rawMSA

rawMSA predicts protein structural features by directly processing multiple sequence alignments (MSAs) with deep neural networks to infer secondary structure, relative solvent accessibility, and inter-residue contact maps.

Key Features:

• Direct MSA input: The entire multiple sequence alignment is used as input to the network without computing sequence profiles or other pre-calculated features.
• NLP-inspired embeddings: Amino acids are mapped into an adaptively learned continuous space inspired by natural language processing techniques.
• Deep neural network architecture: A deep learning model processes the embedded MSAs to predict structural features.
• PSSM-independent predictions: Does not rely on position-specific scoring matrices or manually designed features derived from MSAs.
• Prediction targets: Designed to predict secondary structure, relative solvent accessibility, and inter-residue contact maps.
• Benchmark performance: Demonstrated improved accuracy versus PSSM-based methods for secondary structure and solvent accessibility and comparable contact-map performance in CASP12 and CASP13.
• Training data: Models were trained and benchmarked on a comprehensive dataset of proteins.

Scientific Applications:

• Secondary structure prediction: Predicts local backbone secondary structure states from raw MSAs.
• Relative solvent accessibility prediction: Predicts residue-level solvent exposure from evolutionary information in MSAs.
• Inter-residue contact-map prediction: Predicts pairwise residue contacts to support tertiary structure modeling.
• Evolutionary representation for structure prediction: Provides an alternative representation of evolutionary information for de novo protein structure prediction tasks.

Methodology:

Amino acids in MSAs are mapped into adaptively learned continuous embeddings inspired by NLP, the full MSAs are input to a deep neural network, and models were trained and benchmarked on a comprehensive protein dataset with evaluation on secondary structure, relative solvent accessibility, and inter-residue contact maps including CASP12 and CASP13.