Kernel-elastic autoencoder for molecular design.

Area of Science:

• Computational chemistry
• Artificial intelligence
• Drug discovery

Background:

• Generative models are crucial for de novo molecular design.
• Existing models like Variational Autoencoders (VAEs) have limitations in balancing generation quality and reconstruction accuracy.
• Transformer architectures offer powerful sequence modeling capabilities.

Purpose of the Study:

• To introduce the kernel-elastic autoencoder (KAE), a self-supervised generative model for enhanced molecular design.
• To evaluate the performance of KAE's novel loss functions (m-MMD and weighted reconstruction) against traditional methods.
• To demonstrate KAE's effectiveness in conditional generation and predicting molecular binding affinities.

Main Methods:

• Implementation of the kernel-elastic autoencoder (KAE) based on the transformer architecture.
• Utilizing modified maximum mean discrepancy (m-MMD) and weighted reconstruction loss functions.
• Integration with conditional generation for constrained optimization tasks.
• Validation using AutoDock Vina and Glide scores for binding affinity prediction.

Main Results:

• KAE demonstrates significantly improved generative performance compared to VAEs and standard MMD using m-MMD loss.
• The weighted reconstruction loss enables simultaneous valid generation and accurate reconstruction.
• KAE achieves state-of-the-art results in conditional generation and constrained optimization.
• Generated molecules exhibit superior binding affinities in docking applications compared to training data candidates.

Conclusions:

• KAE represents a significant advancement in self-supervised generative models for molecular design.
• The novel loss functions provide a unique balance between generative flexibility and reconstruction fidelity.
• KAE shows broad applicability beyond molecular design, with potential in diverse generative tasks.