Medea: an omics AI agent for therapeutic discovery that performs verified, long-horizon reasoning across protein contexts, cell states, and personalized patient omics data.
Medea: an omics AI agent for therapeutic discovery that performs verified, long-horizon reasoning across protein contexts, cell states, and personalized patient omics data.
Medea validates each decision and output against data and tool constraints as the analysis unfolds.
2️⃣ MEDEA is a verification-aware omics agent designed for long-horizon therapeutic reasoning.
What is unique:
• Context + integrity verification while building the research plan (catches context slips early)
• Pre-run and post-run checks around tool execution (an analysis can run without errors and still be wrong)
• Study relevance + evidence-strength screening for literature (not just retrieval)
• Multi-round consensus that reconciles tool outputs, literature, and the LLM, with calibrated abstention when evidence is insufficient
• A rich tool space with single-cell and bulk transcriptomics FMs, genetic dependency maps, protein and tissue networks, pathway and ontology resources, and interpretable ML models
3️⃣ We evaluate Medea across 5,679 analyses in three open-ended domains:
🔹 Target identification across cell type contexts
🔹 Synthetic lethality reasoning in cancer cell lines
🔹 Immunotherapy response prediction from patient transcriptomes
4️⃣ Target discovery fails when biology is averaged. Medea reasons at cell-type resolution, integrating:
• single-cell atlases
• disease genetics
• drug evidence
5️⃣ Synthetic lethality asks a subtle question: Do two gene perturbations together reduce viability more than expected from either alone?
Medea treats this as an evidence synthesis task, combining:
• CRISPR dependency data
• pathway context
• literature support
And it abstains when evidence is weak.
6️⃣ Across seven cancer cell lines, Medea outperforms LLMs in identifying true synthetic lethal interactions.
Medea not only predicts more positives but also corrects LLM errors and avoids confident false claims.
7️⃣ Immunotherapy response prediction from personal transcriptomes
Medea: an omics AI agent for therapeutic discovery that performs verified, long-horizon reasoning across protein contexts, cell states, and personalized patient omics data.
Medea validates each decision and output against data and tool constraints as the analysis unfolds.
Pengwei Sui Michelle L. Shanghua Gao
👉 Medea: https://lnkd.in/eGNy5Q_D
👉 Open science: https://lnkd.in/eNfp3D3B
👉 Paper: https://lnkd.in/efe4-98b
2️⃣ MEDEA is a verification-aware omics agent designed for long-horizon therapeutic reasoning.
What is unique:
• Context + integrity verification while building the research plan (catches context slips early)
• Pre-run and post-run checks around tool execution (an analysis can run without errors and still be wrong)
• Study relevance + evidence-strength screening for literature (not just retrieval)
• Multi-round consensus that reconciles tool outputs, literature, and the LLM, with calibrated abstention when evidence is insufficient
• A rich tool space with single-cell and bulk transcriptomics FMs, genetic dependency maps, protein and tissue networks, pathway and ontology resources, and interpretable ML models
3️⃣ We evaluate Medea across 5,679 analyses in three open-ended domains:
🔹 Target identification across cell type contexts
🔹 Synthetic lethality reasoning in cancer cell lines
🔹 Immunotherapy response prediction from patient transcriptomes
4️⃣ Target discovery fails when biology is averaged. Medea reasons at cell-type resolution, integrating:
• single-cell atlases
• disease genetics
• drug evidence
5️⃣ Synthetic lethality asks a subtle question: Do two gene perturbations together reduce viability more than expected from either alone?
Medea treats this as an evidence synthesis task, combining:
• CRISPR dependency data
• pathway context
• literature support
And it abstains when evidence is weak.
6️⃣ Across seven cancer cell lines, Medea outperforms LLMs in identifying true synthetic lethal interactions.
Medea not only predicts more positives but also corrects LLM errors and avoids confident false claims.
7️⃣ Immunotherapy response prediction from personal transcriptomes
Predicting immunotherapy response requires reasoning over:
• tumor transcriptomes
• immune programs
• clinical context
8️⃣ Medea works where heuristics fail. Across TMB-high/low and inflamed/non-inflamed tumors, Medea:
• outperforms LLMs
• rescues errors from ML models
• adapts reasoning to subgroup biology
Congrats to @sui67713 @_michellemli @GaoShanghua Wanxiang Shen Valentina Giunchiglia Andrew Shen Yepeng Huang Zhenglun Kong
https://lnkd.in/eGNy5Q_D
Open science: https://lnkd.in/eNfp3D3B
https://lnkd.in/efe4-98b
Harvard Medical School Harvard University Harvard Medical School Department of Biomedical Informatics Broad Institute of MIT and Harvard Kempner Institute at Harvard University Harvard Data Science Initiative