DeepGraphMolGen, a multi-objective, computational strategy for generating molecules with desirable properties: a...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Generating novel molecules with specific interaction properties is a complex challenge.
• Existing methods may struggle with inaccurate experimental binding data.
• Developing in silico tools for molecular design is crucial for accelerating drug discovery.

Purpose of the Study:

• To develop a robust computational method for generating novel molecules with desired interaction and drug-like properties.
• To address the challenge of potentially erroneous experimental binding data.
• To create a generally applicable in silico platform for molecular design.

Main Methods:

• Utilized graph convolution networks (GCNs) to learn interaction binding models from experimental data.
• Employed a robust loss function to account for potential errors in property scores.
• Applied reinforcement learning with a graph convolution policy for multi-objective optimization, including drug likeness and synthetic accessibility.
• Validated the method using small molecules binding to dopamine transporters.

Main Results:

• Successfully generated novel, chemically valid molecules with high drug-likeness scores.
• Demonstrated the method's ability to optimize for specific binding profiles, such as targeting dopamine transporters selectively.
• Extended the approach to multi-objective reward functions for complex property optimization.
• The generated molecules, while unusual, exhibited desired binding characteristics.

Conclusions:

• The developed method offers a powerful approach for in silico molecular generation with tailored properties.
• The use of GCNs and reinforcement learning provides a robust framework for multi-objective optimization in drug discovery.
• This technique is broadly applicable for designing molecules with desirable characteristics for various therapeutic targets.