CLPred

CLPred applies a bidirectional recurrent neural network with long short-term memory (BLSTM) to predict protein crystallizability from amino acid sequences for crystallography-based X-ray diffraction.

Key Features:

• BLSTM architecture: Uses a bidirectional recurrent neural network with long short-term memory (BLSTM) to model sequence information.
• Sequence-only input: Operates on amino acid sequence data without requiring structural inputs for primary predictions.
• Long-range k-mer interactions: Captures long-range interaction patterns between k-mers of amino acids to inform crystallizability predictions.
• Non-local inter-peptide modeling: Models non-local, long-range inter-peptide interactions relevant to crystal formation.
• Diffraction-quality prediction: Predicts propensity to form diffraction-quality crystals suitable for X-ray diffraction experiments.
• Benchmark performance: Outperforms existing deep-learning predictors and most sequence-based crystallization prediction tools across three independent test sets.
• Improved over window-based networks: Demonstrates higher predictive accuracy compared to previous window-based neural network approaches.
• Feature integration: Performance can be enhanced by incorporating pre-extracted evolutionary, structural, and physicochemical characteristics.
• Validation on specific proteins: Validated with case studies on the Sox transcription factor family and non-structural proteins of the Zika virus.

Scientific Applications:

• Target prioritization for crystallography: Prioritizes protein targets for crystallography-based X-ray diffraction experiments based on predicted crystallizability.
• Analysis of transcription factors: Assesses crystallization propensity of Sox transcription factor family members.
• Viral protein characterization: Evaluates crystallizability of non-structural proteins of the Zika virus to inform structural studies.
• Support for structural biology studies: Guides selection of sequence-based targets for experimental structure determination.

Methodology:

Implements a bidirectional recurrent neural network with long short-term memory (BLSTM) operating on sequence-only inputs to capture long-range k-mer and non-local inter-peptide interactions; evaluated on three independent test sets and validated using case studies on Sox transcription factors and Zika virus non-structural proteins, with optional integration of pre-extracted evolutionary, structural, and physicochemical features.