Abstract 1928: Prediction of neoadjuvant treatment outcomes with multimodal data integration in breast cancer

Neoadjuvant chemotherapy (NAC) is the standard of care for selected patients with high-risk early-stage breast cancer with pathologic complete response (pCR) being the most prominent predictor of favorable outcomes. Here, we sought to study the predictive capacity of integrating orthogonal diagnostic measures on predicting pCR relative to standard clinicopathologic features.

We developed a computational model integrating radiology and pathology images, and tumor genomics to automatically predict pCR from multimodal data. We present an interim analysis on a cohort of 957 patients with at least one available pre-NAC data modality. The baseline AUC for pCR prediction by a trained and tested logistic regression model on 857 patients using standard clinicopathologic features including receptor subtype, demographic information, and stage was 0.77. MR images were input into a convolutional neural network (CNN) and a radiomics model.

The trained CNN and radiomics models using selected images of 576 patients with pre-NAC MR images achieved AUCs of 0.65 and 0.60 on 164 hold-out test cases, respectively.

We trained a multiple instance learning-based weakly supervised learning (MIL-WSL) model using 537,762 extracted tiles from whole slide images (WSI) of digital histopathology scans from 522 patients. The MIL-WSL model achieved AUC of 0.63 for pCR prediction on a hold-out test set of pre-NAC biopsies from 239 patients. A feature based classifier trained on 76 cases using tumor genomic features such as mutational burden, microsatellite instability, fraction genome altered, ploidy, purity, mutation and copy number alterations in selected genes achieved an AUC of 0.72 on 83 hold-out test cases.

We then combined unimodal radiology, histopathology, and genomic predictions in a deterministic manner. This multimodal combination on an independent 68-patient test set achieved an AUC of 0.84, indicating increased power to resolve pCR than any modality alone, and over clinicopathologic baseline.

Together, we present approaches to train models end-to-end using tensor fusion networks and attention-gating combined with MIL. Automated multimodal methods are here shown to improve prediction over established clinical parameters alone, motivating our ongoing efforts to refine and improve the model so as to achieve higher levels of efficiency. We anticipate these interim results will be further improved through refinement of input features and increasing the number of patients included in the final validation cohort.

Citation Format: Pegah Khosravi, Elizabeth J. Sutton, Justin Jee, Timothy Dalfonso, Christopher J. Fong, Doori Rose, Edaise M. Da Silva, Armaan Kohli, David Joon Ho, Mehnaj S. Ahmed, Danny Martinez, Anika Begum, Elizabeth Zakszewski, Andrew Aukerman, Yanis Tazi, Katja Pinker-Domenig, Sarah Eskreis-Winkler, Atif J. Khan, Edi Brogi, Elizabeth Morris, Sarat Chandarlapaty, George Plitas, Simon Powell, Monica Morrow, Larry Norton, Jianjiong Gao, Mark Robson, Hong Zhang, Sohrab Shah, Pedram Razavi, MSK-MIND Consortium. Prediction of neoadjuvant treatment outcomes with multimodal data integration in breast cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 1928.

Abstract LB149: Luna Pathology: An integrated open source platform for computational pathology research

