By embedding the target properties in the latent space together with molecular structures, they were able to sample molecules with desired properties from specific regions of the latent space. As a result, their latent representation consists of a combination of regions for molecular properties and structures.
for a conditional-VAE, which conditions the encoded representation on specific properties. (36) traded the vanilla VAE used by Gómez-Bombarelli et al. For example, researchers have employed graph representations that enforce chemical rules to generate molecules with specifically tailored properties. More recent methods improve on the shortcomings of this work, particularly the use of SMILES representations as inputs. By sampling from regions that were far enough from the training data, they were able to produce a large fraction of novel molecules with a high drug-likeness. (32) learned a mapping between text-string representations of molecules, in simplified molecular-input line-entry system (SMILES) format, and a continuous latent space. These ab initio approaches have had great success in advancing modern computational materials discovery. (4) This allows for computational simulation of materials and their properties, thus ensuring only the most promising candidate materials need to be synthesized experimentally. To narrow down the search space of candidate materials, researchers often employ ab initio methods, such as Density Functional Theory (DFT). (3) The vast space of possible chemical compositions and aforementioned cost in characterizing structures experimentally makes it impossible to fully explore the composition-structure space. The ability to predict material structure from basic first-principles information, such as the chemical composition, is a long outstanding goal that has yet to be achieved. A key part of that process is the characterization of materials structure, upon which properties and, therefore, applications are heavily dependent. (1,2) Materials research ultimately aims to employ new materials for functional applications. Experimental research has long been the backbone of materials science and discovery, but the cost, from both financial and time perspectives, creates a bottleneck in the “design-to-device” workflow.