Last week ended on an open question: researchers had mutated nearly every base of bacteriophage ΦX174, the first genome ever sequenced, and the AI models built to predict exactly that kind of effect struggled. This week follows that question outward, away from the model and toward the patient, and the pattern that emerges is consistent. DeepMind released an atlas scoring all nine billion possible single-nucleotide changes in the human genome, with a second genome-scale model landing days later. Then we ran a DNA language model on the ΦX174 benchmark itself — same rows, same metric, against the same 34 protein predictors.
The benchmark result is more interesting than either a win or a loss. Evo2, which has never seen a protein sequence or a multiple-sequence alignment, ranks 16th of 35 — ahead of every ESM-family model in the comparison, behind the alignment- and structure-aware specialists. It also produces something no protein model can: a score at the nucleotide level, in a genome where genes overlap so densely that three of them have no unambiguous positions at all. That is the honest shape of the prediction side right now. Coverage is enormous, validation is tiny, and the only place the two properly meet is a 5,386-base phage.
Which makes the third story the one that decides whether any of it matters. Three clinical deployments reported this week — a decision-support tool in lung cancer, a randomized mammography-screening trial, an agentic eye clinic in Beijing — and in all three the limiting factor was not model accuracy. It was where the output sits in the workflow, whether the clinician clicks through to it, and whether anyone chose an endpoint that tracks patient benefit rather than algorithmic skill. One of them recovered its physician usage rate from 3.8% to 23% by cutting the interaction down to three clicks. Below the three focus stories, Also This Week collects the rest: what cancer treatment does to healthy tissue, four new accounts of where tumours hide, a drug-delivery collection that has moved from organ to organelle, and genomes read end to end.
DeepMind has released AlphaGenome, a model that predicts the functional consequences of every possible single-nucleotide change across the human genome — roughly nine billion of them. For each variant it estimates effects on gene expression, splicing and chromatin, and the whole thing is pitched as a reference atlas: a lookup table for interpreting the non-coding variation that dominates every association study and that clinical genetics has never had a good story for.
The scale is the headline, and it deserves a moment. Nine billion is not a sample of interesting positions. It is the complete space of single-letter changes you could make to a human genome, scored. Nothing in molecular biology has been annotated at that granularity before, and the non-coding fraction is exactly where the field has been stuck — a GWAS hit lands in a regulatory desert, and the honest answer to “which base is doing the work, and how?” has usually been that nobody knows. It is worth reading alongside a Nature Genetics review published a day earlier, which works through the mechanisms by which small non-coding variants in promoters, enhancers and silencers actually cause monogenic disease — a reminder that “affects expression” covers a lot of distinct causal routes, and that systematically discovering such variants remains an open methodological problem rather than a solved lookup.
It did not arrive alone. The same week, “Predicting genome-wide functional constraints with GPN-Star” describes a genomic language model with a phylogeny-aware architecture that reads whole-genome multiple-alignment data, estimating functional constraint and variant effects across the entire genome and across species. It is positioned as a more scalable and flexible replacement for conservation scores. The two models are doing different things — GPN-Star grounds “this position matters” in evolutionary constraint, AlphaGenome predicts a molecular readout from sequence — but they are aimed at the same clinical bottleneck, and they arrived within days of each other.
Which raises the question this issue keeps circling. A predicted effect on expression is not a prediction of disease. A variant can shift expression measurably and cause nothing; it can cause plenty through a mechanism the model does not represent. And there is no ground truth at anything like nine billion. The largest complete experimental mutagenesis of an entire genome that exists covers about 44,500 variants — in a 5,386-base bacteriophage. That is the gap.
Genome-scale variant scoring is now real and there are at least two credible implementations of it. What has not scaled is the measurement side, and that asymmetry decides how these atlases should be used: as a prioritisation layer that narrows where to look, not as an answer. The test case is the next story — the one genome where a complete measured answer key exists.
