Driving discoveries in biology with AI-powered DNA and RNA sequencing

Decoding the signals of life in real time

Behind breakthroughs like this lies one of the hardest problems in computational biology: basecalling. Every DNA or RNA molecule that passes through a nanopore produces an electrical signal, sampled at 5,000 times per second per pore. This signal must be transcribed into individual nucleotides (represented as the letters A, C, G, and T). With thousands of pores running simultaneously, the challenge is equivalent to transcribing the conversations of an entire town, the size of Oxford, in real time.

Turning raw nanopore signal into accurate DNA and RNA sequences is only possible with advanced machine learning. Dorado, our basecaller, builds on powerful neural network architectures such as long short-term memory (LSTM) networks — originally designed to address the limitations of the first recurrent networks — and transformers, first developed for natural language processing. We adapted these architectures to nanopore data and trained them specifically for the basecalling task, enabling them to process hundreds of millions of signal samples per second with high accuracy.

Our basecalling models are designed from the ground up for efficient execution on NVIDIA accelerated computing. According to scaling laws, increasing model size and complexity typically improves accuracy, but it also raises computational cost and reduces throughput. By focusing on GPU-optimised architectures and kernels (the code that is used to program GPUs), we reduce this overhead — allowing us to capture the accuracy gains predicted by scaling while still maintaining speed.

The rate of progress has been dramatic. As Dr Rasmus Kirkegaard of Aalborg University, Denmark, noted:

‘a decade ago, we were thrilled to get a few long reads with 15% errors. Today, we get well below 1% errors from a single DNA strand and can process more than a trillion bases a day on a single GPU’.

Accuracy and completeness

After the DNA or RNA sequence has been decoded in real time, machine learning continues to drive the analysis, from identifying variants to assembling genomes and detecting epigenetic signals that explain phenotypes.

Our machine learning models can extract far more than canonical bases alone because Oxford Nanopore technology directly sequences native DNA and RNA molecules — no amplification nor enzymatic conversion needed. The platform can measure complex variants with high accuracy and read epigenetic modifications, such as methylation, directly from the raw signal. The result is a detailed portrait of genome function and regulation.

The potential clinical impact of real-time, accurate nanopore sequencing is profound and immediate. For example, researchers at the Broad Institute, US, Dana-Farber Cancer Institute, US, and the German Cancer Research Center (DKFZ) have demonstrated that genome-wide DNA methylation profiling with nanopore sequencing can rapidly classify acute leukaemia samples and resolve subtypes — even reclassifying some ambiguous cases — providing clinically actionable results within hoursⁱ. Plus, recently, clinicians at Guy’s and St Thomas’ NHS Foundation Trust, UK, demonstrated how rapid metagenomic sequencing can identify respiratory pathogens and detect antimicrobial resistance within hours instead of the days taken by current clinical testsⁱⁱ. With a rapid sequencing workflow, treatment plans can be targeted to the infection from the beginning and provide actionable data to inform public health plans early during emerging infectious disease outbreaks.

By sequencing native molecules and embedding machine learning into every stage of analysis, Oxford Nanopore technology delivers rapid, information-rich insights that are helping answer some of the most complex biological problems.

Future innovation

At Oxford Nanopore, we are applying machine learning not only to interpret signals, but also to design the nanopore proteins themselves and the enzymes that drive DNA and RNA through the sequencing pores. This lets us engineer pores with greater stability and richer signal properties than ever before.

We are also using AI to open entirely new avenues, including the direct detection of metabolites and even the de novo sequencing of proteins. By embedding machine learning into both the detectors and the data they generate, the Oxford Nanopore platform continually evolves — expanding the scope of what can be read directly from biology.

The next frontier is connecting genomes to phenotypes: linking genetic variation and epigenetic changes directly to traits and diseases. With nanopore technology streaming native DNA and RNA data in real time — and machine learning models extracting meaning on the spot — this connection is moving from aspiration to routine practice.

Disclaimer: Oxford Nanopore Technologies products are not intended for use for health assessment or to diagnose, treat, mitigate, cure, or prevent any disease or condition.