HiFun

HiFun predicts protein functions from amino acid sequences without relying on sequence homology by using a deep-learning self-attention model that interprets amino acid sequences as a "protein language" to enable annotation in metagenomics and metatranscriptomics.

Key Features:

• Homology Independent Annotation: Predicts functions for novel proteins with low or no similarity to existing protein databases.
• Protein-language Modeling: Interprets amino acid sequences as a "protein language" to generate representations relevant to function.
• Deep-Learning Self-Attention: Employs a self-attention mechanism within a deep-learning framework to capture latent structural features related to function.
• Benchmarking and Validation: Evaluated using datasets and metrics from the Critical Assessment of Functional Annotation (CAFA) 3 challenge.
• Large-Scale Application: Applied to annotate over 2 million previously unknown proteins across diverse ecological niches.
• Novel Motif Discovery: Facilitated discovery of novel motifs within the UHGP-50 catalog.

Scientific Applications:

• Metagenomics: Functional annotation of novel proteins recovered from metagenomic sequencing datasets.
• Metatranscriptomics: Annotation of proteins derived from metatranscriptomic data to inform gene expression studies.
• Environmental Microbiology and Microbial Ecology: Investigating microbial diversity, adaptation mechanisms, and functional potential across ecological niches.
• Study of Organisms with Limited Genomic Data: Enabling functional inference for proteins from organisms lacking homologous sequences or comprehensive genomic references.

Methodology:

HiFun reassembles protein sequences into a "protein language" and utilizes a self-attention deep-learning model to identify function-related structural features for annotation.