Limitations of databases and opportunities for Artificial Intelligence

Limitation of mass spectrometric databases Gas chromatography-mass spectrometry coupling (GC-MS) has become the standard method for many analytical questions [1]. This is attributed not only to its high reproducibility [2], excellent separation performance, and wide range of applications [3] but also to the commonly used electron impact ionization (EI). This method produces highly characteristic, and, most importantly, comparable mass spectra [4], which have been cataloged in databases since the 1970s [5]. These standardized databases, some containing over 1 million entries [6,7], allow for the identification of many known substances within seconds. With the increasing requirements within the chemical industry, the pressure on analytical methods is also rising. More frequently, holistic and explorative analyses must be conducted, requiring not only the elucidation of known compounds but, ideally, unknown compounds as well. While databases supply the structure of compounds known by mass spectrometry in non-target screening (NTS) [8], for mass spectrometry unknown compounds can only be elucidated with considerable effort [1,9]. Considering the range of possibilities for these substances, it becomes evident that databases can hardly cover this space. Currently, over 100 million substances are structurally known to humans [10], while more than 100 billion substances are chemically accessible [11-13]. Even without factoring in that most naturally occurring compounds lack structural information, this fundamental discrepancy of at least five orders of magnitude leads to one conclusion: MS databases are insufficient for effective NTS and, consequently, for current analytical needs and challenges! Since the number of structurally known and accessible substances is growing significantly faster than the size of mass spectrometry databases, achieving improved coverage of the required chemical space is unattainable, even with the greatest effort or the aid of simulated spectra [14]. Given the immense variety of possibilities that exist for small molecules (>>10⁶⁰), the only conclusion is that even the best databases will always stay incomplete.

Spectra interpretation for the derivation of structures

An alternative approach to the structural elucidation of unknowns arises from how experts have been evaluating EI mass spectra for decades: by systematically interpreting the existing fragmentation patterns. Although this direct approach can vary significantly depending on the application and the evaluators' experience, it typically consists of five core steps [1,3]:

1. Definition of a possibility space based on preliminary sample information.
2. Determination of the molecular mass from the spectrum.
3. Interpretation of the main fragments to derive substructures.
4. Formation of structural hypotheses from substructures and existing knowledge.
5. Re-examine the hypothesis, for example, through additional chemical and analytical information.

Thereby, steps three through five are repeated until a plausible structural hypothesis is identified, considering all available information. The benefit of this process, namely the de-novo generation of molecular proposals based on analytical data, is that it is capable of elucidating even fully unknown substances. As this process is particularly time-consuming and requires exceptional expertise, the question arises whether this process, namely the de-novo structure elucidation, can also be replicated by software. The concept of an expert system for structure elucidation has been around since the 1960s, with the Dendral project [15], which is still considered one of the pioneering works in artificial intelligence (AI). Although Dendral made significant contributions to modern chemoinformatic, its final applications have always been limited to specialized areas [16]. However, while constructing algorithms for this task by hand was inefficient at that time and due to its high complexity, machine learning models and AI systems have developed significantly since then. For instance, powerful large language models (LLMs) trained on billions of texts can perform numerous tasks with humanlike accuracy [17]. However, when faced with structure elucidation, current models [18,19] struggle with even the simplest spectra or molecules. This is because, although LLMs excel at generating human language, their accuracy is often significantly lower when tackling mathematical or scientific problems [20,21]. Therefore, to effectively support the process of de-novo structure elucidation with AI systems, it is necessary to develop problem-specific standalone models [22].

AI for the interpretation of GC-EI mass spectra

To develop an AI capable of accurately evaluating mass spectra, established model architectures were altered, adapted, and optimized for the unique characteristics of EI mass spectra and molecules over a five-year research project. The final class of Metis (Molecules from ElecTron Ionization Spectra) models directly reads spectra and constructs plausible molecular structures as Smiles strings based on existing m/z values and intensities. The most powerful model can generate structures up to 650 Da and encompasses most elements relevant to the application (H, B, C, N, O, F, Si, P, S, Cl, Br, I). The models were trained on a variety of publicly available and specially generated training data, with multiple tools developed to standardize and clean data, thereby eliminating low-quality spectra. The final model was trained using low-resolution, i.e., nominal, spectra of approximately 300,000 different molecular structures.

The data and training were designed so that this AI has a certain tolerance for measurement distortions and background signals, allowing it to meaningfully predict molecules across a wide range of classes and application areas. Metis takes 1 to 5 seconds to generate 5 to 25 different structure proposals (hardware: Nvidia L40), enabling the automatic analysis even of complex chromatograms within a few minutes. To evaluate the model's performance and accuracy, the AI was tested on over 20,000 distinct spectra. To simulate the task of identifying unknown substances as accurately as possible, the underlying molecular structures were first removed from the utilized training data. For each of these unknown spectra, five predictions (Top5) and a confidence value were generated to assess the model's accuracy. For most spectra, the first of these predictions (Top1) is remarkably close to the correct structure, with a mean Tanimoto similarity (ECFP-1024 [23]) of 64 %. Notably, the interpretation AI provides a completely correct structure prediction as the first prediction in 37 % of all cases. In 55 % of cases, the top1 prediction is a structural isomer of the correct molecule. Thus, the model impressively demonstrates that detailed structural information can be extracted directly and automatically from the underlying mass spectra. Regarding the first 5 hypotheses (Top5) from the model, the correct assignment is present in 51 % of all cases, while the Tanimoto similarity of the best solution also rises to 73 %.

