An AI model combines digitized breast-biopsy slides with clinical data to estimate pathological complete response after pre-surgery therapy: no invasive cancer remains in the breast and regional lymph nodes. In a retrospective evaluation across nine independent cohorts of 1,412 patients, it achieved a pooled AUROC of 0.79. The study is described in an arXiv preprint submitted Oct. 2.
NEO works in two stages. MORPHEUS first estimates gene expression—roughly, which genes are active and how strongly—from routine biopsy slides stained with hematoxylin and eosin (H&E). It learned this mapping from slides paired with RNA measurements from 8,742 patients across 32 cancer types. NEO then combines inferred gene activity with eight clinical variables; researchers developed this response model using 1,080 breast-cancer patients in five cohorts. It needs no separate gene-expression test when it runs.
AUROC measures how often a model ranks a patient who reached the outcome above one who did not; 0.79 is not 79% accuracy. Scores ranged from 0.658 to 0.880 across external cohorts. In a smaller comparison involving 237 patients, NEO scored 0.868 compared with 0.648 for pathologist-assessed Ki-67, a marker of cell growth. The study did not compare NEO with genomic response assays.
All nine evaluation cohorts were retrospective and observational. The authors say results alone cannot guide treatment decisions; prospective research is needed to show whether the estimate improves care or outcomes. They report that a silent-deployment study has begun enrollment, where the model runs in the background without affecting treatment.
Sources
- Park et al., “Transcriptome-informed multi-modal AI for predicting neoadjuvant therapy response from breast cancer biopsies,” arXiv preprint
- National Cancer Institute, Breast Cancer Treatment PDQ
- Mikael Häggström via Wikimedia Commons, CC0 1.0 (illustrative normal tissue, not an image from the study)
