CAi Copilot: Reducing Operational Workload in Molecular Design through Intent-Driven Agentic Workflows

The paper addresses a central challenge in early-stage molecular design: "Early-stage molecular design is an iterative process, not just a task of generating molecules. Researchers turn broad goals into design strategies, refine candidates, assess many properties, and gather evidence before synthesis and tests. While AI methods can generate molecules, optimize multiple goals, predict properties, dock compounds, and account for synthesis, these functions are spread across specialized tools. Experts must still coordinate each step, judge interim results, and integrate evidence. The authors argue that strong individual models, therefore, do not solve the workflow-level problem, because current molecular design methods lack a systematic framework that translates high-level research intent into coordinated execution and traceable, candidate-level evidence to support experts."

The authors formulate the setting as an intent-to-evidence molecular design workflow execution problem and present CAi Copilot, an expert-oriented agent described as "a three-layer agent for molecular design. CAi links intent grounding, result-driven control, code execution, and diverse tools. The framework unifies molecular generation, validation, assessment, filtering, ranking, and evidence reports."

The task is formalized as:

q = (I, C, Γ, T), where I is the design intent, C is the molecular context, Γ holds the design constraints and evidence needs, and T is the available molecular tool library.

The agent maintains a workflow plan Pt and history ht, selecting and running molecular operations at each step:

at ∼ π(· ht, Pt), ot+1 = Exec(at; T), Pt+1 = Update(Pt, ot+1).

The output joins the execution trajectory τ with candidate-level evidence:

E = (mi, zi, ri, ρi), Y = (E, τ, L),

where mi is the candidate molecule, zi stores computed evidence, ri gives the candidate rank, and ρi links each reported result to the supporting execution record. L stores missing evidence and failed operations. A successful run must satisfy intent satisfaction (S I), trajectory validity (S τ), and evidence grounding (S E): These aspects provide graded evidence of overall performance, whereas strict success requires complete task-level output matching.

CAi Copilot separates intent grounding, workflow control, and molecular execution across three linked layers:

Given intent I, molecular context C, and requirements Γ, the Research Interface Layer identifies the design goal, available inputs, required tests, ranking rules, and missing inputs, building an initial plan P0 = (V0, D0) where V0 contains executable steps and D0 records links between them. Each step has the form vⱼ = (gⱼ, uⱼ, xⱼ, pⱼ, eⱼ), where gⱼ is the local goal, uⱼ is the molecular operation, xⱼ holds required inputs, pⱼ gives input needs, and eⱼ defines expected evidence. The plan records links between steps, input needs, and expected evidence rather than only a tool list.

This layer controls the closed-loop trajectory, where the next operation depends on the current plan, prior steps, and candidate evidence gathered so far. It maintains an evidence gap Δt+1 = Req(Γ) Cov(Et+1): "This evidence-based control links generation to repeated optimization and screening. Invalid or duplicate candidates may trigger cleaning or regeneration, whereas constraint violations may trigger filtering or refinement. Partial results remain available... Execution thus dynamically responds to intermediate molecular outcomes rather than following a predetermined, fixed tool sequence."

The Execution Substrate grounds agent decisions in molecular computation. Each tool Tk follows an execution contract κk = (Xk, Yk, Prek, Execk), which explicitly exposes input readiness and downstream tool failures to the reasoning layer. The substrate provides tools for molecular generation and optimization, evaluation metrics, reusable utilities, and domain workflow skills.

Evidence is built across all three layers: "Each observation updates the candidate record (mi, zi, ri, ρi), where ρi links a reported result to the supporting action and observation... If the available context or tools cannot meet a need, the workflow reports that limit. The final output contains ranked candidates, computational evidence, execution provenance, and known limits for expert review."

The evaluation addresses four research questions: (RQ1) reliability of converting intent into valid execution and grounded outcomes; (RQ2) which workflow stages limit reliable execution; (RQ3) competitiveness under independent benchmarks; (RQ4) molecule-level analysis in a target-specific case.

