SPROF-GO

SPROF-GO predicts protein functions directly from amino acid sequences using pretrained language model embeddings, self-attention pooling, and homology-based label diffusion without relying on structural or network data.

Key Features:

• Pretrained Language Model Integration: A pretrained language model extracts informative sequence embeddings that capture intricate patterns within protein sequences.
• Self-Attention Pooling Mechanism: Self-attention pooling focuses on crucial residues within sequences to identify regions most informative for function prediction.
• Homology-Based Label Diffusion: Incorporates homology information and label diffusion algorithms that consider overlapping communities of proteins to refine function predictions without direct structural or network data.
• Performance Superiority: Demonstrated improvements of over 14.5%, 27.3%, and 10.1% in the area under the precision-recall curve across three sub-ontology test sets compared to state-of-the-art sequence-based and network-based methods.
• Generalization Capabilities: Generalizes across non-homologous proteins and across species not encountered during training.
• Visualization of Attention Mechanisms: Provides attention-based visualizations to indicate sequence domains contributing most to predicted functions.

Scientific Applications:

• Elucidating Disease Mechanisms: Supports functional characterization of proteins relevant to disease mechanisms by predicting likely functions from sequence alone.
• Identifying Potential Drug Targets: Aids identification and prioritization of candidate drug targets through predicted protein functions.
• Functional Annotation Across Diverse Species: Enables annotation of proteins in species or datasets lacking structural or network information by relying solely on sequence data.

Methodology:

Uses a pretrained language model to extract sequence embeddings, applies self-attention pooling to weight informative residues, and refines predictions via homology-based label diffusion across overlapping protein communities.