Structure-informed machine learning for drug discovery: a task-centric perspective.

Area of Science:

• Computational chemistry
• Structural biology
• Artificial intelligence in drug discovery

Background:

• Protein structure prediction advances enable structure-based drug discovery.
• Deep learning models incorporate spatial constraints for target-specific ligand design.

Purpose of the Study:

• Synthesize structure-aware molecular modeling from a task-centric viewpoint.
• Review AI-driven drug design, focusing on structure-conditioned molecular generation.
• Provide insights into next-generation molecular modeling.

Main Methods:

• Focus on binding pocket identification, interaction prediction, pose estimation, and complex modeling.
• Highlight evolution from traditional docking to geometry-informed deep learning.
• Classify generative approaches: sequence-based, fragment-based, graph-based, and 3D coordinate-based (diffusion models).

Main Results:

• Geometry-informed deep learning architectures encode protein structures effectively.
• Diffusion and point cloud models show promise for synthesizing pocket-complementary molecules.
• Co-folding models unify protein folding and ligand binding prediction.

Conclusions:

• Structural knowledge is reshaping AI-driven drug design.
• Challenges include data scarcity, generalization, and multi-objective control.
• Future directions focus on scalable, interpretable, and physically plausible generation pipelines.