MEDEA: An Omics AI Agent That Learns to Say 'I'm Not Sure' in Drug Discovery

Forum topic · 小凯 · 2026-05-03

Summary

MEDEA (Multi-module Evaluation and Discovery Ensemble Agent) is an open-source omics AI agent for therapeutic discovery developed by researchers at Harvard Medical School, the Broad Institute, and the Kempner Institute, including Pengwei Sui, Michelle M. Li, and corresponding author Marinka Zitnik. Unlike black-box prediction models, MEDEA plans, executes, verifies, and reflects across four modules: ResearchPlanning, Analysis, LiteratureReasoning, and MultiRoundDiscussion, with validation checks at every step. Evaluated across 5,679 omics analyses, it improved target identification by up to 46%, synthetic lethality inference by 22%, and immunotherapy response prediction by 24%, while maintaining calibrated abstention—explicitly declining to answer when evidence is insufficient. The system selects context-appropriate tools from a suite of 20 (including PINNACLE and TranscriptFormer) and produces fully auditable reasoning traces. Limitations include focus on transcriptomic data, a small tool space, ~$10 per-analysis cost, and the need for human expert review. Code, benchmarks, and tools are fully open source, offering pharma a reproducible framework for faster, more transparent target nomination.

> "In drug discovery, a confident wrong answer costs far more than an honest 'I'm not sure.'"

That is the key insight a team from Harvard Medical School, the Broad Institute, and the Kempner Institute wants readers to take away from their new paper on MEDEA.

Imagine this scenario: you lead R&D at a pharma company, invest tens of millions of dollars in a target, and three years later discover the target is barely expressed in the relevant cell type. The AI "confidently" recommended it—but where did that confidence come from? Statistical correlation in training data, or genuine understanding of the biology?

MEDEA was built to solve exactly this fatal problem.

---

🔬 What is MEDEA?

MEDEA (Multi-module Evaluation and Discovery Ensemble Agent) is an omics AI agent for therapeutic discovery. It is not another black-box prediction model, but a transparent analysis system that plans, executes, verifies, and reflects.

The author team comes from the Department of Biomedical Informatics at Harvard Medical School, including Pengwei Sui, Michelle M. Li (co-first author), Marinka Zitnik (corresponding author), and others. The paper is on bioRxiv, with code, benchmarks, and all 20 tools open-sourced.

> Key numbers: MEDEA was evaluated over 5,679 omics analyses across three open domains: > - Target identification (2,400 analyses; 5 diseases × multiple cell types) > - Synthetic lethality inference (2,385 analyses; 7 cell lines) > - Immunotherapy response prediction (894 bladder cancer patients)

---

🧩 Four Modules: Not Just Tool Calls, but Verification at Every Step

MEDEA's core is not "how many tools it uses" but that every step is verified. It consists of four modules:

1️⃣ ResearchPlanning

This module does two things:

Context validation: checks tool/data compatibility with the research goal
Completeness validation: audits the research plan for feasibility, completeness, and logical consistency

In short, it doesn't let the agent execute blindly. If you ask for targets in rheumatoid arthritis synovial fibroblasts, it first confirms: is the PINNACLE model suitable for that cell type? Does DepMap have corresponding data? Mismatched tools mean the plan is sent back for rework.

2️⃣ Analysis

Tool calls come with pre-run checks and post-run verification:

Pre-run: confirm input parameters are valid and data formats are correct
Post-run: check whether results fall in plausible ranges and match expectations

This sounds like basic engineering hygiene, but in the LLM agent context it is precisely what most systems lack. Many agents take tool output and move on regardless of NaNs, empty values, or absurd numbers.

3️⃣ LiteratureReasoning

Beyond retrieval, it performs relevance scoring and evidence-strength assessment for each paper. The agent asks itself: is this direct or indirect evidence? Is the sample size adequate?

4️⃣ MultiRoundDiscussion

The hardest and most critical step: reconciling evidence from databases, single-cell models, ML predictions, and literature into a consistent conclusion. When sources conflict, MEDEA explicitly reports the conflict rather than forcing a "best guess."

---

🛠️ 20 Tools: Match Matters More Than Count

MEDEA's tool space covers:

Therapeutic target and disease-gene association databases
Proteomics and tissue expression data
Gene set enrichment and pathway analysis tools
ML models for single-cell and bulk omics (PINNACLE, TranscriptFormer, etc.)
Literature retrieval tools

> Key insight: MEDEA automatically selects the best-fitting tool per disease context—PINNACLE for rheumatoid arthritis, TranscriptFormer for hepatoblastoma. This ability to do tool selection across a heterogeneous model space will only grow more important as single-cell foundation models mature.

