Exploring the GDB-13 chemical space using deep generative models.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Drug Discovery

Background:

• Recurrent neural networks (RNNs) are increasingly used for generating novel molecular structures.
• Sampling large chemical spaces is crucial for discovering new molecules with desired properties.
• Evaluating the learning efficiency of generative models is essential for their practical application.

Purpose of the Study:

• To train and evaluate a recurrent neural network (RNN) model for sampling chemical space using molecular string representations (SMILES).
• To develop and apply methods for assessing the quality and learning capacity of the trained molecular generative model.
• To analyze the characteristics of the chemical space sampled by the RNN model.

Main Methods:

• Training an RNN model on a subset of the GDB-13 molecular database using SMILES representations.
• Assessing model training quality using negative log-likelihood plots.
• Quantifying model learning using a mathematical model based on the coupon collector problem.

Main Results:

• A model trained on 1 million structures (0.1% of GDB-13) reproduced 68.9% of the database when sampling 2 billion molecules.
• The developed methods provide a quantitative assessment of the model's learning progress.
• Analysis revealed that complex molecules with multiple rings and heteroatoms are more challenging to sample due to SMILES syntax limitations.

Conclusions:

• RNNs are capable of learning and sampling large portions of chemical space effectively.
• The proposed methods, including negative log-likelihood plots and the coupon collector problem model, can benchmark molecular generative models.
• SMILES syntax presents challenges for sampling complex molecular architectures, highlighting areas for future model improvement.