Medicines by Algorithm: A Russian AI System Is Reshaping Drug Design
Russian researchers have introduced an AI system capable of designing medicines from a plain-text request. The technology automates the entire drug discovery cycle – from interpreting a scientific task written in everyday language to delivering a validated list of candidate molecules.
A Digital Laboratory
Imagine describing an ideal drug to a researcher: “We need a molecule that blocks this protein, does not affect others, dissolves well, and does not harm the liver.” Within hours, you receive ready-made chemical formulas that have already been screened for toxicity and synthesizability. It sounds like science fiction. Yet this is already happening in real laboratories in Russia.
In 2025, at the global conference on computational linguistics EMNLP held in China, Russian scientists presented a distinctive design. Graduate researcher Gleb Solovyov from ITMO University introduced MADD (Multi-Agent Drug Discovery), a system capable of launching a highly complex molecular discovery process from a request written in natural language. The project emerged from a collaboration between ITMO researchers and specialists from Sber’s Applied Artificial Intelligence Center. Its defining feature is full-cycle automation – from semantic analysis of a scientific problem to the delivery of a vetted list of candidate compounds.
The researchers effectively built a digital laboratory controlled by artificial intelligence. At its core lies a multi-agent architecture: instead of relying on a single monolithic algorithm that can fail at any stage, the system operates as a team of four virtual “specialists.” One interprets the researcher’s request, another selects appropriate methods, a third generates molecular structures, and a fourth evaluates them against strict criteria. The system employs multiple large language models, including GPT-4o, Gemini, Llama, and GigaChat, dynamically choosing the most suitable model for each subtask.
Precision by Design
Traditional drug discovery resembles searching for a needle in a haystack, where the haystack consists of millions of possible molecular combinations. MADD follows a different logic. Rather than searching existing databases, it constructs the “needles” themselves, tailored precisely to the requested parameters.
The key measurable outcome is accuracy. In 79.8% of cases, the system correctly interprets the request and produces relevant molecules – nearly five times the performance of a leading foreign counterpart, ChemAgent, which demonstrated a 16.4% success rate. More importantly, MADD has already proposed concrete candidate molecules targeting Parkinson’s disease, Alzheimer’s disease, lung cancer, multiple sclerosis, and other complex conditions. Five of these candidates feature novel mechanisms of action aimed at proteins that had previously been considered unreachable.
For Russian science and the domestic pharmaceutical industry, the significance of such a tool is difficult to overstate. It dramatically lowers the entry barrier to drug design.
The benefits for research institutions are equally substantial. Academic teams and small biotech startups that lack the budget for expensive commercial software or high-throughput screening gain access to a powerful and, critically, open tool. MADD’s source code is publicly available, aligning with the core principles of modern open science.
Chemistry students, bioinformaticians, and pharmacologists can train in an almost game-like environment, learning how to formulate research tasks and immediately observe how AI translates them into concrete molecular structures. This approach accelerates the training of a new generation of “digital” chemists.
For large pharmaceutical companies, the integration of such systems enables a strategic redistribution of resources. Costly laboratory time and experimental capacity can be focused on the most promising candidates preselected by AI, rather than consumed by broad exploratory screening.
Knowledge as the Primary Product
Russian software and scientific methodologies in the field of AI for Science are gradually gaining visibility on the global stage. MADD stands out as a product that competes not through bold claims, but through openness, architectural design, and verifiable results presented at a respected international venue.
Open-source availability allows laboratories worldwide to adopt and refine the system, fostering a user community and increasing the international citation footprint of Russian developers.
The project strengthens the reputation of Russian schools in chemical informatics and advanced AI. This, in turn, attracts attention from global pharmaceutical corporations and research centers interested in outsourcing, joint development, and collaborative research with Russian teams.
Such initiatives also create an environment in which young scientists can work at the forefront of global science without leaving Russia. Their expertise and experience then become a foundation for new startups and international collaborations.
