A method for assessing the predictivity of drug-specific biomarkers before interventional clinical data is available.

Abstract

e15170 Background: A key goal in precision oncology is to develop predictive biomarkers that can identify patients likely to benefit from a given treatment. The main hurdle for generating such biomarkers is the availability of datasets that match predictive features for each patient (clinical, omics etc.) with treatment and individual outcome data. Such datasets are hard to obtain for approved drugs, and nonexistent for drugs in development. In contrast, there is ample data matching predictive features with survival data, but methods that use such data to develop and validate drug-specific predictive biomarkers are lacking. Here, we introduce a method termed the Surrogate Clinical Outcome Test (SCOT), which assesses, using non-interventional data, whether a drug-specific biomarker will be beneficial in clinical settings. Together with our previously published ENLIGHT platform, this enables early development and validation of predictive biomarkers. Methods: The SCOT test assumes that a patient population that exhibits low expression of the drug target(s) can serve as a surrogate for patients receiving the investigated drug, in the sense that survival dependency on the biomarker score would be similar to that of patients with the same features who actually receive the drug. Hence, a good predictive marker will be characterized by a negative hazard ratio in this population. To ensure that the marker is drug-specific and not merely prognostic, cases for which the same is true also in the population with high target expression should not pass the test. We studied 6 targeted drugs for which interventional clinical data is available. We used SCOT to classify two known prognostic biomarkers, as well as drug-specific transcriptomic biomarkers developed by our ENLIGHT algorithm, and compared the classifications to marker predictivities on the interventional clinical datasets. Results: Of the 6 drug-specific biomarkers developed by ENLIGHT, SCOT classified 4 as predictive of response to the respective drug, one as prognostic, and one as neither. The classification of 4 of these 6 biomarkers was confirmed by the interventional data (67% accuracy). The two prognostic markers - KI67 and Proliferation Index - were identified by SCOT as such. That is, SCOT found an association between these two markers and survival, regardless of target expression, for targets of the 6 drugs tested. Conclusions: Our work presents a novel method to test predictive markers for virtually any targeted cancer drug using big non-interventional clinical data. A key feature of this method is that it offers an independent validation of predictive markers for drugs, based on human clinical data, even before the drug enters clinical development. Combined with our ENLIGHT platform for generating predictive biomarkers in the absence of direct response data, this creates a powerful pipeline for early biomarker development and validation.