T4SEpp

T4SEpp predicts bacterial type IV secreted effectors (T4SEs) from protein sequences using full-length embeddings from pre-trained protein language models combined with machine learning to support studies of bacterial pathogenicity.

Key Features:

• Integration of Protein Language Models: Uses full-length embedding features from six distinct pre-trained protein language models to capture sequence and contextual information.
• Homolog Search Module: Searches for full-length homologs of known T4SEs, including signal sequences and effector domains, to identify candidate effectors.
• Signal Sequence Feature Fine-tuning: Fine-tunes a machine learning model specifically on signal sequence features to enhance detection of secretion signals.
• Performance-Driven Model Integration: Selects and incorporates the three best-performing pre-trained protein language models for final prediction.
• High Accuracy and Specificity: Demonstrated approximately 0.98 accuracy and approximately 0.99 specificity on an independent validation dataset in comparative assessments.

Scientific Applications:

• Discovery of novel T4SEs: Enables identification of candidate type IV secreted effectors from bacterial proteomes, exemplified by prediction of 13 T4SEs from Helicobacter pylori including CagA and 12 additional candidates.
• Host–pathogen interaction inference: Supports prediction of effector proteins with potential human protein interactions, with eleven predicted H. pylori effectors inferred to interact with human proteins.
• Pathogenicity and therapeutic research: Facilitates studies of bacterial-host interactions and pathogenic processes that inform development of targeted therapeutic strategies against bacterial infections.

Methodology:

Computational steps include extraction of full-length embedding features from six pre-trained protein language models, integration of those embeddings with machine learning models, a homolog search for full-length matches including signal sequences and effector domains, fine-tuning on signal sequence features, selection of the three best-performing pre-trained models for integration, and validation on an independent dataset yielding reported accuracy and specificity metrics.