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 at Molecular Reality

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.
Neural-network experiments and plans of an American nanopore company
This article is about the work of one company. For the field in general, see Machine learning on nanopore signals.

Machine learning at Molecular Reality covers the experiments of Molecular Reality Corporation in classifying nanopore current signals with neural networks, and its proposals for larger models trained on data from its sensors. Its one result to date came in April 2026, when the company trained a Conformer, a network design taken from speech recognition, on a public dataset of DNA barcodes read by a solid-state nanopore. The Conformer beat a retrained QuipuNet CNN on a held-out-experiment split, 85.4% vs 81.0% (unpublished).[1][2] The code is public.

The company's plans go much further: a "nanopore world model" trained on streamed sensor data, built from a joint embedding predictive architecture (JEPA) and layers from the Mamba family, and RLMF, in which artificial intelligence systems would learn from molecular measurements. Its own published guide calls these "proposals, not products".[3]

Background

Kent Kemmish has written that "every so often, someone who understands the deep technical underpinnings of machine learning models takes a look at the nanopore signal processing literature for me and reports back". The first such review was in 2018: "I was told that none of the best work at that time was yet making use of frontier capabilities."[4] By 2026 the company's materials made the opposite case in one line: "The machine learning that cracked speech now works on ionic current."[2]

The argument rests on a resemblance. A nanopore produces a one-dimensional time series of current, sampled tens or hundreds of thousands of times a second, in which each molecule leaves a short event whose shape carries its identity; speech recognition turns a similar stream of samples into a sequence of labels. Basecalling, the conversion of current into DNA sequence in commercial nanopore sequencing, has used neural networks for years. The company's interest is in the other branch of the field, single-molecule sensing, where a pore detects and sorts molecules without reading them base by base.

The QuipuNet experiment

Data

The company's test bed is a dataset published in 2016 by Bell and Keyser at the University of Cambridge. They built carrier molecules of DNA with a three-bit barcode, formed by the presence or absence of hairpin loops at three positions, and a site that binds a target protein. Each passage through a nanopore gives an event in which the bits appear as extra dips in the current; the eight possible barcodes allow up to eight targets to be tested in one sample.[5]

In 2018 the same group published QuipuNet, a convolutional neural network (CNN) that classified the events. It assigned barcodes with a precision of 0.946, and because it could classify ambiguous and folded events that the earlier method discarded, it produced about five times as much usable data from the same experiments. Its paper compares the extra dips to "the knots on a string used in the Inca Quipu system".[6][a] The company's website hosts a viewer for the dataset, an oscilloscope for about 84,000 events in which the barcode regions of each trace can be shaded.[2]

Method

The work was committed to a public repository, ai-experiments, between April 19 and April 21, 2026. Its description reads "Experiments replicating existing nanopore/speech applications of novel AI methods with open-source implementations".[1] It contains three steps: a modern copy of QuipuNet's original demonstration, a simplified rerun of the speech experiment from the paper that introduced the Conformer, and a Conformer trained on the barcode events. The Conformer, introduced for speech recognition in 2020, combines the attention layers of a transformer with convolutions; the smallest model in the original paper had about 10 million parameters.[7] The company's model had three Conformer blocks 32 units wide, a projection of each current sample into eight dimensions, and an average over time feeding the classifier. It was trained for 60 epochs with QuipuNet's own augmentations, which rescale each trace, add noise and stretch some traces in time.[1]

The split decided the result. The test set held out one complete nanopore experiment for each of the eight barcodes, and the development set used for tuning was drawn by experiment as well, so that both models were scored on pores they had not seen during training. QuipuNet was retrained under the same conditions.[1] The repository's README summarizes:

On a grouped-dev, held-out-experiment split, a Conformer classifier outperformed the original QuipuNet CNN overall (85.4% vs 81.0% test accuracy), while both models showed a pronounced failure mode on barcode 101, suggesting that this class is especially sensitive to cross-experiment shift rather than the issue being specific to one architecture.

ai-experiments, README

The improvement cut the error rate from 19.0 percent to 14.6 percent, about a quarter fewer errors. Both figures are lower than published accuracies for the same kind of data: QuipuNet's paper reported a precision of 0.946 on its own test set, and a 2025 conference paper using a Mamba network reported 97.1 percent accuracy on what it called the Oxford Nanopore Technologies dataset.[6][8] The company describes the model as the first Conformer it knows of on single-molecule sensing traces. Conformers had already been used for basecalling nanopore DNA and RNA reads.[9][10] The result has not been published; company records describe a paper in preparation.[2]

