The current assays for assessing HER2 were “fit-for-purpose” (over 25 years ago) to identify HER2
overexpression/gene amplification and are not built to subclassify unamplified HER2 expression in cancer.
Applying these legacy assays results in high inter-rater disagreement rates for HER2-negative vs. HER2-low
cases (up to 75% discordance for 0 vs. 1+/2+). With the emergence of new therapy options for HER2-low cancer
– namely trastuzumab deruxtecan (T-DXd) – a reliable companion diagnostic is needed for patients. The
incorrect application of these HER2 legacy assays has led to controversial findings in the field, including the
paradoxical 30% objective response rate seen in IHC=0 breast cancer in the DAISY trial. This proposal aims to
accurately stratify patients with HER2-low/negative breast cancer using our analytic, high-sensitivity HER2 (HS-
HER2) assay to develop algorithms for the selection of the most appropriate therapy for breast cancer patients.
Specifically, the proposal has two aims:
Aim 1 will offer insight on clinically measurable pathologic & spatial parameters that influence the “bystander
effect” associated with these newer, linker cleavable antibody drug conjugates like T-DXd. Specifically, this aim
focuses on quantifying the spatial heterogeneity and HER2 expression patterns alongside HER2 tumor
concentration to ultimately predict T-DXd response. I hypothesize that subclassifying HER2 expression in breast
cancer with quantitative spatial heterogeneity metrics for “cluster”-type and single-cell “mosaic”-type patterns will
further explain why only ~30% of IHC=0 benefit from T-DXd in the DAISY trial (although ~70% of IHC=0 cases
have detectable attomole/mm2 HER2 concentrations by our HS-HER2 assay). I propose to investigate large
retrospective and prospective breast cancer cohorts with our HS-HER2 assay, as well as cohorts of HER2-low
breast cancer from ongoing T-DXd clinical trials.
Aim 2 seeks to explore if an H&E-based "virtual HER2 stain" can predict low HER2 expression status. Our
lab and others have illustrated that HER2 gene amplification is associated with detectable morphological
changes from the H&E image alone. I hypothesize that HER2-low breast cancer will also have distinct
morphologic associations that can be discovered. The study proposes to build a dataset of HS-HER2 and H&E
aligned, adjacent serial section whole slide images. Then, train state-of-the-art computer vision architectures on
H&E images with a multi-task, multi-image pre-training and mixed-supervision multiple instance learning strategy
to predict low HER2 expression status (defined by the HS-HER2 assay in attomole/mm2). This aim will reveal
H&E morphologic features of breast cancer associated with low HER2 expression and T-DXd response.
Overall, this proposal addresses the critical unmet clinical need for accurately assessing HER2 expression
in breast cancer patients and proposes two innovative aims to develop algorithms for T-DXd patient selection.