Abstract 1680: Assessment of statistical power in one mouse per treatment design for preclinical anticancer agent PDX large scale drug screens

Background: Patient-derived tumor xenograft (PDX) models are increasingly used to evaluate the effectiveness of preclinical anticancer agents. To test many anticancer agents simultaneously a large scale drug screen can be utilized with a one mouse per treatment (1 × 1) design. With this approach, typically only treatments found to be effective are used in further studies. We investigated the rates of “false negatives” where drugs are being incorrectly found ineffective in initial screens and potentially not considered again. We focused on modifiable parameters which could increase the statistical power (rate of true positives) of this design based on recent PDX lung experiments. Methods: We used PDX drug screen studies from our lab as a reference for tumor growth rate and mouse variation. Studies included 43 non-small cell lung cancer PDX experiments testing a total of 14 different anti-cancer agents. Each experiment included on average 6 replicates per group (531 total mice), from 25 unique PDX models. In each experiment PDX models were established from patient tumor fragments that were implanted at the flank of immunodeficient mouse hosts. Xenograft tumor fragments were expanded into mouse replicates to test with anti-cancer agents. The standard protocol was treatment with agents at doses with reported in vivo antitumor effects. Tumor size was measured twice weekly. This presented us a distribution of treatment effect sizes, and across mouse variation. We assessed the statistical power of the 1 × 1 design under different settings to determine if/when the design would be appropriate. Settings included; modifying the treatment effect size, mouse variation, and follow up schedule. The estimated treatment effect sizes were divided at the tertiles which we refer to as small, medium, and large. Mouse variation was assessed at the median value (average variation) and at the first quartile (small variation). We assumed a typical measurement schedule to be twice a week for four weeks, and a more intense schedule as three times a week. We used a relaxed 0.2 alpha level when calculating the power rates. Results: Treatments with a large effect have a 98% statistical power under the assumption of average variance and typical measurement schedule. For medium and small treatment effects the statistical power are 67% and 41% respectively. A more intense measurement schedule and small variation increases the statistical power to 99%, 70% and 43%, depending on the effect size. Conclusion: In contrast to large effect sizes which can be detected easily under a 1 × 1 design, the medium and small effect sizes have a large chance of being rejected. A treatment with a median effect under optimal circumstance will have a 30% change of being rejected, and a small treatment effect will have a 57% chance of being rejected. In conclusion a 1 × 1 design is appropriate only when there is a belief the that treatment is very effective. Citation Format: Jessica Weiss, Nhu-An Pham, Melania Pintilie, Ming Li, Ming Tsao, Geoffrey Liu, Wei Xu. Assessment of statistical power in one mouse per treatment design for preclinical anticancer agent PDX large scale drug screens [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 1680.