HyperCys: A Structure- and Sequence-Based Predictor of Hyper-Reactive Druggable Cysteines.

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Cysteine residues, with their reactive thiol groups, are frequently targeted for covalent modification by drugs.
• This modification can enhance drug efficacy and reduce toxicity, but cysteine reactivity and accessibility vary significantly.
• Identifying specific, targetable cysteines is crucial for developing effective targeted covalent inhibitors.

Purpose of the Study:

• To develop a novel ensemble stacked machine learning (ML) model named HyperCys.
• To accurately predict hyper-reactive and druggable cysteine residues in proteins.
• To improve the design of targeted covalent inhibitors with enhanced potency and selectivity.

Main Methods:

• Collected features including pocket, conservation, structural, energy profiles, and physicochemical properties from protein sequences and 3D structures.
• Integrated six ML models (K-nearest neighbors, support vector machine, light gradient boost machine, multi-layer perceptron, random forest, logistic regression) into an ensemble stacked model.
• Evaluated model performance using 10-fold cross-validation, comparing feature group combinations.

Main Results:

• The HyperCys model achieved high performance metrics: 0.784 accuracy, 0.754 F1 score, 0.742 recall, and 0.824 ROC AUC.
• HyperCys demonstrated superior accuracy compared to traditional ML models relying solely on sequence or 3D structural features.
• The study identified optimal feature combinations for predicting hyper-reactive cysteines.

Conclusions:

• HyperCys is an effective tool for identifying potential reactive cysteines in nucleophilic proteins.
• The model contributes significantly to the discovery of new drug targets for covalent inhibition.
• HyperCys facilitates the design of more potent and selective targeted covalent inhibitors, advancing drug development.