Jump to content
[dismiss]
Molecupedia is a parody encyclopedia written by Molecular Reality Corporation. It is not Wikipedia and is not affiliated with the Wikimedia Foundation. About

Machine learning on nanopore signals

Page protected
From Molecupedia, the encyclopedia of very small things
This is the current revision of this page. The present address (URL) is a permanent link to this version.
Use of statistical learning to interpret the ionic current of nanopores
For the company's own experiments and plans, see Machine learning at Molecular Reality.

Machine learning on nanopore signals is the use of statistical and neural-network methods to interpret the ionic current recorded from a nanopore. A nanopore instrument produces a current trace sampled thousands to millions of times a second, in which each molecule that passes through the pore appears as a brief change, a current blockade, buried in noise. Software has to find these events, decide what molecule produced each one and, in nanopore sequencing, turn a continuous series of current levels into a string of bases, a task called basecalling.

Oxford Nanopore's basecallers have been recurrent neural networks since 2016, and in 2026 the company moved its most accurate model to a transformer. For single events, a convolutional network called QuipuNet raised the share of usable events in a public barcode dataset about fivefold in 2018, and two architectures from speech recognition and long-sequence modelling, the Conformer and Mamba, have since been tried on the same recordings. How the data are divided matters as much as the model: QuipuNet read 94.6% of barcodes correctly on the test set chosen by its authors, and 81.0% when Molecular Reality Corporation retrained it on a split that held out whole experiments.

The signal

A recording is a single number per sample: the current through one pore. A typical event lasts from microseconds to milliseconds and changes the current by a few to a few hundred picoamperes. Around it the baseline drifts, the pore occasionally clogs or opens, bubbles come and go, and noise of several kinds sets the smallest change that can be seen. Sampling rates range from a few kilohertz per channel in commercial sequencers, where a motor enzyme slows each strand to a few hundred bases a second, to hundreds of kilohertz or more in solid-state experiments, where molecules cross in microseconds.

No two solid-state pores are quite alike, and the same pore changes over hours. Every analysis therefore has two problems at once: what is the molecule, and what is this pore doing today. Models that learn the second problem while being trained on the first do well on the data they have seen and badly on new pores, a form of the dataset shift familiar from other fields.

Other recordings made alongside the current are much slower. A critique published on Molecular Reality Corporation's website points out that a camera at 30 frames per second takes one frame every 33,000 µs while the amplifier samples every 5 µs, a factor of 6,600, so video of the flow cell describes the state of the instrument and cannot show a molecule's passage.[1]

Finding events

Main article: Event detection (nanopore)

The first step is to separate events from the open-pore baseline. Simple detectors flag stretches where the current falls more than some multiple of the noise below a running baseline. Change-point methods, such as the cumulative-sum algorithm used in the OpenNanopore software of 2012, also split each event into its internal levels, which matters for molecules that pass in several stages.[2] Each event is then described by a handful of numbers: its depth, its duration or dwell time, its integrated charge, its asymmetry and the time since the previous event. Much of nanopore research before 2018 consisted of histograms of these features.

Classifying events

Features and classical learners

Classical machine learning works on the feature table. In 2003 a group at the University of California, Santa Cruz combined hidden Markov models with support-vector machines to tell apart, from blockades in an alpha-hemolysin pore, which base pair closed the end of a DNA hairpin.[3] Random forests and gradient-boosted trees have since become standard baselines. The same critique gives the baseline its due: for "isolated, clean, well-separated pulses", classical features fed to the gradient-boosting library XGBoost "are the baseline and frequently the winner", and deep networks earn their place on "the hard stuff real samples throw at you", such as overlapping events, drifting baselines and fouled pores.[1]

QuipuNet

Main article: QuipuNet
A schematic current trace of one event: a deep drop with a spike at the start, three small spikes or gaps standing for three bits, a stretch for the binding site, and a spike at the end
Schematic of a barcoded DNA carrier's event, as described for the QuipuNet data: a marker at the start, three bits, a sensing region where a protein may be bound, and a marker at the end

The best-known public benchmark comes from the Keyser group at the University of Cambridge. In 2016 it built DNA carriers whose structure encoded a three-bit barcode in the presence or absence of groups of small DNA hairpins, with a second region that could bind a target protein. Driven through a solid-state pore, each carrier produced an event in which the hairpins appeared as small extra spikes, and an algorithm assigned the barcode with 94% accuracy.[4]

