VAE-Sim: A Novel Molecular Similarity Measure Based on a Variational Autoencoder

Area of Science:

• Cheminformatics
• Machine Learning
• Computational Chemistry

Background:

• Molecular similarity is crucial in cheminformatics but lacks a universal definition.
• Existing fingerprint encodings yield varied similarity results, necessitating new approaches.

Purpose of the Study:

• To introduce a novel variational autoencoder (VAE) model for unsupervised molecular similarity.
• To develop a rapid and accurate method for calculating molecular similarity.

Main Methods:

• A variational autoencoder (VAE) neural network was designed with a bottleneck layer for latent vector representation.
• The VAE was trained on a large dataset of over six million druglike molecules and natural products.
• Vector distances in the VAE's latent space were utilized as a novel molecular similarity metric.

Main Results:

• The VAE successfully learned molecular representations and enabled rapid similarity calculations.
• The VAE-derived similarity metric demonstrated effectiveness in addressing typical cheminformatics similarity challenges.
• This approach offers a new, efficient tool for unsupervised molecular similarity assessment.

Conclusions:

• The developed VAE provides a powerful and efficient method for unsupervised molecular similarity.
• This novel approach enhances cheminformatics by offering a robust and rapid similarity metric.
• The VAE-based method represents a significant advancement in computational chemistry and drug discovery.