Gene expression signatures (GES) derived from digital histology to predict pathologic complete response (pCR) to neoadjuvant chemotherapy (CT) in ISPY2 and other trial/real world cohorts.

Abstract

587 Background: GES predictive of response to therapy across multiple breast cancer subtypes are commercially available or in development. Deep learning models can predict GES from digital histology, and may serve as a lower-cost alternative immediately available at the time of biopsy. Methods: Transformer-based models trained to predict 38 distinct breast cancer signatures from pathology (all with Pearson correlation > 0.5 versus true GES) were previously developed using cases from The Cancer Genome Atlas. These models were applied to digital H&E from pre-treatment biopsies from HER2- cases treated with CT or CT + immunotherapy (IO) from the ISPY2 trial. The histology-derived GES most predictive of pCR in ISPY2 (as per area under the ROC curve [AUROC]) was tested in two external neoadjuvant cohorts - a subset of a trial from Yale of durvalumab + CT (NCT02489448) with TIL annotations, and patients receiving standard of care CT at University of Chicago. AUROC significance was assessed with 1000x bootstrapping, with Benjamini Hochberg correction applied in ISPY2 to account for testing multiple GES models. Tertiles of predicted expression calculated in ISPY2 defined groups with low, medium, and high likelihood of pCR; these cutoffs were tested in the external cohorts. Results: Accuracy for pCR prediction was tested in 578 patients from seven arms of ISPY2 – with breakdown by treatment and hormone receptor (HR) status shown in Table. A histology model for a GES defined by estrogen regulated genes (Oh et al, JCO 2006) – including proliferation, apoptosis, and interferon-response genes – predicted pCR with the highest AUROC (0.794) in ISPY2, and outperformed a logistic regression fit on grade, HR status, and tumor / nodal stage (AUROC 0.705, p for comparison 0.0001). Tertiles of predicted expression for this GES (computed in ISPY2) identified groups with low / high pCR rates which were robust to treatment, HR status, and consistent in validation cohorts (Table). This digital signature (AUROC 0.737) compared favorably to pathologist TIL annotation (AUROC 0.664) from the external Yale cohort. An explainability tool demonstrated that patterns of lymphocytic infiltrate and poor differentiation contributed to high signature predictions from histology. Conclusions: A digital histology-derived GES consistently identifies patients at low / high likelihood of pCR with neoadjuvant CT or CT + IO, and may improve treatment personalization. Subgroup n AUROC p % pCR (low expression) % pCR (mid expression) % pCR (high expression) ISPY2 579 0.794 2 x 10 -28 7.6 26.7 58.6  ISPY2, CT + IO 459 0.810 3 x 10 -26 8.0 28.2 64.1  ISPY2, CT only 120 0.726 0.001 6.4 20.0 36.8  ISPY2, HR- 239 0.704 4 x 10 -7 14.8 29.2 58.4  ISPY2, HR+ 340 0.817 1 x 10 -15 6.5 24.8 59.0 UChicago HR- 151 0.746 3 x 10 -7 11.1 27.3 55.1 UChicago HR+ 63 0.847 1 x 10 -5 5.5 12.0 70.0 Yale HR- 41 0.737 0.005 0.0 50.0 61.5