Drug discovery requires coordinated decisions across multiple stages — target validation, structural analysis, molecular design, developability assessment and synthetic feasibility. The authors present TRACEDD (Tool-grounded Reasoning and Agentic Coordination for Explainable Drug Design), a framework developed to address limitations of isolated computational methods by integrating LLM-driven orchestration with validated domain tools. TRACEDD prioritizes explainability and scientific rigor by ensuring each major decision is grounded in explicit tool invocations and intermediate evidence.
A central contribution of the work is a tool-first multi-agent architecture in which large language models (LLMs) orchestrate validated computational tools rather than supplant them. The design positions specialized agents to perform domain tasks and rely on established computational resources. This approach preserves domain-specific methods while leveraging LLMs for high-level coordination and decision sequencing. The architecture is intended to reflect the distributed expertise of discovery teams rather than producing opaque single-model outputs.
TRACEDD implements a multi-agent system that follows a Reason-Act-Observe loop. Agents mirror expert roles and perform tasks such as target validation, druggability assessment, molecular generation, lead optimization, ADMET evaluation, literature evidence integration and retrosynthesis. Within the loop, reasoning agents determine which tools to invoke and why; acting agents call the tools and generate outputs; observation captures results and evidence that feed subsequent reasoning steps. This structure aims to make decision paths transparent and traceable, enabling users to inspect tool calls and intermediate outputs.
The framework demonstrates an end-to-end workflow that adapts to the realities of real-world data. When experimental protein structures are available, TRACEDD retrieves them for downstream analysis. When structures are not available, the system can invoke predictive tools such as AlphaFold to generate models that enable pocket detection and design. The workflow integrates druggability assessment, de novo molecular generation, and downstream developability and synthetic feasibility checks, allowing the pipeline to respond to missing data or variable inputs while maintaining an evidentiary chain for decisions.
To illustrate capabilities, the authors applied TRACEDD to JAK2 as a representative target. The system retrieved experimental protein structures where possible and used AlphaFold for structure prediction when needed. It identified druggable pockets and performed de novo molecular generation. Known JAK2 inhibitors were used to define design hypotheses and guide the molecular generation strategy. The case study exemplifies how TRACEDD links specific tool outputs (for example, pocket identification or structure models) to subsequent generation and optimization steps, preserving a record of the decision-making process.
Molecular generation in the TRACEDD workflow is driven by reinforcement learning guided by multiple reward and prioritization signals. These signals include docking scores, predicted pIC50, physicochemical properties and ADMET predictions. Known inhibitors and design hypotheses inform generation objectives. The framework uses these measurable outputs to prioritize candidates and to provide an interpretable basis for why certain molecules are advanced in the pipeline. Retrosynthesis and synthetic feasibility assessments are included as downstream agents to ensure candidate practicality.
An explicit goal of TRACEDD is to enable transparent, adaptable and human-verifiable molecular design workflows. By coupling agentic orchestration with domain-specific computational tools, the framework links each major decision to explicit tool invocations and intermediate evidence. This auditability supports inspection and verification by human experts and is intended to increase trust and reproducibility relative to black-box approaches that do not record tool-level evidence.
The article is presented as a bioRxiv preprint and has not been peer reviewed. All authors are employed at Tata Consultancy Services Ltd.; this is disclosed as a competing interest. The manuscript describes the framework, architecture and a representative JAK2 application; detailed performance metrics, benchmarking against alternatives, or broader validation results beyond the described workflow were not reported in the abstracted source material provided here. The framework is framed as a foundation for more reliable AI-assisted drug discovery rather than a completed, fully validated commercial pipeline.