KFC Server

KFC Server predicts binding hot spots within protein–protein and protein–DNA interfaces to identify residues that contribute significantly to the binding free energy of an interface.

Key Features:

• Knowledge-based model: Employs the Knowledge-based FADE and Contacts (KFC) model to characterize each residue's local structural environment and compare it against known experimental hot-spot data.
• Predictive models: Implements two advanced classifiers, KFC2a and KFC2b, developed to improve predictive accuracy relative to the original KFC model.
• Balanced training data: Trains on a balanced dataset containing equal numbers of hot spot and non-hot spot residues to reduce bias in predictions.
• Feature engineering: Generates 47 structural features describing residue environments, with KFC2a using eight features focused on solvent accessible surface area and local plasticity and KFC2b using seven features with two overlapping features between models.
• Machine learning algorithm: Constructs both KFC2 models using support vector machines (SVM) for classification of hot spot versus non-hot spot residues.
• Performance metrics: Compared against Robetta, FOLDEF, HotPoint, MINERVA, and the original KFC model, KFC2a achieved a true positive rate (TPR) of 0.85 with a higher false positive rate, while KFC2b achieved TPR = 0.62 and false positive rate (FPR) = 0.15 and the best balance between accuracy and FPR.

Scientific Applications:

• Structural biology: Identification of hot-spot residues to interpret protein interaction mechanisms and the energetic contributions of individual interface residues.
• Functional analysis: Insight into protein function by pinpointing residues that disproportionately affect binding free energy within interfaces.
• Therapeutic targeting: Prioritization of potential sites for therapeutic intervention by highlighting residues critical for binding affinity.

Methodology:

Uses the Knowledge-based FADE and Contacts (KFC) framework to characterize local residue environments, generates 47 structural features and selects feature subsets for KFC2a (8 features emphasizing solvent accessible surface area and local plasticity) and KFC2b (7 features with two overlaps), trains support vector machine classifiers on a balanced hot spot/non-hot spot dataset, and evaluates performance relative to Robetta, FOLDEF, HotPoint, MINERVA, and the original KFC model.