Mambapore

Mambapore is the name that Kent Kemmish gave, in a message to members of EQBio-OS on March 29, 2026, to the method of a 2025 conference paper that sorts the current signals of a solid-state nanopore with a neural network of the Mamba type.^[1] The paper's authors, who work at Southeast University in Nanjing, China, did not use the name. "They don't call it that in the paper ... but I'm calling it that, because I'm a verbal anarchist", Kemmish wrote.^[1]

"Mamba-Based Deep Learning Method for Solid-State Nanopore Signal Classification" is a five-page paper presented at the 4th International Conference on Biomedical and Intelligent Systems, held in Bologna, Italy, from April 11 to 13, 2025, and published open access on August 6, 2025.^[2] Nanopore events differ in length, because a molecule may cross the pore quickly or slowly, so the authors padded each signal with zeros to a fixed length, smoothed it with a moving average and a Gaussian filter, and cut it into patches that a Mamba encoder reads. Each encoder pairs a Mamba layer with a multi-head attention layer. According to the paper, the shallower encoders pick out local features within each patch, as a convolutional layer would, and the deeper ones keep the relations across the whole signal.^[2]

Mamba, introduced in December 2023, carries a compressed state along a sequence and chooses at each step what to keep, so that its cost grows in proportion to the length of the sequence, where the cost of attention grows with the square of it.^[3] The model was trained in PyTorch on a computer with an NVIDIA RTX 4090D graphics processor, for 200 epochs with a batch size of 512.^[2]

The paper tested the network on a set of signals in eight classes, labelled 000 to 111, with 52,525 samples for training, 2,189 for validation and 3,464 for testing. It reported an accuracy of 97.1 percent, a recall of 97.0 percent and an F1 score of 96.8 percent. Of the methods it compared, the best of five classical ones, gradient boosting, reached 89.7 percent accuracy; the neural baselines reached 94.6 percent (QuipuNet), 94.8 percent (S2Snet) and 96.2 percent (T-S2Net). Its error analysis looks at classes 001 and 010: many 001 samples were labelled 010, and both classes are short, with blurred transitions. The authors write that the model's handling of short signals needs improvement.^[2]

Kemmish's post, "Mambapore", begins with a habit: "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." When he first asked, in 2018, he was told that "none of the best work at that time was yet making use of frontier capabilities". In 2026 the company's designated AI lead repeated the review and found the paper, which Kemmish calls "very recent" and says "does seem to be using 'frontier' techniques". He describes Mamba as a design "which combines the best features of both Recursive[sic] Neural Networks and Transformers", and he thanks the authors for not coining a name of their own: "bless them for not neologizing their method".^[1]

The coinage follows the company's habit of attaching "pore" to a word, as in demonpore, juxtapore, CoronaPore and JeoPOREdy!.^[1]

The paper expands "ONT" as Oxford Nanopore Technologies and says that its data were "generated by an Oxford Nanopore Technologies (ONT) device".^[2] Kemmish objected. The model, he wrote, "isn't public domain", and the paper says little about its data: the authors "don't disclose what device was used", and "digging further, it seems that their dataset is from Bell & Keyser (2016), which ... used an Axopatch device."^[1]

The paper's own reference for the dataset is the 2018 QuipuNet paper, whose authors used "a previously published data set on multiplexed single-molecule protein sensing", that of Bell and Keyser (2016).^[4][5] The records are of DNA carriers with a three-bit barcode, driven through a nanopore one at a time. A 2024 preprint on the same data says the set holds 55,989 samples, with 52,525 allocated to training.^[6] The Mamba paper's training count is the same, and its test count, 3,464, is the remainder.^[2] The company shows the same events in its own viewer, QuipuNet.^[7]

The post turns to a plan. "I see our whole mission as a Corps is to basically start over with data collection", Kemmish wrote. "Navigating confusing proprietary cobweb-strewn datasets without curation and standardization is not the way." It ends: "We'll very soon have those datasets".^[1] The company's own tests on the same barcode data, with whole experiments held out of training, gave a Conformer accuracy of 85.4 percent and a retrained QuipuNet accuracy of 81.0 percent, lower than the published figures. They are described in Machine learning at Molecular Reality.^[7]

• Machine learning on nanopore signals
• Molecular Streaming Corps
• The Nanopore World Model