Prospective validation of a radiomics-based multi-modal predictive model for metastatic non-small cell lung cancer patients treated with PD-1 immunotherapy

Summary PD-(L)1 checkpoint inhibitors are an effective class of oncology therapeutics, where some patients with non-small cell lung, bladder cancer and melanoma demonstrate durable responses and significant increases in overall survival. This form of therapy, however, is accompanied by several significant disadvantages including; unpredictable and low activity (15-20% response-

rates) for most patients, high therapy cost ($150,000 per year) and serious immune-related adverse events that often result in hospitalization or stoppage of therapy. There have been a variety of approaches attempting to accurately predict response to PD-(L)1 immunotherapy such as PD-L1 immunohistochemistry staining, tissue or liquid-biopsy based genomic signatures such

as TMB (tumor mutational burden), and multiplex immunohistochemistry / immunofluorescence. None of these approaches have yielded results that materially improve patient selection. Onc.AI is developing the first universal radiomics-based multi-modal biomarker solution to predict response to PD-(L)1 immunotherapy. By leveraging novel imaging biomarkers in combination

with proteomic, genomic and clinical features, all tumors comprising a patient’s tumor burden can be interrogated through the application of advanced Machine Learning and computer vision techniques on CT scans alongside routine lab tests and patient characteristics. This approach avoids the shortcomings of tissue or blood-based biomarkers, which are not able to accurately

capture and account for variable intra/inter tumor heterogeneity, diversity and immunogenicity that characterizes metastatic cancer. Achieving accurate prediction of PD-(L)1 ICI therapy response will enable a host of high-impact clinical applications such as -- identification of hyper- progressing patients (patients where tumor growth accelerates as a result of PD-(L)1 ICI);

improved selection of patients (predicted non-responders) who would benefit from chemotherapy or additional immunotherapy added to a PD-(L)1 ICI backbone; appropriately de-escalating combination therapy to minimize toxicity for predicted responders (switching patient from PD- (L)1/chemo to PD-(L)1 monotherapy); and informing continuation of PD-(L)1 therapy at the 12 to

24 month mark based on radiomic signatures more advanced than tumor shrinkage. Each of these applications has the potential to either significantly improve patient outcomes and/or optimize the health-economic profile of PD-(L)1 treatment. Onc.AI’s proposal has three specific aims: Aim 1: Validate predictive models in a multi-institutional prospective clinical study.

We will validate the efficacy of radiomics-based multi-modal models in predicting response to PD-1/PD-L1 ICI therapy in EGFR/ALK mutation-negative stage IV NSCLC patients and in adjudicating concurrent chemotherapy. Aim 2: Evaluate key performance characteristics and biological basis of radiomics-based

predictive models. We will evaluate our production candidate models on secondary outcome measures (duration of response, progression-free survival, time-to-progression, overall survival) and investigate associations with key imaging, demographic, and molecular pathways, using prospectively collected harmonized data from Aim 1.

Aim 3: Enhance existing models and develop new models to support new clinical use cases. Independently from the models validated in Aim 1, we will evaluate feasibility of incorporating tumor mutation burden (TMB) into our predictive model using prospectively collected data and investigate a novel modeling approach and clinical use case - early response

assessment using serial imaging and delta radiomics.