Bacterial classification with convolutional neural networks based on different data reduction layers.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate DNA sequence classification is crucial for understanding bacterial taxonomy and function.
• Traditional methods face challenges with high dimensionality and computational efficiency.
• Convolutional Neural Networks (CNNs) show promise but require optimized sequence representations.

Purpose of the Study:

• To develop a more efficient and accurate method for classifying bacterial DNA sequences using CNNs.
• To address the dimensionality problem in CNNs by introducing a novel custom layer.
• To improve the speed and accuracy of bacterial classification at various taxonomic levels.

Main Methods:

• Utilized Frequency Chaos Game Representation (FCGR) to convert DNA sequences into image-like representations.
• Employed a pre-trained CNN architecture for image classification.
• Replaced traditional pooling layers with a Random Projection (RP) layer with an activation function for feature reduction.

Main Results:

• The proposed CNN with FCGR and RP achieved high accuracy in classifying bacterial DNA sequences.
• The Random Projection layer effectively reduced feature dimensionality while preserving important data.
• A trade-off between classification accuracy and processing time was observed with the RP-based CNN.

Conclusions:

• The integration of FCGR and a custom RP layer offers an efficient and accurate approach for bacterial DNA sequence classification.
• This method provides a viable alternative to traditional pooling layers in CNNs for genomic data.
• Further optimization of the RP layer parameters could potentially balance accuracy and processing speed.