TRACEDD: Tool-grounded Reasoning and Agentic Coordination for Explainable Drug Design

It establishes a tool-first multi-agent architecture in which LLMs plan tasks, select tools, integrate heterogeneous outputs, and articulate decision rationales, while quantitative prediction remains with domain-validated computational tools such as structure prediction, docking, ADMET modeling, and retrosynthesis planning. Relative to treating LLMs as end-to-end black-box predictors or using computational modules in isolation, the work explicitly separates orchestration from prediction and keeps records of tool calls, intermediate outputs, and decision rationales, forming an auditable link from scientific objective to design recommendation. Primarily architectural description and system implementation; the text reports 26 specialized tools spanning structural biology, cheminformatics, ADMET prediction, and retrosynthesis, with a table of representative tools and their inputs and outputs, presented as a functional demonstration rather than large-scale benchmarking.

It implements a multi-agent system mirroring expert discovery teams, with role-specialized agents operating over a shared execution state through a Reason-Act-Observe loop, coordinated by a supervisory orchestrator that decomposes tasks and manages dependency resolution, iteration, and convergence. Unlike a rigidly predefined pipeline, agent sequencing evolves based on intermediate results and emerging constraints, enabling context-aware adaptation of the workflow. The JAK2 case illustrates the loop in operation, including target mapping and structure selection criteria recorded in the reasoning log; evidence comes from the case demonstration and system description.

In the JAK2 case it demonstrates an end-to-end workflow: retrieving 145 crystallographic structures and selecting the high-resolution complex 3UGC (1.34 Å), automatically invoking AlphaFold when experimental structures are absent, identifying the ATP-binding region as the most promising pocket, and using REINVENT4 with transfer learning and reinforcement learning for multi-property molecular generation, with docking scores or predicted pIC50 and physicochemical and ADMET properties serving as reward and prioritization signals. It connects target validation, hypothesis generation, structure preparation, molecular design, optimization, candidate prioritization, and synthesis planning into a continuous evidence-driven workflow, rather than evaluating the platform solely by the number of molecules generated. The case reports specific structure identifiers, resolution, structure counts, and named tools, accompanied by workflow figures; this is a representative use case for functional validation.

It treats hypothesis generation as an explicit intermediate layer between data acquisition and molecular design, deriving from known potent JAK2 inhibitors constraints such as moderate lipophilicity, scaffolds occupying the ATP-binding cleft and extending into adjacent hydrophobic regions, heterocyclic cores mimicking adenine and forming hinge-region hydrogen bonds with residues such as Leu932 and Glu930, and engagement of hydrophobic pockets near Met929 for selectivity, then converting these into property ranges and design constraints for downstream generation. Rather than treating molecular generation as an unconstrained optimization problem, the work first constructs a mechanistically informed search space from structural, chemical, biological, and developability evidence to guide optimization and candidate selection. Based on analysis of known potent inhibitors from curated sources such as ChEMBL and comparison of ligand scaffolds with conserved kinase binding modes, with specific residues and scaffold classes named (for example pyrazolo[1,5-a]pyrimidines and pyrrolo[2,3-d]pyrimidines); this is evidence synthesis within the case.

The framework targets early-stage, structure-guided drug design settings where a target identifier is available, structural resources are accessible or structure prediction is feasible, and decision provenance with human oversight is desired; the text uses JAK2, a kinase with extensive structural and bioactivity information, as the representative case and notes that when experimental structures are absent the workflow transitions to AlphaFold-based models within the same pipeline. Its design intent is to keep reasoning, computation, and decision-making distinct yet tightly connected, so that human experts remain essential for interpreting biological relevance, assessing structural uncertainty, and selecting candidates for experimental validation.

The text states that the current study focuses on representative use cases and therefore emphasizes functional capability rather than large-scale benchmarking, and that future work aims to establish quantitative evaluation frameworks covering task completion, reasoning consistency, tool-selection accuracy, recovery from execution failures, and agreement with experimental observations. A careful reader might continue to watch how stable the workflow is across more targets and more complex data conditions, how generated and prioritized candidates fare under experimental validation, and how local deployment models compare with cloud models on specialized cheminformatics reasoning. In addition, this is a preprint, and within the present reading scope the specific numerical details in figures and supplementary material are summarized as described in the text.