Leveraging Multimodal Signatures to Personalize Rectal Cancer Radiotherapy Choice - ABSTRACT Colorectal cancer, with rectal cancer comprising approximately 1/3 of all colorectal cases, is the third most common cancer diagnosed and the second most common cause of cancer-related death in the United States. In 2023, there was an estimated 46, 050 new cases of rectal cancer, and almost 27,000 deaths, or about 60% of the patients dying from this disease. Additionally, there is an alarming increase in the diagnosis of rectal cancer in younger patients (<50 years old). Until recently, the established standard of care for locally advanced rectal cancer (LARC) involved pre-operative long course chemoradiation (LCRT) followed by total mesorectal excision (TME) and adjuvant chemotherapy. However, not all patients derive equal benefit from this approach, evidenced by a relatively low complete pathologic response (pCR) rate to LCRT alone, ranging from 15-27%. More recent strategies have aimed to reduce recurrence rates by applying total neoadjuvant therapy (TNT) to increase the rates of complete clinical response (cCR), allowing patients to avoid surgery (non-operative management, NOM). To increase the complete response (CR, including cCR and pCR) rate, which correlates with better outcomes, and to explore a more cost-effective approach to NOM, neoadjuvant short-course radiation (SCRT) followed by chemotherapy (TNT) is emerging as a promising strategy, offering greater patient convenience, cost- effectiveness, and efficient resource utilization. Studies have demonstrated that this regimen can achieve a CR rate twice as high as that achieved with LCRT alone. However, SCRT has been associated with a higher failure rate in some instances, supporting the notion that there is no universal solution for all LARC patients. Therefore, there is an urgent need to develop, on a per-individual basis, a reliable method to predict whether LCRT or SCRT will offer the highest likelihood of achieving CR, enabling NOM as well as the highest cure rates, for LARC patients. Our goal is to develop a powerful and clinically ready signature, applying both imaging and a unique class of genetic biomarkers, that will allow physicians and their patients to identify the best personalized treatment approach in LARC, either LCRT or SCRT, as measured by achieving a CR. To achieve this goal, we will apply insights into LARC characterization derived from medical imaging, as well as apply novel patient-specific germline genetic biomarkers. To this end, we will build an interpretable radiomics pipeline, consisting of a CNN feature extractor on multi-modal images, a superior multi-objective feature selection algorithm and a model interpreter. In addition, we will develop predictive genetic signatures of response to LCRT versus SCRT in LARC using a large panel of microRNA-based germline biomarkers we have previously shown predicting radiation response. Finally, we will develop tiered fusion models that combine the image and germline signatures to predict the response likelihood, estimate treatment effects, and investigate individualized treatment rules to suggest the treatment type with the highest response probability, assisting physicians and patients in treatment selection.
