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

Forum topic · 小凯 · 2026-08-27

Summary

MAP (Mechanism-Aware knowledge-driven Perturbation prediction), developed by Shanghai Jiao Tong University's Zhang Ya and Xie Weidi teams with Harvard Medical School and published in Nature Machine Intelligence (August 2026), is the first single-cell perturbation model capable of zero-shot prediction of cellular responses to entirely new, uncharacterized drugs. Instead of treating drugs as discrete IDs, MAP maps molecular structures (SMILES), protein sequences, and mechanism-of-action (MOA) text into a unified embedding space via contrastive learning, backed by MAP-KG—a medical knowledge graph built from 14 databases with 187,000 drug entities, 22,000 gene entities, and 694,000 manually verified drug-gene mechanistic relations. On the Tahoe-100M dataset, MAP outperforms 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—raising effective virtual screening hit rates from roughly 10% to 31%. In an A549 lung cancer virtual screening of 58 candidates, 4 of 5 approved anticancer drugs ranked in the top 15. The paper proposes 'knowledge scaling' as a third scaling path beyond data and parameters, while noting limitations in MOA text quality, rare mechanisms, cross-species, and dose-response modeling.

MAP: A Knowledge-Driven Virtual Cell That Predicts Drugs No One Has Tested

On August 26, 2026, *Nature Machine Intelligence* published MAP (Mechanism-Aware knowledge-driven Perturbation prediction) by Zhang Ya and Xie Weidi's teams at Shanghai Jiao Tong University's School of AI, in collaboration with Harvard Medical School. The paper achieves something deceptively simple but fundamental: it is the first single-cell perturbation model that can predict cellular responses to brand-new, uncharacterized drugs in a zero-shot setting—moving "virtual cells" from fitting existing experiments to predicting interventions that have never been observed.

Key points

The core problem: Prior single-cell perturbation models treat drugs as discrete category IDs. Trained on Perturb-seq data covering only hundreds to thousands of drugs—versus an estimated 10^60 druggable chemical space—they learn a lookup function (drug ID → cell response), not a causal chain (structure → mechanism → response). Generalization to unseen drugs is near zero.
The solution — MAP-KG: A medical knowledge graph integrating 14 databases (DrugBank, PubChem, ChEMBL, STITCH, IntAct, STRING, Ensembl, HGNC, BindingDB, BioGRID, DrugCentral, TTD, PrimeKG, UniProt):

| Category | Count | Sources | |---|---|---| | Drug entities | 187,000 | DrugBank, PubChem, ChEMBL, etc. | | Gene entities | 22,000 | Ensembl, HGNC, etc. | | Drug–gene mechanistic relations | 694,000 | STITCH, IntAct, STRING, etc. | | Multimodal data | SMILES / protein sequences / MOA text | linked to entities |

Compared to PrimeKG, DRKG, CKG, and Hetionet, MAP-KG is distinguished not by relation count (it has the fewest) but by the most drug entities plus the highest relation precision—all 694,000 relations are manually verified mechanistic relationships (inhibitor, activator, regulator, etc.), not co-occurrence edges.

Model design: Molecular structure + protein sequences + MOA text are mapped into a unified embedding space via contrastive learning. Given only a novel drug's SMILES and known MOA text, the model can predict its perturbation effect on cells.

Benchmark results on Tahoe-100M

Tahoe-100M is a giga-scale single-cell perturbation atlas covering ~100 million cells across hundreds of drugs and multiple cancer cell lines.

Unseen cell–drug combinations (drug repurposing scenario): MAP's Top-50 differential expression gene (DEG) Pearson correlation beats the CRISP baseline by 13.3 percentage points.
Unseen drugs (the drug is removed from knowledge graph nodes too; only molecular structure is given): Top-50 DEG correlation improves 12.2 points, and gene regulation direction accuracy improves 21.0 points.
Business translation: if a traditional model picks ~10% true positives in virtual screening, MAP picks ~31%.

A549 virtual screening case

On the A549 non-small-cell lung cancer cell line, MAP ranked 58 candidate molecules (approved anticancer drugs + clinical + preclinical compounds) using only structure + MOA text, with no wet-lab results:

4 of the 5 approved anticancer drugs landed in the model's top 15 (80% recall vs. a random baseline of ~8.6%).
GSEA pathway enrichment showed predicted expression changes align well with relevant biological pathways.

The authors caution: "These results preliminarily demonstrate MAP's potential in virtual screening. Moving from computational prediction to real drug discovery still requires further cell experiments, animal studies, and clinical validation."

'Knowledge scaling': a third scaling path

Corresponding author Xie Weidi's key claim: as knowledge graph quality and scale grow, model generalization keeps improving—suggesting that life-science AI may have a "knowledge scaling" path alongside data scaling and parameter scaling. Structured biomedical knowledge acts as external memory (analogous to RAG in LLMs), allowing models to compensate for scarce data with knowledge density.

If this paradigm holds, industry implications include:

Data companies (Schrödinger, Recursion, Insilico Medicine, Isomorphic Labs) see their wet-lab data moats partially diluted;
Knowledge companies (Elsevier, Clarivate, etc.) become new infrastructure;
Pharma firms (Pfizer, Roche, AZ, Hengrui, BeiGene) must decide whether to build or buy knowledge graphs.

Limitations

1. MOA text quality varies: ~30% of the 187,000 drugs have auto-extracted + manually verified MOA text (est. 5–15% error rate); the remaining 70% is fully manual but lags 2–3 years behind. 2. Rare mechanisms under-covered: relation density skews toward well-studied drugs and pathways (oncology targets, metabolic enzymes). 3. Cross-species not explicitly modeled: no bridging model for mouse/rat/monkey MOA mapping. 4. Dose-response ignored: the model maps structure → response without a dose dimension.

Context and commercialization

MAP arrives in a busy August 2026 for AI biology: Anthropic's LLM-designed protein binders, Biohub's cancer "ultra-conserved" state findings, Moderna's intismeran autogene Phase III, Goodfire's Silico + Pleiades platform, and SequenTx's RL-designed sequential drug combinations (*Nature Machine Intelligence*, doi:10.1038/s42256-026-01192-1).
Co-development unit OneX Intelligence (co-founded by Xie Weidi) positions MAP alongside its earlier DeepRare system (*Nature* 2025, disease identification/clinical reasoning) as layers of a "biological intelligence foundation model" stack—potentially combining patient phenotype + genetics (DeepRare) with drug candidate + mechanism prediction (MAP) into an AI-assisted precision prescribing system.
For Chinese AI pharma companies (Insilico with Rentosertib in Phase III, XtalPi, BioMap, DP Technology, etc.), suggested actions: adopt the three-modality (SMILES + protein sequence + MOA text) representation, build or license a MAP-KG-scale knowledge graph (open-source PrimeKG is a less precise alternative), and establish dry–wet experiment closed loops.

What to watch in the next 12 months

Whether MAP's team releases weights/APIs;
Whether 3+ AI pharma companies independently replicate 21%+ zero-shot direction-accuracy gains on internal benchmarks;
Whether 1–2 MAP-recommended virtual screening candidates enter wet-lab validation with preclinical data.

Any of these would turn "knowledge scaling" from a paper slide into an industry roadmap.

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 Intelligence), Zhao Ziheng (SJTU), Zhang Xiaoman (Harvard Medical School postdoc → SJTU associate professor)
Corresponding authors: Prof. Zhang Ya, Assoc. Prof. Xie Weidi (SJTU School of AI)
SJTU official release; OneX Intelligence release; Sohu Tech, Zhihu coverage (2026-08-26)