AutoDiscovery generated a hypothesis about immune activity in lobular breast cancer, and the team tested it using independent data and tumor tissue
AutoDiscovery generated a hypothesis about immune activity in lobular breast cancer, and the team tested it using independent data and tumor tissue
On 27 August, the Ai2 research institute and Providence Swedish Cancer Center reported on a study in which a language model searched breast cancer data for testable research questions. Brief comments from an oncologist narrowed the search, after which the team tested one of the identified signals in an independent database and tissue samples.
Lobular cancer is a form of breast cancer. The study compared it with the more common ductal cancer. The open TCGA database contained data from 1 097 cases, including tumor type, mutations, gene activity, and treatment outcomes.
Using these data, AutoDiscovery formulates hypotheses, plans analyses, writes executable Python code, and runs the calculations. The system ranks results more highly when the analysis produces a larger change in its initial assessment of a hypothesis. Researchers receive the hypothesis, analysis, and code so that they can reproduce the test. In the SetScope study, two coding errors changed the results enough that the author withdrew them.
In the first run, the model received only descriptions of the tables and tested 100 hypotheses. The authors considered the results unsuitable for clinical interpretation. An oncologist therefore commented on 26 positively rated results, identifying conclusions that lacked clinical relevance and points that required clarification. In the second run, the system tested 500 hypotheses. A hypothesis proposing higher activity of the PDCD1 and CD274 genes in lobular cancer ranked third among the 13 results for which the analysis produced the largest increase in the model's assessment.
The PDCD1 gene encodes the PD-1 receptor, while CD274 encodes PD-L1, a protein that binds to this receptor. This interaction helps suppress T-cell attacks on tumors. In the independent METABRIC database, the authors selected primary tumors that were hormone-sensitive and did not overexpress the HER2 protein. The sample included 696 ductal and 74 lobular tumors. Both genes showed slightly higher activity in the lobular tumors. The researchers then used staining that distinguishes several cell types within a single sample to examine 12 lobular cancer samples and six normal tissue samples. The cells surrounding the lobular tumors included more PD-1-expressing CD8 and CD4 T-cells.
Oncologist Kelly Paulson described this sequence in the partnership announcement:
“AutoDiscovery helped us identify a promising signal that we might not otherwise have investigated. We then tested it in additional datasets and in the laboratory.”
The 2023 GELATO study had already examined the combination of carboplatin and PD-L1 blockade in metastatic lobular cancer. The authors of the current study propose evaluating immunotherapy specifically for lobular cancer in future trials. In their workflow, comments from a physician guided the search, while an independent database and tissue samples provided two separate ways to test the resulting signal.