DLF-Sul

DLF-Sul predicts S-sulfinylation sites in proteins to identify cysteine sulfinylation positions and support analysis of posttranslational modification impacts on cellular processes such as signal transduction, redox homeostasis, and neuronal transmission.

Key Features:

• Multi-module deep learning: Integrates multiple neural network modules to identify S-sulfinylation sites from protein sequences.
• Input features: Extracts binary encoding, BLOSUM62, and amino acid index features as foundational representations.
• BiLSTM sequential features: Uses a bidirectional long short-term memory (BiLSTM) network to derive sequential features that capture dependencies in protein sequences.
• Multi-head self-attention: Applies a multi-head self-attention mechanism to filter irrelevant attributes while preserving critical information.
• Residual connections: Incorporates residual connections to reduce information loss during processing.
• CNN local feature extraction: Employs convolutional neural networks (CNNs) to extract local deep features from protein sequences.
• Fully connected classifier: Uses fully connected layers to map processed features to S-sulfinylation labels.
• Independent test performance: Reported metrics on an independent test set are sensitivity 91.80%, specificity 92.36%, accuracy 92.08%, Matthews correlation coefficient (MCC) 0.8416, and area under the curve (AUC) 96.40%.

Scientific Applications:

• S-sulfinylation site identification: Predicts potential S-sulfinylation (cysteine sulfinylation) positions in protein sequences.
• Posttranslational modification analysis: Facilitates study of S-sulfinylation roles in regulating protein and cellular functions.
• Signaling and redox biology: Enables investigation of S-sulfinylation impacts on signal transduction and redox homeostasis.
• Neurobiology and disease research: Supports exploration of S-sulfinylation involvement in neuronal transmission and the pathology of human diseases.

Methodology:

Extracts binary encoding, BLOSUM62, and amino acid index features; derives sequential features with a BiLSTM; applies multi-head self-attention with residual connections for feature selection; uses CNNs for local deep feature extraction; classifies sites with fully connected layers and evaluates performance on an independent test set.