Advances in Chromatography Using Artificial Intelligence and Machine Learning

Artificial intelligence (AI) is “the theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages” (1). Foodomic domains, including food metabolomics, sensomics, nutrimetabolomics, and food volatilomics, require great analytical efforts to comprehensively capture sample composition and connect it with biological phenomena or functional properties (2–5). Effective workflows start with the pre-processing of raw data generated by multidimensional analytical systems and proceed with dedicated processing to identify or “visualize” chemical patterns linked to, or predictive of, key properties, such as like nutritional quality, sensory profile, authenticity. AI, by definition, supports researchers in this process that translates chemistry into properties or functions. Chiara Cordero and Marco Vincenti from the University of Turin, Italy have learned how to combine their complementary competencies in analytical chemistry and big data analytics to achieve significant advances in food science and health.

AI and machine learning (ML) are increasingly becoming “hot topics” in separation science. Is there a difference between the role of AI and ML in relation to chromatography?

MARCO VINCENTI: AI and ML are increasingly becoming essential in separation science because the large, multidimensional, high-resolution data generated in the chromatographic process can no longer be efficiently handled with classic intuitive representations based on two-dimensional (2D) features selection. There is no clear distinction in the literature between AI and ML except that ML can be regarded as a subchapter of AI. In relation to chromatography, ML generally operates on large, internally-produced experimental data by means of prearranged programming, methods, and instructions, to generate compositional information not obtainable by classic data processing. Beyond these tasks, AI uses multi-source data including external depositories and databases, to achieve complex objectives by means of flexible strategies, for example, to highlight otherwise undetectable data patterns or to optimize specific analytes’ class separations under multiple and simultaneous constraints. In particular, AI procedures involve—to a variable degree—autonomous decision-making steps leading to the final output.

Why is AI or ML necessary in chromatography, generally? Do you think it will become increasingly used in two-dimensional chromatography?

VINCENTI: AI and ML strategies are becoming essential because at the moment because chromatography is consistently combined with 2D chromatography itself, or spectroscopic techinques, or tandem- or high-resolution mass spectrometry. These hyphenations generate high-density, multi-dimensional data for each of the samples that should be compared to one another, as for example in metabolomics. Extracting minimal differences from these data with similar patterns is a task not feasible by human senses, even if helped by selective data plotting. Two-dimensional chromatography adds further complication to this frame, appending the need of accurate data pre-treatment for peak alignment among different samples or injections to correct potential retention times shifts along a two-dimensional plane. Notably, these shifts are not necessarily linear with retention time. Intelligent automated peak recognition algorithms are crucial to operate these corrections.

Chiara, you recently reviewed the AI tools and concepts adopted to exploit the informative potential of comprehensive two-dimensional gas chromatography (GC×GC) in foodomics. Can you tell us more, and what problems AI solves?

CHIARA CORDERO: Comprehensive two-dimensional chromatography, that is, GC×GC or LC×LC, offers unique opportunities to investigate in greater detail the sample’s compositional complexity. When samples are characterized by a high chemical “dimensionality,” as defined by Giddings, their comprehensive investigation should be approached by suitably designed multidimensional systems (6). In particular, when comprehensive 2D-GC or 2D-LC are used, they highlight ordered structures in retention data, for example, homolog series or chemical classes, and support efficient separation of isomers or isobars with great benefits for confident identification and accurate quantification.

If we can connect the detailed chemical profile with functional and nutritional properties or quality of food, as in many foodomic domains, we advance our knowledge and develop more accurate and precise predictive models (2). Volatiles and semi-volatiles have, in GC×GC–MS, the analytical platform of choice for detailed profiling and effective fingerprinting, investigation strategies that are currently realized with the aid of AI tools (7). Image pattern recognition (PR) is efficiently used to track and realign features across many chromatograms and samples, while computer vision (CV) exploits the full data array of chromatographic images to highlight compositional differences even in the presence of unresolved mixtures or confounding phenomena.

What do these tools solve?

CORDERO: In the context of comprehensive 2D chromatography, image PR by template-matching algorithms enables effective and confident tracking of features across many samples even in case of temporal misalignments due to oven temperature (un)stability or carrier gas flow and pressure inconsistencies (8,9). The use of different modulation technologies, such as thermal modulation or differential-flow modulation or changes in column configuration can be also effectively tackled with this AI tool to enable metadata transfer between applications or analytical campaigns (8,10). This process or realignment is particularly challenging for untargeted components where classical approaches of peak tracking fail or require extensive computational time. Moreover, untargeted features could add further knowledge to the interpretation of many phenomena, their reliable mapping using the information provided in full by the analytical system, such as retention times in two dimensions, retention index, detector response, and spectral signature, to support effective re-investigation of samples.

