ResVAE DNA Classification

A hybrid model for DNA sequence multiclassification based on sequence-to-image conversion

DNA sequence classification plays a crucial role in disease diagnosis and prediction in bioinformatics. This project presents ResVAE, a novel hybrid deep learning approach that transforms DNA sequences into images for enhanced classification performance.

Unlike traditional methods that directly process sequence data, our innovative approach converts DNA sequences into visual representations, enabling the application of powerful computer vision techniques to biological sequence analysis.

📄 View Full Paper

Research Overview

The ResVAE model introduces three key innovations that significantly advance DNA sequence classification:

Character Mapping: ATGC nucleotides are systematically mapped to specific numerical values, preserving essential base information while enabling downstream image construction. This fundamental transformation bridges the gap between biological sequences and computer vision processing.

Image Construction: We developed two novel visualization approaches - traditional square histograms and innovative circular histograms. The circular approach divides each character into proportional fan-shaped segments within a 200×200 tangent square, creating intuitive spatial representations that better capture DNA sequence structure.

Hybrid Feature Extraction: Our architecture combines Variational Auto-Encoder (VAE) reconstruction capabilities with ResNet34’s powerful feature extraction. This parallel processing approach leverages VAE’s ability to learn intrinsic data distributions while utilizing ResNet’s proven image classification strengths.

Key Results

Our comprehensive evaluation demonstrates remarkable performance improvements:

  • AUC: 98.6% - Best-in-class area under curve performance
  • Accuracy: 95.8% - Superior classification accuracy
  • Significant advantages over traditional methods including CNN (90.1% AUC), AlexNet (92.2% AUC), VGG16 (97.4% AUC), and ResNet34 (97.7% AUC)

The model effectively addresses traditional limitations in handling non-equal-length sequences, significantly expanding application scenarios in bioinformatics research.

Impact and Applications

This research opens new avenues for RNA and protein sequence analysis, providing a foundation for broader bioinformatics applications. The sequence-to-image paradigm represents a fundamental shift in how we approach biological sequence classification, bridging computer vision and computational biology.

Key Contributions:

  • Novel sequence-to-image transformation methodology
  • Hybrid VAE-ResNet architecture for enhanced feature extraction
  • Solution for non-equal-length sequence classification challenges
  • Extensible framework for RNA and protein sequence analysis

The ResVAE model demonstrates how innovative data representation can unlock the full potential of deep learning in bioinformatics, establishing new benchmarks for DNA sequence classification performance.

A small project at Tsinghua Shenzhen International Graduate School