DeLA-Drug

DeLA-Drug generates drug-like small-molecule analogues from a single query compound using a recurrent neural network trained on SMILES from ChEMBL28 to explore chemical space for medicinal chemistry and lead optimization.

Key Features:

• Model architecture: A recurrent neural network (RNN) composed of two long short-term memory (LSTM) layers is used to model SMILES syntax.
• Training data: The model was trained on SMILES strings from over one million compounds extracted from the ChEMBL28 database.
• Sampling with substitutions (SWS): A novel SWS strategy is employed to produce new molecular structures while retaining characteristics of the source compounds.
• Preservation of drug-like properties: Generated analogues are designed to maintain druglikeness and synthetic accessibility inherent to the original bioactive compounds.
• No fine-tuning requirement: The approach eliminates time-consuming fine-tuning procedures for generating focused libraries.
• Focused library generation: Enables rapid production of focused compound libraries suitable for high-throughput screening even with limited starting data.

Scientific Applications:

• De novo drug design: Exploration of local chemical space around known actives for generation of candidate molecules in de novo design workflows.
• Lead optimization: Generation of structural analogues to support medicinal chemistry optimization and scaffold exploration.
• High-throughput screening library creation: Rapid construction of focused libraries tailored for HTS campaigns from single query compounds.
• Target-specific example — CB2R: Demonstrated application to cannabinoid receptor subtype 2 (CB2R), a target implicated in cancer and neurodegeneration.

Methodology:

Computational methods include an RNN with two LSTM layers trained on SMILES from >1 million ChEMBL28 compounds and a "sampling with substitutions" (SWS) strategy; the approach operates without time-consuming fine-tuning.