Bidirectional Molecule Generation with Recurrent Neural Networks.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for molecular design

Background:

• Recurrent neural networks (RNNs) generate novel molecules via Simplified Molecular Input Line Entry System (SMILES) strings.
• Unidirectional RNNs grow SMILES strings from left to right, which is suboptimal due to the nature of molecular structures and SMILES representations.
• Bidirectional generation is motivated by the lack of inherent start/end points in molecules and the non-univocal nature of SMILES.

Purpose of the Study:

• To introduce and evaluate bidirectional generative RNNs for de novo molecular design using SMILES.
• To compare the performance of established bidirectional methods and a novel approach (BIMODAL) against unidirectional RNNs.
• To assess generated molecules based on novelty, scaffold diversity, and chemical-biological relevance.

Main Methods:

• Implementation of two established bidirectional RNN methods for SMILES generation.
• Introduction and implementation of a novel bidirectional method: Bidirectional Molecule Design by Alternate Learning (BIMODAL).
• Comparative analysis of unidirectional forward RNNs against the three bidirectional strategies.

Main Results:

• Bidirectional strategies generally outperform unidirectional RNNs for de novo molecular design using SMILES.
• The BIMODAL method demonstrated superior performance across most evaluated criteria (novelty, diversity, relevance) compared to the unidirectional forward RNN.
• The study provides evidence supporting the effectiveness of bidirectional approaches in generative chemistry.

Conclusions:

• Bidirectional generative RNNs represent a significant advancement for de novo molecular design using SMILES.
• The BIMODAL method offers a promising and effective approach for generating novel, diverse, and relevant molecular structures.
• The findings advocate for the adoption of bidirectional strategies in computational chemistry and drug discovery pipelines.