Human-supervised Agentic AI for Hypothesis Generation and Experimental Assistance in Drug Repurposing

The paper introduces RepurAgent, a human-supervised hierarchical multi-agent AI system that successfully automates and supports the entire drug repurposing lifecycle—from hypothesis generation and experimental design to data analysis and candidate refinement—across diverse disease scenarios including Acute Myeloid Leukemia, COVID-19, and Multiple Sulfatase Deficiency.

The Gist

Imagine you are trying to find a new use for an old tool. Maybe you have a hammer that was designed to drive nails, but you realize it could also be used to crack open a coconut. In the world of medicine, this is called drug repurposing: taking an existing, safe medicine and finding a new disease it can treat.

For a long time, computers have been great at the first step of this process: quickly brainstorming ideas like, "Hey, maybe this heart medicine could help with cancer!" But computers usually stopped there. They couldn't actually help you test the idea, run the lab experiments, or figure out if the results were real or just a fluke.

This paper introduces RepurAgent, a new kind of "super-assistant" that doesn't just brainstorm; it does the whole job from start to finish, with a human boss watching over its shoulder.

The Team: A Digital Kitchen Crew

Think of the drug discovery process like running a busy restaurant kitchen. You can't just have one chef trying to do everything at once. Instead, the authors built a hierarchical team of AI "chefs," led by a Head Chef (the Supervisor):

1. The Head Chef (Supervisor Agent): This is the boss. It doesn't chop vegetables or stir pots. Instead, it looks at the menu, assigns tasks to the other chefs, and makes sure everyone is working together. It keeps a human scientist in the loop, asking for approval before making big moves.
2. The Sous Chefs (Specialized Agents):
   • The Researcher: Scours the internet and scientific libraries for clues (like reading old recipe books).
   • The Predictor: Uses math to guess which ingredients (drugs) will taste good together.
   • The Data Analyst: Looks at the results of the cooking (lab experiments) to see if the dish actually works.
   • The Reporter: Writes down the final recipe and the results so humans can understand them.

These chefs have a memory (they remember what they learned yesterday) and a library (they can instantly look up facts), so they don't make the same mistakes twice.

The Proof: Three Real-World Tests

The team tested this digital kitchen crew in three different "restaurants" (medical scenarios) to see if it could handle the pressure:

1. The Speed Test (Acute Myeloid Leukemia):
   They asked the AI to find the right "flavors" (biological pathways) for a type of blood cancer. In just one hour (less time than it takes to watch a movie), RepurAgent found 97% of the correct answers that a much more complex, expensive AI system had found. It was fast and accurate.

2. The Detective Test (COVID-19):
   They looked back at old data from a search for COVID-19 cures. The AI acted like a smart detective. It didn't just follow a rigid rulebook; it adapted. It found the best candidates and, crucially, spotted "traps" (confounders) that human reviewers had missed. It was like a sous-chef who noticed, "Wait, the salt in this dish is actually from a different batch than we thought," and corrected the recipe before serving.

3. The Needle in a Haystack Test (Multiple Sulfatase Deficiency):
   They threw 5,000 different drugs at the AI and asked it to find the few that might work. The AI sifted through the haystack and pulled out 82 high-quality needles. When real human experts looked at these 82, they said, "Yes, these are the ones we should test next."

The Bottom Line

Before this, computers were like idea machines that could only suggest a direction. With RepurAgent, computers are now co-pilots that can fly the plane, navigate the storm, and land the plane, all while a human captain keeps their hand on the controls.

This system is now open for anyone to use, meaning scientists around the world can start using this "digital kitchen crew" to find new cures faster, cheaper, and smarter than ever before.

Technical Summary

1. Problem Statement

Current computational drug repurposing efforts are predominantly limited to the early stage of rapid hypothesis generation. However, the practical application of drug repurposing requires a comprehensive lifecycle that extends far beyond initial suggestions. This broader scope includes:

• Designing specific experiments.
• Analyzing complex assay data.
• Iteratively refining drug candidates based on experimental feedback.

Existing systems often fail to integrate these disparate stages into a cohesive workflow, lacking the autonomy to adapt to experimental realities or the ability to function as a collaborative partner throughout the entire discovery pipeline.

2. Methodology: RepurAgent

To address these gaps, the authors developed RepurAgent, a hierarchical multi-agent AI system designed to operate across the full drug repurposing lifecycle. The system is built upon the REMEDi4ALL consortium's specific data, tools, and standard operating procedures (SOPs).

Core Architecture:

• Hierarchical Multi-Agent Structure: The system consists of a Supervisor Agent and a Planning Agent that orchestrate four specialized sub-agents:
  1. Research Agent: Handles literature review and knowledge retrieval.
  2. Prediction Agent: Performs computational modeling and candidate scoring.
  3. Data Agent: Manages experimental data processing and analysis.
  4. Report Agent: Synthesizes findings and generates documentation.
• Human-in-the-Loop (HITL) Design: The system operates under human supervision, ensuring that critical decisions and experimental directions are validated by domain experts.
• Advanced AI Capabilities:
  • Episodic Memory: Allows the system to retain context from previous interactions and experimental runs.
  • Retrieval-Augmented Generation (RAG): Grounds the AI's outputs in specific, up-to-date scientific literature and internal databases to reduce hallucinations.
  • Tool Integration: The agents are equipped with specific computational tools and SOPs relevant to drug repurposing.

3. Key Contributions

• Full-Lifecycle Integration: RepurAgent is the first system demonstrated to handle the entire drug repurposing workflow, from initial hypothesis generation to experimental design, data analysis, and candidate refinement.
• Adaptive Experimental Collaboration: Unlike static pipelines, the system acts as an adaptive collaborator, capable of prioritizing compounds and flagging experimental confounders without relying on rigid, predefined thresholds.
• Open-Source Deployment: The system is made available as an open-source tool, deployed at https://repuragent.serve.scilifelab.se/, facilitating reproducibility and community adoption.

4. Results and Validation

The system was validated across three distinct scenarios representing different stages of the repurposing lifecycle:

• Scenario 1: Acute Myeloid Leukemia (Hypothesis Generation)

  • Performance: RepurAgent recovered up to 97% of disease-relevant pathways identified by Google Co-Scientist.
  • Efficiency: The entire workflow was completed within 60 minutes, demonstrating high computational efficiency.

• Scenario 2: Retrospective COVID-19 Antiviral Screen (Experimental Analysis)

  • Role: Acted as an adaptive experimental collaborator.
  • Performance: Prioritized compounds achieving an AUC-ROC of up to 0.98.
  • Critical Insight: Successfully flagged confounders that were missed during manual review, highlighting its ability to detect subtle data anomalies.

• Scenario 3: Multiple Sulfatase Deficiency (Candidate Prioritization)

  • Scale: Processed a library of 5,000 compounds.
  • Outcome: Identified and prioritized 82 high-confidence candidates.
  • Validation: These candidates were further corroborated by domain experts, confirming the system's reliability in narrowing down large chemical spaces.

5. Significance

This work represents a paradigm shift in computational drug discovery. By moving beyond simple hypothesis generation to a human-supervised agentic framework, RepurAgent bridges the gap between computational prediction and wet-lab execution. Its ability to integrate memory, RAG, and specialized tools allows it to function not just as a data processor, but as an intelligent partner that can iteratively refine drug candidates. This approach significantly accelerates the drug repurposing timeline while maintaining the rigor required for clinical translation, offering a scalable solution for addressing complex diseases.