TL;DR
ProTact is a novel SE(3)-invariant geometric graph neural network that leverages physics-informed geometric complementarity and trigonometric constraints to accurately predict protein-protein interactions, even with limited evolutionary data. This approach significantly outperforms existing methods and enhances the performance of tools like AlphaFold3, paving the way for more accurate protein complex modeling and drug discovery.
What Problem Does This Paper Address?
Predicting how proteins interact and form complexes is fundamental to understanding biological processes and developing new therapeutics. While deep learning models like AlphaFold have revolutionized single protein structure prediction, accurately modeling protein-protein complexes remains a significant challenge. This is particularly true for cases where evolutionary information, typically derived from Multiple Sequence Alignments (MSAs), is sparse or missing. Such scenarios include engineered proteins, complexes across different species, and antigen-antibody interactions. Existing methods often struggle to capture the intricate geometric and physical constraints governing these interactions, leading to less reliable predictions.
Methods
The core of ProTact is a SE(3)-invariant geometric graph neural network. This architecture is designed to be independent of the protein's orientation and position in space, making it robust to variations. It comprises two main components:
- SE(3)-invariant encoder (SEInvi): This module transforms protein structure and sequence information into rotation and translation-invariant representations. It utilizes spherical harmonics to encode dihedral angles, which are invariant to rotation and translation, ensuring that the learned features are robust.
- Trigonometry module (TrigModule): This module integrates physics-informed geometric complementarity and trigonometric constraints. It processes messages between two proteins, considering factors like excluded volume and saturation effects, to better capture the complementary characteristics of interacting surfaces. This module also incorporates co-evolutionary features derived from MSAs.
ProTact also employs a modulated key point matching algorithm for approximating accurate docking poses and can be used to re-rank predictions from other models like AlphaFold3.
Datasets
ProTact was trained and evaluated on several benchmark datasets:
- DIPS-Plus: A large dataset of protein complexes for interface contact prediction, used for training and validation.
- CASP13&14: A dataset of homodimers and heterodimers with publicly available PDB structures, used for evaluating performance on challenging targets.
- SAbDab: A database of antibody structures, used for fine-tuning and testing in scarce MSA scenarios, particularly for antigen-antibody complexes.
- BlindTest Dataset: A newly collected dataset of heterodimers and homodimers published between January 2023 and June 2023, used for a more comprehensive and unbiased comparison.
Key Insights and Ideas
- Physics-informed inductive biases are crucial: Integrating geometric complementarity and trigonometric constraints directly into the model acts as powerful "inductive biases," significantly improving the accuracy of inter-chain contact predictions, especially when evolutionary information is limited.
- SE(3)-invariance is key for robustness: Making the model invariant to rotation and translation is essential for handling diverse protein structures and orientations without losing predictive power.
- ProTact complements existing tools: ProTact doesn't just replace existing methods; it enhances them. By providing accurate contact predictions, it can be used to re-rank the outputs of powerful models like AlphaFold3, leading to better overall predictions.
- MSA scarcity is not a deal-breaker: ProTact demonstrates remarkable performance even in low-MSA contexts, which is a significant step forward for predicting interactions in challenging biological systems.
Key Results
ProTact consistently outperforms state-of-the-art sequence-based and structure-based methods on benchmark datasets.
- Contact Prediction: Achieved a 31.63% relative improvement in Precision@10 on CASP13&14 targets and a 31.94% relative gain on DIPS-Plus datasets.
- Low-MSA Performance: When used to re-score AlphaFold3 predictions, ProTact showed over 30.48% improvement in low-MSA contexts.
- Docking Accuracy: ProTact's predictions, when used with the Kabsch algorithm, resulted in average RMSD values that surpassed competing methods like GLINTER and DeepInteract. For example, on the HIV protease complex (PDB ID: 1SDU), ProTact accurately predicted both interfaces, a task that challenged traditional methods.
- Antigen-Antibody Complexes: ProTact showed a notable advantage in predicting antigen-antibody complexes, achieving significant improvements in RMSD compared to leading models, even with very limited MSA information. For the SARS-CoV-2 Receptor Binding Domain complex, ProTact achieved an RMSD of 3.08 Å, significantly outperforming GLINTER and DeepInteract.
Limitations and Future Work
- Limitations: The authors acknowledge that ProTact has limitations. One such limitation is the incorporation of protein language models and additional geometric and physical constraints.
- Future Work: The authors suggest several avenues for future research:
- Integrating protein language models and other geometric/physical constraints to further enhance accuracy and efficiency.
- Extending ProTact to handle more complex multi-protein systems and dynamic interactions.
- Exploring the application of ProTact in drug discovery and antibody design.
Summary updated