Significance of scores generated by a cancer therapy matching engine for patient outcomes.

Abstract

e15099 Background: Oncologists need effective precision medicine tools to navigate the myriad of therapeutic options to best address each patient’s unique tumor profile. Described here is a digital solution, which analyzes tumor data (NGS, other), scores and ranks known and novel matching combinations of cancer drugs with molecular-level precision for decision support. Methods: All data was acquired through the published supplement of the I-PREDICT study (NCT 02534675). Patients with available progression-free survival (PFS) and sequencing data who received <3 matched non-experimental cancer drugs in combination were evaluated (n=77). To determine each patient’s score (represented as %), the system used expert-curated content integrated within an artificial intelligence (AI) reasoning framework that computed how well the therapy each patient received matched their tumor’s genomic profile. The dataset was binarized at 36 possible thresholds, which split the data into higher vs. lower score bins. Relationship with PFS and the degree of separation between bins were evaluated with Kaplan-Meier plots and assessed in terms of p-value, hazard ratios and related confidence intervals (Table). To prevent sampling bias at high and low thresholding extremes of this dataset, a limit was imposed of >25 patients per bin. Results: Significant separation between high and low scoring bins was observed in >73% of evaluable thresholds. The mean p-value was 0.044 (range, 0.011-0.13). The hazard ratio was consistently ̃2 (mean, 1.81, range, 1.50-2.08). A similar level of statistical significance was observed for all thresholds up to and including bracketing of 60% (Table). Despite the small size of this dataset and the disparity between sample counts per bin seen at the end of the range, significant p-values of <0.05 were achieved. Ex., a 68-year-old female with an ampullary adenocarcinoma harboring ERBB2, CDK6 and other alterations, received a combination of pertuzumab and trastuzumab scoring 56%. The PFS was 745+ days. However, had palbociclib, ribociclib, or abemaciclib been added to this combination, the resulting 3-drug options would have scored even higher due to also targeting CDK6. Conclusions: Our study indicates that therapies with higher scores were predictive of better PFS, while lower scores were predictive of worse PFS. The results will be validated on larger datasets but are already demonstrative of the utility of this system in providing digital guidance to oncologists in selecting the most suitable treatment. [Table: see text]