Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MADD: Multi-Agent Drug Discovery Orchestra

Gleb V. Solovev, Alina B. Zhidkovskaya, Anastasia Orlova, Nina Gubina, Anastasia Vepreva, Rodion Golovinskii, Ilya Tonkii, Ivan Dubrovsky, Ivan Gurev, Dmitry Gilemkhanov, Denis Chistiakov, Timur A. Aliev, Ivan Poddiakov, Galina Zubkova, Ekaterina V. Skorb, Vladimir Vinogradov, Alexander Boukhanovsky, Nikolay Nikitin, Andrei Dmitrenko, Anna Kalyuzhnaya, Andrey Savchenko

TL;DR

The paper addresses the bottleneck of hit identification in early drug discovery by introducing MADD, a four-agent orchestration framework that maps natural-language queries to end-to-end hit-identification pipelines. It combines LLM-based reasoning with specialized generative and predictive tools, yielding superior end-to-end performance across seven disease cases and outperforming existing LLM-based systems. A new benchmark comprising query-molecule pairs and docking scores for over 3 million compounds supports rigorous evaluation and benchmarking. The work demonstrates AI-first drug design capabilities, introduces open benchmarks, and discusses practical limitations and future directions for wet-lab validation and broader generalization.

Abstract

Hit identification is a central challenge in early drug discovery, traditionally requiring substantial experimental resources. Recent advances in artificial intelligence, particularly large language models (LLMs), have enabled virtual screening methods that reduce costs and improve efficiency. However, the growing complexity of these tools has limited their accessibility to wet-lab researchers. Multi-agent systems offer a promising solution by combining the interpretability of LLMs with the precision of specialized models and tools. In this work, we present MADD, a multi-agent system that builds and executes customized hit identification pipelines from natural language queries. MADD employs four coordinated agents to handle key subtasks in de novo compound generation and screening. We evaluate MADD across seven drug discovery cases and demonstrate its superior performance compared to existing LLM-based solutions. Using MADD, we pioneer the application of AI-first drug design to five biological targets and release the identified hit molecules. Finally, we introduce a new benchmark of query-molecule pairs and docking scores for over three million compounds to contribute to the agentic future of drug design.

MADD: Multi-Agent Drug Discovery Orchestra

TL;DR

The paper addresses the bottleneck of hit identification in early drug discovery by introducing MADD, a four-agent orchestration framework that maps natural-language queries to end-to-end hit-identification pipelines. It combines LLM-based reasoning with specialized generative and predictive tools, yielding superior end-to-end performance across seven disease cases and outperforming existing LLM-based systems. A new benchmark comprising query-molecule pairs and docking scores for over 3 million compounds supports rigorous evaluation and benchmarking. The work demonstrates AI-first drug design capabilities, introduces open benchmarks, and discusses practical limitations and future directions for wet-lab validation and broader generalization.

Abstract

Hit identification is a central challenge in early drug discovery, traditionally requiring substantial experimental resources. Recent advances in artificial intelligence, particularly large language models (LLMs), have enabled virtual screening methods that reduce costs and improve efficiency. However, the growing complexity of these tools has limited their accessibility to wet-lab researchers. Multi-agent systems offer a promising solution by combining the interpretability of LLMs with the precision of specialized models and tools. In this work, we present MADD, a multi-agent system that builds and executes customized hit identification pipelines from natural language queries. MADD employs four coordinated agents to handle key subtasks in de novo compound generation and screening. We evaluate MADD across seven drug discovery cases and demonstrate its superior performance compared to existing LLM-based solutions. Using MADD, we pioneer the application of AI-first drug design to five biological targets and release the identified hit molecules. Finally, we introduce a new benchmark of query-molecule pairs and docking scores for over three million compounds to contribute to the agentic future of drug design.

Paper Structure

This paper contains 80 sections, 4 equations, 18 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. MADD orchestra
  4. Agents
  5. Integrated tools
  6. Experimental studies
  7. Benchmark creation
  8. Comparison of underlying LLMs
  9. Ablation study
  10. Analysis of generative models
  11. Comparison with existing solutions
  12. End-to-end hit identification with MADD
  13. AutoML features of MADD
  14. Case studies
  15. Discussion
  16. ...and 65 more sections

Figures (18)

  • Figure 1: Overview of MADD architecture.
  • Figure 2: Overview of the proposed benchmark
  • Figure 3: Comparison of drug discovery approaches using average metrics across considered diseases. The filtration groups (GR) are based on molecule properties such as Docking score and IC50 (GR1), SA (GR2), BRENK (GR3), and the latter groups additionally include SurehEMBL, Glaxo, PAINS (GR4), and QED (GR5).
  • Figure 4: Comparison of F1 score for the MADD automatically created and trained ML pipeline and manually pre-trained models for IC50 predictions.
  • Figure 5: Results of case studies: a). Alzheimer's disease case study; "ChEMBL" - experimentally validated GSK-3$\beta$ inhibitors from ChEMBL, "MADD" - molecules generated by our approach. b). Thrombocytopenia case study; "SYK-FBRL" - molecules generated by SYK-FBRL approach, "MADD" - molecules generated by our approach.
  • ...and 13 more figures