The algorithm was precise but selective: it labelled fewer than a fifth of the events. In 2018 Karolis Misiunas, Niklas Ermann and Ulrich Keyser trained a one-dimensional convolutional network, QuipuNet, on the raw events. The name refers to the knotted cords of the Inca quipu, which the authors cite because the information in each event sits in extra current drops "much like the knots on a string". Trained on 52,525 events and tested on 3,464 from independent experiments, it read the barcode with 94.6% precision while labelling every event, against 93.7% precision on 18.2% of the events for the original algorithm and 97.8% on 44% for human experts. It roughly quintupled the number of usable events.[5] The code and data are public, and the dataset has become a standard test for new architectures.

Viruses, peptides and proteins

In 2018 a group at Osaka University classified single influenza virions from their pulses in a silicon nitride pore, with 68% accuracy for a single virion and more than 95% when the pulses of more than 20 virions were combined.[6] In 2021 the same approach, which its authors called an artificially intelligent nanopore, told apart four coronaviruses of similar size and detected SARS-CoV-2 in saliva with 90% sensitivity and 96% specificity in five minutes of measurement.[7] Deep networks have also been trained to classify peptides from their blockades in biological pores.[8] See Virus detection with nanopores.

Basecalling

Main article: Basecalling

In strand sequencing, the current at any moment depends on several bases in the pore at once, so a level identifies a short run of bases, a k-mer, and the sequence must be decoded from the order of levels. Early decoders were hidden Markov models. In 2012 one such model decoded measurements in which each current level reflected three base pairs, with a potential accuracy of about 98%, and in 2015 a general toolkit for analysing nanopore data with such models was published.[9][10]

Neural networks replaced them. Oxford Nanopore says its basecallers have been recurrent neural networks since 2016, most recently long short-term memory networks.[11] Academic tools of the same period included DeepNano, a recurrent network for MinION reads, and Chiron, which translated the raw signal directly into bases.[12][13] Most such networks are trained with connectionist temporal classification, a loss function introduced for speech recognition in 2006 that lets a network learn from a sequence of labels without knowing where in the signal each one falls.[14] A benchmark of basecallers in 2019 found that training on data from the organism being sequenced improved consensus accuracy, mostly by removing errors at methylated motifs, and that larger networks were more accurate but slower.[15]

Speech recognition supplied the next architecture as well. The Conformer, a transformer with convolution modules introduced for speech in 2020, was applied to nanopore basecalling the same year in CATCaller, and a transformer with gated convolutions, GCRTcall, was built for RNA in 2024.[16][17][18] In March 2026 Oxford Nanopore described the switch of its most accurate model from long short-term memory to a transformer, with convolutional layers in front and a decoding head of the same family as connectionist temporal classification.[11]

Modified bases are found in the same signal. Methylated cytosine changes the current slightly, and in 2017 a statistical model that compared the measured levels with those expected for methylated and unmethylated sequence called the modification from ordinary sequencing runs.[19]

Decisions in real time

A model fast enough to run during the measurement can act on what it sees. In selective sequencing, introduced in 2016 and often called Read Until or adaptive sampling, software compares the first part of each strand's signal with the sequences of interest and, if the molecule is unwanted, reverses the voltage on that pore to eject it.[20] Later tools mapped the raw signal directly or used fast basecalling to choose molecules across whole genomes.[21][22] Instruments that choose their next measurement in this way are described under active sensing.

Newer architectures

Mamba

Mamba, a state-space architecture introduced in 2023, processes long sequences in time proportional to their length, which suits current traces of thousands to millions of samples.[23] In 2025 a group at Southeast University in Nanjing combined Mamba blocks with multi-head attention to classify solid-state nanopore events into eight barcode classes, reporting 97.1% accuracy.[24][a] Molecular Reality's founder gave the method a name its authors did not: "They don't call it that in the paper ... but I'm calling it that, because I'm a verbal anarchist." See Mambapore.[25]

Self-supervised learning

Labels are the scarce resource. Every labelled event needs a known sample, and the label records only what was put in the cis chamber; any given pulse may have come from something else in the sample. Self-supervised methods learn representations from unlabelled traces first and use labels only at the end. NanoSSL, published at the end of 2025, used an attention-based self-supervised model to identify proteins from nanopore signals.[26]

