COVID-DeepPredictor

COVID-DeepPredictor predicts and identifies SARS-CoV-2 and other pathogenic virus genomic sequences using deep learning on alignment-free k-mer–based sequence embeddings for viral classification.

Key Features:

• Deep Learning Framework: Employs a Recurrent Neural Network (RNN) architecture with Long Short-Term Memory (LSTM) units for sequence classification.
• Alignment-Free k-mer Processing: Uses a k-mer technique to generate Bag-of-Descriptors (BoDs) and refine them into Bag-of-Unique-Descriptors (BoUDs) to produce embedded representations of viral sequences.
• Validation and Testing: Evaluated with K-fold cross-validation achieving 100% accuracy on the validation dataset and test dataset accuracies ranging from 99.51% to 99.94%.
• Comparative Analysis: Compared against Linear Discriminant Analysis, Random Forests, and Gradient Boosting Method and reported superior accuracy.
• Optimization Studies: Includes comparative studies to optimize the k value for k-mers, to compare BoDs versus BoUDs, and to compare performance against Nucleotide BLAST.

Scientific Applications:

• Pathogen identification: Classifies sequences from SARS-CoV-2, SARS-CoV-1, MERS-CoV, Ebola, Dengue, and Influenza for pathogen identification.
• Early diagnosis and surveillance: Supports early diagnostic classification and surveillance of viral sequences based on genomic information.
• Monitoring viral evolution: Facilitates monitoring of virus evolution through sequence-based classification.
• Alignment-free analyses: Applicable in research settings where traditional sequence alignment methods are impractical.

Methodology:

Uses an RNN with LSTM units; an alignment-free k-mer approach to generate BoDs and BoUDs and produce embedded sequence representations; K-fold cross-validation for validation; comparative evaluations against Linear Discriminant Analysis, Random Forests, Gradient Boosting Method, and Nucleotide BLAST; and optimization studies to select k and compare BoDs versus BoUDs.