exposomic space and enables the evaluation of analytical coverage across hundred thousands of candidate structures. Importantly, through the distance threshold, the framework does not assume that all database entries are equally relevant or equally measurable. Instead, the predicted measurability landscape itself can be used to identify regions of chemical space that are either analytically accessible or systematically underrepresented.

The major bias of the combination of available analytical standard data with chemical space databases is that compounds tend to cluster within regions already compatible with the established analytical methods in NTA. This creates a feedback loop in which measurable compounds are preferentially investigated, further reinforcing the apparent boundaries of accessible chemical space.

The framework helps expose these biases rather than conceal them. By projecting large chemical inventories into predicted retention and ionization domains, it becomes possible to visualize where analytical blind spots occur and where future standards, methods, or acquisition strategies may be needed.

Ultimately, the study emphasizes that measurable chemical space should not be interpreted as synonymous with total exposomic space. Instead, it represents the subset of chemistry accessible under specific analytical conditions. Recognizing this distinction is essential for interpreting NTA results realistically and for avoiding overconfidence in apparent analytical comprehensiveness.

To what extent do chromatographic conditions, such as stationary phase and gradient profile, versus ionization parameters dominate the boundaries of “measurable space”?

The boundaries of measurable chemical space in LC–ESI–HRMS are shaped by both chromatographic and ionization constraints, but their relative influence can vary substantially depending on the analytical workflow and the chemical domain being considered.

Chromatographic conditions primarily determine whether compounds fall within a practically observable retention window. Parameters such as stationary phase chemistry, mobile phase composition, gradient slope, and pH strongly influence analyte retention and separation selectivity. Compounds that elute too early may suffer from ion suppression and poor retention, whereas excessively retained compounds may never elute efficiently within the acquisition window.

Ionization parameters, on the other hand, govern signal generation once compounds reach the electrospray source. Even compounds with favorable chromatographic behavior may remain effectively invisible if their ionization efficiency is insufficient under the selected source conditions. Factors such as solvent composition, additive chemistry, source voltage, and droplet desolvation all contribute to this response.

One of the strengths of the framework is that it models retention-related and ionization-related behavior separately before integrating them into a unified measurability landscape. This separation makes it possible to distinguish whether analytical blind spots originate primarily from chromatographic exclusion or from ionization inefficiency.

For example, a compound class may exhibit favorable predicted ionization efficiencies but consistently poor retention under reversed-phase conditions, suggesting that alternative chromatographic strategies such as hydrophilic interaction liquid chromatography (HILIC) or mixed-mode separations may improve coverage. Conversely, compounds with acceptable retention but low predicted ionization efficiency may benefit more from source optimization, polarity switching, or alternative ionization techniques, such as atmospheric pressure chemical ionization ( APCI).

Rather than evaluating methods solely through the number of detected features, the framework enables comparison in terms of the regions of chemical space each analytical configuration is predicted to access, providing a more systematic route for understanding and optimizing analytical selectivity in NTA workflows.

How can this framework be operationalized to guide method development or selection of orthogonal LC–ESI–HRMS setups in routine NTA workflows, and what are the limitations when scaling to new/specific context applications?

One of the most practical aspects of the framework is its ability to evaluate analytical coverage before experimental measurements are performed. In routine NTA workflows, this creates opportunities for more rational method development by allowing users to predict which regions of chemical space are likely to be captured—or missed—under specific LC–ESI–HRMS conditions.

For example, the framework can be used to compare orthogonal chromatographic setups by projecting large chemical inventories into predicted retention and ionization domains. This makes it possible to identify complementary analytical windows among reversed-phase, HILIC, or mixed-mode separations, as well as between positive and negative ionization modes. Rather than relying exclusively on empirical trial-and-error experimentation, analysts can prioritize workflows predicted to maximize overall chemical coverage. In exposomics and environmental screening, where analytical comprehensiveness is increasingly important, this type of predictive prioritization can significantly improve workflow efficiency.

However, scalability may introduce specific limitations to the approach. Among these, chemical space coverage prediction may suffer from a reduced chemical diversity in the standards used as anchors to the sample space of interest, causing certain chemical domains to remain sparsely represented, especially highly polar, unstable, or emerging contaminants.

How does this approach benefit chromatographers in practice?

For chromatographers, one of the key benefits of the framework is that it provides a quantitative way to evaluate the achievable coverage for a specific analytical selectivity before extensive experimental work is performed. Traditionally, method development in non-targeted LC–ESI–HRMS has relied heavily on iterative optimization and empirical experience. While effective, this process can be resource-intensive and may still leave important regions of chemical space unexplored.

The framework introduces a more predictive strategy by estimating which compounds are likely to be retained, ionized, and ultimately detected under a given analytical configuration if occurring in the analysed sample. This allows chromatographers to move beyond simply maximizing feature counts and instead evaluate how effectively a method samples chemically relevant space.

Another important benefit is transparency. Because the framework is based on chemically interpretable similarity relationships and measurable analytical properties, the resulting predictions can often be rationalized in terms familiar to chromatographers, such as polarity, functional group chemistry, and retention behavior.

More broadly, the approach contributes to a shift in NTA from feature-centric analysis toward chemical space-centric analysis. Instead of asking only “What did we detect?”, the framework encourages analysts to also ask “What was the chemical space realistically accessible under this method?” This distinction is critical for interpreting non-targeted datasets and for understanding the analytical biases inherent to any LC–ESI–HRMS workflow.

By integrating predictive modeling with chromatographic knowledge, analysts can make more informed decisions about method design, analytical coverage, and the interpretation of complex exposomics datasets.

References

1. Renai, L., Turkina, V., Chojnacka, A., Gargano, A. F. G. & Samanipour, S. Measurable Feature Prediction for Estimating Chemical Space Coverage in LC–ESI–HRMS Nontargeted Analysis. Anal. Chem. 2026, 98, 7637–7643. DOI:https://doi.org/10.1021/acs.analchem.5c07705
2. Hulleman, T.; Turkina, V.; O’Brien, J. W.; Chojnacka, A.; Thomas, K. V.; Samanipour, S. Critical assessment of the chemical space covered by LC–HRMS non-targeted analysis. Environ. Sci. Technol. 2023, 57, 14101– 14112. DOI: https://doi.org/10.1021/acs.est.3c03606
3. Williams, A. J.; Grulke, C. M.; Edwards, J.; McEachran, A. D.; Mansouri, K.; Baker, N. C.; Patlewicz, G.; Shah, I.; Wambaugh, J. F.; Judson, R. S.The CompTox Chemistry Dashboard: a community data resource for environmental chemistry. J. Cheminform. 2017, 9, 61. DOI: https://doi.org/10.1186/s13321-017-0247-6

Biography

Lapo Renai is a Marie Skłodowska-Curie Postdoctoral Fellow at the Van 't Hoff Institute for Molecular Sciences, University of Amsterdam. His research combines chromatography, high-resolution mass spectrometry, and chemometrics to improve non-targeted analysis and exposomics. His current work focuses on understanding and expanding the measurable chemical space of LC–HRMS workflows through computational and analytical innovations for environmental monitoring.