AI Assistant for Chemists: Innopolis Develops a System to Identify Molecules
Innopolis University is developing a digital assistant for chemists that analyzes mass spectra, the distinctive “fingerprints” scientists use to determine what a substance contains. The system proposes the composition of a sample and suggests possible molecular structures. The team is now testing the technology with leading chemistry institutes of the Russian Academy of Sciences.
Identifying an unknown substance is a challenge researchers face in chemistry, pharmaceuticals and the petrochemical industry. To determine which atoms make up a molecule and how they are connected, specialists must analyze experimental data, compare possible structures and test hypotheses. The process can take hours, or even several days in more complex cases.
Researchers at Innopolis University’s Laboratory of Artificial Intelligence in Chemistry are developing an AI assistant to speed up this work. The system analyzes mass spectra, proposes possible structures for chemical compounds and helps researchers identify the most likely candidates. The team presented preliminary results at an exhibition of AI solutions during the Innopolis Digital Days forum, held on October 8–9, 2026.
How AI Identifies Molecules
Mass spectrometry provides information about the composition of a substance. In tandem mass spectrometry, molecules are first separated according to their mass-to-charge ratio and then subjected to additional activation. The resulting fragments produce a characteristic pattern. This mass spectrum can reveal clues about the structure of the original compound.
But a single molecular formula can correspond to several different structures, so interpreting a spectrum requires chemical knowledge as well as computational analysis. Researchers must determine the substance’s composition, identify structural features and establish which candidate structure best explains the experimental data.
The Innopolis system carries out this analysis in several stages. First, it predicts the most likely molecular formula. At the current stage of development, it identifies the correct formula in about 80% of cases. A neural network then reconstructs the compound’s structural features, creating a kind of molecular “fingerprint” that captures key aspects of its architecture. When this stage succeeds, the match between the predicted features and reference data exceeds 90%. In some tests using spectra supplied by research partners, the matching accuracy reached 93% or higher. These figures reflect specific tests, rather than performance across all possible samples.
An Integrated System
Tools already exist for predicting molecular formulas, analyzing structural features and searching for candidate compounds. What sets the Innopolis system apart is that it brings all these stages together on a single platform.
As Ivan Golov, a developer at the laboratory, explained, the team has built a digital assistant to support the entire analysis process. A researcher uploads an experimental spectrum, and the system works through the steps in sequence, generating results, proposing chemical hypotheses and presenting the findings in a format researchers can readily use.
The aim is not to replace human specialists, but to narrow the range of candidates chemists must investigate using conventional research methods.
Benefits for Science and Industry
The technology could be useful in laboratories that investigate complex mixtures, analyze the products of chemical reactions and search for compounds with specific properties. In pharmaceutical research, tools of this kind can help with the preliminary identification of substances and the selection of promising candidates for further study. In the chemical and petrochemical industries, they could speed up sample analysis and the processing of large volumes of experimental data.
One promising application is high-throughput screening, in which researchers need to process large numbers of samples quickly and identify those that warrant closer investigation. Automating the initial analysis could substantially reduce time spent on routine tasks, freeing specialists to focus on other work.
Scale and Future Prospects
The team estimates that the potential user base includes more than 7,600 companies in the chemical and petrochemical industries, as well as around 130,000 scientists and researchers. The developers estimate that the system could automate 240 million experiments a year and generate economic benefits of up to 1.5 billion rubles (about $18 million) by reducing the time required to process data.
The team plans to complete the project’s first stage in 2026. In 2027, the researchers intend to integrate the system with the Science and Innovation domain of the government information system used to track research and development projects.
