Janice Goh1, Junqi Chen1, Grigory Nesvijevski1
1Bioinformatics Institute (BII), Agency for Science, Technology and Research (A*STAR)
Objectives: Cell line derived xenograft (CDX) mouse models, established by engrafting commercial cell lines into an immunodeficient mouse model, are widely used for validation of drug efficacy prior to clinical trials. However, such studies are resource intensive. In vitro assays such as 2D cell cultures, on the other hand, can be done in a low cost, high throughput fashion, but its potency read outs are often insufficient to inform and design subsequent animal experiments. Since the advent of large scale cell line screens such as Genomics of Drug Sensitivity in Cancer (GDSC)¹, multiple efforts using machine learning approaches have been made to predict in vivo drug efficacy from in vitro data. However, these efforts have yet to consider a more mechanistic understanding of variables that are known to affect the output such as the drug dose, treatment and experimental duration. The use of pharmacokinetic (PK) indices which combines in vitro drug efficacy with pharmacokinetic profiles of the tested drugs can help us account for these dynamic processes in vivo. Therefore, using these publicly accessible databases and information, we thus aim to come up with a pan cancer translational PK index framework that allows us to predict in vivo drug efficacy across CDX models. Methods: Mouse PK and CDX data was first compiled from literature using plotdigitizer. In vitro cell line potency data in the form of inhibitory concentration to achieve 50% max efficacy (IC50) was taken from GDSC 1 and 2¹ studies for the respective matched cell lines used to establish CDX models. In vivo end points were calculated as tumor growth inhibition (TGI) of control vs treated over the total duration of the experiment. One and two compartment PK models were built from mouse PK data and validated using visual predictive check. Subsequently, mouse PK was simulated according to treatment regimens used in CDX studies and combined with in vitro cell line data to calculate PK indices. Area under the curve (AUC) as a ratio to IC50 (AUC/IC50), AUC above IC50 (AUC>IC50), Area under the effect curve (AUEC)² and average drug concentration over the whole experimental duration (Cavg/IC50) were tested as PK indices alongside IC50. A 80/20 training/testing split was then used to train and validate a four parameter log logistic model and prediction accuracy of each PK index was tested using bootstrap of 1000 iterations. Results: A total of 16 CDX studies with 17 unique drugs were identified from literature. Of which, 6 drugs had matching cell line and CDX data, giving us a total of 17 unique data points. Cavg/IC50 gave the lowest mean square error (MSE) of all the methods (mean 1.63*10^-4 +/- 3.14*10^-5 standard deviation (SD) training, 3.50*10^-4 +/- 1.94*10^-4 testing). AUC/IC50 performed next best (1.87*10^-4 +/- 4.59*10^-5 training, 3.99*10^-4 +/- 2.52*10^-4 testing. AUC>IC50 was the worst performing where no trend line could be reasonably fitted. Using Cavg/IC50, we then proposed a minimum threshold ratio for in vivo efficacy of 50% of max observed TGI as 485.8 +/-258.7 standard error. Conclusions: Using in vitro data, mouse PK and CDX data compiled from literature, we were able to derive PK indices and test their predictivity against observed tumor growth inhibition values from preclinical CDX models. We found that Cavg/IC50 gave the best predictiveness and this will be further validated with more drug classes and cell line types in future work. This in vitro to in vivo prediction pipeline is promising in helping us to better prioritize novel drugs and appropriate doses for testing in animals, thereby accelerating drug discovery and saving resources.
1. Yang, W. et al. Genomics of Drug Sensitivity in Cancer (GDSC): a resource for therapeutic biomarker discovery in cancer cells. Nucleic Acids Res 41, D955–D961 (2013). 2. Chen, C., Lavezzi, S. M. & Iavarone, L. The area under the effect curve as an efficacy determinant for anti-infectives. CPT Pharmacometrics Syst Pharmacol 11, 1029–1044 (2022).
Reference: PAGE 33 (2025) Abstr 11448 [www.page-meeting.org/?abstract=11448]
Poster: Drug/Disease Modelling - Oncology