DDcCNN

DDcCNN predicts protein solubility from amino acid sequences using sequence feature fusion and a deep dual-channel convolutional neural network for improved solubility classification.

Key Features:

• Sequence Feature Fusion: CD-HIT is used to reduce redundancy and four subsequences are constructed from each protein sequence: one global subsequence (the entire protein) and three local subsequences generated by a sliding-window technique governed by specific rules.
• Feature Extraction: A G-gap algorithm extracts features from the subsequences to form a mixed matrix input for one channel, while features from the SCRATCH tool are provided to the other channel with a single convolutional layer.
• Deep Dual-Channel Convolutional Neural Network (DDcCNN): The architecture comprises two parallel channels—one with three convolutional layers and one with a single convolution—whose outputs are concatenated and passed to a hidden layer to produce the solubility prediction.

Scientific Applications:

• Protein solubility screening for pharmaceutical research: Enables preselection of soluble targets and guides solubility modification efforts to reduce production costs and improve efficiency in drug development.

Methodology:

Redundancy reduction with CD-HIT; construction of one global and three local subsequences via sliding window; feature extraction using the G-gap algorithm and SCRATCH; processing by a dual-channel CNN (three-layer and single-layer convolutional branches) with concatenation and a hidden layer for prediction; evaluated in comparative experiments against GCNN, LCNN, and PCNN reporting accuracy 77.82%, MCC 0.57, sensitivity 76.13%, and specificity 79.32%.