Lab Newsletter — September 7, 2026: Reading the Genome, Base by Base

Here is a fact that’s easy to forget: you never actually read a genome. A sequencer doesn’t hand you A, C, G, T — it hands you a noisy electrical trace as a molecule squeezes through a pore, or billions of short fragments each stamped with errors. Somewhere between that raw physical signal and a clean list of an individual’s genetic variants sits an algorithm doing an enormous amount of inference. This week we’ve mapped the cell in layers, in circuits, and by designing its parts — all of which quietly assume you already have the genome. Today is about the layer underneath all of it: reading the DNA itself, and how deep learning became the thing that does the reading.

🖼️ Variant calling as a vision problem

The reframing that reset the field came from DeepVariant (Poplin … DePristo, Nature Biotechnology, 2018). It names the difficulty exactly: “despite rapid advances in sequencing technologies, accurately calling genetic variants present in an individual genome from billions of short, errorful sequence reads remains challenging.” The trick was to stop treating it as a statistics problem and start treating it as seeing: the authors showed “a deep convolutional neural network can call genetic variation in aligned next-generation sequencing read data by learning statistical relationships between images of read pileups around putative variant and true genotype calls.” Turn the stacked reads into a picture, and let a CNN — the same architecture that reads microscopy — spot the variant. The payoff was blunt: DeepVariant “outperforms existing state-of-the-art tools,” and “the learned model generalizes across genome builds and mammalian species,” so a project on any organism can borrow the wealth of human ground-truth data.

〰️ Reading the raw signal, end to end

DeepVariant starts from reads that a basecaller already produced — and that basecalling step is itself an AI problem, especially for nanopore sequencing. Chiron (Teng … Coin, GigaScience, 2018) tackled it head-on: nanopore sequencing “offers faster and cheaper sequencing than other approaches. However, accurately deciphering the DNA sequence from the noisy and complex electrical signal is challenging.” Chiron was “the first deep learning model to achieve end-to-end basecalling and directly translate the raw signal to DNA sequence without the error-prone segmentation step” — no hand-built intermediate stages, just signal in and bases out. Remarkably, “trained with only a small set of 4,000 reads … our model provides state-of-the-art basecalling accuracy, even on previously unseen species,” at “more than 2,000 bases per second” on a desktop GPU. This is the true first mile of reading a genome: the physical molecule becomes text.

🎲 Calling variants when the reads are 5–15% wrong

Short reads are accurate but stumble in repetitive regions; single-molecule long reads reach those regions but are individually noisy. Clairvoyante (Luo … Schatz, Nature Communications, 2019) confronted that trade-off directly: variant identification “is particularly difficult for single molecule sequencing, which has a per-nucleotide error rate of ~5-15%.” Their answer was “a multi-task five-layer convolutional neural network model for predicting variant type (SNP or indel), zygosity, alternative allele and indel length from aligned reads” — one network, several questions at once. And it earned its keep: the authors “present 3,135 variants that are missed using Illumina but supported independently by both PacBio and Oxford Nanopore reads.” Genome you couldn’t see before, made visible — and the tool is open source.

🗺️ Into the genome’s dark regions

Long reads have a superpower beyond reach: they carry phase — which variants sit together on the same chromosome copy. PEPPER-Margin-DeepVariant (Shafin … Paten, Nature Methods, 2021) built that into the caller. Long-read sequencing, they note, “has the potential to transform variant detection by reaching currently difficult-to-map regions and routinely linking together adjacent variations to enable read-based phasing,” and their “haplotype-aware variant calling pipeline … produces state-of-the-art variant calling results with nanopore data.” Crucially, it “produces high-quality single-nucleotide variants in segmental duplications and low-mappability regions where short-read-based genotyping fails” — the parts of the genome that short reads simply cannot resolve. This is how AI helps finish a genome, not just the easy 90%.

🧮 A transformer to polish the reads

The transformer — the architecture behind modern language models — turns out to read DNA too. DeepConsensus (Baid … Carroll, Nature Biotechnology, 2023) applied it to PacBio HiFi sequencing, where a molecule is read in circles and those passes are merged into one accurate consensus. DeepConsensus “uses an alignment-based loss to train a gap-aware transformer-encoder for sequence correction,” and the gains are concrete: “compared to pbccs, DeepConsensus reduces read errors by 42%,” lifting the yield of the highest-quality reads — “at Q30 by 27% and at Q40 by 90%.” Better reads flow straight downstream into better variant calls: a reminder that reading a genome is a pipeline, and AI now improves it at every stage.

📏 The referee: prove it on the hard parts

A field full of clever callers needs a blind, shared test — and genomics built a good one. precisionFDA Truth Challenge V2 (Olson … Zook, Cell Genomics, 2022) “aimed to assess the state of the art of variant calling in challenging genomic regions.” It was properly adversarial: “20 challenge participants applied their variant-calling pipelines and submitted 64 variant call sets” across Illumina, PacBio HiFi and Oxford Nanopore, all scored “following best practices for benchmarking small variants with updated Genome in a Bottle benchmark sets.” Two findings ring familiar to this digest: “graph-based and machine learning methods scoring best for short-read and long-read datasets, respectively,” and “with machine learning approaches, combining multiple sequencing technologies performed particularly well.” It’s the same prove-it discipline we admire everywhere — a public truth set that tells you honestly how far there still is to go.

🧬 Why it’s our kind of problem

Every story we tell rests on this one. A genome language model reads a sequence; a variant-effect predictor asks what a mutation does; a regulatory circuit or a serious virtual cell assumes a genotype it can trust — and all of it depends on the genome and its variants being read correctly in the first place. AI is what does that reading, from noisy signal up. There’s a pattern here we keep meeting: reframe the raw problem so a general architecture fits. DeepVariant makes variant calling a vision task; DeepConsensus makes read correction a transformer task — the same move as seeing molecules as pictures or spectra as language. And the way these tools travel is our ethos exactly: DeepVariant, Chiron, Clairvoyante, PEPPER-Margin-DeepVariant and DeepConsensus are open source; precisionFDA and Genome in a Bottle are the public truth sets — the same publish-the-model-and-the-test spirit behind the BioImage Model Zoo and BioEngine. A variant caller you can call like a service, benchmarked against a shared standard: that’s the foundation the whole omics stack is built on.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (The X/Twitter sweep was skipped again — our news API is out of credits and a Grok-based replacement is wired, awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.