HERMES: Holographic Equivariant neuRal network model for Mutational Effect and Stability prediction (2407.06703v1)

Abstract: Predicting the stability and fitness effects of amino acid mutations in proteins is a cornerstone of biological discovery and engineering. Various experimental techniques have been developed to measure mutational effects, providing us with extensive datasets across a diverse range of proteins. By training on these data, traditional computational modeling and more recent machine learning approaches have advanced significantly in predicting mutational effects. Here, we introduce HERMES, a 3D rotationally equivariant structure-based neural network model for mutational effect and stability prediction. Pre-trained to predict amino acid propensity from its surrounding 3D structure, HERMES can be fine-tuned for mutational effects using our open-source code. We present a suite of HERMES models, pre-trained with different strategies, and fine-tuned to predict the stability effect of mutations. Benchmarking against other models shows that HERMES often outperforms or matches their performance in predicting mutational effect on stability, binding, and fitness. HERMES offers versatile tools for evaluating mutational effects and can be fine-tuned for specific predictive objectives.

• The paper introduces HERMES, a novel 3D rotation-equivariant model that predicts protein mutation stability using holographic encodings.
• Its architecture leverages efficient tensor operations and removal of invariant skip connections to achieve a computational gain of about 2.75x.
• Benchmarking reveals that fine-tuned HERMES attains Pearson correlations up to ~0.8, underscoring its value in protein engineering.

HERMES: Holographic Equivariant Neural Network Model for Mutational Effect and Stability Prediction

The paper introduces HERMES, a Holographic Equivariant Neural Network (HCNN) designed to predict the stability and fitness effects of amino acid mutations in proteins. Given the significance of accurately modeling these effects for biological discovery and engineering, HERMES offers a novel approach based on 3D rotational equivariant neural networks.

Key Components and Methodologies

HERMES is built upon the Holographic Convolutional Neural Network (HCNN) architecture, which utilizes 3D rotation-equivariant layers to better manage the spatial complexity of protein structures. The model is pre-trained to predict amino acid identity from atomic neighborhoods within a 10Å radius. This pre-training uses a variety of structural features, including atom type, partial charge, and solvent accessible surface area (SASA), which are projected onto an orthonormal Zernike Fourier basis, creating what the authors term "holographic encodings."

The pre-trained HCNN is subsequently fine-tuned to predict the stability effects of mutations (∆∆G), with the model empowered to identify amino acid propensities as influenced by local structural contexts. This is critical for tasks like predicting the impact of mutations on protein stability, binding affinity, and overall fitness.

Performance and Benchmarks

The authors benchmark HERMES against contemporary models, including RaSP, ESM-1v, and ProteinMPNN, across several datasets:

1. Protein stability (∆∆G): Zero-shot performance of HERMES (r ≈ 0.4/0.5) was significantly enhanced after fine-tuning with data resulting in Pearson correlations up to ∼0.8. Notably, the fine-tuned model HERMES PR with 0.50Å noise outperformed other configurations.
2. Binding affinity (∆∆G binding): Performance was generally lower compared to stability predictions, with the best-performing HERMES model achieving correlations of ∼0.34. This highlights the potential need for binding-specific fine-tuning strategies.
3. Deep Mutational Scanning (DMS): On 25 DMS studies, HERMES models, particularly the fine-tuned ones, showed a notable increase in performance. However, sequence-based models like ESM-1v and DeepSequence generally outperformed structure-based models.

Structural Robustness and Efficiency

One notable improvement in HERMES lies in its architecture: with the removal of certain invariant skip connections and the use of efficient tensor product operations, it achieves a computational efficiency gain of about 2.75x. Furthermore, training HERMES with Gaussian noise on the 3D coordinates of protein structures was shown to enhance the robustness of zero-shot predictions.

Implications and Future Directions

The authors present several critical implications:

• Practical Utility: HERMES can be integrated seamlessly into existing computational pipelines due to its efficient architecture and straightforward interface for fine-tuning. Its open-source implementation further enhances accessibility for the research community.
• Theoretical Contributions: The introduction of holographic encodings and SO(3)-equivariant layers represents a meaningful advancement in handling 3D structural data in proteins, potentially paving the way for more nuanced models in structural biology.
• Future Developments: Future work could focus on specialized fine-tuning strategies for binding affinity and other biologically relevant interactions. Additionally, leveraging advanced protein structure prediction models like AlphaFold to generate more accurate training data could further refine HERMES' predictive accuracy.

Conclusion

HERMES emerges as a robust tool in the computational modeling of protein mutational effects, displaying competitive performance across multiple benchmark datasets. Its improvements in both architectural efficiency and practical utility mark it as a valuable contribution to the field of computational biology and protein engineering. As AI models continue to evolve, combining such innovative neural architectures with comprehensive structural datasets promises to advance our understanding and manipulation of protein functions.