ABCRpred

ABCRpred predicts beta-lactamase-mediated ceftazidime resistance from beta-lactamase protein sequences to enable in silico identification of ceftazidime-resistant bacterial variants.

Key Features:

• Dataset: Uses 199 beta-lactamase protein sequences (87 ceftazidime-sensitive, 112 ceftazidime-resistant) sourced from the β-lactamase database.
• Feature generation: Employs Pfeature to generate 9,149 features per beta-lactamase protein sequence.
• Feature selection: Refines the feature set using algorithms implemented in the scikit-learn package to identify predictive features.
• Machine learning models: Implements K-nearest neighbors (KNN), Decision Trees (DT), Random Forests (RF), Gaussian Naive Bayes (GNB), Logistic Regression (LR), Support Vector Classifier (SVC), and XGBoost (XGB) for model development.
• Model performance: Reports Random Forest AUROC of 0.80 on the training dataset and 0.79 on the validation dataset.
• Amino acid composition analysis: Identifies enrichment of non-polar side-chain amino acids in ceftazidime-resistant variants and enrichment of polar side chains and charged residues in sensitive variants.

Scientific Applications:

• Resistance prediction: Predicts beta-lactamase-mediated ceftazidime resistance from protein sequence data.
• Mechanistic insight: Enables comparative analysis of amino acid composition to elucidate biochemical differences between resistant and sensitive beta-lactamases.
• Therapeutic support: Provides sequence-based evidence that can inform development of inhibitors or treatment strategies targeting ceftazidime resistance.

Methodology:

Collected 199 beta-lactamase sequences from the β-lactamase database (87 sensitive, 112 resistant); generated 9,149 features per sequence with Pfeature; performed feature selection using scikit-learn algorithms; trained KNN, DT, RF, GNB, LR, SVC, and XGB models and evaluated performance (RF AUROC: 0.80 training, 0.79 validation); conducted amino acid composition analysis comparing resistant and sensitive variants.