MolGPT: Molecular Generation Using a Transformer-Decoder Model.

Area of Science:

• * Computational chemistry and cheminformatics.
• * Artificial intelligence and machine learning.
• * Pharmaceutical sciences and drug discovery.

Background:

• * Deep learning, particularly transformer models, is increasingly applied to molecular generation for drug design.
• * Generative Pre-training (GPT) models demonstrate success in generating coherent text, inspiring new molecular design approaches.
• * SMILES notation allows molecular representation as strings, enabling natural language processing techniques for molecular design.

Purpose of the Study:

• * To develop and evaluate MolGPT, a transformer-decoder model for *de novo* molecular generation.
• * To assess MolGPT's performance against existing machine learning frameworks for generating valid, unique, and novel molecules.
• * To demonstrate the model's capability for conditional generation of molecules with desired properties and scaffolds.

Main Methods:

• * Training a transformer-decoder model on a next token prediction task with masked self-attention.
• * Utilizing SMILES strings for molecular representation.
• * Implementing conditional generation based on scaffold SMILES and property values.
• * Employing saliency maps for interpretability analysis.

Main Results:

• * MolGPT generates valid, unique, and novel molecules comparable to state-of-the-art methods.
• * The model successfully generates molecules with controlled properties through conditional training.
• * Generation of molecules with specific desired scaffolds and property values was demonstrated.
• * Saliency maps provided insights into the model's generative process interpretability.

Conclusions:

• * MolGPT is a powerful tool for *de novo* molecular generation in drug discovery.
• * Conditional generation capabilities allow for targeted design of molecules with specific characteristics.
• * The model offers a balance of performance, novelty, and interpretability in inverse molecular design.