Generating and screening de novo compounds against given targets using ultrafast deep learning models as core...

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Drug Discovery

Background:

• Deep learning models geometric transformations, showing promise in drug development.
• Generative AI models can navigate chemical space for de novo drug design.
• Public structure databases facilitate AI-driven drug discovery.

Purpose of the Study:

• To develop a computational strategy for generating de novo small molecular compounds for specific biological targets.
• To demonstrate the efficacy of the proposed method in identifying potential drug candidates.
• To explore iterative training and simulation methods for optimizing compound generation.

Main Methods:

• Utilized an accelerated LSTM-Chem (long short-term memory for de novo compounds generation) and dense fully convolutional neural network (DFCNN).
• Integrated docking simulations to predict binding affinities for generated compounds.
• Employed iterative training with selected candidates and molecular dynamics (MD) simulations for refinement.

Main Results:

• Generated a large number of de novo small molecular compounds for six diverse human disorder targets.
• Demonstrated that iterative training significantly increases the likelihood of discovering compounds with higher predicted binding affinities.
• Validated the approach using M protease as a proof-of-concept, with MD and metadynamics simulations supporting compound reliability.

Conclusions:

• The proposed deep learning strategy offers a practical and effective approach for de novo compound generation and binder discovery.
• Iterative refinement and simulation enhance the identification of potent drug candidates.
• Further biochemical validation of top-ranked compounds is recommended to advance drug development efforts.