Unsupervised Reference Modeling of Nanopore Signals for DNA/RNA Modification Detection

Area of Science:

• Genomics
• Bioinformatics
• Molecular Biology

Background:

• Nanopore sequencing detects DNA and RNA chemical modifications via ionic current signals.
• Accurate modification detection is hindered by limited labeled data and experimental variability.

Purpose of the Study:

• To develop a scalable unsupervised framework for de novo DNA and RNA modification discovery.
• To address challenges in modification detection using nanopore sequencing data.

Main Methods:

• A CNN-Transformer variational autoencoder (VAE) learns reference signal distributions from unmodified sequences.
• Large-scale data training utilizes streaming sampling and k-mer-aware soft balancing.
• Modification evidence is generated by scoring nucleotides with VAE reconstruction error and aggregating read-level signals.

Main Results:

• Models trained on unmodified sequences show strong discrimination on synthetic modified oligonucleotides.
• Performance degrades on cell line samples due to biological noise and heterogeneity.
• Site-level anomaly scores reveal patterns corresponding to known modification-enriched regions despite reduced classification accuracy.

Conclusions:

• Large-scale unsupervised reference modeling is feasible for de novo modification detection.
• Translating models from synthetic data to robust genome-wide detection in biological samples presents significant challenges.