VeGA: A Versatile Generative Architecture for Bioactive Molecules across Multiple Therapeutic Targets.

Area of Science:

• Artificial Intelligence
• Cheminformatics
• Drug Discovery

Background:

• De novo molecular design is crucial for identifying novel drug candidates.
• Existing models often struggle with efficiency and performance in data-scarce environments.
• There is a need for lightweight yet powerful generative models in medicinal chemistry.

Purpose of the Study:

• To introduce VeGA, a lightweight decoder-only Transformer model for efficient de novo molecular design.
• To evaluate VeGA's generative performance, particularly in target-specific fine-tuning under data-scarce conditions.
• To demonstrate VeGA's capability in generating novel, chemically realistic molecules for specific pharmacological targets.

Main Methods:

• Developed VeGA, a streamlined decoder-only Transformer architecture.
• Pretrained VeGA on the ChEMBL database.
• Evaluated VeGA on the MOSES benchmark and against state-of-the-art models (S4, R4) across five pharmacological targets using leakage-safe protocols.
• Applied VeGA to the Farnesoid X receptor (FXR) target for case study validation.

Main Results:

• VeGA achieved high validity (96.6%) and novelty (93.6%) on the MOSES benchmark.
• Demonstrated superior performance in target-specific fine-tuning, especially in extremely low-data scenarios (e.g., mTORC1).
• Consistently generated the most novel molecules while maintaining chemical realism compared to S4 and R4 models.
• Successfully generated novel FXR-targeting compounds with validated binding potential via molecular docking.

Conclusions:

• VeGA is an efficient and robust model for de novo molecular design, suitable for resource-limited settings.
• The model shows significant promise for accelerating drug discovery through novel chemotype generation, particularly under challenging data constraints.
• VeGA's open-access availability aims to empower medicinal chemists in designing target-specific molecules.