Any study consists of multiple stages: identifying the problem and setting objectives, formulating and testing hypotheses, designing experiments, analyzing results, and publishing them. Just the literature analysis, extraction and preparation of data, and training generative models to look for new molecules or compounds can take around a month.
In 2025, a team from ITMO’s Institute for Artificial Intelligence developed ChemCoScientist, a digital assistant to chemists. It automates research in chemistry and takes on routine and time-consuming tasks, such as extracting knowledge from chemical articles, generating new chemical compounds or improving existing ones, and predicting their properties. ChemCoScientist can facilitate and accelerate the production of new treatments and materials at research laboratories, pharmaceutical companies, and in chemical production. However, according to its developers, the model, like its counterparts, is only good at simple, separate tasks within a specific domain, for instance, “generate a molecule” or “predict its properties.”
CoScientist is a new solution by the team that can fully automate research in computational chemistry and medicine, including the search for new molecules. The assistant can cover the full cycle – from searching for literature to automatic dataset collection, training models for particular tasks, and multi-criteria filtering of candidate molecules.
At the core of the new tool is a combination of a multiagent system and workflow. The developers used both of these methods to minimize their drawbacks. A multiagent workflow fits a narrow class of tasks and is expensive in development, but it’s more stable, cheaper to use, and predictable, as it follows strict instructions. A multiagent system, on the other hand, has agents that aren’t limited and can decide how to interact to reach a goal; however, they are expensive to use.
“In CoScientist, the multiagent system acts like a manager that automatically generates a multiagent workflow for a specific task in chemistry or medicine. It interacts with a library of typical in-demand roles like a hypothesis generator, a researcher, a data analyst, and a library of applied tools. This allows the system to adapt to various tasks and fields to complete the main steps in a researcher’s work: from data collection and literature analysis to training generative models, evaluating the results, and preparing technical specifications for experimental design work. Next, together with CoScientist, researchers manage the work of agents, assessing their results, forming new hypotheses, and adjusting objectives. With the combination of a multiagent system and a multiagent workflow, we doubled the solution’s efficiency while also bringing down its cost by two times compared to other products that generate multiagent systems,” shares Anna Kalyuzhnaya, the head of the project, a senior researcher at ITMO’s Research Center “Strong AI in Industry.”
Anna Kalyuzhnaya. Photo by ITMO University
To evaluate the efficiency of CoScientist, the team created a benchmark with over 300 complex questions on articles in analytical chemistry, electrical chemistry, biology, and five more adjacent fields. The AI system had to find the answers in published papers and put them out for users. CoScientist completed the task 41% better than top models like Gemini 2.5 Pro and ChatGPT-5.
CoScientist is now being tested in various tasks in the fields of chemistry, nanomaterials, and clinical medicine. For example, it’s used to create new inhibitors to treat multiple sclerosis, considering the activity of molecules against the disease’s target proteins described in literature. In the future, the researchers are planning to expand the functionality of the new AI system to other natural and engineering sciences.
