Can a neural net sort DNA from the nanopore’s raw current?
SquiggleNet, a ResNet-style 1D CNN, classifies Oxford Nanopore reads from ~1 s of raw current faster than DNA translocates, beating base-calling-plus-alignment and exceeding 90% accuracy on human vs bacterial DNA.
Source
SquiggleNet: real-time, direct classification of nanopore signals
Study at a glance
- Design
- Computational / modelling — SquiggleNet: ResNet-style 1D CNN classifying raw Oxford Nanopore electrical signals (squiggles) vs base-calling + alignment; four mixed human/bacterial datasets
- N
- N=4 · Four experimental mixed human/bacterial datasets (HeLa&Zymo plus three Human&Zymo/metagenome sets); 1 s of signal per read
- Population
- Nanopore electrical traces from human (HeLa/human DNA, LINEs) and bacterial (Zymo mock, respiratory metagenome) samples
- Outcome
- Species-class accuracy, TPR/TNR, speed/memory vs alignment, and generalization to unseen bacteria
Structured fields used in claim comparison tables when every cited study has a complete layer.
What they did
Trained a four-block residual 1D CNN on squiggles, compared it with base-calling then Minimap2 alignment, and tested four mixed human/bacterial datasets including HeLa&Zymo (8 Zymo species) and a respiratory metagenome with unseen bacteria.
What they found
With adaptor/barcode noise removed, accuracy reached 89.35% (90% TPR, 86.9% TNR). Human vs bacterial DNA exceeded 90% accuracy; the respiratory metagenome hit 90.8% overall (72.5% TPR, 90.9% TNR, AUROC 0.817) despite unseen species. The model used ~10× less memory than alignment and ran faster than pore translocation.
The limits
What it doesn't show
A binary human-vs-bacteria classifier is not full species ID or clinical diagnostic clearance; accuracy drops when the bacterial fraction is tiny, and training still needs labeled squiggles.
Key terms
- Squiggle
- Time series of ionic current as DNA translocates a nanopore.
- Read until / ejection
- Rejecting off-target molecules mid-pore once classified.
- ResNet-style 1D CNN
- SquiggleNet’s architecture: residual bottleneck blocks over the current trace.
- HeLa&Zymo
- Training mix of HeLa DNA plus an 8-species Zymo bacterial standard.
- TPR / TNR
- True positive (bacterial) and true negative (human) rates.
- LINEs
- Human long interspersed repeats that SquiggleNet still classified as human.
Flashcards
Research intelligence for this paper
See its role on concept claims, tensions it is part of, placement history, and related discoveries.
Quiz yourself
SquiggleNet classifies from:
Common questions
How much signal is needed?
About 1 second of sequencing data per read.
Human vs bacterial accuracy?
Over 90%.
Unseen bacteria in a metagenome?
90.8% overall accuracy (72.5% TPR, 90.9% TNR).
Architecture?
1D convolutions with ResNet-style residual blocks.
More on Gene expression