Building a Drug-Discovery Agent — What Actually Chains, and Where the Wall Is

TL;DR

Agentic drug discovery crossed from demo to architecture. A handful of systems now connect multiple computational stages of early discovery — target → structure → generate → score → optimize → retrosynthesis — inside a common agent framework. (Not yet a validated closed loop where each stage consumes and checks the last.) The honest lesson from building these is that the advance is orchestration, not intelligence: the LLM coordinates the work and holds the memory, while specialized models do the domain science. It plans, decomposes, and selects tools — but it does not independently generate reliable biological or chemical truth. So the real bottleneck is tool fragmentation, state, and provenance, not model IQ. And nearly every result here stops at the same wall: the wins are in silico and predicted, not wet-lab-confirmed. For current systems, the most credible deliverable isn’t a finished molecule — it’s a human-reviewable, mechanism-grounded hypothesis with a traceable evidence chain. Here’s the map, from someone building exactly this for cardiac disease.

The story so far

Drug discovery’s early pipeline is a relay of hard, separate problems: pick a target (a protein worth drugging), find its structure, generate candidate molecules, score how well they bind, optimize the winners against many objectives at once, and plan a retrosynthesis (how you’d actually make them). This sequence is a conceptual map, not a literal linear pipeline — real discovery loops across these stages as structure uncertainty, assay results, selectivity, toxicity, and synthesis constraints send it backward. For decades each leg had its own specialist software and experts, and the hand-offs between them were manual, slow, and lossy.

The bet of “agentic” drug discovery is that a language model can run that relay — read the goal, call the right specialist tool at each leg, carry state across the hand-offs, and keep a record of why. That’s newly plausible for two reasons: the specialist models got good (structure prediction, molecule generation, docking), and LLMs got good enough at tool use to drive them. What follows is a builder’s map of what actually chains today — and the wall every one of these systems hits.

Where we stand

Six systems make the same bet, and it works as an architecture — though each covers a different slice of the chain, and none demonstrates a fully validated stage-to-stage loop:

The right-hand column is the whole point: every metric is computed or benchmark, none wet-lab-confirmed. And the metrics aren’t directly comparable — an in-silico property-filter pass rate, a docking-score change, a Q&A-benchmark accuracy, and a tool-calling completion rate measure very different things; the table maps coverage, not a common yardstick. (Publication note: five are arXiv preprints; only LIDDIA is peer-reviewed — EMNLP 2025.)

The real pattern: orchestration, not intelligence. In all of these, the LLM is the glue, not the predictor. It decomposes the goal into tool calls; specialized models (Pocket2Mol, docking, DeepPurpose, retrosynthesis) do the science. What’s actually substantiated is integration — and you can tell, because the standout systems fix plumbing:

• The tool registry. TxAgent runs over ToolUniverse — 211 biomedical tools from trusted sources (every FDA-approved drug since 1939, Open Targets) — and hits 92.1% on the authors’ DrugPC benchmark, beating GPT-4o [4]. That score isn’t the registry alone (it also takes a fine-tuned model, ToolRAG retrieval, and large-scale synthetic reasoning traces). My builder’s inference: the registry is the durable moat — trusted, maintained scientific tools are harder to reproduce than the surrounding agent logic.
• State across turns. DrugPilot’s “parameterized memory pool” exists solely because state-loss across multi-turn tool calls is the real failure mode — it lifts multi-turn task completion to 64% (on a function-calling benchmark) where naive agents collapse [5].

Blunt takeaway for a builder: invest in the tool registry and the state/provenance layer, and treat the LLM as dispatcher. For today’s systems, integration, state, and provenance bite before model IQ does — that’s where the marginal engineering hour pays off (specialized fine-tuning still matters; it’s just not the first bottleneck).

