Inverse QSAR: Reversing Descriptor-Driven Prediction Pipeline Using Attention-Based Conditional Variational...

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence in drug discovery

Background:

• The inverse quantitative structure-activity relationship (QSAR) problem, which involves identifying chemical structures with desired QSAR-predicted properties, is a significant challenge in drug discovery.
• Existing methods often struggle to generate novel and valid chemical structures efficiently.

Purpose of the Study:

• To develop and validate a novel computational workflow for solving the inverse-QSAR problem.
• To generate novel chemical compounds with specific predicted activities and desirable drug-like properties.

Main Methods:

• A two-step process was employed: first, identifying "seed" descriptor vectors corresponding to target QSAR values, and second, generating chemical structures matching these vectors.
• A new attention-based conditional variational autoencoder neural network was developed for structure generation.
• The workflow was trained and validated using the ChEMBL database.

Main Results:

• The proposed workflow successfully generated novel chemical compounds predicted to possess desired biological activities.
• The generated compounds exhibited acceptable druglikeness and synthetic accessibility.
• Orthogonal validation using pharmacophore modeling and molecular docking confirmed the predicted activity and binding potential of the generated structures.

Conclusions:

• The developed attention-based deep learning model offers an effective solution to the inverse-QSAR problem.
• This approach enables the generation of novel, potentially active, and drug-like compounds.
• The findings support the utility of advanced AI methods in accelerating drug discovery and lead optimization.