CAiMD contains 45 manually curated tasks organized into three 15-task molecular contexts: mutant IDH2 scaffold decoration, JAK-family dual-target design involving JAK1 and JAK2, and PDK1 lead optimization, instantiated from DiffDec examples, JAK2 macrocycle studies, and PDK1 studies using REINVENT 4. Five baselines were compared: ChemCrow (chemistry-specific tool use), Biomni (biomedical research), AgentD (modular drug discovery), and Codex and Hermes (general-purpose tool interaction). All agents used the same DeepSeek V4 Pro-compatible endpoint, prompts, inputs, isolated workspaces, and 2,400s budget.

Key results (Table 1): Performance separates after intent recognition. All systems score 95.56% on intent understanding, while their overall means range from 47.62% to 75.49%. Hermes achieved the best task-decomposition score (79.44%) but the lowest overall mean. CAi achieved the strongest overall performance: CAi improves the valid-tool-output rate over Codex by 32.58 percentage points, from 34.91% to 67.49%. Objective satisfaction increases by 18.07 points, from 66.52% to 84.59%. The final Avg. scores were: ChemCrow 49.22%, Biomni 56.98%, AgentD 49.12%, Codex 69.67%, Hermes 47.62%, CAi 75.49%.

Table 2 provides a complementary capability-level aggregation: "the cohort-level result preserves the main-table ordering: CAi leads Codex by 10.86 points overall and by 18.07 points on task outcome. Agreement between the two aggregation schemes indicates that the lead achieved by CAi is not an artifact of assigning equal weight to all nine diagnostics."

Execution is the largest cross-system differentiator... The clearest separation between CAi and Codex instead occurs during execution: CAi improves input preparation by 9.11 points and valid tool output by 32.58 points. The authors note: "CAi remains limited by coverage and evidence consistency. CAi does not lead in trajectory accuracy, report–tool consistency, or hallucination-free analysis... CAi executes selected operations reliably but may still omit required steps or give less consistent explanations than the strongest general agent."

SMDD-Bench: On 2D Pharmacophore Identification, CAi improved success rate from 20.0% to 24.0%; on Lead Optimization, from 50.0% to 63.3%; on Fragment Assembly, from 0.0% to 1.7%.

LIDDiA: Using a budget-matched comparison on 30 targets with 50 candidate slots, "CAi returns 78.5% valid molecules and fills 39.2 of 50 slots on average. LIDDiA retains 14.4 valid molecules, or 28.7% of the budget. CAi also leads every constraint-level pass rate, including a 28.7% high-quality rate compared with 22.1% for LIDDiA. However, LIDDiA leads NVT, QED, LRF, and macro-averaged Vina, while CAi achieves better synthetic accessibility, micro-averaged Vina, and macro diversity."

MolBench: On all 190 examples, CAi raises MS-1 accuracy from 0.1800 to 0.9600, reaches an editing accuracy of 0.9744 compared with 0.8718 for MolClaw, and matches both MS-2 measures. MolClaw remained stronger on receptor-dependent virtual screening (0.5200 Hit@3 vs. 0.4000 for CAi). For optimization, CAi achieves a larger gain of 0.2263 and a higher success rate of 0.7949, although MolClaw preserves scaffolds more consistently.

Cross-benchmark analysis: "Across CAiMD, SMDD-Bench, LIDDiA, and MolBench, CAi is strongest when workflow decisions must yield valid molecular outputs... Recurring limits include omitted steps in integrated workflows, dependence on structure transfer in receptor-based screening, and reduced scaffold retention under aggressive optimization."

In a 15-task JAK-family dual-target case, CAi returned evaluable molecules for all three design entry points. "In the matched scaffold-based setting... CAi returns 24 selected candidates with a mean Vina score of −8.706 kcal/mol... 18 candidates meet the prespecified −8.0 kcal/mol threshold, whereas only seven of 80 raw baseline molecules meet this threshold. However, no uniform portfolio advantage in QED, toxicity, synthesis, similarity, or selectivity was observed: The workflow enriches candidates for the requested predicted-affinity objective while retaining evidence that makes molecular compromises visible to expert review. The authors caution that Results characterize predicted, selected portfolios and do not establish experimental activity or controlled generator superiority."

