Adaptive basecalling at the edge
Why existing cloud callers fail at clinical point-of-care — and how Nanolix solves it from first principles.
The raw signal problem
Nanopore sequencing measures the disruption in ionic current as a DNA strand passes through a protein pore. The current trace — called the squiggle — is noisy by design. Three sources dominate clinical error rates:
- Homopolymer runs — repeated bases (AAAA, CCCC) produce near-identical current levels. Naive callers systematically miscall run length.
- Modified bases — methylation and other modifications alter the current signal. Clinical isolates carry modification patterns absent from reference genome training data.
- Current noise — thermal fluctuation and pore chemistry variation add stochastic noise to every signal window. Standard callers treat windows independently; context-aware callers reduce this systematically.
In a research lab, you compensate by running longer and averaging across more reads. In a clinical point-of-care setting, you have approximately 40,000 reads and 45 minutes.
The adaptive basecalling engine
Four design decisions that make local, clinical-grade nanopore analysis possible.
Reads are processed as they arrive from the pore — no batch collection, no waiting for run completion. Pathogen calling begins within the first few thousand reads.
The adaptive model conditions on preceding bases and signal context to reduce systematic errors in homopolymer regions and modified-base-rich sequences — the exact failure modes of standard callers on clinical isolates.
Pathogen-optimized calling models are trained on clinical isolates, not reference genomes alone. The library covers bacteria, fungi, and respiratory viruses commonly encountered in hospital microbiology.
Neural network inference runs on on-device GPU or NPU. No data leaves the workstation. CPU fallback is available for facilities without dedicated GPU hardware — with longer run times.
On published benchmarks
Our internal validation results are available to qualified laboratory partners under NDA. We do not publish benchmark numbers without laboratory co-validation context — this is consistent with responsible clinical software practice.
Benchmark numbers published without co-validation context can mislead clinical buyers about real-world performance in their specific patient population and instrument configuration. We prefer direct dialogue with your team.
Request technical documentationThe full processing pipeline
From raw ionic current to clinical report output — every stage runs locally.