Enhancing precision medicine for brain metastases using interpretable and integrative deep learning

Massachusetts General Hospital, Boston MA

Investigators

Abstract

Project Summary: The goal of this Mentored Career Development Award is to facilitate the candidate’s transition to independence as a translational oncologist using deep learning (DL) to study therapeutic resistance in solid tumor oncology. The Candidate is an oncologist at the Massachusetts General Hospital (MGH) with a background in cancer genomics and functional imaging. This K08 Award will allow the Candidate to build upon his burgeoning experience in supervised DL. However, as an MD-only scientist, the Candidate has reached the limits of what he can accomplish as a self-taught programmer. Therefore, to further develop his DL skillsets, the Candidate will be mentored by world leaders in DL (John Guttag; Jayashree Kalpathy-Cramer). This computational expertise is complemented by the Candidate’s scientific advisors: Drs. Gerstner, Iafrate, and Rosen, who are experts in early-phase clinical trials, molecular pathology, and radiology applications of DL, respectively. With this mentorship team, the Candidate will be uniquely positioned to succeed as a physician- scientist using DL to analyze and integrate clinically acquired data to develop personalized treatment paradigms for cancer patients. This Award will be further supported by the unparalleled institutional support and environment offered by the MGH, Broad Institute of Harvard / MIT, and Martinos Center for Biomedical Imaging. Here, the Candidate seeks to build upon a seminal study which demonstrated intracranial efficacy of checkpoint inhibition (ICI) for brain metastases of diverse tumor types (Brastianos & Kim et al., Nature Med, 2023). Despite functional imaging and comprehensive genomic analysis, the Candidate and other groups have not identified a robust predictive biomarker for ICI response. Therefore, given the ability of DL to extract and combine complementary information across high-dimensional modalities, the objective of this Award is to use DL to integrate radiology, histopathology (H&E), and clinico-genomic data to predict ICI response in BM. We will use a multi-national dataset comprised of brain MRI, H&E from resected BM tissues, and clinico-genomic data from 2100 BM patients treated with ICI. The first aim will develop a DL model, using pre-treatment brain MRI, to predict ICI efficacy in BM. The second aim will develop a separate DL model, using H&E, to predict ICI efficacy in BM. The third aim will develop a multi-modal fusion model, integrating MRI, H&E, and clinico-genomic data, to predict ICI efficacy. Performance of each model (MRI-based model vs. H&E-based model vs. multi- modal model) will be compared to current clinical biomarkers (e.g., PD-L1 expression) to determine the most accurate technique. In addition, the Candidate will experiment with state-of-the-art pretraining and data augmentation strategies to maximize model generalizability. These results will lay the groundwork to design a future R01-funded effort to accrue prospectively collected data to validate our multimodal fusion strategy. Through identifying optimal methods for multimodal data fusion, our work supports and highlights NIBIB’s mission of using machine intelligence tools in the clinic to augment clinical decision making.