Fuzz testing molecular representation using deep variational anomaly generation

As efforts to improve the robustness of molecular representations advance, so does the need for methods to test and validate them. We use a Variational Auto-Encoder (VAE), an unsupervised deep learning model, to generate anomalous samples of a well-known molecular string format called SELF-referencIng Embedded Strings (SELFIES). This exercise questions a fundamental SELFIES assumption -- that they are always valid when converted to another string representation, SMILES. Interestingly, we discover that specific regions in the VAE's latent space are particularly effective at generating SELFIES that defy this assumption. This organization of validity in the latent space, which proceeds continuously and radially, helps us better understand the factors affecting the molecular representation's reliability. We propose that the VAE and associated anomaly generation approach offer an effective tool for assessing the robustness of molecular representations. We also explore why some SELFIES strings (version 2.1.1) might be invalid and suggest changes to improve them, aiming to spark further discussion on molecular string representations.