The repository keeps its dead ends. A version of the classifier that weighted time steps by attention survives under a one-line comment, "Gets worse performance", and a commit of April 20 restores "the mean pooling classifier because it works better than attention pooling".[1] The README notes that the QuipuNet replication "doesn't need much in the way of resources" and was run on an ordinary processor, its author, Jade O'Neill, having "an AMD GPU and didn't want to install TensorFlow-DirectML just for this."[1]

Mambapore

Main article: Mambapore

In March 2026 the company's AI lead, O'Neill, asked to survey the field again, found the 2025 Mamba paper, from Southeast University in Nanjing. Kemmish posted it to members under a name the paper does not use: "I'm calling it that, because I'm a verbal anarchist." He objected that the trained model was not public and that the paper said little about its data beyond its being "generated by an Oxford Nanopore Technologies" device; on his reading the data were the Bell and Keyser set, recorded with an Axopatch amplifier. His conclusion was that the company's community should "basically start over with data collection", because "Navigating confusing proprietary cobweb-strewn datasets without curation and standardization is not the way."[4]

Plans

Main articles: The Nanopore World Model and RLMF

The company's plans describe a stack in which a time-series JEPA learns from unlabeled current traces by predicting the representation of a hidden stretch of signal from the rest, with layers that mix Mamba and attention, a Conformer, and later inputs from voltage, instrument telemetry and video of the flow cell. The intended central model is called Maxine's Mind.[2] Training data were to come from the company's own live molecular streams. In the words of its RLMF page, "Even the most intelligent AI systems will need to somehow measure the world of molecules", and the aim is "a public domain nanopore world model".[11]

On August 18, 2026, the company published a critique of its own plan, written in the voice of its mascot, Keltar, headed "Transmission v2" and subtitled "Hardened after senior review". It argues that the model should predict in a learned representation instead of reconstructing raw current, because "A 200 kHz current trace is mostly thermal noise, capacitive ringing, and the amplifier's Bessel filter talking to itself." It warns that a JEPA without safeguards "collapses every signal to one constant vector and reports a beautiful, meaningless loss of ~0." A camera recording the flow cell at 30 frames per second takes one frame every 33,000 microseconds while the amplifier samples every 5, a factor of 6,600, so video can supply context but never the identity of a molecule. Before any rig is built, the page asks that the core claim be tested on free public data, which "falsifies the claim for the price of a pizza". It also proposes that an anomaly seen by one device be trusted only when others confirm it, since "a lone anomaly is a bubble".[12]

See also

Notes

  1. ^ The company's own first signals of this kind came from knot sequencing, in which knots in a nylon filament were drawn through a hole about 2 mm wide.

References

  1. ^ a b c d e f "ai-experiments". GitHub. Retrieved October 25, 2026.
  2. ^ a b c d e "Company records". Molecular Reality Corporation. October 2026.
  3. ^ "Keltar's Field Guide to Molecular Reality". Molecular Reality Corporation. August 2026.
  4. ^ a b Kemmishtree (March 29, 2026). "Mambapore". EQBio-OS.
  5. ^ "Digitally encoded DNA nanostructures for multiplexed, single-molecule protein sensing with nanopores". Nature Nanotechnology 11 (7): 645-651. 2016. doi:10.1038/nnano.2016.50.
  6. ^ a b "QuipuNet: Convolutional Neural Network for Single-Molecule Nanopore Sensing". Nano Letters 18 (6): 4040-4045. 2018. doi:10.1021/acs.nanolett.8b01709.
  7. ^ "Conformer: Convolution-augmented Transformer for Speech Recognition". 2020. pp. 5036-5040. doi:10.21437/Interspeech.2020-3015.
  8. ^ "Mamba-Based Deep Learning Method for Solid-State Nanopore Signal Classification". 2025. pp. 99-103. doi:10.1145/3745034.3745051.
  9. ^ "An End-to-end Oxford Nanopore Basecaller Using Convolution-augmented Transformer". 2020. doi:10.1109/bibm49941.2020.9313290.
  10. ^ "GCRTcall: a transformer based basecaller for nanopore RNA sequencing enhanced by gated convolution and relative position embedding via joint loss training". Frontiers in Genetics. 2024. doi:10.3389/fgene.2024.1443532.
  11. ^ "Reinforcement Learning with Molecular Feedback". Molecular Reality Corporation. 2026.
  12. ^ "Keltar's Critique v2: The Nanopore World Model (Hardened)". Molecular Reality Corporation. August 18, 2026.
Retrieved from "https://molecularreality.com/wiki/Machine_learning_at_Molecular_Reality"
Last edited 1 hour ago by PicoampPedant