This issue's stories share a thread worth pulling on. Goodfire does not build a DNA model - it wraps interpretability around one that already exists and extracts clinical value. AWS does not train biological AI models - it packages more than 40 of them into a managed service with lab access built in. VCHarness takes existing foundation models as raw material and lets an autonomous system figure out how to combine them into something better. Even dnaHNet, the most traditional model paper in the lineup below, is really about what sits between the raw sequence and the model: the segmentation layer that decides how DNA gets represented.

In AI right now, everyone is talking about the harness. The model reasons. The harness - the system wrapped around it - manages tools, search, feedback, and evaluation. It is an idea that started in software engineering and is now showing up in biology. VCHarness is obviously named after it. But you can see the same logic in every story here.

One more thing. This issue includes something new: FIRSTHAND, a section built from my own reporting. Max Jaderberg does not give many interviews, and what he told me about where Isomorphic Labs is heading - in his own words - is available only to BAIO subscribers. I hope to bring more conversations like this to the newsletter.

When patients get their genome sequenced, the results often include variants no one can interpret. About half the 4.2 million known variants in NIH's ClinVar database are “variants of uncertain significance” - a test found something, but no one can say what it means. Goodfire, a San Francisco interpretability lab, and Mayo Clinic think they have cracked it.

Let's back up a bit. Arc Institute's Evo 2 is a DNA model trained to predict the next nucleotide in a sequence across genomes from bacteria to mammals. It was, however, never shown which human mutations cause disease. But by learning what DNA looks like across thousands of species, it internalized the patterns evolution preserved - sequences that stayed the same because organisms that lost them didn't thrive. Goodfire found a way to extract that knowledge: feed the model a stretch of normal DNA, then feed it the same stretch with a mutation, and compare how the model's internal response changes. A small, simple algorithm - trained on variants already known to be harmful or benign - then learns to read those shifts, picking up patterns that distinguish dangerous mutations from harmless ones. In a preprint this approach outperformed every existing method tested.

It also, intriguingly, explains itself. The system identifies which biological features a variant disrupts and writes a plain-language hypothesis about why. For a BRCA1 mutation, the system spotted that a single DNA letter change had collapsed the signals the cell uses to know where to cut and stitch the gene's instructions together. Without those signals, the gene produces garbled output - which is exactly what published lab work had already confirmed.

Now, the results are so far purely computational. And they are strongest for mutations in protein-coding DNA. Anshul Kundaje, professor of genetics and computer science at Stanford, warned that non-coding regulatory variants - mutations in the stretches of DNA that control genes without coding for proteins - remain “a distant frontier” for DNA language models. And James Zou at Stanford told TIME there is “no guarantee” the model uses biological reasoning rather than statistical shortcuts.

Why it matters: About two million ClinVar variants currently sit in diagnostic limbo - found in patients, but uninterpretable. Goodfire is releasing predictions and mechanistic hypotheses for all 4.2 million ClinVar variants through EVEE, a free tool (more on that below). If the accuracy holds up in clinical validation, genetic counselors get a starting point where today they have none.

Did you know? Goodfire was founded in 2024 and valued at $1.25 billion in February. In January, the company used similar interpretability methods on an epigenetics model to identify a novel class of Alzheimer's biomarkers - DNA fragment length patterns - not previously explored for the disease.

AWS launched Amazon Bio Discovery - a managed platform where scientists design antibody candidates using AI, then send them directly to lab partners for physical testing. The platform offers more than 40 biological AI models. AI agents help select models, identify where a drug could bind to a target protein, and rank candidates. Results from the lab route back for the next design cycle.

Memorial Sloan Kettering used Bio Discovery to design nearly 288,000 novel antibody molecules for a pediatric cancer target. The top 100,000 went to lab screening. Of 116 candidates that cleared initial filters, 46 showed confirmed binding - with affinities reaching sub-nanomolar range (affinities tight enough to be clinically relevant). Work that traditionally takes up to a year took weeks, according to AWS.

Lab partners include Ginkgo Bioworks, Twist Bioscience, and A-Alpha Bio. Pricing is outcome-based - you pay per experiment, not for access. Viewing results and adding collaborators is free.

Why it matters: A major cloud provider packages biological AI models, design agents, and physical lab access into a single managed service. BAIO has covered the lab-in-the-loop idea building across startups - Adaptyv (Issue 13), JURA Bio (Issue 10), Medra (Issue 16) - and Ginkgo Bioworks, a lab partner here, ran 36,000 closed-loop experiments with OpenAI (Issue 1).

Did you know? AWS also announced a wet-lab-validated antibody benchmark database with Johns Hopkins' Gray Lab, built into the platform. More benchmarks and a paper are expected later in 2026.

