Source-linked AI summary
Mozi: Governed Autonomy for Drug Discovery LLM Agents
He Cao, Siyu Liu, Fan Zhang, Zijing Liu, Hao Li, Bin Feng, Shengyuan Bai, Leqing Chen, Kai Xie, Yu Li
TL;DR
LLM agents for drug discovery face unconstrained tool use and unreliable long-horizon execution in dependency-heavy workflows. Mozi addresses these barriers with governed supervision and stateful workflow graphs, and reports improved orchestration and competitive candidate-generation performance while retaining HITL oversight and traceability.
Problem
Drug-discovery LLM agents lack reliable governance and long-horizon execution across dependency-heavy, heterogeneous pharmaceutical workflows.
Method
Mozi uses a dual-layer architecture combining role-based Control Plane governance with stateful Workflow Plane skill graphs, validated data contracts, and HITL checkpoints.
Results
Mozi improves PharmaBench classification/MCQ accuracy over the baseline and outperforms general-purpose configurations on the HLE subset in the reported low-cost evaluation.
Takeaways & Limitations
Mozi presents governed, workflow-native orchestration as a way to combine flexible LLM reasoning with more reliable and auditable drug-discovery execution.
Takeaways & Limitations
Mozi remains dependent on changing external tools and data sources, and constrained design does not fully eliminate LLM stochasticity.
Abstract
from arXiv · showhide
Tool-augmented large language model (LLM) agents promise to unify scientific reasoning with computation, yet their deployment in high-stakes domains like drug discovery is bottlenecked by two critical barriers: unconstrained tool-use governance and poor long-horizon reliability. In dependency-heavy pharmaceutical pipelines, autonomous agents often drift into irreproducible trajectories, where early-stage hallucinations multiplicatively compound into downstream failures. To overcome this, we present Mozi, a dual-layer architecture that bridges the flexibility of generative AI with the deterministic rigor of computational biology. Layer A (Control Plane) establishes a governed supervisor--worker hierarchy that enforces role-based tool isolation, limits execution to constrained action spaces, and drives reflection-based replanning. Layer B (Workflow Plane) operationalizes canonical drug discovery stages -- from Target Identification to Lead Optimization -- as stateful, composable skill graphs. This layer integrates strict data contracts and strategic human-in-the-loop (HITL) checkpoints to safeguard scientific validity at high-uncertainty decision boundaries. Operating on the design principle of ``free-form reasoning for safe tasks, structured execution for long-horizon pipelines,'' Mozi provides built-in robustness mechanisms and trace-level audibility to completely mitigate error accumulation. We evaluate Mozi on PharmaBench, a curated benchmark for biomedical agents, demonstrating superior orchestration accuracy over existing baselines. Furthermore, through end-to-end therapeutic case studies, we demonstrate Mozi's ability to navigate massive chemical spaces, enforce stringent toxicity filters, and generate highly competitive in silico candidates, effectively transforming the LLM from a fragile conversationalist into a reliable, governed co-scientist.
1. Introduction
Drug discovery is a complex, costly pipeline requiring orchestration across heterogeneous expertise and data, while existing AI systems remain fragmented and generic agents unstable. Mozi addresses this with governed autonomy, workflow structure, auditable execution, and human intervention.
- Drug discovery spans target identification, hit identification, hit-to-lead, lead optimization, and preclinical or clinical evaluation.The pipeline requires integrating biology, chemistry, pharmacology, and heterogeneous data streams.
- Specialized AI models improve isolated tasks but remain disconnected from interoperable pharmaceutical workflows.LLM agents are proposed as orchestrators, yet generic agents can exhibit tool-use hallucinations and probabilistic instability.
- Mozi combines a governed Control Plane with a Workflow Plane that turns the canonical discovery pipeline into structured, auditable execution.The Control Plane governs actions and recovery, while the Workflow Plane operationalizes discovery stages as executable artifacts.
- Mozi integrates domain-specific tools and HITL checkpoints to support transparent candidate identification and expert intervention at high-uncertainty decisions.Reported tools include functional protein retrieval, structure-based docking, and ADMET prediction.
2. Related Work
Related work spans tool-augmented agents, hierarchical orchestration, governance, workflow graphs, and drug-discovery systems. Mozi positions itself as a workflow-native combination of these capabilities for long-horizon scientific tasks.
- Tool learning: Scientific tool-learning systems address domain-specific action spaces through verified tools, parameterized memory, and structured tool interaction.
- Mozi's positioning: Mozi's architecture combines hierarchical coordination, reflection-driven execution, integrated tools, and standardized access through the MCP Platform.
