Improve hot region prediction by analyzing different machine learning algorithms

Jing Hu, Longwei Zhou, Bo Li, Xiaolong Zhang, Nansheng ChenBMC Bioinformatics 2021 · DOI

TL;DR

This paper tackles the challenge of accurately identifying "hot regions" in protein-protein interactions, crucial for drug design. By combining Gaussian Naïve Bayes for hot spot residue prediction with DBSCAN for clustering these residues into regions, the authors achieve a significant improvement in hot region prediction accuracy, outperforming other machine learning approaches.

What Problem Does This Paper Address?

Proteins interact with each other to carry out essential biological functions. Within these interactions, a small subset of amino acid residues, known as "hot spots," contribute disproportionately to the binding energy. Identifying these hot spots is vital for understanding protein function and for designing drugs that can modulate these interactions. Hot spots often cluster together to form "hot regions" on the protein interface. While experimental methods exist to identify these regions, they are time-consuming and labor-intensive. This paper aims to accelerate this process by developing a computational model using machine learning to predict hot regions, thereby aiding in drug and protein design.

Methods

The study employs a two-stage approach:

  1. Hot Spot Residue Prediction: Several machine learning algorithms were evaluated for their ability to predict individual hot spot residues. These included:
  2. Gaussian Naïve Bayes (GNB)

    • Support Vector Machine (SVM)
    • Xgboost
    • Random Forest (RF)
    • Artificial Neural Network (ANN)

    Feature selection was performed using the mRMR (Max-Relevance and Min-Redundancy) algorithm to identify the most informative features from protein sequence, structure, and energy data.

  3. Hot Region Clustering: Once hot spot residues were predicted, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm was used to group these predicted hot spot residues into hot regions. DBSCAN is well-suited for this task because hot spot residues tend to cluster densely on the protein interface. The algorithm's hyperparameters, "MinPts" (minimum number of points) and "epsilon" (radius), were optimized through grid search.

The core novelty lies in the combination of a high-recall hot spot residue predictor (GNB) with a high-precision clustering algorithm (DBSCAN) to achieve improved hot region prediction.

Datasets

The study utilizes the SKEMPI 2.0 database. This database contains experimental data on changes in binding free energy upon mutation of amino acids in protein-protein interactions. The authors define hot spot residues based on a binding free energy change threshold of 1.0 kcal/mol (greater than 1.0 kcal/mol for hot spots, less than 1.0 kcal/mol for non-hot spots). This threshold was chosen to avoid extreme data imbalance observed with a 2.0 kcal/mol threshold. The dataset consists of 180 protein complexes, yielding 2326 interface residues (1513 non-hot and 813 hot spots). The data was split into training and testing sets.

Key Insights and Ideas

  • The synergy of prediction and clustering is key: Simply predicting hot spot residues accurately isn't enough; effectively grouping them into meaningful "hot regions" is crucial. The paper highlights that a method with high recall for hot spot prediction (like GNB) is beneficial for ensuring that most true hot spots are identified, which then feeds into the clustering step.
  • DBSCAN's suitability for dense clusters: The inherent density of hot spot residues on protein interfaces makes DBSCAN a natural choice for identifying these regions.
  • Amino acid properties matter: The analysis of the SKEMPI 2.0 dataset revealed that amino acids with aromatic side chains (like Tryptophan - TRP) are more likely to be hot spot residues.
  • Feature importance: The mRMR algorithm identified features like RctASA and RctmPI as particularly important for hot spot prediction.

Key Results

  • Hot Spot Residue Prediction:
    • Xgboost achieved the highest accuracy (0.726) and precision (0.672) in predicting hot spot residues.
  • Gaussian Naïve Bayes (GNB) demonstrated the highest recall (0.792) and F-measure (0.657) for hot spot residue prediction.
  • Hot Region Prediction (Combined GNB + DBSCAN):
    • The combined GNB and DBSCAN approach achieved the highest F-measure of 0.809 for hot region prediction, significantly outperforming other methods.
  • Visualizations (Figure 2) showed that the GNB algorithm correctly predicted 15 out of 16 true positive hot spot residues in a specific protein complex (3HQY), with only two non-hot spots being misclassified. This high recall contributed to the effective formation of hot regions.

Limitations and Future Work

  • Limited biological experiments: The authors acknowledge that the current model's performance is primarily evaluated against existing datasets and that its biological relevance needs further validation through more extensive biological experiments.
  • Next steps: The authors suggest collecting and applying more biological data to further verify and refine the model. They also aim to explore the impact of improving the factors that form hot regions on the clustering algorithm's performance.

Summary updated

BibTeX
@article{hu2021improve,
  author = {Jing Hu and Longwei Zhou and Bo Li and Xiaolong Zhang and Nansheng Chen},
  title = {{Improve hot region prediction by analyzing different machine learning algorithms}},
  journal = {BMC Bioinformatics},
  year = {2021},
  doi = {10.1186/s12859-021-04420-0},
}
Figure