BCrystal: an interpretable sequence-based protein crystallization predictor.

Area of Science:

• Structural Biology
• Computational Biology
• Machine Learning

Background:

• X-ray crystallography is the primary method for determining protein structures.
• Current methods for predicting protein crystallization are costly and inefficient, involving extensive trial-and-error.
• Accurate sequence-based predictors are needed to streamline protein structure determination.

Purpose of the Study:

• To develop a novel, accurate, and explainable model for predicting protein crystallization propensity.
• To improve the efficiency and reduce the cost associated with protein structure determination.

Main Methods:

• Developed BCrystal, a machine learning model utilizing an optimized gradient boosting machine (XGBoost).
• Incorporated sequence, structural, and physicochemical features for prediction.
• Employed the SHAP algorithm for model interpretability, identifying key predictive features.

Main Results:

• BCrystal achieved superior performance on three independent test sets, outperforming state-of-the-art methods.
• Achieved an average accuracy of 93.7%, recall of 96.63%, and Matthew's correlation coefficient of 0.868.
• Identified key features: higher relative solvent accessibility positively correlates with crystallizability, while disordered regions and specific tripeptide stretches negatively correlate.

Conclusions:

• BCrystal offers a highly accurate and explainable solution for predicting protein crystallization propensity.
• The model's performance enables effective screening of sequence variants for enhanced crystallizability.
• BCrystal can significantly reduce the experimental burden in structural biology research.