AC-ModNet: Molecular Reverse Design Network Based on Attribute Classification.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Deep generative models are increasingly used for molecular exploration.
• A key application is the reverse design of drug compounds with desired properties like solubility and synthesizability.
• Current generative models struggle to generate molecules within specific attribute ranges.

Purpose of the Study:

• To propose AC-ModNet, a novel model combining Variational Autoencoders (VAE) with Auxiliary Classifier Generative Adversarial Networks (AC-GAN).
• To enable the generation of molecular structures within specific attribute intervals for drug design.
• To evaluate the model's performance against existing methods and assess the drug-likeness of generated compounds.

Main Methods:

• Development and implementation of the AC-ModNet architecture, integrating VAE and AC-GAN.
• Training and evaluation of AC-ModNet using the 250K ZINC dataset.
• Comparative analysis with existing generative models using Fréchet ChemNet Distance (FCD) and Frag metrics.
• Validation of generated molecules' potential in drug design through comparison with PubChem database records.

Main Results:

• AC-ModNet demonstrates superior performance in FCD and Frag evaluation indicators compared to related models.
• Generated molecules show potential applicability in drug design, validated against PubChem data.
• The model effectively generates molecular structures within specified attribute intervals.

Conclusions:

• AC-ModNet offers a significant advancement in deep generative models for drug reverse design.
• The model's ability to control attribute intervals enhances its utility for targeted drug discovery.
• This work provides a new machine learning-driven approach for designing novel drug compounds with desired properties.