Building AI models that simulate how cells respond to perturbations - a key step toward virtual cells - currently takes months of expert work. GenBio AI, a Palo Alto startup co-founded by Nobel laureate David Baker and AI scientist Eric Xing, wants to automate that. In a preprint they describe VCHarness, a system that does it.

Give it a dataset and a task - say, predicting which genes change expression after a gene is silenced with CRISPR. An AI coding agent writes and debugs candidate models, drawing on a library of biological foundation models. A search algorithm decides which designs to refine and which to discard. The loop runs autonomously, evaluating hundreds of candidate architectures in days. Across four cell lines, VCHarness outperformed every expert-designed baseline.

What it found is interesting in its own right. Protein-protein interaction networks kept showing up as the strongest building block regardless of cell type - a pattern human designers had not converged on.

Why it matters: One of the big ideas in AI right now is the “harness” - the system wrapped around a model that manages search, feedback, tools, and evaluation. The model reasons; the harness does the rest. VCHarness applies this to biology. If the harness matters more than any single foundation model, then virtual cell progress does not have to wait for the next bigger model. It can come from building smarter systems that combine existing ones - a different theory of how this field advances.

Did you know? The AI coding agent inside VCHarness is Claude Sonnet 4.6. It writes, debugs, and iterates on model code autonomously - equipped with roughly 100 built-in skills for virtual cell development. GenBio is hiring.

DNA foundation models face a trade-off. To process long genomes efficiently, most split the sequence into fixed chunks - but those cuts often land in the middle of biologically meaningful units. Process one letter at a time and computing costs are punishing.

A team from the University of Toronto, Arc Institute, and Cartesia AI offers a third option: let the model learn where to split. Their model, dnaHNet, figures out its own segmentation during training. The researchers told the first stage to group letters in threes - matching the triplet structure of the genetic code - but the model figured out where each group should start and end, landing on codon boundaries, where one amino acid instruction ends and the next begins. No one told it what a codon is. A second stage, on its own, learned to mark where genes begin and the gaps between them.

Why it matters: dnaHNet runs three to six times faster using a fraction of the memory compared to StripedHyena2, the architecture behind Evo 2. But so far it has only been trained on bacterial genomes, far simpler than those of animals and plants.

Did you know? dnaHNet is built on H-Net, an architecture from Albert Gu's company Cartesia AI. Gu also created Mamba, the architecture family that Evo 2's StripedHyena2 descends from - making dnaHNet and Evo 2 architectural cousins from the same lineage.

Sometimes my reporting for Ny Teknik and my work on BAIO converge. I recently sat down with Max Jaderberg, who runs Isomorphic Labs - the Alphabet-owned company that spun out of DeepMind to, in their words, “solve” all disease using AI. The full conversation is in Swedish at Ny Teknik. Here I have picked out five things he said - in his own words - that I think BAIO readers will find worth knowing.

1. The timeline is wild - and they mean it. Demis Hassabis has said Isomorphic could solve all diseases within a decade. Jaderberg does not back away: “Our mission is to solve all disease. It's kind of a crazy mission, but we say it with a straight face because we really believe that this is possible.” His pragmatic hedge: even if they have not done it in ten years, he is certain they will see a clear path. “I do believe we will be able to see that at that point in time for sure.”

2. They admit the data is not there yet. When I asked about the data challenge, Jaderberg was blunt: “Do we already have the data for this? Almost certainly no. But the data might not exist today, but that doesn't mean we can't create it.” He pointed to new data-generating technologies and robotics entering wet labs over the next five years.

3. They have generative design tools the public has not seen. The IsoDDE paper is “the tip of the iceberg,” Jaderberg said. Internally, Isomorphic runs “very powerful generative models, proprietary models, agents” on drug programs that have stalled for a decade or more. “Some of these problems are things that people have worked on for 10 years, 20 years sometimes, and we put our IsoDDE on this and in our first round of results we've made more progress than anyone's made in that time period before.”

4. Novartis expanded because of what they saw. After giving Isomorphic some of their hardest problems, Novartis wanted more. “They've had a front row seat to see how the progress has been going. And just a year into the collaboration - yeah, okay, we want to expand the collaboration even further because we have this conviction.”

5. The whole industry needs to mobilize. Jaderberg is candid that Isomorphic alone cannot do it: “This is going to take lots and lots of breakthroughs, not just from us, but for many people. We're going to have to rally the whole industry, maybe make new industries to actually achieve this mission.”

EVEE - the Evo Variant Effect Explorer - is a free tool from Goodfire (see our news story above) that lets you look up any of 4.2 million ClinVar variants and see an AI-generated pathogenicity prediction plus a plain-language hypothesis about why the variant matters. That includes roughly two million variants of uncertain significance - the ones currently sitting in diagnostic limbo.

☑️ AI Versus Eroom's Law.

☑️ Novartis CEO Vas Narasimhan joins Anthropic board.

☑️ An AI system passed peer review. The scientific community isn't ready.