MADD: Multi-Agent Drug Discovery Orchestra (2511.08217v1)

Abstract: Hit identification is a central challenge in early drug discovery, traditionally requiring substantial experimental resources. Recent advances in artificial intelligence, particularly 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.

• The paper introduces a modular multi-agent system, MADD, that automates hit identification in drug discovery using LLM orchestration.
• It integrates specialized agents for query decomposition, molecule generation, property prediction, and summarization, enhancing process clarity.
• Experimental results demonstrate high accuracy (≈79.8%) and notable novelty in drug candidate generation across disease-specific benchmarks.

Multi-Agent Drug Discovery Orchestra: Architecture, Benchmarking, and Empirical Validation

Introduction and Motivation

The MADD framework addresses the challenge of automating end-to-end hit identification in early-stage drug discovery using multi-agent systems coordinated by LLMs. Existing AI-driven methods, while promising, often lack comprehensive workflow integration and accessibility for wet-lab researchers. MADD explicitly combines decompositional query analysis, target-adaptive molecule generation, property prediction (e.g., docking score, IC50), and result summarization into a unified, modular agent pipeline. The hypothesis motivating this work is that distributing subtasks among specialized agents improves both accuracy and flexibility compared to single-agent or monolithic LLM architectures.

System Architecture and Agent Specialization

MADD consists of four core agents: Decomposer, Orchestrator, Summarizer, and Chat. Each is responsible for a distinct stage in the drug discovery workflow.

• Decomposer decomposes user queries into atomic subtasks, ensuring that complex or ambiguous instructions are handled via clarification and context enrichment.
• Orchestrator constructs executable action plans, invokes generative models (LSTM-GAN, transformer-based CVAE), property predictors, and third-party cheminformatics tools (RDKit, docking frameworks).
• Summarizer aggregates validated outputs into structured, chemically intelligible answers for the user.
• Chat Agent mediates user interaction, solicits missing information, and provides guidance on task specification and system capabilities.

The architecture promotes error isolation and modularity, mitigating failure propagation and enabling scalable pipeline customization.

Figure 1: Overview of MADD architecture, illustrating agent-oriented orchestration and modular tool integration.

Integrated Toolchain and Generative Modeling

MADD's toolset spans generative models (LSTM-GAN for fast, low-resource generation and transformer-CVAE for high-fidelity, conditional molecule design), property prediction (AutoML pipelines leveraging FEDOT, ensemble gradient boosters, random forests), and classical cheminformatics for filtering (e.g., drug-likeness via QED, synthetic accessibility, PAINS, Brenk filters). Pretraining is conducted per-disease on ∼500k ChEMBL compounds with explicit calculation of binding affinity and physicochemical descriptors.

Automated model training is available via the AutoML-DL agent for new disease targets, supporting iterative refinement based on loss minimization and property prediction error, thus avoiding manual hyperparameter design.

Figure 2: Schematic of the conditional transformer-VAE architecture for molecule generation.

Benchmarking: Novel Drug Discovery Dataset

Conventional chemistry datasets are insufficient for multi-agent benchmarking due to their limited complexity and lack of sequential task structure. MADD introduces a new benchmark comprising:

• 245 natural-language user queries of varying complexity (single to multi-step, expert-generated and LLM-augmented).
• Datasets of generated molecule structures per disease-task, with annotated docking scores and physicochemical properties for over 3 million compounds across seven therapeutic indications.

  Figure 3: Overview of the proposed benchmark, showing dataset organization by disease and query/task complexity.

Experimental Evaluation

Orchestration: LLM Selection and Prompt Engineering

Multiple LLMs were evaluated for orchestrator role. Llama-3.1-70b with optimized system prompts achieved 92.3% orchestrator accuracy, outperforming newer and more expensive models, and was selected for all agent implementations.

Ablation Studies: Architectures and Pipeline Accuracy

Systematic ablation of agent roles reveals that conflation of responsibilities, particularly removal of the dedicated Summarizer, leads to inconsistent response delivery and severe drops in overall pipeline accuracy. MADD attains 79.8% final accuracy on the hardest queries (Dataset L), with dedicated roles outperforming all reduced architectures.