- Orchestration: Prior scientific-agent work uses organizational roles such as researchers, reviewers, planners, validators, and evaluators to coordinate specialized tasks.
- Tool governance and safety: Governance research emphasizes capability control, sandboxing, audit trails, access management, and safety frameworks for tool-enabled agents.
- Workflow graphs: Graph-based orchestration and state machines add explicit control flow and recoverability to LLM-driven scientific workflows.
- Drug-discovery pipelines: Drug-discovery agents increasingly execute end-to-end pipelines, refine molecules, tune generative models, or navigate chemical space.
3. Problem Formulation
Mozi formalizes drug discovery as governed sequential decision-making over heterogeneous tools and dependent scientific artifacts. Its objective is an auditable trajectory that preserves correctness, compliance, and reproducibility across long-horizon workflows.
- 3. Problem Formulation: Drug discovery is formulated as governed sequential decision-making over a heterogeneous tool environment with costly and side-effectful actions.
- 3.1 Formalizing the Scientific Agent Workflow: The execution model represents a discrete-horizon agent process as the tuple M = ⟨S, T, A, Π⟩.The formulation specifies state, tools, actions, and execution policy over time.
- 3.1 Formalizing the Scientific Agent Workflow: The hybrid state combines unstructured context with scientific artifacts and external environment states, whose synchronization is critical for valid execution.Artifacts include target lists, PDB identifiers, binding pockets, and generated molecular structures.
- 3.1 Formalizing the Scientific Agent Workflow: The tool environment includes read, compute, and write tools, while actions may generate thoughts, invoke parameterized tools, or terminate execution.
- 3.2 The Governance Challenge: Constrained Exploration: The Governance Problem is the mismatch between unbounded LLM generation and strict laboratory constraints, including parameter validity and role-based permissions.The policy must restrict actions to a feasible region without severely limiting reasoning flexibility.
- 3.3 Long-Horizon Reliability and State Dependency: Long-horizon discovery contains stage dependencies in which each stage's output supplies the next stage's input, so early errors can invalidate downstream computation.The pipeline runs from Target Identification through Hit Identification, Hit-to-Lead, and Lead Optimization.
- 3.3 Long-Horizon Reliability and State Dependency: Mozi therefore discretizes execution into verifiable stages and applies quality gates before transitioning between pipeline phases.
- 3.4 Objective: Auditable Trajectory Generation: The target is an auditable trajectory satisfying correctness, compliance, and reproducibility through validated evidence, governance constraints, and replayable intermediate artifacts.This objective motivates separating high-level governance in Layer A from standardized workflow execution in Layer B.
4. System Architecture
Mozi separates flexible reasoning from governed, stateful execution through a dual-layer architecture. Layer A controls planning and tool access, while Layer B encodes drug-discovery workflows with contracts, artifact tracking, and HITL gates.
- Architectural Overview: Mozi uses a dual-layer topology: Layer A handles unstructured reasoning context, while Layer B provides governed skill-graph execution for complex workflows.The separation addresses trajectory drift, accumulated errors, context growth, and local loops in long-horizon drug-discovery tasks.
- Architectural Overview: An intent router classifies requests as knowledge retrieval, single-stage tasks, or end-to-end workflows and dispatches them accordingly.End-to-end requests trigger supervisor orchestration across ordered skill graphs such as Target Identification and Hit Identification.
- Layer A: Governance and Orchestration: Layer A uses a supervisor-worker hierarchy with minimal-step planning, bounded execution, reflection after each step, and replanning when needed.The supervisor can stop when sufficient information is available or update the plan with new steps.
- Layer A: Governance and Orchestration: Role-based hard-coded tool filtering isolates workers from inappropriate resources, with Strict Mode for production and Permissive Mode for debugging.Strict Mode physically restricts each worker’s tool list, preventing role overlap and protecting resources such as high-cost docking clusters.
- Layer B: Artifact-Centric Skill Graphs: Layer B represents drug-discovery stages as composable stateful skill graphs that support internal state, parallel branches, persistence, and schema enforcement.Specialized graphs cover canonical stages including Target Identification, while hit-generation paths can combine generative modeling and virtual screening.
- Layer B: Artifact-Centric Skill Graphs: Format adapters validate node inputs and outputs, while hybrid state management tracks rolling context summaries, artifact files, and provenance.HITL nodes can pause execution for approval, rejection, parameter correction, or rollback at critical decision boundaries.
5. Evaluation
Mozi is evaluated on PharmaBench and long-horizon therapeutic case studies spanning tool-heavy prediction, scientific reasoning, and execution governance. The results report improved aggregate classification/MCQ accuracy, model-scaling effects, and trace-level robustness, while also identifying tool-selection and invocation weaknesses.
