Transformer-Decoder GPT Models for Generating Virtual Screening Libraries of HMG-Coenzyme A Reductase Inhibitors:...

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Enzyme inhibition studies

Background:

• HMG-Coenzyme A reductase (HMGCR) is a key target for cholesterol-lowering drugs.
• Developing novel HMGCR inhibitors requires efficient methods for exploring chemical space.
• Deep learning models offer potential for de novo drug design.

Purpose of the Study:

• To generate libraries of novel HMGCR inhibitors using attention-based decoder models.
• To optimize model pretraining and fine-tuning strategies for robust molecular library generation.
• To identify molecular properties and model parameters that yield desirable inhibitor candidates.

Main Methods:

• Attention-based deep neural networks were pretrained on the ZINC15 database and fine-tuned on HMGCR inhibitors.
• Model architecture (number of layers) and fine-tuning parameters (temperature, prompt length) were systematically varied.
• Generated libraries were screened using predicted IC50 values, docking scores, druglikeness, and similarity to known inhibitors.

Main Results:

• Models with 50/50 or 25/75% pretraining/fine-tuning ratios, non-zero temperature, and shorter prompts yielded the most robust libraries.
• Predicted IC50 values correlated well with docking scores and similarity to existing statins.
• 42% of generated molecules were classified as statin-like, with rosuvastatin-like molecules showing the best predicted efficacy.

Conclusions:

• Attention-based decoder models can effectively generate novel HMGCR inhibitor libraries.
• Optimized deep learning approaches can guide the discovery of potent and druglike molecules.
• This strategy accelerates the identification of promising drug candidates for HMGCR-related conditions.