PremPS: Predicting the impact of missense mutations on protein stability

Yuting Chen, Haoyu Lu, Ning Zhang, Zefeng Zhu, Shuqin Wang, Minghui LiPLOS Computational Biology 2020 · DOI

TL;DR

Predicting how single-point mutations affect protein stability is crucial for understanding diseases and designing new proteins. This paper introduces PremPS, a novel computational method that significantly improves the accuracy of these predictions, particularly for mutations that increase protein stability, a long-standing challenge. PremPS achieves this by using a balanced training dataset and a refined set of evolutionary and structural features.

What Problem Does This Paper Address?

Proteins are the workhorses of biology, and their function is intimately linked to their stability. Small changes in their amino acid sequence, known as missense mutations, can alter this stability, leading to dysfunction and disease. While many computational methods exist to predict these changes, they often struggle with a key issue: most available experimental data is heavily skewed towards mutations that decrease protein stability. This bias leads to models that are much better at predicting destabilizing mutations than stabilizing ones. Furthermore, existing methods often show inconsistent performance across different datasets. The authors aim to develop a more accurate and balanced predictor that can reliably assess the impact of both stabilizing and destabilizing mutations.

Methods

PremPS is a machine learning approach that utilizes a random forest regression model. Its novelty lies in two main areas:

  • Balanced Training Data: Instead of relying on skewed experimental datasets, PremPS is trained on a carefully curated dataset with an equal number of stabilizing and destabilizing mutations. This is achieved by modeling reverse mutations for existing destabilizing mutations.
  • Feature Engineering: The model employs a concise set of ten features, categorized into evolutionary and structural properties. These include:
  • PSSM scores: Derived from sequence alignments to capture evolutionary conservation.
  • Amino acid conservation changes (ACS): Quantifying how conservation shifts after a mutation.
  • Hydrophobicity differences (ΔOMH): Measuring changes in a residue's tendency to avoid water.
  • Solvent Accessible Surface Area (SASA): Assessing how exposed a residue is to the solvent.
  • Buried residue fractions: The proportion of specific residue types (aromatic, charged, leucine) found in the protein core.
  • Local residue composition: The count of hydrophobic and charged amino acids in the vicinity of the mutation site.

The authors also explored Support Vector Machines (SVM) and eXtreme Gradient Boosting (XGBoost) but found that the random forest model performed best.

Datasets

  • Training Dataset: The primary training dataset, referred to as S5296, was derived from the ProTherm database and the S2648 dataset. It was expanded to include reverse mutations, resulting in 5,296 single mutations.
  • Testing Datasets: Eight established benchmark datasets (S350, S605, S1925, S134, p53, Ssym, S250, S2000) were used for initial validation and comparison with other methods.
  • Independent Test Set: A combined independent test set, S921, was created by removing redundant and overlapping mutations from the benchmark datasets to ensure a more rigorous evaluation.

Key Insights and Ideas

  • The Stabilizing Mutation Problem is Real: The paper strongly reinforces the long-recognized difficulty in accurately predicting stabilizing mutations due to data bias.
  • Balanced Data is Key: Training on a dataset with an equal representation of stabilizing and destabilizing mutations is crucial for achieving balanced predictive performance.
  • Concise Feature Set is Sufficient: PremPS demonstrates that a relatively small set of well-chosen features can be highly effective, avoiding the complexity of methods using hundreds of features.
  • Robustness Across Structures: The method shows good performance even when using different structural representations of the same protein, indicating a degree of robustness against minor structural variations. However, low-resolution structures and template-based models with low sequence identity can impact accuracy.
  • Web Server for Accessibility: The availability of a user-friendly web server significantly lowers the barrier for researchers to utilize PremPS for their own studies.

Key Results

  • Improved Accuracy: PremPS consistently outperforms other state-of-the-art methods across multiple benchmark datasets, particularly in predicting stabilizing mutations.
  • Low Bias: The method exhibits a significantly lower bias towards predicting destabilizing mutations compared to previous approaches.
  • Cross-Validation Performance: Various cross-validation strategies (CV1-CV5) showed robust performance, with Pearson correlation coefficients (R) generally above 0.73 and RMSE values around 1.23 kcal mol⁻¹.
  • Independent Test Set Performance: On the independent test set S921, PremPS achieved an R of 0.78 for all mutations, 0.72 for destabilizing mutations, and 0.60 for stabilizing mutations, outperforming other methods.
  • Speed: PremPS is computationally efficient, taking approximately four minutes for a single mutation on a protein of ~300 residues, and only 0.4 seconds for each additional mutation.

Limitations and Future Work

  • Structural Resolution: The paper notes that prediction accuracy decreases with lower structural resolution, suggesting that structures with a resolution of 3Å or better are preferable.
  • Model Quality: The accuracy of predictions is also dependent on the quality of modeled structures, with lower quality models leading to less accurate predictions.
  • Monomeric vs. Multimeric Structures: While the method performs well on monomeric structures, the analysis of multimeric complexes is limited by data availability, though initial results suggest a slight preference for monomeric structures.
  • Future Work: The authors suggest that further improvements could be made by incorporating more features or exploring advanced modeling techniques. The current work lays a strong foundation for future development in accurate and balanced prediction of mutation effects on protein stability.

Summary updated

BibTeX
@article{chen2020premps,
  author = {Yuting Chen and Haoyu Lu and Ning Zhang and Zefeng Zhu and Shuqin Wang and Minghui Li},
  title = {{PremPS: Predicting the impact of missense mutations on protein stability}},
  journal = {PLOS Computational Biology},
  year = {2020},
  doi = {10.1371/journal.pcbi.1008543},
}
Figure