An AI agent now predicts what happens inside a cancer cell when a single gene gets switched off, and its creators actually checked the guesses against real patient tumors instead of a list of known cancer genes.
The system, called CASCADE, runs on precomputed gene regulatory networks (maps of which genes control which other genes) and takes requests through MCP, the same protocol used to plug AI agents into outside tools. Rather than asking whether CASCADE's predicted genes sounded like plausible cancer genes, the researchers tested whether it got the direction of change right, using gene copy-number amplification in tumors as a stand-in for the opposite of switching a gene off. For MYC, a well-studied cancer gene, CASCADE's predictions matched real tumor expression in breast, colon, and stomach cancer samples 90.0%, 72.0%, and 85.7% of the time, and held up in an independent patient dataset, METABRIC, at 87.2%.
The catch is that against curated gene-set databases already built from MYC and E2F biology, CASCADE did not predict anything scientists did not already know; it just clearly beat a coin flip. Extended to fifteen more genes, the accuracy got patchy: genes that drive cell division mostly held up, while genes that define a cell's identity, plus one gene called CCND2, consistently failed for reasons the authors call a hedged guess rather than a finding.
The more telling number might be buried further down the paper: asked to translate plain English requests into CASCADE's actual tool calls, a small local language model got it exactly right 71.4% of the time and a larger model 85.7%, and both models kept defaulting to the wrong kind of gene edit whenever a request was ambiguous, a bug the team could not fix because the trigger condition never came up in testing.