A joint embedding predictive architecture (JEPA) learns by predicting the internal representation of a hidden part of its input from the visible part, without reconstructing the hidden samples themselves. The approach was proposed by Yann LeCun and first demonstrated on images in 2023; versions for electrocardiograms and for general time series followed.[27][28][29] The method has a known failure, which the critique on Molecular Reality's website describes: without safeguards the predictor "collapses every signal to one constant vector and reports a beautiful, meaningless loss of ~0".[1]

At Molecular Reality

Main article: Machine learning at Molecular Reality

Molecular Reality Corporation publishes its experiments in a public repository. On a split that held out whole experiments, a Conformer classifier trained on the QuipuNet barcode data reached 85.4% test accuracy, against 81.0% for the original QuipuNet network retrained under the same conditions, and both models did worst on the barcode 101, which the repository's notes attribute to that class's sensitivity to differences between experiments. The result is unpublished.[30][b]

Beyond that experiment, the company proposes pretraining a JEPA on its own recordings, a backbone that mixes Mamba, attention and Conformer layers, and what it calls reinforcement learning with molecular feedback, a loop of "machines asking molecular questions and receiving molecular answers". Its public guide describes the nanopore world model that would result as "a planned public domain model, trained on our streamed data", and adds: "The model itself is still ahead of us."[31] The critique of these plans on its website recommends a cheaper first step, to "falsify the core ML claim on free public data for ~$50 of GPU before building a rig", and warns that coarse-grained simulation "smooths out the very ionic currents you need".[1]

See also

Notes

  1. ^ The paper describes its data as an "ONT barcode dataset" that "is generated by an Oxford Nanopore Technologies (ONT) device", with eight barcode classes labelled 000 to 111. Its training and test sets hold 52,525 and 3,464 events, the same numbers as the QuipuNet split of the Cambridge barcode recordings.[24][5]
  2. ^ The 81.0% is the company's re-run of the QuipuNet network on its own split; the 2018 paper reported 94.6% precision on its test set.

