PROTAC Modeling Breakthrough: PRosettaC Outperforms AlphaFold3

According to Nature, a comprehensive benchmarking study published in Scientific Reports demonstrates that PRosettaC outperforms AlphaFold3 in predicting PROTAC-mediated ternary complex geometries. The research analyzed 36 experimentally determined ternary complex crystal structures identified through a systematic search of the Protein Data Bank conducted on September 3, 2023. PRosettaC, a specialized Rosetta-based protocol, generated up to 1,000 models per system using enhanced sampling, while AlphaFold3 predictions were limited to five models per complex due to server constraints. Evaluation using DockQ v2 scoring revealed consistent, though modest, performance advantages for PRosettaC across multiple systems, suggesting it may offer more reliable structural predictions for PROTAC-focused modeling applications. This comparative analysis provides crucial guidance for computational tool selection in targeted protein degradation research.

The Specialized Tool Advantage in Complex Molecular Systems

What makes PRosettaC’s performance particularly noteworthy is how it exemplifies a broader trend in computational biology: specialized, purpose-built tools often outperform generalist AI systems for specific applications. While AlphaFold3 represents a remarkable achievement in generalized protein structure prediction, PRosettaC’s design specifically for PROTAC ternary complexes gives it inherent advantages. The methodology incorporates geometric constraints derived from known warhead binding modes and uses structure-guided assembly rather than pure sequence-based prediction. This domain-specific knowledge allows PRosettaC to navigate the complex energy landscapes of ternary complex formation more effectively than a general-purpose predictor, even one as sophisticated as AlphaFold3.

The Unique Challenges of PROTAC Ternary Complex Modeling

Modeling PROTAC-mediated complexes presents distinct challenges that go beyond traditional protein-ligand docking. A ternary complex involves three components: the target protein, the E3 ubiquitin ligase, and the bifunctional PROTAC molecule itself. The PROTAC acts as a molecular bridge, inducing proximity between proteins that wouldn’t normally interact. This creates a dynamic system where the linker flexibility, protein-protein interface formation, and induced conformational changes all contribute to the final structure. Traditional docking approaches struggle with these systems because they must account for both the protein-ligand interactions and the novel protein-protein interface created by the PROTAC-induced proximity.

Implications for Accelerated Drug Development

The improved accuracy of PRosettaC could significantly impact the early stages of targeted protein degradation drug discovery. PROTACs represent one of the most promising new modalities in pharmaceutical research, with the potential to target proteins previously considered “undruggable.” However, rational design of these molecules has been hampered by the difficulty in predicting whether a particular PROTAC will successfully form the required ternary complex. More reliable computational predictions could reduce the number of synthetic cycles needed to optimize PROTAC candidates, potentially shortening development timelines and reducing costs. The ability to accurately model these complexes before synthesis could also enable more sophisticated linker design and warhead optimization strategies.

Critical Methodological Considerations and Limitations

While the results are promising, several important limitations warrant consideration. The study’s focus on crystal structures from the Protein Data Bank means the benchmarking reflects static, solid-state conformations rather than dynamic solution behavior. The molecular dynamics simulations included in the study provide some insight into transient states, but the field still lacks comprehensive experimental data on ternary complex dynamics. Additionally, the exclusion of larger scaffold proteins like cullin ring ligases due to AlphaFold3’s input constraints means the benchmarking doesn’t fully represent the biological context where these complexes operate. The E3 ubiquitin ligase machinery typically involves multiple components beyond the recognition domains included in the study.

Future Directions and Integration Opportunities

The most exciting potential lies in integrating these complementary approaches rather than viewing them as competitors. Future iterations could combine AlphaFold3’s remarkable ability to predict protein structures from sequence with PRosettaC’s specialized knowledge of PROTAC-induced complex formation. The open-source nature of PRosettaC’s implementation facilitates such integration and community-driven improvements. As more experimental ternary complex structures become available—particularly from cryo-EM studies that can capture more dynamic states—both methods will likely improve. The field would also benefit from standardized benchmarking datasets and metrics specifically designed for ternary complex evaluation, moving beyond adaptations of traditional protein-protein docking scores.

Broader Impact on Computational Drug Discovery

This research highlights a crucial evolution in computational structural biology: the shift from general-purpose prediction tools to specialized methods optimized for specific biological contexts. As targeted protein degradation expands beyond PROTACs to include molecular glues, LYTACs, and other emerging modalities, we can expect similar specialization in computational approaches. The pharmaceutical industry’s growing investment in targeted degradation platforms—with multiple companies now advancing PROTAC candidates through clinical trials—creates strong demand for reliable computational tools that can accelerate and de-risk these programs. The modest but consistent performance advantage demonstrated by PRosettaC suggests that domain-specific knowledge, when properly encoded into computational methods, can provide meaningful advantages even against state-of-the-art general AI systems.