Scientists at ITMO University and Sberbank Develop Multi-Agent AI System for Rapid Development of Medical Products

To develop new drugs, scientists search for “hits” – molecules with specific biological activity and suitable physicochemical properties. Typically, chemists “manually” select molecules from databases, synthesize them in the laboratory, and test for biological activity. However, this method has significant disadvantages: scientists have to spend a lot of time and resources, as well as choose from a limited number of known molecules.

It is possible to significantly reduce the cost and duration of expensive experiments by using AI tools to complete some of the work. However, many of these tools have notable disadvantages. For example, domain-specific LLMs (e.g., LlaSMol, X-LoRA-Gemma, ChemDFM) can only perform a single task (e.g., will generate molecules but cannot check their properties) and therefore often produce useless molecules. Conversely, single-agent systems, in which one general AI model is responsible for all tasks, may operate inconsistently and make mistakes (ChemAgent, ChemCrow, CACTUS). 

Scientists from the Research Center “Strong AI in Industry,” the Center for AI in Chemistry, and the Infochemistry Scientific Center at ITMO University have developed a multi-agent system that fully automates the process of discovering new drug molecules based on queries in natural language. MADD (Multi-Agent Drug Discovery) consists of four AI agents that perform different tasks sequentially, such as analyzing a text query from a researcher, selecting the appropriate algorithms, generating molecules and calculating their properties, and compiling the results into a unified report. Each of the agents operates using large language models like GPT-4o, Gemini 2.5 Flash-Lite, Llama-3.1-70b, GigaChat, and others.

The system checks each molecule according to five criteria: biological activity, binding affinity, synthetic accessibility, drug similarity and lack of toxicity. The model selects appropriate molecules with high accuracy – 79.8%, which is significantly higher than the results of its foreign analogue ChemAgent, whose accuracy reaches only 16.4%.

Chemists tested the model and used it to identify new promising drug molecules to be used in developing treatments for seven diseases: Alzheimer’s and Parkinson’s, multiple sclerosis, lung cancer, thrombocytopenia, dyslipidemia (a lipid metabolism disorder), and drug resistance in cancer therapy. In particular, this was the first time that chemists succeeded in generating high-potential molecules that simultaneously target five proteins associated with disease development, which had not been possible before. Some of the discovered molecules surpass existing analogues in terms of biological activity, binding affinity, and synthetic accessibility. To further investigate these molecules, the scientists will synthesize these molecules in the lab conditions and test them in real-life experiments.

“The Infochemistry Scientific Center worked on cases related to dyslipidemia, Parkinson’s disease, and drug resistance in cancer therapy. For each of these cases, we developed models to predict the biological activity of molecules and created a separate module for calculating binding affinity, a parameter that shows how a protein will blind to small molecules,”  says Rodion Golovinsky, an engineer at the Research Center “Strong AI in Industry.”

“We are among the first in the world to demonstrate the effectiveness of multi-agent systems for early-stage drug development. The team from the Center for AI in Chemistry prepared half of the cases, which focused on developing molecules against Alzheimer’s disease, multiple sclerosis, lung cancer, and thrombocytopenia. To achieve this, we analyzed scientific literature on the relevant diseases and targets, prepared datasets, trained predictive models, evaluated the quality of predictions and created tools for the language agents,” says Andrey Dmitrienko, the head of ITMO’s Center for AI in Chemistry.

The model is completely open-source: the code and data are published on GitHub, and can be tested on Hugging Face. The tool will be useful for academic researchers and small biotech startups that lack the resources for expensive platforms, chemists in pharmaceutical companies, students and teachers.

”Our model not only generates new molecules, but also fully automates the entire search cycle – from analyzing a user’s text query to producing a list of prospective candidates with calculated properties. In our tests, the model surpassed all existing counterparts in terms of accuracy and efficiency. In addition, the system is capable of generalizing information to solve unfamiliar tasks without relying on prior examples: MADD successfully identified molecules for drugs against thrombocytopenia, even though this case was not considered while the system were in development,” emphasizes Anna Kalyuzhnaya, a senior researcher at the Research Center “Strong AI in Industry.”

The paper’s first author, PhD student Gleb Solovyov, presented the team’s work at the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP) – the most prestigious event in its field. The conference took place in China November 4-9. Participants presented papers on dialogue systems and named entity recognition (NER), and discussed topics such as AI privacy, ethics in computational linguistics, and the impact of large language models on society.

Translated by Anna Butko