- PharmaBench: PharmaBench contains 88 tasks spanning Target Identification to Preclinical Research, combining TDC quantitative tasks, HLE reasoning tasks, and auxiliary biomedical checks.The protocol uses Exact Match where applicable and manual verification for complex HLE reasoning traces.
- PharmaBench Results: Mozi improves overall classification/MCQ accuracy over the baseline on the reported PharmaBench runs.The supplied results summarize the aggregate comparison without providing the underlying table values.
- HLE Results: On HLE drug-discovery tasks, Mozi with open-source models outperforms Gemini-2.5-Pro with web search, while more capable base models increase overall agent accuracy.The HLE subset contains 28 expert-written questions emphasizing scientific reasoning, experimental design, and tool-augmented behavior.
- Limitations: The architecture’s remaining limitations include difficulty selecting tools from massive registries, repeated failed calls that create local loops, and parameter hallucinations in open-source-model computation agents.Reported hallucinations include misaligned argument names, invented parameters, and misunderstandings of required data types.
- Evaluation Scope: The benchmark’s aggregate accuracy conflates scientific priors, tool-interface robustness, and orchestration correctness across tool-heavy and knowledge-heavy tasks.The authors therefore complement benchmark results with long-horizon case studies focused on stage-wise artifact generation and governance behavior.
- Long-Horizon Case Studies: Across Crohn’s disease, Parkinson’s disease, and sepsis demonstrations, Mozi adapts strategies, contains localized failures, corrects toxicity penalties, and uses human checkpoints at uncertain decisions.The cases cover diverse physiological mechanisms and pharmacological constraints; Parkinson’s screening processed 377,760 compounds in 35 minutes using a parallelized prefilter.
6. Limitations and Risk Analysis
Mozi’s governance mechanisms reduce but do not eliminate risks arising from external dependencies, residual LLM stochasticity, and human oversight costs. Its in silico benchmarking also remains distinct from physiological validation and motivates stronger uncertainty support before wet-lab work.
- External dependencies: Mozi depends on heterogeneous external tools and data sources whose updates, version drift, availability, and implementation changes can affect reproducibility.The risk is especially relevant to long-horizon workflows spanning multiple tool invocations.
- Residual stochasticity: Constrained actions, schemas, and execution budgets reduce brittle behavior but do not eliminate stochasticity in reasoning, tool selection, or result synthesis.Identical inputs may still produce divergent intermediate artifacts or decisions, particularly during early-stage exploration.
- HITL trade-offs: HITL checkpoints improve reliability at high-impact decisions but increase user burden and reduce full automation.The limitation concerns the trade-off between intervention and automation rather than the value of checkpoints themselves.
- Validation boundary: Benchmarking against clinical compounds reflects algorithmic outputs rather than physiological realities.The paper therefore distinguishes computational comparison from biological validation.
- Future safeguards: Mozi’s future iterations require explicit uncertainty quantification and cross-method consensus scoring before expensive wet-lab syntheses.These additions are intended to better inform HITL checkpoints.
7. Conclusion and Future Work
Mozi unifies a governed multi-agent control plane with structured skill graphs for long-horizon drug discovery. The framework aims to improve reliability, auditability, and controllability while preserving flexible short-horizon reasoning, with future work targeting learned routing, stronger governance, hybrid memory, and standardized evaluation.
- Conclusion: Mozi unifies a governed multi-agent control plane with structured skill graphs for long-horizon scientific tasks.This workflow-native design separates flexible agent reasoning from explicit workflow structure, state, and governance.
- Conclusion: The dual-layer architecture aims to improve reliability, auditability, and controllability without sacrificing broad scientific competence.Its design preserves free-form reasoning for short-horizon interactions while structuring multi-stage workflows.
- Future Work: Future work includes learned routing policies, reusable skills induced from execution traces, machine-checkable governance, stronger sandboxing, and hybrid memory.The proposed memory combines explicit workflow state with long-term case-based retrieval.
- Future Work: Standardized evaluation suites targeting governed, workflow-native scientific agents remain an open direction.Such suites are proposed to support evaluation of this agent design space.
A.1 Database and Search Tooling
Mozi federates biomedical databases and computational tools through MCP while enforcing role-based tool isolation. Its strict mode limits workers to permitted tool categories, with permissive access reserved for debugging.
- Tooling: Mozi integrates biomedical databases and computational tools through MCP, covering protein, chemical, genomic, clinical, docking, cheminformatics, prediction, and generative chemistry functions.The listed components include databases such as UniProt and PDB, plus tools such as AutoDock Vina, RDKit, Open Babel, ADMET predictors, and Reinvent-style workflows.
