Abstract 35: Molecular optimization of phase III trial failed anticancer drugs using target affinity and toxicity-centered multiple properties reinforcement learning

Abstract

Abstract Despite the potential of anticancer drugs in the early stages of drug development, many clinical trials have failed at phase III because of low efficacy and/or high toxicity profiles. To improve such issues, effective lead optimization based on machine learning (ML) is needed. Many ML-based lead optimization models have been proposed, but they are capable of optimizing only a single property of a drug; therefore, multiple models are required and would likely be significantly different from the original drug as processed by multiple models. Our contributions are as follows: 1) We devised a new deep learning architecture based on a long short-term memory on a convolutional neural network for predicting binding affinity score (AS) and absorption, distribution, metabolism, excretion, and toxicity (ADMET) prediction. 2) We adapted the Molecule Deep Q-Networks (MoleDQN) to our new reward combinations so that a single model can consider multiple properties such as quantitative estimate of drug-likeness (QED), AS, synthetic accessibility scores (SAS), and toxicity scores. 3) Binding affinity estimator and ADMET estimator are used to indicate drug efficacy and toxicities. Iniparib, brivanib, and rebimastat, which failed in phase III clinical trials, were used for this analysis, and the top two optimized candidates of each drug with an improved score were selected. Iniparib was developed as a PARP inhibitor, but QED and AS were not predicted as a potential drug in our model (QED, 0.51; SAS, 2.23; AS, 5.02). This may be related to the low efficacy reported in clinical trials. Thus, iniparib was optimized for improved QED and AS to increase efficacy. Optimized candidates OPT-INI-01 (QED, 0.85; SAS, 3.33; AS, 7.77) and OPT-INI-02 (QED, 0.83; SAS, 3.15; AS, 7.93) were qualified with QED higher than 0.8 and demonstrated AS higher than 7.0, which is equivalent to Kd < 100 nM. Brivanib (QED, 0.56; SAS, 3.42; AS, 6.13), which targets VEGFR2, failed in clinical studies with low efficacy and unexpected toxicity (heart toxicity score, hERG, 0.77). To improve the efficacy of brivanib, hERG scores were added to the optimization properties to produce less toxic molecules. BRIV-OPT-01 (QED, 0.61; SAS, 2.80; AS, 7.79, hERG, 0.45) and BRIV-OPT-02 (QED, 0.60; SAS, 3.01; AS, 7.86; hERG, 0.41) were optimized and showed less than 0.5 in hERG score with AS higher than 7.0. Rebimastat (QED, 0.26; SAS, 3.95; AS, 7.44) targeting MMPs did not show satisfactory clinical results due to skin hypersensitivity (skin toxicity score, STS, 0.74). Thus, optimization was performed with focus on the STS and AS. REB-OPT-01 (QED, 0.35; SAS, 4.48; AS, 8.44, STS, 0.49) and REB-OPT-02 (QED, 0.52; SAS, 4.33; AS, 7.80; STS, 0.53) showed less than 0.55 in STS and better overall results than that of rebimastat. In conclusion, this work may improve and accelerate drug development through ML methods that use differential weights for different properties such as efficacy and toxicity, and it may make it easily accessible and less expensive to test new drug candidates in the real world. Citation Format: Sungsoo Park, Yoon Ho Ko, Bora Lee, Bonggun Shin, Bo Ram Beck. Molecular optimization of phase III trial failed anticancer drugs using target affinity and toxicity-centered multiple properties reinforcement learning [abstract]. In: Proceedings of the AACR Special Conference on Advancing Precision Medicine Drug Development: Incorporation of Real-World Data and Other Novel Strategies; Jan 9-12, 2020; San Diego, CA. Philadelphia (PA): AACR; Clin Cancer Res 2020;26(12_Suppl_1):Abstract nr 35.