Deep-AVPiden

Deep-AVPiden predicts antiviral peptides from amino acid sequences using a deep learning Temporal Convolutional Network to accelerate discovery of antiviral therapeutics.

Key Features:

• Temporal Convolutional Network (TCN) Architecture: Employs a TCN to capture temporal dependencies and patterns in amino acid sequences.
• Word Embedding Layer: Transforms amino acid sequences into dense vector representations for model input.
• Spatial Dropout Layer: Applies spatial dropout to reduce overfitting by dropping entire feature maps during training.
• Two TCN Blocks: Processes embedded sequences through two TCN blocks to extract sequence features.
• Global Average Pooling Layer: Aggregates features across sequence length to reduce dimensionality while preserving information.
• Dense and Dropout Layers: Uses dense layers with dropout to add non-linearity and mitigate overfitting in prediction stages.
• Deep-AVPiden (DS) Variant: Implements point-wise separable convolutions in a computationally efficient variant to reduce computational load while maintaining performance.

Scientific Applications:

• Initial screening of therapeutic molecules: Enables rapid in silico screening of candidate antiviral peptides from sequence data.
• Identification of novel AVPs in proteomes: Applied to discover putative antiviral peptides within the proteomes of plants, mammals, and fishes.
• Experimental follow-up: Provides candidates for chemical synthesis and laboratory testing to validate antiviral activity.

Methodology:

Model architecture comprises a word embedding layer, spatial dropout, two TCN blocks, global average pooling, and dense/dropout layers; the DS variant uses point-wise separable convolutions; models were evaluated achieving 90% and 88% accuracy for the original and DS variants respectively with 90% precision for both, validated by Student's t-test and compared against existing state-of-the-art classifiers.