JANUS: Parallel Tempered Genetic Algorithm Guided by Deep Neural Networks for Inverse Molecular Design

Abstract

Inverse molecular design, i.e., designing molecules with specific target properties, can be posed as an optimization problem. High-dimensional optimization tasks in the natural sciences are commonly tackled via population-based metaheuristic optimization algorithms such as evolutionary algorithms. However, expensive property evaluation, which is often required, can limit the widespread use of such approaches as the associated cost can become prohibitive. Herein, we present JANUS, a genetic algorithm that is inspired by parallel tempering. It propagates two populations, one for exploration and another for exploitation, improving optimization by reducing expensive property evaluations. Additionally, JANUS is augmented by a deep neural network that approximates molecular properties via active learning for enhanced sampling of the chemical space. Our method uses the SELFIES molecular representation and the STONED algorithm for the efficient generation of structures, and outperforms other generative models in common inverse molecular design tasks achieving state-of-the-art performance.Preprint. Under review.

A molecular generative model with genetic algorithm and tree search for cancer samples

“…GA-based generative models showed good performance in previous studies, but required large searching time [27,26]. In our experiment, we showed that MCTS is helpful to reduce the search time, and ST and GA contribute to increasing the reward and both drug-likeness and uniqueness, respectively.…”

“…7. JANUS-C [27] was compared, which uses SELFIES [51], a trainable classifier, and GA, but not a deep generative model. Although [51] does not use a deep generative model, they train a DNN-based classifier during iterations, which is designed to distinguish high potential molecules for achieving a high reward.…”

“…Therefore, we compared each model's performance with the limited number of samplings. GA-based methods, GEGL and JANUS-C [26,27], • means results after 50 iterations. Thus, the number of valid samplings is about 12,000.…”

“…Thus, when we measured performances, we changed the number of iterations and samplings by considering the features of each method (Table 1). For example, because JANUS-C [27] is only based on GA, the numbers of mutation and crossover were maintained as those in their implementation, which would be a time burden to the fast search. For GEGL [26], the number of samplings was set as twice of FasterGTS's to maximize their searching property, which relies on the large samplings, and the number of training iterations was equal to or more than FasterGTS's.…”

“…CogMol [20] was designed to generate drugs for COVID-19, and it used the controlled latent attribute space sampling method [21] instead of RL. Also, Monte Carlo tree search (MCTS) [22,23,24] and a genetic algorithm (GA) [25,26,27] were employed to create a molecular generative model. Especially, [26] developed a generative model using GA and primary queues.…”