Automated Generation of Novel Fragments Using Screening Data, a Dual SMILES Autoencoder, Transfer Learning and Syntax...

Area of Science:

• Computational Chemistry
• Medicinal Chemistry
• Artificial Intelligence in Drug Discovery

Background:

• Fragment-based hit identification (FBHI) offers greater chemical space coverage than high-throughput screening but relies heavily on library design.
• Designing effective fragment libraries involves optimizing solubility, stability, diversity, and synthetic tractability, which is a time-consuming process.

Purpose of the Study:

• To develop and evaluate an autoencoder model for de novo fragment library design.
• To improve the quality and diversity of generated fragments by simultaneously reconstructing SMILES and chemical fingerprints.

Main Methods:

• A recurrent neural network-based autoencoder model was trained on a large dataset of commercial fragments.
• Transfer learning was used to train a classifier on frequent hitter fragments.
• A particle swarm optimization approach was employed to sample and evaluate generated molecules.
• A syntax-correction procedure was implemented during training to enhance generative performance.

Main Results:

• The dual autoencoder model generated valid SMILES with improved features, including higher synthetic accessibility and a novel Feature Complexity (FeCo) score.
• Generative performance was enhanced by incorporating a syntax-correction procedure during training.
• The model successfully generated a library of novel candidate privileged fragments.

Conclusions:

• Artificial neural networks, specifically dual autoencoders, are effective tools for fragment library design.
• The developed model improves the quality and relevance of generated fragments for drug discovery.
• This approach accelerates the design of novel fragment libraries with desirable properties.