The HPLC 2026 Symposium, held in Indianapolis this year, wrapped up on June 11. The premier conference featured around 800 attendees, cutting-edge separation science sessions, an impressive poster selection and over 30 vendor exhibitions.
Given the show’s niche focus, LC-MS was a reoccurring theme. The technique is increasingly being used across a variety of analytical sub-disciplines to the point where the question is not how sensitive is it, but rather, how sensitive can we make it? This is especially true for the analysis of PFAS chemicals where analysts are measuring at the nanogram per liter level—which is the equivalent of finding one drop of water in four to five Olympic-sized swimming pools.
At HPLC 2026, there were many posters pertaining to PFAS analysis. These, as well as new products and additional trends, will be covered in an article published on June 30. This article focuses on three highly engaging oral sessions that all center on a common theme: AI and LC-MS.
LC-MS/MS and the living brain metabolome
Caitlin Cain, from the Robert Kennedy Lab at the University of Michigan, gave a well-attended and well-received presentation on how advances in LC-MS-based metabolomics coupled with sophisticated data analysis approaches, are helping researchers better understand the chemical complexity of the living brain and the neurochemical mechanisms underlying substance use disorders.
Using highly optimized LC-MS/MS methods, Cain and team were able to identify about 500 compounds from a chromatogram containing nearly 3,700 detected features. To investigate the large number of unidentified signals, the team applied a peak-shape similarity approach known as lack-of-fit clustering. This method identifies features with similar retention times and chromatographic peak shapes that likely originate from the same metabolite through isotopes, adducts, or in-source fragmentation. The analysis revealed that the majority of detected features were degenerate signals rather than unique metabolites. Approximately 500 additional compounds appeared to represent genuine but currently unidentified metabolites, highlighting the need for improved tandem mass spectral libraries and continued advances in LC-MS instrumentation.
The identified metabolites demonstrated the complexity of brain chemistry, encompassing compound classes ranging from amino acids and organic acids to lipids and lipid-like molecules. Building on this analytical foundation, the researchers applied untargeted metabolomics to investigate vulnerability to substance use disorders.
The study employed selectively bred high-responder and low-responder rats, animal models that differ in novelty-seeking behavior and susceptibility to addiction-like traits. High responders display greater exploratory behavior, impulsivity and addiction vulnerability, whereas low responders exhibit reduced exploration and greater anxiety-like characteristics.
Using LC-MS metabolomics and chemometric analysis, the researchers compared baseline brain metabolite profiles between the two phenotypes. Multivariate analysis revealed clear metabolic differences associated with both phenotype and sex. Statistical analysis identified 64 metabolites that distinguished high responders from low responders, with pathway analysis pointing to amino acid metabolism as a key differentiating factor. Because many amino acids function as neurotransmitters or neurotransmitter precursors, these findings suggest a neurochemical basis for addiction vulnerability.
Several metabolites associated with anxiety-related behaviors were also identified. Low-responder animals exhibited elevated levels of corticosterone, as well as changes in endocannabinoid signaling. One particularly interesting finding involved a lipid species—LPA 18:1—which was detected almost exclusively in low-responder animals and has previously been associated with anxiety-like behaviors in rodents.
Cain and team also examined the effects of acute cocaine exposure. Comparing baseline and post-cocaine samples revealed broad shifts in brain metabolism, with 230 metabolites significantly altered following drug administration. Pathway analysis indicated widespread effects on neurotransmission-related processes. Several metabolites associated with neural health and cognitive function declined after cocaine exposure, including N-acetylneuraminic acid and docosenamide. In contrast, kynuric acid, a metabolite associated with neurotoxicity and aging processes, increased markedly following cocaine administration.
Lastly, the researchers explored voluntary cocaine self-administration, a model that more closely resembles human addictive behavior. Longitudinal metabolomic analysis demonstrated substantial neurochemical remodeling over a three-week self-administration period. Changes were particularly evident in antioxidant pathways. Glutathione, the brain's primary antioxidant, initially increased in response to cocaine exposure but later declined as oxidative stress accumulated. Related metabolites, including N-acetylcysteine and carnitine, also became depleted over time, suggesting progressive impairment of antioxidant defenses and increasing neurotoxicity.
