Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• De novo drug design utilizes computational methods to create novel molecules with high affinity for biological targets.
• Recurrent neural networks (RNNs) offer a promising approach for generating molecular structures.

Purpose of the Study:

• To demonstrate the efficacy of RNNs as generative models for molecular structures in drug design.
• To explore fine-tuning RNNs with active compounds for targeted drug discovery.
• To evaluate the model's performance against specific pathogens.

Main Methods:

• Training recurrent neural networks as generative models for molecular structures, analogous to natural language processing.
• Fine-tuning the trained models with small datasets of known active molecules against specific biological targets.
• Assessing the correlation between properties of generated molecules and training data.
• Evaluating model performance by comparing generated molecules against hold-out test sets for Staphylococcus aureus and Plasmodium falciparum.

Main Results:

• Generated molecules exhibited properties highly correlated with the training dataset.
• The model successfully reproduced 14% of test molecules against Staphylococcus aureus and 28% against Plasmodium falciparum.
• The RNN model, when combined with a scoring function, can facilitate the entire de novo drug design process.

Conclusions:

• Recurrent neural networks are effective generative models for de novo drug design.
• Fine-tuning RNNs enhances their ability to generate molecules active against specific targets.
• This approach accelerates the generation of novel molecular libraries for drug discovery.