DEEPSMP

DEEPSMP predicts ectodomain shedding events of membrane proteins from protein sequence using an end-to-end deep learning model to identify cleavage and sorting signals relevant to drug target and biomarker research.

Key Features:

• Deep Neural Network Architecture: DEEPSMP employs an end-to-end deep neural network integrating long short-term memory (LSTM) units and an attention mechanism to capture sequence patterns associated with ectodomain shedding.
• Sequence Information Utilization: Encodes evolutionary profiles from protein sequences using Position-Specific Iterated BLAST (PSI-BLAST) against the UniRef50 database.
• Attention Mechanism: Incorporates an attention mechanism to identify sorting signals irrespective of their position in the sequence.
• Regularization and Ensemble Learning: Applies dropout, L2 regularization, and bagging ensemble learning to mitigate overfitting during training.
• Output Layers: Produces predictions via a fully connected dense layer followed by a softmax output layer.
• Performance Metrics: Reported cross-validation accuracy is 81.19% with an AUC of 0.835, and on independent test datasets accuracy 83.14%, sensitivity 84.08%, and specificity 81.63%.

Scientific Applications:

• Membrane protein processing analysis: Predicts ectodomain shedding events to study membrane protein maturation and turnover.
• Drug target prioritization: Identifies membrane proteins with shedding events that may influence therapeutic targeting.
• Biomarker discovery: Supports identification of shedding-related biomarkers for disease research.

Methodology:

Evolutionary profiles are encoded via PSI-BLAST searches against UniRef50, processed through LSTM units with memory cells and an attention mechanism, passed to a fully connected dense layer with a softmax output, and trained with dropout, L2 regularization, and bagging ensemble learning.