CNN_FunBar: Advanced Learning Technique for Fungi ITS Region Classification.

Area of Science:

• Mycology
• Bioinformatics
• Computational Biology

Background:

• Fungal species identification from metagenomic data presents significant challenges.
• The Internal Transcribed Spacer (ITS) region is a key DNA marker for fungal taxonomy.
• Deep learning offers advanced pattern recognition for large biological datasets.

Purpose of the Study:

• To introduce CNN_FunBar, a convolutional neural network (CNN) for classifying fungal ITS sequences.
• To evaluate the impact of various parameters (kernel size, k-mer size, etc.) on CNN classification performance across taxonomic levels.
• To compare CNN_FunBar against existing machine learning algorithms and fungal identification software.

Main Methods:

• Development and application of CNN_FunBar using UNITE+INSDC reference datasets.
• Systematic assessment of CNN model parameters, including convolution kernel size, filter numbers, and k-mer size.
• Analysis of classification performance across all taxonomic ranks (species to phylum) using balanced datasets.

Main Results:

• CNN models achieved >93% average accuracy in classifying fungal ITS sequences.
• Optimal performance was observed with 6-mer frequency features on balanced datasets (500 sequences/category).
• CNN_FunBar demonstrated superior performance compared to machine learning methods (SVM, KNN, Naïve-Bayes, Random Forest) and existing tools (funbarRF, Mothur, RDP Classifier, SINTAX).

Conclusions:

• CNN_FunBar provides a highly accurate and efficient method for fungal taxonomy classification.
• Deep learning approaches, like CNN_FunBar, are effective for analyzing large-scale metagenomic data.
• This study offers a valuable tool for advancing fungal identification in complex biological samples.