AI Agents for Drug Discovery

Drug discovery is one of the most expensive and time-consuming endeavors in science. On average, bringing a new drug to market takes over a decade and costs upward of $2 billion. A central bottleneck is the iterative, trial-and-error nature of the early discovery phase: a chemist proposes a molecule, synthesizes it, tests it, interprets the results, and then redesigns. Each cycle can take weeks.

AI has already transformed individual steps of this workflow — property prediction, generative molecular design, docking — but these tools still operate in isolation, requiring human intervention at every handoff. The next frontier is AI agents: systems that can reason, plan, use tools, and execute multi-step workflows autonomously. In this post, we discuss what AI agents are, how they apply to drug discovery, and where our lab is pushing the boundaries.

What Is an AI Agent?

An AI agent is a system that perceives its environment, maintains a goal, selects actions, and executes them — iterating until the goal is achieved. Unlike a standard ML model that maps an input to an output in a single forward pass, an agent operates in a loop:

Observe — receive information from the environment (assay results, molecular properties, literature)
Reason — think about what to do next (often using a large language model as the "brain")
Act — call a tool or generate an output (run a docking simulation, propose a new SMILES, query a database)
Reflect — evaluate the result and update the plan

This loop mirrors what a skilled medicinal chemist does mentally — and recent work has shown that LLM-based agents can replicate and even surpass human-level decision-making in constrained drug design scenarios.

The Drug Discovery Pipeline as an Agent Workflow

Consider the standard hit-to-lead optimization task. A medicinal chemist starts with a hit compound from a screening campaign and needs to improve its potency, selectivity, and ADMET profile simultaneously. This requires:

Generating structurally diverse analogs
Predicting their binding affinity and off-target effects
Checking synthetic accessibility
Prioritizing candidates for synthesis
Interpreting experimental feedback to guide the next round

Each of these steps is a tool call. An AI agent equipped with a molecular generator, a property predictor (e.g., MolMap), a docking engine, and a retrosynthesis planner can execute this entire workflow — continuously, without rest, at scale.

Key Components of a Drug Discovery Agent

1. The Reasoning Core (LLM)

Modern agents use large language models as their central planner. Models like GPT-4, Claude, or domain-fine-tuned variants can understand scientific instructions, decompose complex tasks into sub-goals, and select the right tool for each step. The LLM is the "scientist" — it doesn't run experiments itself, but it decides what experiment to run next.

2. Molecular Tools

Agents need a toolkit to interact with molecular space. Typical tools include:

Molecular generators — variational autoencoders, diffusion models, or fragment-based methods to propose new structures
Property predictors — models like MolMap for ADMET, toxicity, and bioactivity prediction
Docking software — AutoDock Vina, Glide, or ML-based docking for binding pose estimation
Retrosynthesis planners — ASKCOS, AizynthFinder to check synthetic feasibility
Literature search — retrieval-augmented generation (RAG) over PubChem, ChEMBL, PubMed

3. Memory and State

A key challenge is that LLMs are stateless — each call is independent. Drug discovery agents need a memory layer that tracks which molecules have been explored, what experiments have been run, and what the current best candidates are. Vector databases and structured experiment logs serve this purpose.

4. Closed-Loop Feedback

The most powerful agents incorporate real experimental feedback. Our Leadmaster platform, for instance, creates a closed loop between computational design and wet-lab synthesis and testing. The agent proposes a compound, the lab synthesizes it, the assay result is fed back, and the agent updates its model of structure-activity relationships in real time.

# Simplified pseudocode for a drug discovery agent loop
agent = DrugDiscoveryAgent(
    llm=Claude3(),
    tools=[MolMapPredictor(), DockingEngine(), RetroSynthesisPlanner()],
    memory=VectorMemory(),
    goal="Optimize IC50 of hit compound X against target Y"
)

while not agent.goal_achieved():
    observation = agent.observe()          # Get current state
    plan = agent.reason(observation)       # LLM decides next step
    action = agent.select_tool(plan)       # Pick appropriate tool
    result = action.execute()             # Run the tool
    agent.memory.store(result)             # Remember what happened
    agent.reflect(result)                 # Update understanding

Medea: Our Omics AI Agent

Our lab recently published Medea (Sui, Shen et al., bioRxiv 2026), an omics AI agent for therapeutic discovery. Medea integrates multi-omics data interpretation, target identification, and compound prioritization into a unified agentic framework. It can accept natural language queries like "What are the most promising druggable targets in triple-negative breast cancer based on TCGA expression data?" and autonomously execute a multi-step analysis pipeline to answer them.

This represents a paradigm shift: instead of bioinformaticians running rigid pipelines, researchers interact with an intelligent system that reasons about biology and proposes actionable therapeutic hypotheses.

Challenges and Open Problems

Despite rapid progress, several hard problems remain:

Hallucination — LLMs can confidently propose chemically invalid or synthetically impossible structures. Robust tool-use grounding is essential.
Reward design — defining a good "reward" for drug-like molecules is non-trivial. Potency, selectivity, ADMET, and novelty are often in tension.
Wet-lab latency — even with AI agents, synthesis and testing take days to weeks, creating a slow feedback signal that limits how quickly agents can learn.
Interpretability — understanding why an agent proposed a particular molecule is critical for regulatory acceptance and scientific trust.

Where We're Going

At AIDDPM Lab, our vision is an autonomous drug discovery platform — one where a researcher can specify a disease target and a set of constraints, and the system runs the entire early-discovery process: hypothesis generation, virtual screening, lead optimization, and synthesis planning. The human remains in the loop for key decisions, but the agent handles the combinatorial search problem that is otherwise intractable at human timescales.

We are actively recruiting a postdoctoral researcher to lead this direction. If you are excited about the intersection of LLMs, agentic AI, and drug discovery, we'd love to hear from you.

AI Agents for Drug Discovery: Closing the Loop from Hypothesis to Molecule