Going deeper — how the pieces chain, node by node

• Target ID / disease biology → knowledge graphs, not LLM recall. AbbVie mined >53M documents into a 911-million-edge KG and used a genetic-evidence gate to surface repurposing candidates — e.g. Ruxolitinib for Carney Complex [7]. (Separately, human genetic support has been associated with roughly 2–3× greater clinical success across historical drug programs [8] — an observational base rate that varies by disease area, not odds you add to one candidate by applying the gate.) Healx’s RL path-mining pruned evidence paths by 85–95% to make expert review feasible, then checked its generated evidence chains against the transcriptional response to two compounds (Sulindac, Ibudilast) that had already improved behavior and cognition in Fragile X mice — the KG’s contribution was the plausible mechanism-hypothesis, not proposing or validating the drugs [9]. Target ID’s deliverable is a readable evidence chain, not a black-box score.
• Target→compound interaction → interpretable fusion. DrugAgent fuses ML predictions + KG paths + literature for drug-target interaction, beating the non-reasoning baseline by 45% in F1 (0.355 → 0.514 — a real lift, still a modest absolute) and emitting human-readable rationales (“3-hop connection via LYN,” a kinase gene) [10].
• Structure & affinity → dedicated models (AlphaFold/Boltz/DiffDock — covered in the structure-prediction chapter); the agent’s job here is prioritization, where active learning is the well-validated lever for cutting docking cost: MolPAL recovers 87.9% of the top-50k docking-ranked compounds after actually docking only 2.4% of a 100M library — it learns to reproduce the docking ranking, not measured activity [11].
• Lead optimization → the most agent-native step. It’s an iterative multi-objective search, exactly what a reasoner-evaluator-memory loop is for — the auditable optimization platform (six agents, demonstrated on the kinase AKT1 — the target, not the platform’s name) improved its average Vina docking score by 31% in the multi-agent setting vs a single agent. But it’s a computed docking score, not measured affinity, and it came at a cost: drug-likeness (QED) and synthetic accessibility got worse in the trade — a reminder that “better on the primary objective” isn’t “better molecule” [6].

The validation wall. Here’s the discipline the “autonomous drug discovery” headlines skip. Almost every number above is in silico and predicted — “31% better docking score,” “37.9% in-silico success,” “92.1% benchmark accuracy” are computed or benchmark metrics, none are wet-lab-confirmed potency from an autonomous campaign. The one genuinely wet-lab-validated result here is a KG evidence-generation check (Fragile X [9]), not an end-to-end agentic discovery. The honest mirror is the physical lab: Cooper’s autonomous mobile robots run make → test → decide across three chemistry areas — but “the choice of chemistry was not autonomous”; humans still frame the problem [12]. And the auditable-platform paper’s most useful finding isn’t its 31% — it’s that unguided LLM runs finish fastest but leave their reasoning unverifiable [6]. That is the crux: the reproducibility-and-transparency critique [13] is why an agent’s output has to be an auditable hypothesis, not an oracle. The wet lab remains the arbiter.

What’s next

Overhyped: “virtual pharma,” “autonomous drug discovery,” and “months to hours” — these describe demos and internal benchmarks, not shipped drugs; the field’s own review is the first to even survey deployed impact [14], and every paper’s numbers stop at the validation boundary.

Not hyped enough: the unglamorous wins that actually compound — a tool registry you can trust (ToolUniverse), evidence-chain pruning that makes agentic hypotheses reviewable (85–95%), active-learning screening that cuts docking cost by orders of magnitude, and RAG + agents that keep pace with fresh assay data without re-fine-tuning [15]. And one actionable, study-specific finding: in the auditable-optimization study, the multi-agent configuration did better on the primary objective, while the single-agent configuration produced a more balanced overall profile [6] — worth confirming across more systems before treating it as a design law.

