Matrix-based vector representations in neural networks for classifying molecular biology data

Area of Science:

• Machine Learning
• Bioinformatics

Background:

• Accurate classification relies on appropriate classifier selection.
• Standard classifiers like support vector machines have limitations.
• Neural networks (NNs), especially CNNs and transformers, excel at image data processing.

Purpose of the Study:

• To propose novel NN-based alternatives to standard classifiers.
• To adapt NNs for 1D vector data classification by converting it to 2D matrix representations.
• To leverage NNs pre-trained on large-scale image datasets.

Main Methods:

• Exploration of methods to transform 1D vector data into 2D matrix representations.
• Introduction of a novel data restructuring technique using Wigner transforms.
• Comparison of the proposed method with existing literature approaches.

Main Results:

• Demonstrated effectiveness and robustness of the NN-based approach.
• Consistent strong performance across diverse benchmark datasets.
• Successful application in peptide and DNA barcoding classification tasks.

Conclusions:

• The proposed matrix representation technique enables effective use of image-trained NNs for vector data classification.
• This approach offers a powerful alternative to traditional classifiers.
• The methodology shows promise for various classification challenges in bioinformatics and beyond.