The study and results demonstrate the success of LC-MS-based metabolomics, combined with advanced chemometric workflows, for unprecedented insight into living brain chemistry. Cain said she believes these approaches will ultimately help to bridge the gap between brain metabolism and behavior.
AI-augmented LC-MS in metabolomics
Driven largely by advances in high-resolution mass spectrometry (MS), metabolomics is a growing discipline, finding its place and importance in the broader landscape beside genomics and proteomics.
Despite these advances, first-generation untargeted metabolomics faces significant limitations, including challenges related to instrument drift, batch effects, declining sensitivity caused by matrix contamination and more.
In his presentation, Guowang Xu, a researcher from the Chinese Academy of Sciences, suggested that artificial intelligence and large-scale data infrastructure provide a path toward next-generation metabolomics. Rather than relying solely on experimental workflows, Xu said future metabolomics platforms will increasingly combine analytical measurements with automated data processing, machine learning, standardized databases and intelligent decision-making systems.
He outlined four defining characteristics of this next-generation framework: standardization, automation, pipeline-based data processing and AI-assisted analysis throughout the entire workflow.
Xu stressed that even the “most advanced analytical instrumentation cannot compensate for poor sample collection and handling” practices. His laboratory has collaborated with clinical partners to develop SOPs for serum, plasma, urine, and other biological samples, as well as quality control markers capable of assessing sample suitability for metabolomics analysis.
The laboratory is also developing highly automated analytical workflows. These systems integrate sample storage, freeze-drying, centrifugation, preparation, sealing, and UPLC-MS analysis into a robotic platform capable of operating with minimal human intervention.
To address data analysis challenges like different laboratories obtaining different biological conclusions when analyzing the same metabolomics dataset, Xu and team are developing an integrated software platform that performs feature extraction, metabolite annotation, pathway analysis, biomarker discovery, and patient subgroup classification within a standardized workflow.
The laboratory has also assembled a database containing approximately 50,000 metabolites and is using molecular networking strategies to improve structural characterization of unknown compounds.
Xu concluded that the future of metabolomics will depend on combining advanced chromatography and mass spectrometry with AI, automation, large-scale databases and interdisciplinary expertise.
Like many other disciplines, as metabolomics expands to more and more applications, laboratories will need scientists who bridge analytical chemistry, biology, informatics and data science to fully realize the field’s potential.
AI‑powered chromatographic method development
In his presentation, “From Experiments to Algorithms: AI‑Powered Digital Transformation of Chromatographic Method Development,” Pankaj Aggarwal outlined how pharmaceutical laboratories are applying machine learning, statistical modeling, and digital analytics across the drug development lifecycle to improve chromatographic method development, robustness, and long-term performance monitoring.
Aggarwal, an analytical chemist at Merck, described a project in which his team analyzed large historical datasets containing tens of thousands of compounds. Initial efforts used classical machine learning approaches, including support vector machine (SVM) models, to predict retention times. These models achieved strong predictive performance, with R² values around 0.85, while demonstrating that smaller, chemically similar clusters of compounds could perform nearly as well as larger global datasets.
The team then moved on to more advanced modeling approaches. Graph neural networks outperformed the earlier SVM models, increasing predictive accuracy to approximately R² = 0.9. To further reduce experimental requirements, researchers applied machine learning techniques. By training a model on one chromatographic condition and adapting it to a new condition, the team reduced the amount of new experimental data needed from roughly 64,000 compounds to as few as 80. According to Aggarwal, these models were capable of distinguishing subtle structural differences, including compounds differing by a single functional group.
But a subsequent challenge arose: incorporating chromatographic conditions directly into predictive models. Rather than relying solely on molecular descriptors, the researchers built datasets spanning multiple columns and mobile-phase conditions. Using approximately 500 carefully selected analytes designed to represent chemical diversity, the team developed models capable of predicting optimal screening conditions. Experimental validation showed that model-recommended conditions successfully identified effective separations, suggesting that machine learning could significantly reduce chromatographic screening workloads and eliminate unnecessary experimental runs.
Aggarwal concluded that no single modeling approach is sufficient across the entire pharmaceutical workflow. Instead, different tools are appropriate at different stages, ranging from machine learning for rapid screening, to Bayesian and statistical methods for robust development, to digital monitoring systems for lifecycle management. Ultimately, success depends not on big data alone, but on systematically generated, high-quality data combined with close collaboration between analytical scientists and data scientists.