PROST

PROST detects remote and global protein homologs by comparing iDCT-compressed embeddings produced from the ESM-1b protein language model to improve homolog detection and functional inference.

Key Features:

• Language Model Utilization: Uses the ESM-1b protein language model to generate numeric embeddings that represent protein sequences beyond simple sequence identity.
• Data Compression via iDCT Quantization: Employs discrete cosine transforms (iDCT) to compress protein embeddings while preserving essential information for comparison.
• Enhanced Detection at Lower Sequence Identity Levels: Computes distances between pairs of protein sequence embeddings to identify homologs within the low sequence-identity "twilight zone."
• Global Homology Detection: Focuses on global structural and functional similarity detection rather than local alignment, aiding analyses of allosteric and broad structural relationships.
• Efficiency in Large-Scale Comparisons: Enables whole-genome or proteome-scale comparisons with linear runtime performance.

Scientific Applications:

• Remote Homolog Detection: Increases the number of detected remote homologs across diverse protein sequences by operating effectively at low sequence identity.
• Function Identification: Improves functional assignment and provides new insights for proteins previously lacking assigned functions.
• Large-Scale Comparative Genomics/Proteomics: Supports comprehensive whole-genome and proteome comparisons for broad comparative analyses.

Methodology:

Generates ESM-1b embeddings for protein sequences, applies iDCT (discrete cosine transform) quantization to compress those embeddings, and computes pairwise distances between compressed embeddings for global homology detection.