Comparative Study of Deep Generative Models on Chemical Space Coverage (v18)

Abstract

<p>In recent years, deep molecular generative models have emerged as novel methods for <i>de novo</i> molecular design. Thanks to the rapid advance of deep learning techniques, deep learning architectures such as recurrent neural networks, generative autoencoders, and adversarial networks, to give a few examples, have been employed for constructing generative models. However, so far the metrics used to evaluate these deep generative models are not discriminative enough to separate the performance of various state-of-the-art generative models. This work presents a novel metric for evaluating deep molecular generative models; this new metric is based on the chemical space coverage of a reference database, and compares not only the molecular structures, but also the ring systems and functional groups, reproduced from a reference dataset of 1M structures. In this study, the performance of 7 different molecular generative models was compared by calculating their structure and substructure coverage of the GDB-13 database while using a 1M subset of GDB-13 for training. Our study shows that the performance of various generative models varies significantly using the benchmarking metrics introduced herein, such that generalization capability of the generative model can be clearly differentiated. Additionally, the coverage of ring systems and functional groups existing in GDB-13 was also compared between the models. Our study provides a useful new metric that can be used for evaluating and comparing generative models.</p>

De novo drug design using reinforcement learning with graph-based deep generative models

“…We have shown our model is able to perform well in several tasks most notably promoting the generation of DRD2 active molecules. While favouring certain properties, our RL framework also improves other performance metrics including increasing the percentage of valid and properly terminated molecules, reaching validity rates comparable to that of state-of-the-art models (Brown et al, 2019;Polykovskiy et al, 2020;Zhang et al, 2021).…”

“…The main drawback of the proposed model is the amount of time and computational power needed to pre-train the underlying GraphINVENT model (a few days on an NVIDIA Tesla K80); however, this is on par with that needed for other state-of-the-art molecular DGMs (Zhang et al, 2021), and only has to be done once per dataset. After pre-training, fine-tuning the model with RL is comparatively quick and requires only between 10 − 40 minutes, where scoring the model is the main bottleneck.…”

Generative Deep Learning for Targeted Compound Design

“…Cieplinski et al proposed to better represent real discovery problems by using docking as a benchmark of the different methods of goal-directed generation. Zhang et al improved on their earlier work, measuring the coverage of chemical space by generative models, by also evaluating the coverage of functional groups and ring systems. Furthermore, the team provided results for various, recently introduced, generative model architectures allowing for their comparison as well as providing useful baselines for future works.…”