Abstract X-ray crystallography is a central technique for high-resolution protein structure determination; however, the production of diffraction-quality crystals remains a major bottleneck due to high experimental cost and low success rates. Computational prediction of protein crystallization propensity offers a promising strategy to prioritize targets and reduce unnecessary experimental screening. While numerous computational predictors have been proposed, many rely primarily on sequence-derived features or provide limited interpretability, thereby offering little practical guidance for experimental design. Here, we present Shoebill, an interpretable protein crystallization propensity predictor that integrates AlphaFold2 (AF2)-derived structural descriptors with an XGBoost framework to assess whether a protein—assuming successful expression and purification—is likely to form diffraction-quality crystals. Shoebill leverages a comprehensive feature set extracted directly from AF2-predicted structures, capturing complementary structural information beyond sequence alone, including structural disorder, AF2 confidence metrics, molecular geometry, and surface physicochemical properties. On our primary independent benchmark, Shoebill outperforms existing nondeep-learning crystallization propensity predictors while maintaining balanced sensitivity and specificity, improving the area under the receiver operating characteristic curve from 0.700 to 0.804 and more than doubling the Matthews correlation coefficient from 0.123 to 0.297 relative to DCFCrystal, the best-performing nondeep-learning method. Although the deep learning-based predictor SADeepcry achieves higher overall predictive accuracy, Shoebill provides feature-level explanations for individual predictions through SHAP analysis, highlighting biologically meaningful features associated with crystallization propensity that may help guide rational strategies for construct optimization. The Shoebill source code and a user-friendly web server are publicly available at https://github.com/KJ-Liao/Shoebill.
Liao et al. (Thu,) studied this question.