Shoebill: an interpretable AlphaFold2-informed predictor of protein crystallization propensity using XGBoost

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• X-ray crystallography is key for high-resolution protein structures but faces challenges with crystal formation.
• Current computational methods often lack interpretability or rely solely on sequence data.
• Predicting crystallization propensity computationally can streamline experimental screening and reduce costs.

Purpose of the Study:

• To develop an interpretable computational predictor for protein crystallization propensity.
• To integrate structural information from AlphaFold2 predictions into crystallization propensity assessment.
• To provide practical guidance for experimental protein structure determination.

Main Methods:

• Developed Shoebill, integrating AlphaFold2 (AF2)-derived structural descriptors with an XGBoost framework.
• Extracted comprehensive features from AF2-predicted structures, including disorder, confidence metrics, geometry, and surface properties.
• Utilized SHAP analysis for feature-level explanations of predictions.

Main Results:

• Shoebill demonstrated superior performance over existing non-deep-learning predictors on an independent benchmark.
• Achieved an improved area under the receiver operating characteristic curve (0.804 vs. 0.700) and Matthews correlation coefficient (0.297 vs. 0.123).
• Provided interpretable insights into features influencing crystallization propensity.

Conclusions:

• Shoebill offers an accurate and interpretable method for predicting protein crystallization propensity.
• The tool leverages structural insights from AlphaFold2, enhancing predictive power.
• Interpretability aids in guiding rational strategies for optimizing protein constructs for crystallization.