Seq-SetNet

Seq-SetNet leverages multiple sequence alignments (MSAs) to predict residue-level structural properties such as secondary structure and backbone torsion angles.

Key Features:

• Encoding Module: Processes each component homologue in an MSA individually and encodes residue-specific mutations, insertions, deletions, and long-distance residue correlations into context-specific features for every residue.
• Aggregation Module: Consolidates encoded features from all homologues and transforms the aggregated representation into residue-level structural properties for the query protein.
• Automatic Feature Learning: Learns effective features directly from MSAs within the encoding module, removing the need for manual feature engineering.
• Symmetric Aggregation Functions: Uses order-invariant (symmetric) aggregation functions to ensure predictions are independent of the sequence order of homologues in the MSA.

Scientific Applications:

• Residue-level structural prediction: Predicts secondary structure and backbone torsion angles of protein residues directly from MSAs.
• Protein folding, function, and interaction studies: Provides residue-resolved structural information to support analyses of protein folding, molecular function, and protein–protein or protein–ligand interactions.

Methodology:

Applies an explicit "encoding and aggregation" strategy that individually encodes each homologue in an MSA, aggregates those encoded features with symmetric functions, and transforms the aggregated features into residue-level structural properties, with feature representations learned automatically within the encoding module.