Abstract B008: Machine-learning enabled quantification of colocalized multiplex IHC signals with spectral overlap
Waleed Tahir, Yibo Zhang, Jun Zhang, Jacqueline Brosnan-Cashman, Robert Egger, Justin Lee- Cancer Research
- Oncology
Abstract
Background: Immunohistochemistry (IHC) is the gold standard for detecting protein biomarkers in cancer. In breast cancer, several nuclear biomarkers are relevant for guiding treatment decisions, including estrogen receptor (ER) and progesterone receptor (PR), indicative of response to endocrine therapy, and the proliferation marker Ki67, indicative of response to CDK inhibition1. However, these nuclear markers are detected using chromogens that colocalize and spectrally overlap with hematoxylin (HTX), the standard nuclear counterstain for IHC. Thus, the accurate detection and quantification of nuclear antigens is challenging. To overcome this limitation of IHC-based biomarker assays, we developed a method to identify colocalized signals after staining with chromophores with spectral overlap.
Methods: Multiplex IHC (mIHC) using primary antibodies against Ki67, PR, and ER was performed on a breast cancer tissue microarray (N=65 cores). Antigens were detected by dabsyl-, tamra-, and Cy5-conjugated secondary antibodies, respectively, and slides were counterstained with HTX. The cores were scanned with brightfield imaging (Aperio AT2 and 3DHistech microscopes). The same cores were imaged on a custom brightfield microscope using an LED light source and 10 nm wide band-pass filters at 405, 450, 520, 550, 589, 632 and 660 nm. Images were spectrally unmixed using absorption spectra of the different chromophores and the background estimated based on the images themselves and aligned with the brightfield images. Pathologist-trained individuals collected exhaustive manual annotations in regions of interest for each marker and HTX using unmixed channels. A convolutional neural network (CNN) was trained using the annotations for each marker and HTX and applied to the WSI. Predicted cell locations were aggregated across the four channels to obtain an exhaustive “super-annotation” co-expression map of the WSI. Super-annotations were then used to train a CNN to detect ER/PR/Ki67/HTX colocalization in the aligned RBG image of the same slide. Predictions from the super-annotation model were compared to ER/PR/Ki67/HTX detection done via HALO (Indica Labs).
Results: Our multi-spectral imaging pipeline allowed for comprehensive annotations of all nuclei positive for at least one of the three markers or HTX. Using our super-annotation model, the count error of nuclei that are positive for HTX and one, two, or three markers was reduced by 7.8%, 6.4%, and 0.03%, respectively, compared to HALO.
Conclusions: We have developed a method to identify colocalized antigens after mIHC using chromogens with spectral overlap. Related work to extend this approach to membranous biomarkers is ongoing and will enable additional novel analyses (e.g., describing the precise spatial organization of PD-L1+ tumor cells relative to CD8+ tumor-infiltrating lymphocytes). The approach that we describe here has great potential to facilitate the utilization of mIHC for the clinical detection of biomarkers in oncology.
References: 1Hacking, SM, et al. Cancers. 2022 14:6439
Citation Format: Waleed Tahir, Yibo Zhang, Jun Zhang, Jacqueline Brosnan-Cashman, Robert Egger, Justin Lee. Machine-learning enabled quantification of colocalized multiplex IHC signals with spectral overlap [abstract]. In: Proceedings of the AACR-NCI-EORTC Virtual International Conference on Molecular Targets and Cancer Therapeutics; 2023 Oct 11-15; Boston, MA. Philadelphia (PA): AACR; Mol Cancer Ther 2023;22(12 Suppl):Abstract nr B008.