A team led by Jinghao Feng and Weidi Xie published a framework in Nature Machine Intelligence this week that predicts how a single cell responds to a drug that has never been profiled at that cell level. The system is called MAP. It built a knowledge graph spanning 187,000 drugs, 23,000 genes, and 694,000 mechanistic relationships. It uses that graph to make zero-shot predictions for compounds with no prior experimental data.
The bottleneck this hits is the cost of drug screening. A pharma company that wants to know how a new molecule affects a lung cancer cell line today has to run the experiment. That means growing cells, dosing them, sequencing the transcriptome, and waiting. For every candidate molecule in a pipeline, that is a separate wet lab campaign. MAP removes that step for the early rounds. You feed it the drug's structure and its known mechanism. It predicts the transcriptional response.
The validation is concrete. In virtual screening of approved anti-cancer drugs against A-549 cells, a non-small-cell lung cancer line, MAP correctly prioritized four out of five approved drugs. It did this without ever seeing those drugs in training. The model generalizes because it is not memorizing drug identifiers. It is reading the mechanism.
The knowledge graph is the key component. It unifies 14 public databases into one structure. Drug-target interactions. Pathway memberships. Molecular mechanisms. Textual descriptions of how a drug works at the protein level. The model aligns all of that into a shared embedding space. Two drugs that hit the same pathway end up close together in that space, even if their chemical structures look nothing alike.
What this changes for a screening pipeline is the order of operations. Today, a company screens 10,000 compounds in the lab and sequences the top 100. With MAP, they can run all 10,000 in silico first. The top 50 by predicted response go to the lab. The sequencing cost drops by an order of magnitude. The time from target to candidate drops with it.
The limitation is clear. The predictions are at the transcriptome level. They tell you which genes go up and down. They do not tell you whether the drug will work in a human patient. That still requires the wet lab. But the wet lab now starts with a shorter list and a clearer hypothesis about why each drug should work.