Computer vision is implemented in commercial software platforms with tools using simplified grids or tiles or peak regions that explore the 2D or 3D array of data and facilitate the identification of compositional differences between samples and samples’ classes CV, when supported by image PR, enables augmented visualization and gives access to the chemical information encrypted in the data array (11).

Does AI help in the "translation between languages" when applied to foodomics applications when applied to foodomics applications?

CORDERO: The intriguing application of AI as a sensomics-based expert system (SEBES) capable of predicting key aroma signatures of food without human olfaction has been successfully realized for many foods (12–14). It is referred to as an AI smelling machine because it captures key food odorant patterns, and by their accurate quantification resembles the aroma identity of food in an objective and unbiased way. GC×GC is the core of the analytical platform since it enables efficient separation of odorants from the bulk of interfering volatiles and provides suitable method sensitivity to achieve sub-parts per billion (ppb) levels for the most potent aroma active compounds. The approach of AI smelling truly enables “translation between languages” by means of translating chemical patterns in olfactory qualities.

Any advice to chromatographers who want to incorporate AI to improve their methods?

VINCENTI: Several ML softwares are freely available on the internet at increasingly “user-friendly levels”. These softwares may be very useful for the elaboration of chromatographic data produced at laboratory level, provided that the user has basic knowledge on handling R, Phyton, or MATLAB languages and a deeper knowledge on the theory of the different ML methods and their pertinence to the problems to be solved. The ability to exploit these powerful ML methods rapidly increases with practice. More dedicated AI softwares for 2D chromatography, possibly combined with mass spectrometry, are generally proprietary and relatively expensive, but are recommended when the queries’ complexity of chromatographic analyses increase substantially.

Can you summarize the role of AI and ML in chromatography?

CORDERO and VINCENTI: From our viewpoint, AI and ML appear to be mandatory to make significant progress in the comprehension of complex phenomena; in other words, to profitably investigate analytes at the the boundaries between chemistry and biology (15). Currently AI tools are mostly accessible to specialized computer scientists and bioinformatics, and more user-friendly software platforms should be made available to chemists and biologists— and even provided with sufficient ML knowledge—to boost rapid progresses in crucial applications, such as food science and health. Once validated by a wider community of differently specialized scientists, these manageable AI tools will be more easily accepted in a variety of research and industrial laboratories, where the analytical data frequently trigger the decisions and drive the strategies.

References

(1) The Oxford Dictionary of Phrase and Fable, Oxford University Press, 2005. DOI: 10.1093/acref/9780198609810.001.0001

(2) Wishart D.S. Metabolomics: Applications to Food Science and Nutrition Research. Trends Food Sci. Technol. 2008, 19 (19), 482–493. DOI: 10.1016/j.tifs.2008.03.003

(3) Dunkel, A.; Steinhaus, M.; Kotthoff, M.; Nowak, B.; Krautwurst, D.; et al. Nature’s Chemical Signatures in Human Olfaction: A Foodborne Perspective for Future Biotechnology. Angew. Chemie - Int. 2014, 53 (28), 7124–7143. DOI: 10.1002/anie.201309508

(4) Ulaszewska M.; Weinert C. H; Trimigno A.; Portmann, R.; Andres Lacueva C. et al. Nutrimetabolomics: An Integrative Action for Metabolomic Analyses in Human Nutritional Studies. Mol. Nutr. Food Res. 2019, 63 (1), 1800384. DOI: 10.1002/mnfr.201800384

(5) Lytou, A. E.; Panagou, E. Z.; Nychas, G.-J. E.; Volatilomics For Food Quality and Authentication. Curr. Opin. Food Sci. 2019, 28, 88–95. DOI: 10.1016/j.cofs.2019.10.003

(6) Giddings, J. C. Sample Dimensionality: A Predictor of Order-disorder in Component Peak Distribution in Multidimensional Separation. J. Chromatogr. A. 1995, 703 (1–2), 3–15. DOI: 10.1016/0021-9673(95)00249-M

(7) Caratti, A.; Squara, S.; Bicchi, C.; Liberto, E.; Vincenti, M.; et al. Boosting Comprehensive Two-dimensional Chromatography with Artificial Intelligence: Application to Food-omics. TrAC Trends Anal. Chem. 2024, 174, 117669. DOI: 10.1016/j.trac.2024.117669