Pathologic review of tissue samples is a crucial step in cancer diagnosis and treatment planning. In recent years, quantitative analysis, including artificial intelligence (AI) techniques, have been applied to facilitate the evaluation of histopathologic images. Research in computational pathology comes with numerous engineering challenges: from management of whole slide images (WSIs) and their metadata, collection of expert annotations, data manipulation and feature extraction, to AI model training and evaluation of the results. These challenges span multiple departments and sets of expertise; there is a lack of an efficient system that helps manage and process histologic images and supports AI analysis. In particular, computational oncology requires significant compute resources. The need for an integrated platform with systematic processes to manage the job parameters and complex workflows that interface with multidisciplinary teams is critical for successful experiments and reproducible research. To address these challenges, we developed Luna Pathology, an open-source library and platform for computational pathology research. With flexibility, scalability, and FAIR data principles in mind, Luna Pathology was designed to provide 1) reusable workflows and modular tools for end-to-end pathology analysis 2) FAIR datasets for reproducible research and 3) a collaborative open-source research platform. Namely, Luna Pathology features support data management, annotation support, image processing, image and annotation visualization, feature extraction, and human-in-the-loop machine learning. A description of the Luna Pathology platform, its supporting computational infrastructure, and its integration within the standard computational pathology will be presented. Luna Pathology meets data processing and management needs, as well data analysis experiments and reproducible AI research on a flexible data platform. Its workflows have been deployed in colorectal, neuroblastoma and ovarian cancer projects at Memorial Sloan Kettering Cancer Center (MSK) to derive new insights from ~10,000 WSIs. We additionally present a view of the Luna Pathology platform as part of the institutional effort to bridge computational research and clinical care at MSK. The integrated open platform brings together a multidisciplinary team of pathologists, computational biologists, and bioinformatics engineers to more effectively collaborate on iterative large-scale computational pathology research.

Citation Format: Doori Rose, Andrew Aukerman, Druv Patel, Arfath Pasha, Armaan Kohli, Ignacio Vázquez-García, Kevin Boehm, Rami Vanguri, Matthew Shabet, Vassiliki Mancoridis, Elizabeth Zakszewski, Benjamin Gross, Christopher Fong, Pegah Khosravi, Sohrab Shah, JianJiong Gao. Luna Pathology: An integrated open source platform for computational pathology research [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr LB149.

Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer

Patients with high-grade serous ovarian cancer suffer poor prognosis and variable response to treatment. Known prognostic factors for this disease include homologous recombination deficiency status, age, pathological stage and residual disease status after debulking surgery. Recent work has highlighted important prognostic information captured in computed tomography and histopathological specimens, which can be exploited through machine learning. However, little is known about the capacity of combining features from these disparate sources to improve prediction of treatment response. Here, we assembled a multimodal dataset of 444 patients with primarily late-stage high-grade serous ovarian cancer and discovered quantitative features, such as tumor nuclear size on staining with hematoxylin and eosin and omental texture on contrast-enhanced computed tomography, associated with prognosis. We found that these features contributed complementary prognostic information relative to one another and clinicogenomic features. By fusing histopathological, radiologic and clinicogenomic machine-learning models, we demonstrate a promising path toward improved risk stratification of patients with cancer through multimodal data integration.

Kaiso (ZBTB33) subcellular partitioning functionally links LC3A/B, the tumor microenvironment, and breast cancer survival

The use of digital pathology for the histomorphologic profiling of pathological specimens is expanding the precision and specificity of quantitative tissue analysis at an unprecedented scale; thus, enabling the discovery of new and functionally relevant histological features of both predictive and prognostic significance. In this study, we apply quantitative automated image processing and computational methods to profile the subcellular distribution of the multi-functional transcriptional regulator, Kaiso (ZBTB33), in the tumors of a large racially diverse breast cancer cohort from a designated health disparities region in the United States. Multiplex multivariate analysis of the association of Kaiso’s subcellular distribution with other breast cancer biomarkers reveals novel functional and predictive linkages between Kaiso and the autophagy-related proteins, LC3A/B, that are associated with features of the tumor immune microenvironment, survival, and race. These findings identify effective modalities of Kaiso biomarker assessment and uncover unanticipated insights into Kaiso’s role in breast cancer progression.

Through automated image analysis, Singhal et al quantify nuclear versus cytoplasmic distribution of the Kaiso transcription factor in breast cancer patient tissue. They find that Kaiso distribution correlates with breast cancer subtype and overall survival, and discover a link between cytoplasmic Kaiso and autophagy marker LC3.

