Machine learning

Machine learning

Patch-clamp and single-channel electrophysiology concepts are relevant to nanopore measurements. Source: Wikimedia Commons, CC BY-SA 4.0.

Machine learning in nanopore research refers to computational methods that learn patterns from raw electrical signals, basecalled reads, alignments, and quality-control data. It is important because nanopore measurements are long, noisy, context-dependent, and often produced in real time.

Why it matters

Nanopore devices measure changes in ionic current or related signals as molecules pass through or interact with a pore. The same molecule can produce different signal traces depending on pore chemistry, motor speed, temperature, sample preparation, electronics, and neighboring bases. Machine learning helps connect these variable signals to biological or chemical interpretation.

Key uses include:

• Basecalling - converting raw signal into DNA or RNA sequence.
• Modified-base detection - identifying methylation and other nucleotide modifications directly from signal.
• Signal segmentation - detecting events, stalls, adapters, pores, and read boundaries.
• Quality control - predicting read reliability, pore health, contamination, and run performance.
• Adaptive sampling - deciding in real time whether to continue sequencing a molecule.
• Variant and consensus analysis - combining long reads and model-based error profiles.

Representative tasks

Task	Typical input	Example output	Why it is difficult
Basecalling	Raw current signal or POD5/FAST5 signal	FASTQ reads with quality scores	Signal speed varies and neighboring bases affect current levels.
Modified-base detection	Signal plus reference alignment	5mC, 5hmC, m6A or other modification calls	Modified and unmodified bases can have subtle, context-dependent signal differences.
Adaptive sampling	Streaming signal or early basecalled sequence	Accept/reject decision for a molecule	The decision must be fast enough to act during sequencing.
Event detection	Time-series current trace	Candidate translocation or binding events	Noise, drift, pore variation, and short events can obscure boundaries.
Run monitoring	Sequencing summary, read length, pore state, quality metrics	Alerts and quality summaries	Metrics change over time and vary by protocol.

Notable examples

• DeepNano used recurrent neural networks for early open-source MinION basecalling and showed how learned models could reduce nanopore read errors.
• MinCall explored end-to-end convolutional neural-network basecalling with CTC-style sequence modeling.
• Bonito, Guppy, and Dorado represent modern neural-network basecalling directions used around Oxford Nanopore workflows.
• Dynamic-pooling basecallers such as Heron and Osprey address variable molecule translocation speed by adapting pooling behavior in neural networks.
• Hardware-aware basecalling projects such as RUBICON study how to make neural basecallers faster and less memory-intensive.
• Modified-base tools such as Nanopolish-style signal models and newer neural approaches support direct methylation or modification analysis without bisulfite conversion.
• Adaptive sampling tools use rapid signal or sequence analysis to enrich targets or deplete unwanted molecules during a run.

Good practice

Machine-learning results should be interpreted with experimental context. A model trained on one pore chemistry, kit, organism, base modification, or software version may not transfer perfectly to another. Reports should state the model version, chemistry, training or calibration data, reference genome, thresholds, and validation method where possible.

Common pitfalls

• Treating model output as ground truth without orthogonal validation.
• Comparing tools across different pore chemistries or basecaller versions without controls.
• Ignoring calibration, class imbalance, and false-positive rates in modified-base analysis.
• Reporting only read-level accuracy while ignoring consensus accuracy or downstream task accuracy.
• Using proprietary model names without recording software versions and parameters.

Selected papers and resources

• Boza V, Brejova B, Vinar T. DeepNano: Deep Recurrent Neural Networks for Base Calling in MinION Nanopore Reads. 2016.
• Wick RR, Judd LM, Holt KE. Performance of neural network basecalling tools for Oxford Nanopore sequencing. Genome Biology. 2019.
• Miculinic N, Ratkovic M, Sikic M. MinCall: MinION end-to-end convolutional deep learning basecaller. 2019.
• Kovaka S, Fan Y, Ni B, Timp W, Schatz MC. Targeted nanopore sequencing by real-time mapping of raw electrical signal with UNCALLED. Nature Biotechnology. 2021.
• Boza V, Peresini P, Brejova B, Vinar T. Dynamic Pooling Improves Nanopore Base Calling Accuracy. 2021.
• Singh G et al. RUBICON: A Framework for Designing Efficient Deep Learning-Based Genomic Basecallers. 2022.
• Wang Z et al. Adapting nanopore sequencing basecalling models for modification detection via incremental learning and anomaly detection. Nature Communications. 2024.
• Oxford Nanopore Technologies. Dorado basecaller and related tools.

Related pages

• Nanopore sequencing
• Bioinformatics
• Data analysis
• Event detection
• Noise and filtering