Artificial Intelligence to PRedicting Efficacy of Drugs and radiation in Head and Neck Cancer through Digital pathology and radiology Imaging Combined to Tumor Genomics (AI-PREDICT)

Project goals:

To develop a machine-learning model that predicts the probability of disease control based on a multi-omics approach. This will be achieved through the following specific aims:

1. Harmonize standard-care clinical data with radiomics, pathomics and genomics data.
2. Establish AI-based multi-omic predictive models in a retrospective cohort of LA-HNC.
3. Establish AI-based multi-omic predictive models in prospective cohorts of LA-HNC from on-going investigator-initiated clinical trials.

Summary:

Context: The efficacy of radiotherapy and systemic therapy in locally advanced head and neck cancer (LA-HNC) varies widely across tumors, and patient-specific clinical, radiological and biological factors are thought to drive these individual treatment responses. Because our current understanding of these patient-specific factors is poor, the current approach in HNC treatment is population-based, whereby all patients are prescribed standard radiation doses and systemic therapy regimen. Selection of patients for treatment de-intensification or intensification is limited by the lack of biomarkers guiding individualisation of treatment. The use of multi-modality functional imaging in HNC, along with the expansion of our current on-going head and neck biobank, is an opportunity for detailed multi-omic HNC tumor characterization, including rare tumors of the head and neck. We hypothesize that novel artificial intelligence (AI) methods integrating multi-omic data would uncover non-invasive biomarkers that could guide individualized decision- making, based on expected outcomes.

Research plan: This study will first explore retrospective data from 209 patients with LA-HNC treated with radiotherapy currently in our HNC biobank. Retrospective data includes: 1- standard clinical data, longitudinal cancer control outcomes (clinical), 2- fresh frozen tissues, formalin fixed paraffin embedded tissue, germline samples, plasma and peripheral blood mononuclear cells (genomics and pathomics); 3- pre-treatment multi-modality imaging and daily volumetric imaging during radiation (radiomics). The prospective component of the study will involve collection of samples from at least 3 on-going investigator-initiated trials: SHORT-OPC (NCT04178174), ORATOR2 (NCT03210103), and VOCAL (NCT03759431). These studies will additionally involve high quality prospective clinical data acquisition (including quality of life) and serial liquid biopsies for dynamic assessment of circulating tumor cell, circulating HPV DNA, exosomes and microRNA.

Cyberstructures will be merged through the creation of data lakes combining data from various sources (electronic charts, picture archiving and communication system, biobank information management system, digitized histological slides). Machine-learning algorithms including supervised, semi-supervised and unsupervised neural networks will be used for features selection and predictive performance of cancer control. Prediction models will be constructed by combining genomics, radiomics and pathomics to known clinical attributes for prediction of recurrence both as binary classification (recurrence vs. no recurrence) and time-to-event endpoint.

MOHCC value: The clinical integration of multi-omic predictive algorithms in HNC could transform head and neck oncological care delivery in a permanent and disruptive by allowing individualized data-driven decision- making, based on expected outcomes. Ultimately this would translate into accelerating discoveries in the field of oncology through elicitation of novel correlations. The clinical trials would lead to inclusion of high quality prospective biobanking from at least 300 new patients, with spatial and temporal evaluation of response by sequential blood samples and radiological imaging.