- Tool Governance: Tools are mapped to source servers and categorized into search, computation, and filesystem types for access control.In strict mode, each worker sees only its permitted category; permissive mode exposes all tools for debugging.
A.4 Tooling Description
Mozi separates read-only biomedical retrieval from computational and transformative tools, using structured outputs, provenance, validation, and bounded execution to support auditable long-horizon workflows.
- Tool Categories: Mozi divides MCP tooling into database-and-search and computation categories to govern retrieval separately from side-effectful computation.The separation supports fine-grained governance, provenance tracking, and control over computational side effects.
- Database and Search: Database and search tools provide read-only access for entity grounding and evidence collection without computational side effects.They support structured retrieval across protein, chemical, genomic, clinical, pathway, interaction, variant, structural, and disease–target resources.
- Database and Search: Retrieval tools return structured records with identifiers, database provenance, and version information for reliable cross-database reasoning.This design avoids dependence on unstructured scraping or heuristic parsing while maintaining auditability.
- Computation: Computation tools perform explicit transformations or model-based inference and are governed separately because they may incur cost and produce derived artifacts.Examples include docking, molecular representation conversion, pocket detection, ADMET prediction, and generative chemistry.
- Computation: Supervised computation uses bounded budgets, structured schemas, and result validation to mitigate failure propagation and uncontrolled tool use.Raw artifacts are retained externally in execution traces, while oversized outputs are summarized according to policy.
- Design Rationale: Separating retrieval from expensive or side-effectful computation enables role-based isolation, auditing, debugging, and reproducible stage-wise execution.The rationale directly connects tool categorization to long-horizon workflow control.
A.5 Server Descriptions
Table 4 catalogs MCP server descriptions used in Mozi, with each server assigned a coherent toolset and strategic role in long-horizon drug discovery workflows.
- Table 4 presents MCP server descriptions used in the system.
- The continued table sections extend the MCP server descriptions across additional entries.
B.1 Domain-Aware Tool Selection in Large Tool Suites
Mozi controls domain-aware tool selection through an internal search–execution pattern, structured output handling, and stateful DAGs spanning canonical drug discovery stages.
- Tool selection: The system uses a two-step internal pattern to manage tool selection when many search tools are available.
- Tool selection: The alternation constraint reduces repeated searching without execution and encourages explicit, auditable tool choices.
- Output handling: Long tool outputs are summarized before reuse, prioritizing filepaths, URLs, and key numeric values for downstream evidence reference.
- Workflow graphs: The four canonical drug discovery stages are implemented as directed acyclic graphs with explicit state management, error containment, and strict data contracts.
- Target Identification: Target Identification aggregates clinical and literature evidence, maps targets to standardized identifiers, and retrieves structural templates for downstream simulation.
- Hit Identification: Hit Identification combines de novo generative design with deep-learning virtual screening to expand chemical diversity across complementary discovery tracks.
- Hit-to-Lead: Hit-to-Lead expands fewer than 10 hits to approximately 200 molecules, then applies staged pharmacochemistry, pharmacokinetic, toxicity, and interaction filters.
- Lead Optimization: Lead Optimization uses multi-objective reinforcement learning to balance drug-likeness, synthetic accessibility, and metabolic liabilities before ranking candidates for synthesis.
C. Cases
Mozi’s therapeutic case studies illustrate governed, long-horizon workflows that combine HITL decisions, error containment, chemical-space exploration, and iterative candidate optimization. The traces also emphasize that favorable computational profiles remain unverified hypotheses requiring empirical validation.
- Case studies: The case studies provide step-by-step execution traces across Crohn’s disease, Parkinson’s disease, and sepsis therapeutic workflows.
- Governed orchestration: The traces highlight intent routing, HITL decision boundaries, and robust error containment as mechanisms used during multi-day computational pipelines.
- Scope boundary: The execution traces are fully autonomous in silico workflows, and their favorable computational proxy metrics require rigorous in vitro and in vivo validation.
- Crohn’s disease: Crohn’s disease target retrieval identified 25 potential protein targets, filtered 20 for evaluation, and produced a prioritized top-five list before structural selection.
- Crohn’s disease: The Crohn’s disease workflow generated 49 molecules, clustered them to 20 representative hits, locally contained docking failures, and produced a top score of -9.0 kcal/mol.
- Hit-to-Lead: The hit-to-lead workflow expanded confirmed hits through R-group exploration and scaffold hopping before applying multi-stage pharmacochemistry filtering.
- Lead Optimization: Lead optimization refined validated leads with reinforcement learning and selected final candidates using multi-parameter objectives and docking validation.