Accurate multistage prediction of protein crystallization propensity using deep-cascade forest with sequence-based...

Area of Science:

• Biophysics
• Computational Biology
• Structural Biology

Background:

• X-ray crystallography is vital for determining protein structures at the atomic level.
• Predicting protein crystallization propensity is crucial for experimental success in structural biology.
• Current methods face challenges, especially with difficult-to-crystallize proteins like membrane proteins.

Purpose of the Study:

• To develop a novel machine learning pipeline for predicting protein crystallization propensity.
• To create accurate predictors (DCFCrystal and MDCFCrystal) to guide experimental design.
• To improve the success rate of X-ray crystallography experiments.

Main Methods:

• A machine learning pipeline utilizing a deep-cascade forest (DCF) model.
• Incorporation of multiple sequence-based features, including a novel pseudo-predicted hybrid solvent accessibility (PsePHSA) feature.
• Development of a multistage predictor (DCFCrystal) and a single-stage predictor (MDCFCrystal) for general and membrane proteins, respectively.

Main Results:

• DCFCrystal and MDCFCrystal significantly outperformed existing state-of-the-art predictors, increasing Matthew's correlation coefficient by 199.7% and 77.8%, respectively.
• The DCF model's efficiency and the sensitivity of sequence-based features, particularly PsePHSA, contribute to improved prediction accuracy.
• New crystal-dataset constructions enhanced model training with comprehensive crystallization knowledge.

Conclusions:

• The developed DCF-based pipeline and predictors (DCFCrystal, MDCFCrystal) offer a powerful tool for predicting protein crystallization propensity.
• These predictors can guide experimental efforts, potentially accelerating protein structure determination.
• The novel PsePHSA feature and improved datasets represent significant advancements in computational structural biology.