This demonstrates that specifically adapted AI models can interpret completely unknown mass spectra accurately across a wide range of applications. Metis complements database solutions, significantly enhancing the computer-aided evaluation process. Due to its architecture, it can correctly generate structures even for completely unknown substances, regardless of whether these substances have ever been identified or synthesized before. With this, the software supports questions in the area of NTS- and GC-MS-based structure elucidation while increasing the elucidation rate and reducing the time required.

Conclusion and outlook

Even though the use of this MS interpretation AI is already significantly accelerating existing workflows, mass spectrometry has inherent limitations. Despite the high performance of EI-MS, many molecules cannot be fully distinguished based on their mass spectra. This is particularly true for related structural isomers, such as regioisomers, whose spectra often differ only slightly in the intensities of individual m/z values. To achieve optimal semi-automatic solutions, the presented AI software will be enhanced with additional data sources in the upcoming years. A model for the incorporation of retention indices is currently being developed to differentiate between isomers, while other analytical and structural-chemical data will also be integrated in subsequential steps. Thereby, an interaction with the AI model will be possible to iteratively refine hypotheses in a feedback loop. The model is already capable of learning from specific applicational data, whereby the model is capable of recognizing characteristics of the respective chemistry based on hundred to a thousand data points. For the first time, the described AI software provides the opportunity to quickly and reliably elucidate unknown sample components, thereby offering deeper insights into the molecules' surroundings – for a world without unknowns.

Additional information: https://cheminnovation.de/assets/video/video.mp4

Contact

Philipp M. Pflüger
ChemInnovation GmbH
Münster, Germany
www.cheminnovation.de
[email protected]

References

[1] Hübschmann, H.-J. (2015). Handbook of GC‐MS. Wiley.

[2] Kelly, K., Bell S. (2018). Evaluation of the reproducibility and repeatability of GCMS retention indices and mass spectra of novel psychoactive substances. Forensic Chemistry. DOI: 10.1016/j.forc.2017.11.002.

[3] Sparkman, O. D. et al. (2011). Gas Chromatography and Mass Spectrometry. Elsevier.

[4] Maciel, E. V. S. et al. (2022). Electron ionization mass spectrometry: Quo vadis? Electrophoresis. DOI: 10.1002/elps.202100392.

[5] Heller, S. R. (1999). The History of the NIST/EPA/NIH Mass Spectral Database. Today's Chemist at Work. 8(2), 45–46, 49–50. http://www.hellers.com/steve/resume/Todays-chemist.pdf

[6] Wiley Registry of Mass Spectral Data 2023. Wiley Science Solutions. https://sciencesolutions.wiley.com/solutions/technique/gc-ms/wiley-registry-of-mass-spectral-data/.

[7] NIST/EPA/NIH Mass Spectral Library (2023). https://www.nist.gov/programs-projects/electron-ionization-library-component-nistepanih-mass-spectral-library-and-nist-gc.

[8] Hollender, J. et al. (2023). NORMAN guidance on suspect and non-target screening in environmental monitoring. Environmental Sciences Europe. DOI: 10.1186/s12302-023-00779-4.

[9] Hufsky, F. et al. (2012). De novo analysis of electron impact mass spectra using fragmentation trees. Analytica Chimica Acta. DOI: 10.1016/j.aca.2012.06.021.

[10] Kim, S. et al. (2025). PubChem 2025 update. Nucleic Acids Research. DOI: 10.1093/nar/gkae1059.

[11] Sadybekov, A. A. et al. (2022). Synthon-based ligand discovery in virtual libraries of over 11 billion compounds. Nature. DOI: 10.1038/s41586-021-04220-9.

[12] Warr, W. A. et al. (2022). Exploration of Ultralarge Compound Collections for Drug Discovery. Journal of Chemical Information and Modeling. DOI: 10.1021/acs.jcim.2c00224.

[13] Korn, M. et al. (2023). Navigating large chemical spaces in early-phase drug discovery. Current Opinion in Structural Biology. DOI: 10.1016/j.sbi.2023.102578.

[14] Wei, J. N. et al. (2019). Rapid Prediction of Electron-Ionization Mass Spectrometry Using Neural Networks. ACS Central Science. DOI: 10.1021/acscentsci.9b00085.

[15] Buchanan, B. G., Feigenbaum, E. A. (1981) Dentral and Meta-Dentral: Their Applications Dimension, Readings in Artificial Intelligence. DOI: 10.1016/B978-0-934613-03-3.50026-X.

[16] Lindsay, R. K. et al. (1993). DENDRAL: a case study of the first expert system for scientific hypothesis formation. Artificial Intelligence. DOI: 10.1016/0004-3702(93)90068-M.

[17] Lappin S. (2023). Assessing the Strengths and Weaknesses of Large Language Models. Journal of Logic, Language and Information. DOI: 10.1007/s10849-023-09409-x.

[18] Llama (2025). https://www.llama.com/docs/overview/.

[19] Introducing OpenAI o1 (2025). https://openai.com/o1/.

[20] Satpute, A. et al. (2024). Can LLMs Master Math? Investigating Large Language Models on Math Stack Exchange. arXiv. DOI: 10.48550/arXiv.2404.00344.

[21] Ramos, M. C., Collison, C. J., White A. D. (2024). A review of large language models and autonomous agents in chemistry. Chemical Science. DOI: 10.1039/D4SC03921A.

[22] Mirza, A. et al. (2024). Are large language models superhuman chemists? arXiv. DOI: 10.48550/arXiv.2404.01475.

[23] Rogers, D., Hahn, M. (2010). Extended-connectivity fingerprints. Journal of Chemical Information and Modeling. DOI: 10.1021/ci100050t.