Machine learning, docking, or physics for structure prediction of ligand-induced ternary complexes

Area of Science:

• Biochemistry
• Structural Biology
• Computational Chemistry

Background:

• PROTACs and molecular glues are crucial for targeted protein degradation.
• A ternary complex structure (E3 ligase, ligand, protein of interest) is vital for rational degrader design.
• Experimental structure determination faces challenges like conformational flexibility and dynamic interactions.

Purpose of the Study:

• To review recent advances in computational modeling of ternary complexes.
• To critically discuss the predictive power and limitations of current computational methods.
• To identify remaining challenges in structure-based degrader design.

Main Methods:

• Examined multistep computational approaches (e.g., docking).
• Examined single-step deep learning methods for direct complex prediction.
• Assessed methods for structure-based design in the absence of experimental structures.

Main Results:

• Multistep methods face sampling, accuracy, and cost limitations.
• Single-step deep learning methods offer speed but are limited by training data scarcity.
• Both approaches have limitations impacting predictive power for ternary complex modeling.

Conclusions:

• Computational modeling is essential for designing PROTACs and molecular glues.
• Further development is needed to overcome limitations in speed, accuracy, and data availability.
• Addressing these challenges will advance structure-based drug design for targeted protein degradation.