PROJECT SUMMARY/ABSTRACT The overall goal of this proposal is to optimize the use of radiomic and genomic data to develop biomarkers which make clinical predictions that change cancer patient management. While the need for such predictive biomarkers is evident across cancer types, we focus our proposal on the particularly prevalent and damaging condition of recurrent, locally-advanced cervical cancer (LACC). Cervical cancer remains the third most common cancer diagnosis of women, and treatment failure for locally-advanced disease is 30-50% following chemoradiation therapy. There is a pressing need to identify patients at risk for treatment failure to allow for personalized treatment including modified chemoradiation regimens, early escalation of therapy, and clinical trial enrollment. To develop radiogenomic biomarkers for LACC recurrence, this proposal addresses three outstanding methodological needs: limited availability of gene expression data for cancer subtypes, noisy and redundant imaging feature data, and lack of disease-informed, interpretable -omics integration, each addressed in its own specific aim. Aim 1 will use generative adversarial networks (GAN) to augment the small gene expression datasets for all high-risk HPV subtypes. Aim 2 will optimize imaging feature selection using a deep convolutional autoencoder (CAE). Aim 3 will integrate radiogenomic features through a structural equation modeling (SEM) approach incorporating HPV-specific oncogenic mechanisms as latent variables. Together, we expect fulfillment of these aims will create an optimized recurrence biomarker which will out- perform other prediction modalities as well as standard-of-care follow-up imaging. Beyond the specific application to HPV-driven malignancies, our proposal will generate novel tools and methods to integrate any high-dimensional radiogenomic data with hypothesis-driven research findings to improve cancer prediction.

PROJECT NARRATIVE Cancer patients treated with standard-of-care chemoradiation therapy commonly experience disease recurrence, and current clinical testing cannot predict which patients will fail therapy or would benefit from personalized treatment. Patient characteristics which predict treatment outcome may be hidden within the expansive genomic and radiomic data rapidly becoming available, but methods to extract and integrate predictive features from this complex data are lacking. In this proposal we develop new deep learning tools to find and contextualize these features from noisy, fragmented, real-world datasets.