(8) Stilo, F.; Cordero, C.; Bicchi, C.; Peroni, D.; Tao, Q.; Reichenbach S.E. Chromatographic Fingerprinting by Template Matching for Data Collected by Comprehensive Two-Dimensional Gas Chromatography. J. Vis. Exp. 2020, (i63) e61529. DOI: 10.3791/61529

(9) Cialiè Rosso, M.; Stilo, F.; Bicchi, C.; Charron, M.; Rosso, G. et al. Combined Untargeted and Targeted Fingerprinting by Comprehensive Two-Dimensional Gas Chromatography to Track Compositional Changes on Hazelnut Primary Metabolome During Roasting. Appl. Sci. 2021, 11 (2), 525. DOI: 10.3390/app11020525

(10) Stilo, F.; Bicchi, C.; Jimenez-Carvelo, A. M.; Cuadros-Rodriguez, L.; Reichenbach, S. E.; et al. Chromatographic Fingerprinting by Comprehensive Two-dimensional Chromatography: Fundamentals and Tools. TrAC Trends Anal. Chem. 2021, 134, 116133.
DOI: 10.1016/j.trac.2020.116133

(11) Caratti, A.; Squara, S.; Bicchi, C.; Tao, Q.; Geschwender, D.; et al. Augmented Visualization by Computer Vision and Chromatographic Fingerprinting on Comprehensive Two-dimensional Gas Chromatographic Patterns: Unraveling Diagnostic Signatures in Food Volatilome. J. Chromatogr. A. 2023, 1699, 464010. DOI: 10.1016/j.chroma.2023.464010

(12) Nicolotti, L.; Mall, V.; Schieberle, P. Characterization of Key Aroma Compounds in a Commercial Rum and an Australian Red Wine by Means of a New Sensomics- Based Expert System (SEBES) - An Approach to Use Artificial Intelligence in Determining Food Odor Codes. J. Agric. Food Chem. 2019, 67 (14), 4011–4022. DOI: 10.1021/acs.jafc.9b00708

(13) Squara, S.; Caratti, A.; Fina, A.; Liberto, E.; Spigolon, N.; et al. Artificial Intelligence Decision-making Tools Based on Comprehensive Two-dimensional Gas Chromatography Data: The Challenge of Quantitative Volatilomics in Food Quality Assessment. J. Chromatogr. A. 2023, 1700, 464041. DOI: 10.1016/j.chroma.2023.464041

(14) Stilo, F.; Segura Borrego, M. del P.; Bicchi, C.; Barraglinol, S.; Callejon Fernandez, R. M.; et al. Delineating The Extra-virgin Olive Oil Aroma Blueprint by Multiple Headspace Solid Phase Microextraction and Differential-flow Modulated Comprehensive Two-dimensional Gas Chromatography. J. Chromatogr. A. 2021, 1650, 462232. DOI: 10.1016/j.chroma.2021.462232

(15) Hofmann, T. F. Vanishing Boundaries Between Chemistry and Biology: Reflections on Our Jssournal. J. Agric. Food Chem. 2015, 63 (32). DOI: 10.1021/acs.jafc.5b03450

About the Authors

Chiara Cordero is a full professor of food chemistry at the University of Turin (Torino, Italy). Her passion is gas chromatography (1D-2D) thanks to the unique opportunities it offers in foodomics. Research interests focus on the development of instrumental configurations and data processing tools for comprehensive two-dimensional GC in high-resolution profiling and fingerprinting of complex samples. Application domains include food metabolomics and volatilomics, nutrimetabolomics, and sensomics. The goal is to go beyond the current knowledge and explore the chemistry behind biological phenomena. She received the “Leslie S. Ettre Award” in 2008 as a young scientist for “presenting original research in capillary gas chromatography with an emphasis on environmental and food safety”, the “John B. Phillips Award” in 2014 for her research activity in the GC×GC field, and the Scientific Achievement Award in 2022 for her commitment in the GC×GC research community.

Marco Vincenti is a full professor of analytical chemistry at the University of Turin (Torino, Italy). His 40-year career has continuously evolved around organic mass spectrometry and hyphenated chromatographic techniques in different applications. His most recent research interests focus on the development of targeted and untargeted methods for metabolomic studies in clinical, forensic, and agricultural chemistry combined with machine learning data elaboration algorithms. He has authored or coauthored more than 200 peer-reviewed publications and 10 book chapters.