ATGO

ATGO predicts protein Gene Ontology (GO) attributes from protein sequences using self-attention transformer-based pre-trained language models combined with triplet neural networks.

Key Features:

• Pre-trained transformer language models: Employs self-attention transformer-based pre-trained language models on protein sequences to extract discriminative functional patterns from feature embeddings.
• Triplet neural network: Uses a triplet neural network architecture to align functional similarity with embedding feature similarity for refined GO prediction.
• GO domain coverage: Predicts GO attributes across molecular function, biological process, and cellular component.
• Benchmarking and validation: Evaluated on 1,068 non-redundant benchmarking proteins and 3,328 targets from the third Critical Assessment of Protein Function Annotation (CAFA) challenge.
• Integration with homology and network scores: Combines model outputs with network scores and complementary homology-based inferences to improve prediction accuracy.
• Improved prediction accuracy: Demonstrates superior performance over existing state-of-the-art approaches in GO prediction as reported in validation studies.

Scientific Applications:

• Protein function annotation: Enables high-accuracy assignment of GO terms to proteins based on sequence-derived embeddings.
• Proteome-scale annotation: Supports large-scale automated protein function annotation from sequence data.
• Drug target characterization: Facilitates elucidation of protein functions relevant to drug design.
• Functional genomics and proteomics: Improves interpretation of molecular functions, biological processes, and cellular localization in genomics and proteomics analyses.

Methodology:

Pre-trained self-attention transformer language models generate protein sequence feature embeddings; a triplet neural network is trained on these embeddings to align functional similarity with feature similarity; model outputs can be combined with network scores and complementary homology-based inference; methods were evaluated on 1,068 non-redundant benchmarking proteins and 3,328 targets from the third Critical Assessment of Protein Function Annotation (CAFA) challenge.