Protein-protein interface hot spots prediction based on a hybrid feature selection strategy

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Hot spots are crucial interface residues driving protein-protein interaction binding affinity.
• Predicting hot spots requires identifying a compact and relevant subset of features.
• Current feature selection methods face challenges in identifying optimal feature subsets for predictive models.

Purpose of the Study:

• To propose and validate a novel hybrid feature selection strategy for predicting hot spot residues.
• To effectively reduce the feature space for building accurate prediction models.
• To investigate the importance of feature generalization and complementarity in selection.

Main Methods:

• Compared three distinct feature selection methods.
• Developed a hybrid strategy combining decision tree and mRMR (maximum Relevance Minimum Redundancy) feature subsets.
• Employed Pseudo Sequential Forward Selection (PSFS) to build a model with 6 features.
• Evaluated performance on an independent test set using F-measure and recall.

Main Results:

• The proposed hybrid strategy significantly reduced the feature space.
• A 6-feature model achieved comparable or superior predictive performance to state-of-the-art methods (F-measure 0.622, recall 0.821).
• The newly proposed feature CNSV_REL1 was identified as important for prediction; feature complementarity is crucial.

Conclusions:

• The novel hybrid feature selection strategy is effective for optimizing feature subsets in hot spot prediction models.
• Feature generalization and complementarity are critical considerations for feature selection methods.
• The proposed CNSV_REL1 feature offers an effective alternative for hot spot prediction; a webserver is available.