Abstract 396: Beyond thresholds in precision oncology: Use of probability of benefit as function of continuous biomarker levels leads to better patient care

The current threshold-based precision medicine paradigm stems in part from the ‘all-or-nothing9 nature of regulatory labeling: a lung cancer patient can access pembrolizumab on-label only if tumor mutational burden (TMB) is at least 10 mutations/megabase, and crizotinib only if at least 15% of their tumor cells have ALK rearrangements (Pembrolizumab and Crizotinib package inserts). However, the drug-centric question - defining the patients most likely to respond to the drug – does not necessarily optimize patient care: in this case, choosing the best drug for a patient with nonzero levels of TMB and ALK. Furthermore, the current paradigm requires the drug sponsor to choose a threshold between a low value, to provide access to as many potentially benefitting patients as possible, and a high value, to maximize probability of success of the pivotal trial. We hypothesize that modeling probability of benefit (PoB) from each drug as a function of continuous biomarker levels (CBLs), and then giving the patient the treatment with the larger PoB not only eliminates the sponsor threshold selection dilemma, it provides better patient care. To test this hypothesis, we developed a simulation study. Given two drug/biomarker pairs, DX/BX and DY/BY, a ‘training9 population of 100 virtual patients responsive to DX, DY, or neither DX nor DY was generated using a ‘ground truth9 model of PoB from DX or DY as a function of patient CBLs of BX and BY, which were drawn from a prespecified random distribution. Logistic regression models were trained to the patients9 benefit outcomes to quantify PoBs as functions of CBLs. The threshold paradigm was simulated by choosing an ‘altruistic9 optimal biomarker threshold (OBT) for each drug which minimizes the net error rate (NER = False Positive + False Negative rate). Then a ‘test9 population of 1000 patients was generated using the same ground truth model used to generate the training population. We then applied the trained PoB model to give each virtual test patient the drug with the highest PoB based on CBLs and applied the threshold method to give each virtual test patient a drug only if the corresponding CBL > OBT. Finally, we compared the NER between PoB and OBT methods. NERs were estimated for three methods: OBT yielded a NER of 20.4%. Maximum univariate PoB, where PoB(DX) is a function of CBL BX alone and PoB(DY) is a function of CBL BY alone, yielded a NER of 14.8%. Maximum bivariate PoB, where PoB(DX) and PoB(DY) are functions of both CBLs BX and BY, yielded a NER of 10.7%. Our analysis predicts that of 100 patients treated according to the OBT method, 20 would either not benefit from the drug they were given or did not get a drug that would have benefited them. In contrast only 10-15 patients would be incorrectly treated using the PoB method, representing a 25-50% lower error rate than OBT. We believe the PoB method warrants consideration in the future of precision oncology. Citation Format: Cameron McBride, Dean Bottino. Beyond thresholds in precision oncology: Use of probability of benefit as function of continuous biomarker levels leads to better patient care [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 396.