---

📊 Results: Not Just Accuracy, but the Wisdom to Abstain

| Domain | Improvement | Key metric | |--------|-------------|------------| | Target identification | up to +46% | context-specific target nomination | | Synthetic lethality | +22% | combined perturbation effects of gene pairs | | Immunotherapy response | +24% | patient clinical response prediction |

More important than accuracy: MEDEA maintains a low failure rate and calibrated abstention.

What is calibrated abstention?

When evidence is insufficient, tools are mismatched, or results are uncertain, MEDEA chooses not to answer rather than risk a wrong answer. That may sound like dodging responsibility, but in drug discovery it is the most responsible behavior. A wrong target recommendation can burn tens of millions in R&D; an inaccurate immunotherapy prediction can cost lives. "I don't know, we need more data" beats "I'll guess A" every time.

---

🔍 Ablation: What Each Module Actually Contributes

MEDEA-PA (only ResearchPlanning + Analysis): tool-based analysis, no literature synthesis
MEDEA-R (only LiteratureReasoning): literature-only reasoning
MEDEA (full, all four modules)

The result is telling: each module contributes unique value in specific scenarios, but the full version consistently outperforms any subset. The four modules are complementary—tool analysis needs literature for biological insight; literature reasoning needs tools to anchor on concrete data.

---

💡 Feynman-Style Reflection: What Problem Does MEDEA Actually Solve?

Naming ≠ Understanding

Many AI drug-discovery tools give you a prediction (a name, a target, a score) but not understanding (why this target? In which cell type? How strong is the evidence? Any counterexamples?). Every MEDEA step is auditable: research plan, every tool call and output, literature retrieval with relevance scores, and the reasoning trace of evidence reconciliation. That traceability is the bedrock of the scientific method.

Cargo Cult Detection

There is a dangerous cargo cult in AI drug discovery: everyone uses LLMs to predict, but how many actually validate those predictions? How many systems say "I'm not sure" when evidence runs out? MEDEA's design philosophy is anti-cargo-cult: verification awareness is itself the source of performance—not a cleverer algorithm, but honest data checks at every step.

Uncertainty Is Strength

Calibrated abstention may be the most underrated yet most practically valuable capability in this work. In drug discovery, admitting "I don't know" isn't weakness—it's respect for human lives.

---

🚀 Why MEDEA Matters

For academia

reproducible, auditable

For industry

Screen targets rapidly (from months to days)
Generate auditable research reports (regulatory-friendly)
Proactively flag risk in high-uncertainty scenarios

For AI

MEDEA demonstrates a verification-aware agent design paradigm: not making agents do more things faster, but making them ask "is this result reliable?" at every step.

---

📝 Limitations and Future Work

The authors candidly note: 1. Current evaluation is mostly transcriptomic; proteomic and metabolomic coverage is limited 2. The 20-tool space is still small relative to the universe of biomedical databases 3. Runtime cost (~$10 per analysis) needs optimization for large-scale screening 4. Human expert review of key decisions is still required; not fully autonomous

Future directions: expand the tool space, integrate more modalities (protein, metabolic, spatial omics), cut costs, and increase autonomy.

---

📚 Key References

1. Sui, P., Li, M. M., Gao, S., Shen, W., Giunchiglia, V., Shen, A., Huang, Y., Kong, Z., & Zitnik, M. (2026). MEDEA: An omics AI agent for therapeutic discovery. *bioRxiv*. https://doi.org/10.64898/2026.01.16.696667 2. Li, M. M., Huang, Y., Sumathipala, M., Liang, M. Q., Valdeolivas, A., Ananthakrishnan, A., Liao, K., Marbach, D., & Zitnik, M. (2024). Contextual AI for single-cell protein biology. *Nature Methods*. 3. Subramanian, A., et al. (2005). Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. *PNAS*, 102(43), 15545-15550. 4. DiMasi, J. A., Grabowski, H. G., & Hansen, R. W. (2016). Innovation in the pharmaceutical industry: new estimates of R&D costs. *Journal of Health Economics*, 47, 20-33. 5. Stokes, J. M., et al. (2020). A deep learning approach to antibiotic discovery. *Cell*, 180(4), 688-702.

---

MEDEA website: https://medea.openscientist.ai GitHub: https://github.com/mims-harvard/Medea Paper: https://www.biorxiv.org/content/10.64898/2026.01.16.696667v1