| Resumo: | Recent advances in applying deep generative learning to molecular design have led to a large number of novel approaches to the targeted generation of molecules towards specific features and applications. In this work, we expand on the latent space navigation approach, where molecules are optimized by operating in their latent representation inside a deep auto-encoder, by introducing multi-objective evolutionary algorithms (MOEAs), and benchmarking the proposed framework on several objectives from recent literature. Using several case studies from literature, we show that our proposed method is capable of controlling abstract chemical properties, is competitive with other state-of-the-art methods and can perform relevant tasks such as optimizing a predefined molecule while maintaining a similarity threshold. Also, MOEAs allow to generate molecules with a good level of diversity, which is a desired feature.
|
|---|