The paper situates CAi relative to constraint-aware molecular generation (pocket-conditioned methods, reaction-based generators) and scientific workflow agents (ChemCrow, Biomni, Co-Scientist, LIDDIA, MolClaw), noting a unified workflow for constructing diverse molecular design alternatives from open-ended expert intent remains underexplored.

The conclusion states: "We formulate this process as an intent-to-evidence workflow execution and present CAi Copilot. The Research Interface Layer, Agent Reasoning Layer, and Execution Substrate work together to plan, adapt, and run molecular workflows. Experiments show gains in valid molecular outputs and objective satisfaction... CAi thus turns open-ended design intent into traceable candidate-level evidence for expert review. Experimental validation and final chemical judgment remain outside the computational workflow."

The paper lists three main contributions: (1) formulating early-stage molecular design as an intent-to-evidence workflow execution problem, shifting the output from an isolated molecule set to a traceable workflow record with candidate-level evidence; (2) presenting CAi Copilot, a three-layer agent for molecular design that unifies molecular generation, validation, assessment, filtering, ranking, and evidence reports; and (3) evaluating CAi on 45 design tasks, further benchmarks, and 15 target-specific cases, showing strong intent-to-evidence performance while identifying long-horizon execution as a remaining limit. Code is available at https://github.com/datamllab/CAi copilot.

Improvements to AI systems:

• Replace fixed tool pipelines with a three-layer architecture that separates intent grounding, evidence-driven reasoning, and molecular execution. The intent layer translates high-level goals into explicit workflow steps with required inputs, expected evidence, and dependency links.

• Add an evidence-gap control loop: maintain a coverage set of collected evidence and compare it against required design constraints; trigger regeneration, filtering, cleaning, or refinement whenever invalid molecules, duplicates, or constraint violations appear.

• Formalize each tool with an execution contract exposing input readiness, preconditions, outputs, and failure modes, so the agent can prepare missing inputs, detect downstream tool failures, and adapt without crashing or hallucinating.

• Implement provenance tracking: every candidate molecule must be linked to the exact action and observation that produced its computed evidence, and any missing evidence or failed operation must be explicitly reported rather than omitted.

• Make the output a structured workflow record containing ranked candidates, their evidence, the execution trajectory, and known limitations, so experts can review both results and reasoning.

• Strengthen long-horizon reliability by adding plan-coverage checks, step-omission detection, trajectory-validation metrics, and consistency checks between reported analysis and tool outputs.

• Add scaffold-retention constraints and structure-transfer fallbacks for receptor-based screening, reducing known weaknesses in aggressive optimization and receptor-dependent tasks.

• Evaluate with multiple aggregation schemes to avoid metric artifacts, and use capability-level diagnostics to identify which workflow stages still bottleneck performance.

What the improved AI system can do:

• Accept an open-ended molecular design intent and automatically produce a concrete, executable plan with dependencies, required inputs, and expected evidence per step.

• Dynamically adjust its workflow based on intermediate molecular results, cleaning invalid or duplicate candidates, refining constraint-violating molecules, and regenerating when evidence gaps remain.

• Run diverse tools — generation, validation, property prediction, docking, synthesis assessment, filtering, ranking — through uniform contracts, recovering gracefully from missing inputs or tool failures.

• Deliver a transparent, expert-reviewable report: for each candidate, show the molecule, its rank, all computed evidence, and the precise provenance linking each result to the underlying execution step.

• Explicitly state missing evidence and failed operations, reducing hallucinated completions and unsupported claims in molecular design reports.

• Improve valid-molecular-output rates, objective satisfaction, and high-quality candidate yields compared to general-purpose and chemistry-specific agents, while making trade-offs among affinity, scaffold retention, synthesis, and toxicity visible to experts.

• Work across benchmarks and real target families, reliably converting broad research intent into traceable candidate-level evidence ready for expert judgment.