Comparative Study of Deep Generative Models on Chemical Space Coverage

Abstract

In recent years, deep molecular generative models have emerged as promising methods for de novo molecular design. Thanks to the rapid advance of deep learning techniques, deep learning architectures such as recurrent neural networks, variational autoencoders, and adversarial networks have been successfully employed for constructing generative models. Recently, quite a few metrics have been proposed to evaluate these deep generative models. However, many of these metrics cannot evaluate the chemical space coverage of sampled molecules. This work presents a novel and complementary metric for evaluating deep molecular generative models. The metric is based on the chemical space coverage of a reference datasetGDB-13. The performance of seven different molecular generative models was compared by calculating what fraction of the structures, ring systems, and functional groups could be reproduced from the largely unseen reference set when using only a small fraction of GDB-13 for training. The results show that the performance of the generative models studied varies significantly using the benchmark metrics introduced herein, such that the generalization capabilities of the generative models can be clearly differentiated. In addition, the coverages of GDB-13 ring systems and functional groups were compared between the models. Our study provides a useful new metric that can be used for evaluating and comparing generative models.

Augmented Hill-Climb increases reinforcement learning efficiency for language-based de novo molecule generation

“…Despite improvement in the optimization ability by AHC, it is irrelevant if the resulting de novo structures are invalid or implausible (e.g., incorrect valences, unstable or idiosyncratic functional groups or strained ring systems). The chemistry generated by RNNs has been evaluated previously [ 3 , 23 , 33 , 98 , 99 ] and has usually been considered reasonable with respect to overall topology, fragments, substructures and property space. On the other hand, a comparison of chemistry between AHC and REINVENT is complicated by the scoring function and its suitability for an objective e.g., greater optimization may actually lead to unreasonable chemistry due to scoring function exploitation rather than as a function of the RL strategy.…”

Exploring Graph Traversal Algorithms in Graph-Based Molecular Generation

“…To evaluate the models, we considered the following metrics: Percent validity and uniqueness of sampled structures, Training and validation losses, Chemical space coverage, Fraction molecules reproduced from reference set, Fraction ring systems reproduced from reference set, Fraction functional groups reproduced from reference set, Shape analysis, , NPR1 and NPR2 distributions (aka “PMI plots”). Above, the term reference set refers to the combined training, testing, and validation sets for each COCONUT subset. The validity was evaluated for 1M samples, while uniqueness and coverage were evaluated only for the valid samples.…”

Rechargeable Batteries of the Future—The State of the Art from a BATTERY 2030+ Perspective

“…[105] A range of algorithms have been tried out with varying level of success, among which most popular and versatile have been, variational auto encoders [106] and generative adversarial networks. [107] Both generative schemes can operate on image (pixels and voxels), [108][109][110] graph (nodes and edges) [111][112][113] (e.g., SMILES) and fingerprint-vector based representations [114] Figure 5. Cross-section through the negative electrode, the SEI, and the electrolyte.…”