SPROF

SPROF predicts protein sequence profiles from 3D structural data by learning from two-dimensional (2D) residue distance maps using an image captioning learning framework to inform protein design.

Key Features:

• Integration with Neural Networks: Integrates energy-based and fragment-based methods with neural networks and builds on SPIN2, which achieved a sequence recovery rate of 34% using one-dimensional (1D) structural properties.
• 3D Structure Representation via 2D Maps: Represents three-dimensional (3D) protein structures as two-dimensional (2D) maps of pairwise residue distances.
• Image Captioning Learning Framework: Employs an image captioning learning framework to predict protein sequence profiles from 2D distance maps.
• Improved Sequence Recovery Rate: Achieves a sequence recovery rate of 39.8% on an independent test set, a 5.2% absolute improvement over SPIN2.
• Long-Range Information Learning: Learns long-range spatial information from 2D distance maps, with sequence recovery increasing as the number of neighboring residues in 3D space increases.

Scientific Applications:

• Binding Site Prediction: Aids identification of potential binding sites by leveraging spatial arrangements of residues from 2D distance maps.
• Protein Function Prediction: Provides structural-derived sequence profiles that support prediction of protein function.
• Protein Interaction Prediction: Informs prediction of protein interactions by capturing long-range residue relationships from 2D distance maps.

Methodology:

Uses 2D distance maps derived from pairwise residue distances and a novel network architecture that applies an image captioning learning framework, integrating energy-based and fragment-based methods with neural networks to learn spatial relationships and long-range interactions.