Artificial Intelligence in Chromatography: Advancing Method Development and Data Interpretation

Artificial intelligence (AI) is rapidly reshaping chromatography, offering powerful tools to streamline traditionally labor-intensive processes such as method development, peak detection, and data interpretation. By leveraging large datasets, AI can optimize chromatographic methods, detect subtle patterns in complex profiles, and deliver insights that would be difficult or impossible to achieve manually. Machine learning (ML) approaches are enabling smarter peak deconvolution, reducing false positives, and handling overlapping peaks with greater efficiency than conventional mathematical algorithms. However, integrating AI into real-time monitoring and existing software platforms presents challenges that demand careful consideration. LCGC International spoke to Dave Abramowitz, from the unified product management and product ownership team for chromatography and mass spectrometry software at Thermo Fisher Scientific, about the advantages, challenges, and future potential of AI in chromatography, highlighting how scientists can begin adopting AI-driven tools to improve accuracy, efficiency, and discovery.

How is AI enhancing method development, peak identification, and data interpretation in chromatography workflows?

The capability for AI to utilize large amounts of data for different purposes is really changing chromatography. The large amounts of data now available to analyze using AI are allowing scientists to drive decisions faster with a higher level of confidence. We are gaining the ability to build in-silico predictors for compounds under varying conditions. AI will also drastically change method development, which has historically been done manually and by trial-and-error. AI can optimize method parameters by analyzing large previously run data sets and build methods without the trial-and-error method that is currently used. Peak detection has generally been performed by mathematical algorithms that use first and second derivatives to calculate inflection points and build reproducible chromatographic peak detection. One of the biggest problems about this type of approach is making the algorithms aware of retention drift, matrix effects, compressing/stretching and other factors. The ML approach to peak detection shifts the dependency from a mathematical approach to a learning-engine approach trained on data sets to identify peaks, both obvious and those less-so, by cycling through three steps of matching, selecting and executing rules.

Data interpretation is where AI will deliver the most value in efficiency and scientific insight. The ability that AI has to analyze large datasets and detect patterns that would otherwise not be evident, find new relationships between data elements, and allow scientists to discover something new is most promising.

What are the advantages of using ML over traditional methods in chromatographic peak deconvolution?

ML models can be specifically trained for certain data sets, which is incredibly beneficial for large scale studies such as metabolomics or proteomics by reducing the need for manual curation. ML can quickly make decisions to assess elements like signal quality and peak shape leading to fewer false positives. Specifically with regards to peak deconvolution, ML models are better able to address overlapping and otherwise complex peaks, can process the data much more quickly, and can continuously learn and improve when trained on new and additional data.

What challenges might arise when applying AI to real-time chromatography monitoring, and how would you address them?

Real-time chromatography monitoring is an interesting topic. At face value, allowing an AI agent access to real-time data to make decisions sounds great. The challenges with this lies with the types of decisions that AI can make autonomously. What happens if the AI system goes down—what are the failovers? What about anomalous data points or errors in the model or tolerances of parameters that may result in a bad decision? One solution is to guarantee there is a person in the loop confirming the system is operating as intended and that decisions being made by the system follow protocol and adhere to standards. Putting dashboards and alarms in place to “monitor the monitors,” regularly revisiting the model’s performance, creating test systems to try new models and new parameters to measure against incumbent systems are all ways to address possible challenges. Additional concerns are on expecting high quality real-time data and always having available computational resources to perform the real-time monitoring. Addressing these concerns with IT or your IT service provider is critical to guarantee data quality and resource availability.

In your opinion, how important is data quality and labeling for building robust AI models in chromatographic analysis?

High-quality, well-labeled chromatographic data is critical and is fundamental to build robust AI models to analyze data, learn patterns, identify peaks, and make predictions. High-quality data also allows for a more streamlined path to reproducibility and validation. The time saved through AI-assisted analysis can only be realized if the underlying models are trained on high-quality, well-labeled data.

How would you approach integrating AI models into an existing chromatography software platform?

Carefully. Today, I am a bit hesitant to use AI models for critical parts of the workflow, such as acquisition, but as the technology progresses and the ways to test it get better, we will start using AI for these critical steps. In the meantime, I would focus on areas like peak integration, peak selection and report generation. The quickest wins will likely be in pattern recognition, monitoring trends over time and making decisions based on predicted outcomes. Implementing these models will lead to new insights into pre-existing data while driving decisions for the future. Finding the right feature that leverages will use AI models most effectively and give scientists an added edge is critical.

Can you provide an example of a published study or project where AI significantly improved chromatographic analysis? What was the outcome?

Dimitri Abrahamsson’s team at Stockholm University developed a practical approach using our Q Exactive Orbitrap mass to predict molecular structures based on how compounds partition between different solvents and water (1). Their neural network achieved about 70% accuracy in predicting functional groups for unknown compounds without analytical standards, addressing a major challenge in environmental analysis. This method helps identify thousands of compounds that would typically remain uncharacterized in complex samples, expanding our understanding of chemical exposures in environmental and biological matrices.

What role does feature selection play in developing AI models for analyzing chromatographic datasets, and how would you implement it?

Feature selection is crucial in finding the right mix of features—there are too many to analyze, and the noise becomes inseparable from the signal, too few and the model can produce garbage data. It is a process learned through experience and adjusting until the model consistently produces results that are as good as those produced with non-AI analysis, meaning that the results must be achieved quickly, accurately, and concisely. With the right features included, AI models could see through the noise of complex data sets and highlight the truly informative peak characteristics, spectral patterns, and retention time relationships.

Do you think AI could eventually replace human analysts in chromatographic interpretation? Why or why not?

I do not think that AI will replace human analysts in chromatographic interpretation, but I do think it will reduce the number of humans needed. AI will reduce the effort taken to interpret every chromatogram and remove the simplest forms of analysis, leaving those the system has trouble with for humans to interpret.

What practical advice would you give analytical chemists who want to start integrating AI-driven data analysis tools into their day-to-day workflows?

Whether using AI or not, you are ultimately responsible for the quality and accuracy of your work—you can't blame an AI agent when something is falsely reported or incorrect. Trusting AI models implicitly is dangerous—labs can leverage the technology, but they should always verify the results. Ensure the data used to train any models is highly curated and accurate. The concept of "garbage in, garbage out” applies here—if a model is trained on poor data, it will produce poor results. Labs should not try to implement AI all at once, but rather, they should implement AI in steps. Be imaginative in using the AI tools available to you. AI doesn't have to be used everywhere. It can be used in specific areas to great effect, saving hours of effort and hopefully allowing labs to make new discoveries.

Join us at noon EST all this week as Dave Abramowitz dives deeper on a variety of topics pertaining to AI in chromatography through a series of short video interviews on this website.

References

1. Abrahamsson, D.; Brueck, C. L.; Prasse, C. et al. Extracting Structural Information from Physicochemical Property Measurements Using Machine Learning─A New Approach for Structure Elucidation in Non-targeted Analysis. Environ. Sci. Technol. 2023, 57 (40), 14827-14838. DOI: DOI: 10.1021/acs.est.3c03003