Radiomics of MRI for pretreatment prediction of pathologic complete response, tumor regression grade, and neoadjuvant rectal score in patients with locally advanced rectal cancer undergoing neoadjuvant chemoradiation: an international multicenter study

OBJECTIVE

To investigate whether pretreatment MRI-based radiomics of locally advanced rectal cancer (LARC) and/or the surrounding mesorectal compartment (MC) can predict pathologic complete response (pCR), neoadjuvant rectal (NAR) score, and tumor regression grade (TRG).

METHODS

One hundred thirty-two consecutive patients with LARC who underwent neoadjuvant chemoradiation and total mesorectal excision (TME) were retrospectively collected from 2 centers in the USA and Italy. The primary tumor and surrounding MC were segmented on the best available T2-weighted sequence (axial, coronal, or sagittal). Three thousand one hundred ninety radiomic features were extracted using a python package. The most salient radiomic features as well as MRI parameter and clinical-based features were selected using recursive feature elimination. A logistic regression classifier was built to distinguish between any 2 binned categories in the considered endpoints: pCR, NAR, and TRG. Repeated k-fold validation was performed and AUCs calculated.

RESULTS

There were 24, 87, and 21 T4, T3, and T2 LARCs, respectively (median age 63 years, 32 to 86). For NAR and TRG, the best classification performance was obtained using both the tumor and MC segmentations. The AUCs for classifying NAR 0 versus 2, pCR, and TRG 0/1 versus 2/3 were 0.66 (95% CI, 0.60-0.71), 0.80 (95% CI, 0.74-0.85), and 0.80 (95% CI, 0.77-0.82), respectively.

CONCLUSION

Radiomics of pretreatment MRIs can predict pCR, TRG, and NAR score in patients with LARC undergoing neoadjuvant treatment and TME with moderate accuracy despite extremely heterogenous image data. Both the tumor and MC contain important prognostic information.

KEY POINTS

• Machine learning of rectal cancer on images from the pretreatment MRI can predict important patient outcomes with moderate accuracy. • The tumor and the tissue around it both contain important prognostic information.

Subcellular partitioning of Kaiso (ZBTB33) as a biomarker to predict overall breast cancer survival

3534

Background: The epigenetic transcriptional regulator, Kaiso (ZBTB33) has been identified as a member of the C2H2 zinc finger proteins containing a BTB/POZ -zinc finger family of transcription factors that are implicated in development of cancer. Although, our understanding of clinical relevance of subcellular distribution (cytoplasmic/nuclear) Kaiso in the growth and survival of human Breast cancer (BC) is limited. Methods: We examined a cohort of 555 BC patients who underwent surgery for their primary BC in Greenville, NC using AI and SM approach. Results: The sub-classification BC shows, cytoplasmic Kaiso is differentially enriched in ER- BC (p=0.001) compared nuclear Kaiso (p=0.8) and is significantly enriched in the more aggressive classes LumB (p=0.0017), HER2+ (p=0.05) and TNBC (p=6.1e-07) with respect to LumA BC patients. Additionally, the survival analysis of different compartments of Kaiso demonstrates that high cytoplasmic Kaiso (HR = 16.29 (7.6 – 34.8), p = 5.5e−13) is much more predictive of poor survival compared to nuclear Kaiso (HR = 2.83 (2.02 – 3.8), p = 6.1e−11). At gene expression level, ZBTB33 mRNA levels do not correlate with either nuclear (Spearman correlation: -0.03157, p= 0.7267) or cytoplasmic levels (Spearman correlation: -0.03526, p= 0.6962) of Kaiso. Surprisingly, ZBTB33 mRNA abundance is predictive of poor overall BC survival as demonstrated in two independent publicly available BC cohorts Metabric (HR = 2.14 (1.49 − 3.08), p = 2.7e−05) and Gyorffy B et al. (HR = 1.81 (1.55 − 2.12), p = 2.5e−14). Nuclear and cytoplasmic levels of Kaiso do not show significant differences based on race p=0.27 and p=0.1 respectively. Conclusions: Our data suggest subcellular distribution of high Kaiso is associated with poor prognosis of BC survival and subcellular localizations of Kaiso may play differential biological roles in BC prognosis.