PaleAle 5.0

PaleAle 5.0 predicts protein relative solvent accessibility (RSA) from sequence using deep learning to infer residue exposure when structural data or homology transfer are unavailable.

Key Features:

• Deep Learning Architecture: Employs a deep neural network that integrates stacks of bidirectional recurrent neural networks (RNNs) with convolutional layers to model sequence context.
• Clipped Encoding Method: Uses a "clipped" encoding method to represent protein sequence information for input to the model.
• Long-range Interaction Modeling: Captures long-range interactions within protein sequences that are important for accurate RSA prediction.
• Multi-class Predictions: Produces two-class, three-class, and four-class RSA predictions from sequence.
• Reported Performance: Achieves over 80% accuracy for two-class RSA prediction and is reported to surpass benchmarked predictors.

Scientific Applications:

• Protein structure prediction: Provides residue-level solvent accessibility estimates to support protein structure prediction research.
• Structural inference from sequence: Enables inference of structural features when experimental structures or homology transfer are unavailable.
• Drug discovery: Supplies residue exposure information that can inform target and ligand analysis in drug discovery contexts.
• Enzyme design: Informs enzyme engineering by indicating residue accessibility relevant to activity and stability considerations.
• Study of disease mechanisms: Aids investigation of molecular disease mechanisms by highlighting residue exposure that can affect function.

Methodology:

PaleAle 5.0 applies a deep neural network combining stacks of bidirectional RNNs and convolutional layers, uses a "clipped" encoding method for sequence representation, models long-range sequence interactions, and outputs two-, three-, and four-class RSA predictions with reported >80% accuracy for two-class.