Shanghai Jiao Tong University's MAP Enables Zero-Shot Prediction of Unprofiled Drugs, Boosting Drug Repurposing Hit Rates from 10% to 21%

Forum topic · 小凯 · 2026-08-27

Summary

Researchers at Shanghai Jiao Tong University's School of Artificial Intelligence, together with Harvard Medical School and OneX Intelligence, published MAP (Mechanism-Aware knowledge-driven Perturbation prediction) in Nature Machine Intelligence on August 26, 2026. MAP is the first single-cell perturbation prediction model capable of zero-shot prediction of cellular responses to entirely new, uncharacterized drugs. The team built MAP-KG, a medical knowledge graph integrating 14 biological databases containing 187,000 drug entities, 22,000 gene entities, and 694,000 manually curated drug-gene mechanism relations. Unlike prior models that treat drugs as discrete IDs, MAP maps molecular structures (SMILES), protein sequences, and mechanism-of-action text into a unified embedding space via contrastive learning. On the Tahoe-100M dataset, MAP outperformed the CRISP baseline by 13.3 percentage points (Top-50 DEG Pearson correlation) on unseen cell-drug combinations and by 12.2 points on unseen drugs, with a 21.0-point gain in gene regulation direction accuracy. In an A549 lung cancer virtual screening of 58 candidates, MAP ranked 4 of 5 approved anticancer drugs in its top 15. The authors propose 'knowledge scaling' as a third scaling path beyond data and parameter scaling for AI in drug discovery.

MAP: Zero-Shot Single-Cell Perturbation Prediction for Unprofiled Drugs

On August 26, 2026, *Nature Machine Intelligence* published MAP (Mechanism-Aware knowledge-driven Perturbation prediction) by the team of Zhang Ya and Xie Weidi (Shanghai Jiao Tong University School of Artificial Intelligence), in collaboration with Harvard Medical School and OneX Intelligence. The paper is the first to enable single-cell perturbation prediction models to predict cellular responses to entirely new, uncharacterized drugs under zero-shot conditions — moving "virtual cells" from fitting existing experiments to predicting never-observed interventions.

Key points

The core problem: Prior single-cell perturbation models treat drugs as discrete category IDs, effectively learning a lookup function. With only hundreds to thousands of experimentally profiled drugs versus an estimated 10^60 druggable chemical space, generalization to unseen drugs was near zero.
MAP-KG knowledge graph: Integrates 14 biological databases (DrugBank, PubChem, ChEMBL, STITCH, IntAct, STRING, Ensembl, HGNC, UniProt, etc.) with:
187,000 drug entities
22,000 gene entities
694,000 manually verified drug-gene mechanism relations (inhibitor, activator, regulator semantics)
Multimodal associations: SMILES structures, protein sequences, MOA text
Method: Instead of discrete drug IDs, MAP maps molecular structure + protein sequence + mechanism-of-action text into a unified embedding space via contrastive learning. Given only a new drug's SMILES and known MOA text, the model can predict its cellular perturbation response.

Benchmark results on Tahoe-100M

Unseen cell-drug combinations (drug repurposing scenario): Top-50 DEG Pearson correlation improved by 13.3 percentage points over the CRISP baseline.
Unseen drugs (drug removed from both training data and knowledge graph nodes): Top-50 DEG correlation improved by 12.2 percentage points; gene regulation direction prediction accuracy improved by 21.0 percentage points.
A549 NSCLC virtual screening: Among 58 candidate molecules, MAP's top-15 rankings included 4 of 5 approved anticancer drugs (80% recall vs. ~8.6% random baseline). GSEA pathway enrichment showed strong consistency with predicted expression changes.

"Knowledge scaling" as a third scaling path

Corresponding author Xie Weidi noted that as the quality and scale of the biomedical knowledge graph increase, model generalization continuously improves — suggesting life-science AI may have a "knowledge scaling" path alongside the established data scaling and parameter scaling paradigms. Structured biomedical knowledge serves as external memory, compensating in data-scarce regions, analogous to how RAG supplements LLMs.

If validated, this reshapes industry economics: wet-lab data advantages of companies like Recursion or Insilico may be diluted, while owners of structured biomedical knowledge bases become new infrastructure, and pharma companies must decide whether to build or buy knowledge graphs.

Limitations

MOA text quality: ~30% of MOA text is auto-extracted (estimated 5–15% error rate); manually written portions lag 2–3 years.
Rare mechanisms undercovered: relation density skews toward well-studied drugs and pathways.
Cross-species: no explicit bridging model for mouse/rat/monkey MOA mappings.
Dose-response not modeled: identical structures at different doses cannot be distinguished.

Industrialization

Co-developer OneX Intelligence (co-founded by Xie Weidi) is building a biological-intelligence foundation model stack: DeepRare (Nature 2025, disease identification and clinical reasoning) plus MAP (intervention prediction), pointing toward integrated AI-assisted precision medicine. The authors caution that computational predictions still require cell, animal, and clinical validation.

What to watch

1. Whether MAP's team releases weights or an API. 2. Whether 3+ AI pharma companies independently reproduce the 21%+ zero-shot direction-accuracy gains. 3. Whether 1–2 MAP-recommended virtual screening candidates enter wet-lab validation with preclinical data.

---

Sources

Nature Machine Intelligence (online 2026-08-26): *A Knowledge-driven Framework for Predicting Single-cell Responses for Unprofiled Drugs*, DOI: 10.1038/s42256-026-01286-w
Paper: https://www.nature.com/articles/s42256-026-01286-w
First authors: Feng Jinghao (SJTU + OneX), Zhao Ziheng (SJTU), Zhang Xiaoman (Harvard Medical School postdoc → SJTU associate professor)
Corresponding authors: Prof. Zhang Ya, Assoc. Prof. Xie Weidi (SJTU School of Artificial Intelligence)
Comparison paper: SequenTx, same issue, doi:10.1038/s42256-026-01192-1