Fuzz Testing Molecular Representation Using Deep Variational Anomaly Generation.

Area of Science:

• Computational chemistry
• Machine learning
• Cheminformatics

Background:

• Molecular representations are crucial for cheminformatics and drug discovery.
• Existing molecular string formats require rigorous validation methods.
• Deep learning models offer new avenues for testing data integrity.

Purpose of the Study:

• To stress-test the robustness of the SELF-referencIng Embedded Strings (SELFIES) molecular format.
• To identify and analyze anomalous SELFIES strings that do not convert to valid chemical structures.
• To explore the utility of unsupervised deep learning for validating molecular representations.

Main Methods:

• Utilized a variational auto-encoder (VAE), an unsupervised deep learning model, to generate anomalous SELFIES strings.
• Analyzed the latent space of the VAE to identify regions producing inconvertible SELFIES.
• Investigated the structural and chemical reasons behind the invalidity of generated SELFIES strings.

Main Results:

• Successfully generated anomalous SELFIES strings that cannot be converted into valid SMILES strings.
• Identified specific regions in the VAE's latent space associated with the generation of invalid SELFIES.
• Discovered that not all SELFIES strings are convertible to valid SMILES, challenging previous assumptions.

Conclusions:

• Variational auto-encoders are effective tools for empirically stress-testing molecular representation robustness.
• The study highlights the need for improved validation protocols for molecular string formats like SELFIES.
• Findings provide insights into SELFIES string invalidity and suggest modifications for enhanced reliability.