(Figure 4, Figure 5)

Figure 4: Visualization of MADD-v2A and MADD-v3 agent architectures.

Figure 5: Visualization of MADD-v2B and MADD-v2C agent architectures.

Hit Generation and Properties Filtering

Drug-candidate generation was compared across generative methods; transformer-CVAE produced top results in three out of six conditions, maintaining high hit rates through stringent filtering (bioactivity, SA, QED, toxicity flags), and consistently outperformed GAN and baseline LLM-based solutions. Only ChemDFM approached comparable multi-property hit rates but lacked controllable novelty or structure-property mapping.

Figure 6: Cross-disease benchmarking of molecule generation methods, reporting filter group hit percentages for each approach.

Case Studies: Experimental Validation

Alzheimer's Disease: MADD-generated GSK-3β inhibitors display higher mean SA and QED compared to ChEMBL controls, as well as improved docking affinity and a Tanimoto diversity of 0.43, indicating substantial chemical novelty.

Thrombocytopenia: Without domain-specific reconfiguration, MADD is able to autonomously train generative and predictive models, achieving IC50 prediction ($R^2 = 0.75$) comparable to SOP RL approaches. Out of 10k generated molecules, 132 satisfy all five property filters, surpassing the efficiency of prior RL pipelines.

Figure 7: F1-score comparison for IC50 prediction between automated and manually trained models.

Figure 8: Properties comparison of molecules generated by MADD (Alzheimer’s, Thrombocytopenia) versus reference approaches.

Hit Novelty and Diversity

MADD maintains 78-73% novelty rates for Alzheimer’s and Thrombocytopenia, measured as generation outside the training set and comparative datasets, confirming its ability to extend chemical space and avoid trivial sampling.

Figure 9: Tanimoto similarity analysis for all generated molecules, indicating high structural diversity.

AutoML Versus Manual Training

AutoML-based generative and predictive pipelines consistently outperform manually curated baselines on IC50 prediction across most diseases, with modest resource requirements (CVAE model generation: ∼8GB GPU, 45min for 10k molecules).

Limitations and Implementation Considerations

• Dataset Dependency: MADD requires user-supplied datasets of molecules and target properties for new disease adaptation; absence of curated data may impair performance and reproducibility.
• Black-Box Workflow: The system currently operates as a closed pipeline; incorporating agent decision logs would facilitate interpretability and post-hoc audit.
• Tool Extension: Expansion with external tools requires code modification, posing barriers for non-expert users despite open-source availability.
• Domain-Specific Filtering: Effective hit identification is contingent on user-defined thresholds and criteria, which may limit accessibility for non-specialists.

  Figure 10: Accuracy and cost breakdown for agent pipelines using different LLMs and system prompts.

Implications, Contradictions, and Future Directions

MADD empirically demonstrates that multi-agent orchestration of specialist tools yields robust automation and high-quality hit generation in early drug discovery tasks. It achieves high accuracy and strong generalization, with minimal manual intervention needed for new targets. However, performance relies on the availability of disease-specific training data, and most validations remain in silico. As such, closing the loop with wet-lab experimental validation and integrating modules for automated data curation and hypothesis generation constitute pressing future work.

The paradigm shift towards agentic systems also presents unique risks, including potential misuse for harmful substance design and job displacement. Proactive filtering and ethical oversight are necessary for responsible deployment in pharmaceutical contexts.

Conclusion

MADD operationalizes a multi-agent strategy for automated hit identification, achieving superior pipeline accuracy and generation novelty compared to contemporary LLM-based and agentic solutions. The introduced benchmark and empirical results substantiate its capability for robust, autonomous drug-candidate design, providing a tractable framework for further extensions into closed-loop drug development with experimental feedback. Integration of open-ended discovery and wet-lab validation remains the critical frontier for the next generation of multi-agent drug discovery systems.