Molecule Sequence Generation with Rebalanced Variational Autoencoder Loss.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Molecule generation utilizes variational autoencoders (VAEs) to create novel molecular structures via latent space embeddings.
• Recurrent neural network-based VAEs face a posterior collapse issue, hindering molecule sequence generation performance.

Purpose of the Study:

• Investigate the causes of posterior collapse in VAE-based molecule sequence generation.
• Propose an effective solution to mitigate posterior collapse and enhance generation quality.

Main Methods:

• Analysis of the posterior collapse problem in VAEs for molecule generation.
• Development and implementation of a novel method to address underestimated reconstruction loss.
• Experimental validation on benchmark datasets.

Main Results:

• Identified underestimated reconstruction loss as the primary factor causing posterior collapse.
• Demonstrated analytical and experimental evidence supporting the findings.
• Achieved competitive reconstruction accuracy and validity scores using the proposed method.

Conclusions:

• The proposed method effectively resolves posterior collapse in VAE molecule generation.
• Improved VAE performance leads to more accurate and valid novel molecule proposals.
• This work advances AI-driven approaches in molecular design and discovery.