MEDEA: An Omics AI Agent for Therapeutic Discovery
> "In drug discovery, a confident wrong answer costs far more than an honest 'I'm not sure.'"
This is the core insight of MEDEA (*Multi-module Evaluation and Discovery Ensemble Agent*), a new paper from Harvard Medical School, the Broad Institute, and the Kempner Institute.
Imagine you lead R&D at a pharma company: you invest tens of millions of dollars in a target, and three years later discover it barely expresses in the target cell type. The AI "confidently" recommended it—but was that confidence from statistical correlation in training data, or genuine understanding of biology? MEDEA was built to solve this exact problem.
What Is MEDEA?
MEDEA is an omics AI agent for therapeutic discovery. It is not another black-box predictor but a transparent analysis system that plans, executes, verifies, and reflects. The author team is from the Department of Biomedical Informatics at Harvard Medical School, including Pengwei Sui, Michelle M. Li (co-first authors), and Marinka Zitnik (corresponding author). The paper is on bioRxiv; code, benchmarks, and all 20 tools are open source.
Key numbers: MEDEA was evaluated in 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
- Context verification: checks tool/data compatibility with the research goal
- Completeness verification: audits the plan's feasibility, completeness, and logical consistency
- 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
- MEDEA-PA (ResearchPlanning + Analysis only): tool-based analysis, no literature synthesis
- MEDEA-R (LiteratureReasoning only): purely literature-based reasoning
- MEDEA (full system)
- Naming ≠ understanding: Most AI drug-discovery tools output predictions without understanding. MEDEA produces a fully auditable report: research plan, every tool call and output, literature relevance scoring, and the evidence-reconciliation reasoning trace.
- Anti cargo-cult: Everyone is using LLMs for prediction; how many actually verify those predictions? MEDEA's philosophy is that verification awareness is itself the source of performance.
- Uncertainty is strength: Calibrated abstention may be the most underrated capability—admitting "I don't know" is respect for life.
- Academia: a reproducible, auditable agent framework for omics analysis, with a solid 5,679-analysis benchmark.
- Industry: faster target triage (months → days), auditable research reports for regulators, proactive risk flagging under high uncertainty.
- AI: a design paradigm for verification-aware agents—not doing more things faster, but asking at every step, "Is this result reliable?"
Four Modules: Not Just Tool Calls—Every Step Is Verified
1. ResearchPlanning
The agent does not blindly execute. If asked to find targets in rheumatoid arthritis synovial fibroblasts, it first confirms: Is the PINNACLE model suited to this cell type? Does DepMap have matching data? Mismatched tools get the plan sent back for rework.
2. Analysis
Pre-run checks (valid parameters, correct data formats) and post-run validation (results within reasonable ranges, consistent with expectations). This sounds like basic engineering hygiene, but most LLM agents skip it—accepting tool output that is NaN, empty, or absurd and moving on regardless.3. LiteratureReasoning
Not just retrieval: each paper gets a relevance score and an evidence-strength assessment. Is the evidence direct or inferential? Is the sample size adequate?4. MultiRoundDiscussion
The hardest step: reconciling evidence from databases, single-cell models, ML predictions, and literature into a consistent conclusion. When sources conflict, MEDEA explicitly reports the conflict instead of forcing a "best guess."20 Tools: Match Matters More than Quantity
The tool space covers:
Key insight: MEDEA selects the best-fitting tool per disease context—PINNACLE for rheumatoid arthritis, TranscriptFormer for hepatoblastoma. This tool selection across a heterogeneous model space will matter more as single-cell foundation models mature.
Results: Accuracy Plus the Wisdom to Abstain
| Domain | Improvement | Key metric | |--------|------------|-----------| | Target identification | up to +46% | context-specific target nomination | | Synthetic lethality | +22% | joint perturbation effects of gene pairs | | Immunotherapy response | +24% | patient clinical response prediction |
More important than accuracy: low failure rate and calibrated abstention. When evidence is insufficient, tools mismatched, or results uncertain, MEDEA chooses *not to answer*. In drug discovery this is the most responsible behavior—an erroneous target recommendation can burn tens of millions of dollars; an inaccurate immunotherapy prediction can affect lives. "I don't know, we need more data" beats "I guess it's A."
Ablation Study
Every module contributes unique value in specific scenarios, but the full version consistently outperforms any subset—the modules are complementary: tool analysis needs literature for biological insight; literature reasoning needs tools to anchor in concrete data.
What Problem Does MEDEA Actually Solve?
Why It Matters
Limitations and Future Work
1. Evaluation is mostly transcriptomic; proteomics and metabolomics coverage is limited 2. The 20-tool space is small relative to the universe of biomedical databases 3. Runtime cost (~$10 per analysis) needs optimization for large-scale screens 4. Human expert review is still required for key decisions—not fully autonomous
Future directions: expand the tool space, integrate more modalities (protein, metabolic, spatial omics), reduce cost, increase autonomy.
References
1. Sui, P., Li, M. M., et al. (2026). MEDEA: An omics AI agent for therapeutic discovery. *bioRxiv*. https://doi.org/10.64898/2026.01.16.696667 2. Li, M. M., et al. (2024). Contextual AI for single-cell protein biology. *Nature Methods*. 3. Subramanian, A., et al. (2005). Gene set enrichment analysis. *PNAS*, 102(43), 15545-15550. 4. DiMasi, J. A., et al. (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(6), 688-702.
MEDEA site: https://medea.openscientist.ai GitHub: https://github.com/mims-harvard/Medea Paper: https://www.biorxiv.org/content/10.64898/2026.01.16.696667v1