WalkIm: Compact image-based encoding for high-performance classification of biological sequences using simple...

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Classifying biological sequences (e.g., viral, metagenomics) is challenging.
• Existing methods often focus on neural network architecture, neglecting input encoding and implementation efficiency.
• Current approaches face accuracy and speed limitations.

Purpose of the Study:

• To introduce WalkIm, an image-based encoding method for biological sequence classification.
• To evaluate WalkIm's accuracy and efficiency compared to existing methods.
• To explore WalkIm's compatibility with optical processing for further speed enhancement.

Main Methods:

• Developed WalkIm, an image-based encoding technique for biological sequences.
• Applied WalkIm to a simple neural network for classifying diverse datasets (viral genomes, metagenomics, metabarcoding).
• Investigated WalkIm's performance with optical implementation of convolutional layers.

Main Results:

• WalkIm achieved competitive accuracy and superior efficiency across various biological sequence datasets.
• The method demonstrated consistent performance without dataset-specific parameter tuning or architecture adjustments.
• Achieved near 100% accuracy for classifying high-mutant datasets like Coronaviruses.
• Significantly reduced network complexity and training time, enabling execution on standard desktops.
• Optical implementation of WalkIm reduced training time by up to 500x.
• Preserved sequence structure under various transformations (reverse complement, complement, reverse).

Conclusions:

• WalkIm offers a highly accurate and efficient solution for biological sequence classification.
• The encoding method simplifies deployment and significantly accelerates training, even with optical acceleration.
• WalkIm is a versatile tool applicable to diverse biological sequence data and transformations.