Basecalling

Basecalling is the conversion of the raw signal of a nanopore sequencer into letters. The signal is the ionic current through one pore, sampled several thousand times a second, and it changes as a strand of DNA or RNA is pulled through the narrowest part of the pore. The software that reads it began with hidden Markov models, moved to recurrent neural networks in 2016, and in 2026 gave its most accurate model a transformer.^[1]

The current at any moment does not belong to one base. In the early MinION pores the signal was influenced mainly by five or six nucleotides in the pore at once, and in the R9 pores mainly by three central ones with a weaker influence from the rest.^[2] A level therefore belongs to a short word, a k-mer, and there are 4,096 six-letter words. In the public table of expected levels for the R9.4 chemistry they run from 58.5 to 119.7 pA, and the current observed for one k-mer has a standard deviation of about 2 pA. The median level is 91.1 pA, and 342 of the 4,096 levels lie within 2 pA of it.^[3] A single level cannot name its k-mer; the sequence is recovered from the way successive levels overlap. For the R10.4.1 chemistry at 400 bases a second the repository lists 262,144 expected levels, one for each of the nine-letter words.^[3]

The strand does not move at a constant speed, since a motor protein steps it through the pore. In 2018 the maker reported an average of 450 bases a second with the current sampled at 4 kHz, about nine samples for each base, but the number varies from one base to the next.^[2] And the signal does not change within a homopolymer, a run of one letter, so its length has to be read from the time it takes. Because the speed is not constant, many deletion errors in MinION reads fall in homopolymers, and one study found 2.6 times as many deletions in sequences that overlap them.^[2]

The first basecallers did not read the raw signal. Until MinKNOW version 1.9 the control software cut the signal into events by calculating a t statistic over two pairs of adjacent sliding windows, and described each event by its mean, standard deviation and duration; the sequence of events is the squiggle. The maker supplied pore models, tables of the mean and spread expected for each k-mer, and the basecaller found the most probable path through states that stood for k-mers. In the hidden Markov model of the open-source Nanocall, each step from one event to the next could be a step of one position, a skip of more, or a stay in the same k-mer, and the best path was found by Viterbi decoding. Skips of more than one position were not allowed, to save time, and homopolymers longer than the k-mer could not be detected.^[2]

In the early days basecalling was a cloud service run for MinION users by Metrichor Ltd., which was discontinued in March 2017; from August 2016 it was also available inside MinKNOW on the local computer.^[2]

DeepNano, published in 2017, was the first basecaller to use recurrent neural networks. It had no k-mer length built in and could use information from further along the strand, and the maker's own network, Nanonet, used bidirectional long short-term memory units in the same way.^[4][2]

Scrappie, a research basecaller from the maker, was the first to take the raw signal, and in Chiron a convolutional network that finds local shapes feeds a recurrent network, whose output is scored by a connectionist temporal classification decoder and turned into bases by a beam search, with no segmentation between them.^[2][5] Connectionist temporal classification, introduced in 2006 for labelling sequences such as speech, lets a network learn from a sequence of labels without knowing where in the signal each one falls.^[6] The maker's training software for its basecallers says that nanopore signals "exhibit long-range dependencies, much like spoken or written language", and its flip-flop models are trained by a technique "inspired by Connectionist Temporal Classification".^[7]

A network learns from signals whose bases are known, and the software that prepares them does it by "remapping signal to reference sequence".^[7] BasecRAWller, trained on E. coli, did much better on new E. coli data than on human data, and when it was trained on human data its accuracy on the two was more alike.^[2] A benchmark of 2019 found that training on the organism being sequenced raised consensus accuracy mostly by removing errors at methylated motifs.^[8]

The maker's present basecaller, Dorado, is "a high-performance, easy-to-use, open source analysis engine for Oxford Nanopore reads". It runs on NVIDIA graphics processors and on Apple silicon, is built on libtorch, and reads the maker's POD5 format.^[9] The research basecaller Bonito carries the advice to use Dorado for anything but model training or method development.^[10]

The model of highest accuracy now has a transformer. In the maker's words, "Switching our SUP model from LSTM to the transformer architecture has improved a wide range of accuracy metrics". A stack of convolutional layers turns the signal into local features, encoder blocks with self-attention follow, and a head that combines connectionist temporal classification with a conditional random field feeds a beam search that picks the most probable k-mer sequence, which is converted to nucleotides. The recurrent networks stay: they "remain highly competitive at smaller compute budgets", and "high accuracy (HAC) basecalling can keep pace with a fully loaded PromethION 24 across a sequencing run". That instrument produces up to 360 million samples a second, from 24 flow cells of up to 3,000 channels each, and its compute unit has four NVIDIA GPUs.^[1]

The README of the MR1 Nanopore Reader says that its data-ingestion software is to decode, structure and log the real-time electrical data "for machine learning basecalling or browser visualization".^[11] The design samples one channel at 100 kilosamples a second, which is 3,600 times fewer samples than the 360 million a second of a PromethION 24. The company's guide is plain about the state of the field: "sequencing full DNA remains the frontier for everyone in this field", and "No solid state pore has done it commercially yet." It adds that the company has "trained machine learning models on public DNA barcode datasets", a task of classifying whole events, which is a different job from reading bases; see Machine learning on nanopore signals and Machine learning at Molecular Reality.^[12]

• Nanopore sequencing
• Current blockade
• Conformer (neural network)