References

  1. ^ a b c d "Keltar's Critique v2: The Nanopore World Model (Hardened)". Molecular Reality Corporation. 2026.
  2. ^ Raillon, C.; Granjon, P.; Graf, M.; Steinbock, L. J.; Radenovic, A. (2012). "Fast and automatic processing of multi-level events in nanopore translocation experiments". Nanoscale 4 (16): 4916-4924. doi:10.1039/c2nr30951c.
  3. ^ Winters-Hilt, S.; Vercoutere, W.; DeGuzman, V. S.; Deamer, D.; Akeson, M.; Haussler, D. (2003). "Highly accurate classification of Watson-Crick basepairs on termini of single DNA molecules". Biophysical Journal 84 (2): 967-976. doi:10.1016/S0006-3495(03)74913-3.
  4. ^ Bell, N. A. W.; Keyser, U. F. (2016). "Digitally encoded DNA nanostructures for multiplexed, single-molecule protein sensing with nanopores". Nature Nanotechnology 11 (7): 645-651. doi:10.1038/nnano.2016.50.
  5. ^ a b Misiunas, K.; Ermann, N.; Keyser, U. F. (2018). "QuipuNet: Convolutional neural network for single-molecule nanopore sensing". Nano Letters 18 (6): 4040-4045. doi:10.1021/acs.nanolett.8b01709.
  6. ^ Arima, A.; et al. (2018). "Selective detections of single-viruses using solid-state nanopores". Scientific Reports 8: 16305. doi:10.1038/s41598-018-34665-4.
  7. ^ Taniguchi, M.; et al. (2021). "Combining machine learning and nanopore construction creates an artificial intelligence nanopore for coronavirus detection". Nature Communications 12: 3726. doi:10.1038/s41467-021-24001-2.
  8. ^ Krantz, B. A. (2025). "Deep learning-based classification of peptide analytes from single-channel nanopore translocation events". PLOS ONE 20 (9): e0324777. doi:10.1371/journal.pone.0324777.
  9. ^ Timp, W.; Comer, J.; Aksimentiev, A. (2012). "DNA base-calling from a nanopore using a Viterbi algorithm". Biophysical Journal 102 (10): L37-L39. doi:10.1016/j.bpj.2012.04.009.
  10. ^ Schreiber, J.; Karplus, K. (2015). "Analysis of nanopore data using hidden Markov models". Bioinformatics 31 (12): 1897-1903. doi:10.1093/bioinformatics/btv046.
  11. ^ a b "Transforming basecalling in genomic sequencing". Oxford Nanopore Technologies. March 18, 2026. Retrieved October 25, 2026.
  12. ^ Boža, V.; Brejová, B.; Vinař, T. (2017). "DeepNano: Deep recurrent neural networks for base calling in MinION nanopore reads". PLOS ONE 12 (6): e0178751. doi:10.1371/journal.pone.0178751.
  13. ^ Teng, H.; et al. (2018). "Chiron: translating nanopore raw signal directly into nucleotide sequence using deep learning". GigaScience 7 (5): giy037. doi:10.1093/gigascience/giy037.
  14. ^ Graves, A.; Fernández, S.; Gomez, F.; Schmidhuber, J. (2006). "Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks". pp. 369-376. doi:10.1145/1143844.1143891.
  15. ^ Wick, R. R.; Judd, L. M.; Holt, K. E. (2019). "Performance of neural network basecalling tools for Oxford Nanopore sequencing". Genome Biology 20: 129. doi:10.1186/s13059-019-1727-y.
  16. ^ Gulati, A.; et al. (2020). "Conformer: Convolution-augmented transformer for speech recognition". pp. 5036-5040. doi:10.21437/Interspeech.2020-3015.
  17. ^ Lv, X.; Chen, Z.; Lu, Y.; Yang, Y. (2020). "An end-to-end Oxford Nanopore basecaller using convolution-augmented transformer". pp. 337-342. doi:10.1109/BIBM49941.2020.9313290.
  18. ^ Li, Q.; Sun, C.; Wang, D.; Lou, J. (2024). "GCRTcall: a transformer based basecaller for nanopore RNA sequencing enhanced by gated convolution and relative position embedding via joint loss training". Frontiers in Genetics 15: 1443532. doi:10.3389/fgene.2024.1443532.
  19. ^ Simpson, J. T.; et al. (2017). "Detecting DNA cytosine methylation using nanopore sequencing". Nature Methods 14 (4): 407-410. doi:10.1038/nmeth.4184.
  20. ^ Loose, M.; Malla, S.; Stout, M. (2016). "Real-time selective sequencing using nanopore technology". Nature Methods 13 (9): 751-754. doi:10.1038/nmeth.3930.
  21. ^ Kovaka, S.; Fan, Y.; Ni, B.; Timp, W.; Schatz, M. C. (2021). "Targeted nanopore sequencing by real-time mapping of raw electrical signal with UNCALLED". Nature Biotechnology 39 (4): 431-441. doi:10.1038/s41587-020-0731-9.
  22. ^ Payne, A.; et al. (2021). "Readfish enables targeted nanopore sequencing of gigabase-sized genomes". Nature Biotechnology 39 (4): 442-450. doi:10.1038/s41587-020-00746-x.
  23. ^ Gu and Dao (2023). "Mamba: Linear-time sequence modeling with selective state spaces". arXiv.
  24. ^ a b Xiao, X.; Tao, C.; Pan, R.; Li, J. (2025). "Mamba-based deep learning method for solid-state nanopore signal classification". pp. 99-103. doi:10.1145/3745034.3745051.
  25. ^ Kemmishtree (March 29, 2026). "Mambapore". EQBio-OS.
  26. ^ Xie, Y.; et al. (2026). "NanoSSL: attention mechanism-based self-supervised learning method for protein identification using nanopores". Bioinformatics 42 (1): btaf657. doi:10.1093/bioinformatics/btaf657.
  27. ^ Assran and et al. (2023). "Self-supervised learning from images with a joint-embedding predictive architecture". arXiv.
  28. ^ Weimann and Conrad (2024). "Self-supervised pre-training with joint-embedding predictive architecture boosts ECG classification performance". arXiv.
  29. ^ Ennadir and Golkar (2025). "Joint embeddings go temporal". arXiv.
  30. ^ "ai-experiments". GitHub. April 2026. Retrieved October 25, 2026.
  31. ^ "Keltar's Field Guide to Molecular Reality". Molecular Reality Corporation. August 2026.
Retrieved from "https://molecularreality.com/wiki/Machine_learning_on_nanopore_signals"
Last edited 7 hours ago by Ionophore