What to watch: the unlock isn’t a smarter planner — it’s closing the loop to the assay with trust intact: agents whose outputs carry a mechanistic evidence chain and are auditable, KG target-ID validated prospectively (not just retrospectively), and the make-test-decide robot loop [12] fused with the in-silico design loop so a prediction becomes an experiment automatically. When an agentic small-molecule campaign produces a wet-lab-confirmed lead through a genuinely connected design→make→test→learn loop — not a benchmark number — this last wall comes down. (For other modalities it is already cracking: the Virtual Lab’s LLM-agent team, under high-level human steering, designed 92 SARS-CoV-2 nanobodies with ESM (a protein language model) + AlphaFold-Multimer + Rosetta (protein-design software) and experimentally confirmed functional binders — two with improved binding to the JN.1/KP.3 variants [16].) And the make and test legs of that loop are themselves starting to close: OpenAI’s wet-lab evaluation reports an agent iterating a molecular-cloning protocol to a ~79× efficiency gain via a novel mechanism, plus agentic high-throughput reaction optimization [17] — a self-reported, not-yet-peer-reviewed signal, and process optimization rather than lead discovery, but a real narrowing of the gap.

Sources

1. Gao et al., “PharmAgents: Building a Virtual Pharma with Large Language Model Agents” (2025) — https://arxiv.org/abs/2503.22164
2. Averly et al., “LIDDIA: Language-based Intelligent Drug Discovery Agent” (EMNLP 2025, main proceedings — the one peer-reviewed system here) — https://aclanthology.org/2025.emnlp-main.603/ · arXiv: https://arxiv.org/abs/2502.13959
3. Pan et al., “FROGENT: An End-to-End Full-process Drug Design Agent” (2025) — https://arxiv.org/abs/2508.10760
4. Gao et al., “TxAgent: An AI Agent for Therapeutic Reasoning Across a Universe of Tools” (Zitnik lab, 2025) — https://arxiv.org/abs/2503.10970
5. Li et al., “DrugPilot: LLM-based Parameterized Reasoning Agent for Drug Discovery” (2025) — https://arxiv.org/abs/2505.13940
6. Unlu et al., “An Auditable Agent Platform For Automated Molecular Optimisation” (2025) — https://arxiv.org/abs/2508.03444
7. Martin et al., “Leveraging a Billion-Edge Knowledge Graph for Drug Re-purposing and Target Prioritization” (AbbVie, bioRxiv, 2022) — https://doi.org/10.1101/2022.12.20.521235
8. Nelson et al., “The support of human genetic evidence for approved drug indications” (Nat. Genet., 2015) — https://doi.org/10.1038/ng.3314
9. Sudhahar et al., “An experimentally validated approach to automated biological evidence generation in drug discovery using knowledge graphs” (Healx, Nat. Commun., 2024) — https://doi.org/10.1038/s41467-024-50024-6
10. Inoue et al., “DrugAgent: Multi-Agent LLM-Based Reasoning for Drug-Target Interaction Prediction” (2024) — https://arxiv.org/abs/2408.13378
11. Graff, Shakhnovich & Coley, “Accelerating High-Throughput Virtual Screening Through Molecular Pool-Based Active Learning” (MolPAL, Chem. Sci., 2021) — https://doi.org/10.1039/d0sc06805e
12. Dai et al., “Autonomous mobile robots for exploratory synthetic chemistry” (Nature, 2024) — https://doi.org/10.1038/s41586-024-08173-7
13. Coveney & Highfield, “Artificial Intelligence Must Be Made More Scientific” (2024) — https://www.ncbi.nlm.nih.gov/pmc/articles/PMC11323241/
14. Seal et al., “AI Agents in Drug Discovery” (review, 2025) — https://arxiv.org/abs/2510.27130
15. Lee et al., “RAG-Enhanced Collaborative LLM Agents for Drug Discovery (CLADD)” (2025) — https://arxiv.org/abs/2502.17506
16. Swanson, Wu, Bulaong, Pak & Zou, “The Virtual Lab: AI Agents Design New SARS-CoV-2 Nanobodies with Experimental Validation” (Zou lab, bioRxiv, 2024; now in Nature, 2025) — https://doi.org/10.1101/2024.11.11.623004
17. OpenAI, “Measuring AI’s capability to accelerate biological research in the wet lab” (with Red Queen Bio & Molecule.one; self-reported, 2026) — https://openai.com/index/accelerating-biological-research-in-the-wet-lab/

Canonical: https://latentcell.ai/posts/agentic-drug-discovery