Small molecule autoencoders: architecture engineering to optimize latent space utility and sustainability

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Autoencoders are vital for molecular embedding in deep learning.
• Current autoencoder architectures are often arbitrarily chosen, impacting latent space quality and sustainability.
• Systematic evaluation of architecture's effect on chemical information is lacking.

Purpose of the Study:

• To systematically analyze how autoencoder architecture influences reconstruction and latent space quality.
• To optimize autoencoder architectures for molecular encoding tasks and energy efficiency.
• To provide insights for sustainable deep learning models in chemistry.

Main Methods:

• Conducted systematic experiments varying autoencoder architectures.
• Evaluated reconstruction performance and latent space quality.
• Quantified energy consumption during training.
• Compared molecular representations like SMILES and SELFIES.

Main Results:

• Optimized architectures maintain quality using 97% less data and 36% less energy.
• SELFIES representation showed lower reconstruction performance than SMILES.
• Training with enumerated SMILES improved latent space quality.

Conclusions:

• Autoencoder architecture significantly impacts molecular embedding utility and sustainability.
• Optimized architectures offer a more sustainable approach to molecular encoding.
• This work provides a framework for designing efficient and effective molecular autoencoders.