Expert-seeded AI hypothesis search finds immune signal in lobular breast cancer

The immune signal in this study did not come from an AI working alone. Ai2's AutoDiscovery is a set of language-model agents that propose a hypothesis about a dataset, write and run Python code to test it, and score how much the result changed the model's own belief. Hypotheses that move that belief the most are ranked as the most surprising. When researchers from Providence and Ai2 pointed it at The Cancer Genome Atlas breast cancer data with only a description of the dataset, the preprint reports that it tested 100 hypotheses at roughly $0.19 each and none was judged clinically useful.

The second run is where the result came from. A breast cancer physician-scientist specializing in immunotherapy annotated the first run's positive hypotheses with short notes, from one word to a few sentences, taking about four hours in total. Fed those notes, AutoDiscovery tested 500 more hypotheses. Many of the positives restated known biology. Among those with the largest belief shifts, the third-ranked one was clinically interesting: invasive lobular carcinoma, a breast cancer subtype long treated as immunologically "cold" and excluded from trials such as KEYNOTE-756, showed higher expression of the immune-checkpoint genes behind PD-1 and PD-L1 than the more common ductal type.

The authors then checked that ranking outside the AI. In METABRIC, a separate European cohort, they compared 74 lobular and 696 ductal tumors of the same hormone-receptor subtype and report a modest but statistically significant increase in both genes. In 12 lobular tumor samples, the standard clinical PD-L1 test was positive in only three, but a more sensitive multiplex immunofluorescence stain found more PD-1-positive T cells than in normal breast tissue, and five of the nine standard-negative tumors still showed many PD-L1-positive T cells.

On that basis the authors argue that lobular breast cancer, including early-stage hormone-receptor-positive disease, should be tested in immunotherapy trials rather than excluded. That is a case for running trials, not evidence that patients benefit. The work is a preprint that has not been peer reviewed, its tissue check covers a dozen tumors, and the authors say they could not validate every surprising hypothesis and expect some to be wrong. They also say expert annotation markedly improved performance compared with the unguided run. Ai2's announcement adds that Providence is now installing AutoDiscovery in its own cloud environment to run on protected research and clinical data.