Benchmarking Deep Learning for PROTAC Ternary Complex Prediction.

Area of Science:

• Biochemistry and Structural Biology
• Drug Discovery and Development
• Computational Biology

Background:

• Proteolysis Targeting Chimeras (PROTACs) utilize the ubiquitin-proteasome system for targeted protein degradation.
• PROTACs consist of an E3 ligase binder, a protein of interest binder, and a linker, forming ternary complexes.

Purpose of the Study:

• To benchmark the accuracy of four computational tools (Chai-1, AlphaFold2, AlphaFold3, Protenix) in predicting PROTAC-induced ternary complex structures.
• To assess the performance of these tools in predicting the relative orientations and positions of the target protein, E3 ligase, and PROTAC molecule.

Main Methods:

• Comparative analysis of ternary complex structures predicted by Chai-1, AlphaFold2, AlphaFold3, and Protenix.
• Evaluation of prediction accuracy using Cα-RMSD metrics for overall complex, protein of interest (POI), E3 ligase, and PROTAC positions.

Main Results:

• All four tools achieved satisfactory overall accuracy for ternary complex prediction (Cα-RMSD < 10 Å).
• Chai-1, AlphaFold3, and Protenix outperformed AlphaFold2, showing superior performance in over 50% of tests.
• Significant challenges persist in accurately predicting the orientation of the POI and E3 ligase, and the precise positioning of the PROTAC molecule.

Conclusions:

• Recent advancements in protein structure prediction tools show promise for modeling PROTAC ternary complexes.
• Accurate prediction of PROTAC ternary complex structures remains challenging, particularly regarding specific component orientations and positions.
• This benchmarking provides insights into current predictive tool capabilities and guides future development for PROTAC-based drug discovery.