TALE

TALE annotates protein functions from high-throughput amino acid sequence data by applying a transformer-based self-attention model with joint sequence–label embedding to assign hierarchical Gene Ontology terms represented on directed graphs.

Key Features:

• Transformer-based Architecture: Employs transformer self-attention to capture global sequence patterns and improve generalization across novel sequences and species.
• Joint Sequence–Label Embedding: Integrates sequence encodings and hierarchical function labels (Gene Ontology terms on directed graphs) into a joint latent space to support prediction of unseen or rarely annotated functions.
• Enhanced Generalizability: Demonstrates superior annotation performance for sequences with low homology and for proteins from novel species or functions absent in training data.
• Integration with Sequence Similarity Methods (TALE+): Combines TALE with sequence similarity-based approaches to increase accuracy using only sequence data and surpasses methods that use additional network information in two of three Gene Ontology categories.

Scientific Applications:

• Protein Function Annotation: Provides high-throughput computational annotation of protein functions based solely on sequence data as an alternative to experimental methods.
• Novel Species and Functions Discovery: Enables exploration of proteins from novel species and detection of rarely annotated or previously unseen functions through improved generalization.

Methodology:

Uses deep learning with transformer architectures (self-attention) to learn global sequence patterns, constructs joint sequence–label embeddings that combine sequence encodings and hierarchical Gene Ontology label embeddings on directed graphs, and can be combined with sequence similarity-based methods in the TALE+ variant.