Last week's issue covered the ΦX174 result: Wei, Li and Lehner measured the fitness effect of essentially every possible mutation in the phage's genome — 16,098 single-nucleotide and 28,380 single-amino-acid substitutions, 44,478 in total — and benchmarked roughly fifteen variant-effect predictors against the real numbers. Every one of those predictors reads protein sequence, or a multiple-sequence alignment built from related proteins. Not one of them looks at DNA.
That is a conspicuous hole, and it is the one this project fills. Evo2 is a genome-scale DNA language model from the Arc Institute and NVIDIA, StripedHyena 2 architecture, pretrained on 8.8 trillion nucleotides across bacteria, archaea and phage. Scoring it is mechanically simple: for each of the 44,478 designed mutations, take the model's log-likelihood of the mutant genome and subtract its log-likelihood of the wild type. A bigger drop means the model finds that specific change more disruptive. Correlate that delta against measured fitness, gene by gene, using the paper's own per-gene Spearman metric — same rows, same statistic, so the number drops straight into the published ranking table.
At the amino-acid level, Evo2 ranks 16th of 35, mean per-gene ρ=0.257, median 0.329, across 28,380 variants. The interesting part is who sits on either side of it. A model that has never been shown a protein sequence or an alignment beats every ESM-family model in the comparison — ESM1v, ESM2 up to 15B parameters, ESM3, ESMC up to 6B — along with Tranception and PoET. It trails the specialists that get to use extra information: the MSA- and structure-aware methods (ESCOTT, GEMME) and the strongest single-sequence protein language models (VESPA, Progen2).
The nucleotide-level result has no precedent to compare against, because no protein model can be scored on raw DNA at all. Across every gene with an unambiguous reading frame, Evo2's predictions correlate with measured fitness at ρ=0.55–0.65, all p<0.001. Genes B, E and K are absent from that list, and the reason is a property of the organism rather than the analysis: ΦX174's genome is packed so tightly that every base in those three genes also sits inside another gene's reading frame, so no position can be attributed to one gene alone.
Two things had to be fixed before any of it could be trusted, and both are the kind that quietly corrupt a result. The lab's actual wild-type genome differs from the canonical NCBI reference at four fixed positions — a genuine strain difference — so the wild type used for scoring was reconstructed from the fitness data itself rather than downloaded. And the source table's residue numbering turned out to be 0-based for two of the eleven genes and 1-based for the other nine; rather than trust the numbering column, every substitution was located by diffing the mutant protein sequence against the correct gene translation. The compute was less elegant: a Colab A100 preempted about nine times over roughly a day of wall-clock GPU time, survived by checkpointing scored variants to disk every two and a half minutes.
The caveat belongs attached to every number above. This is zero-shot in the sense that the fitness measurements were never a training target, which is the same position every protein model in the comparison is in. But ΦX174 was the first DNA genome ever sequenced, and it is almost certainly somewhere in Evo2's pretraining corpus. This is a test of scoring on a known genome, not of generalisation to an unseen one.
It is a concrete data point for the gap the atlas story leaves open. Reading DNA directly, with no protein representation and no alignment, is already competitive with mid-tier protein-specific methods — which suggests the sequence itself carries more of the signal than the field's architecture choices assume. It also shows what an honest benchmark costs: a known answer key, the original paper's exact metric, and a willingness to report 16th place.
Three clinical AI deployments reported this week, in three specialties, on three continents. What makes them worth reading together is that none of them is really about model accuracy — and two of them say so outright.
Start with the one that looks most like a conventional result. A large international real-world study in non-small cell lung cancer tested a multimodal explainable-AI decision-support tool and reported two things: the model outperformed established biomarkers at predicting immunotherapy outcome, and physicians using it made measurably better decisions. The second finding is the rare one. Most clinical-AI papers establish the first and quietly assume the second follows. The accompanying News & Views is more restrained, noting that pooling many patient-level biomarkers across modalities reliably beats any single one, but that the models degrade across sites — leaving generalisability, not accuracy, as the open problem.
The deepest account of why that gap exists comes from Kristina Lång, who spent nearly a decade on AI in mammography before running the MASAI randomized trial. Her retrospective work had been encouraging — very few screen-detected cancers got low AI risk scores, and a substantial share of interval cancers turned out to be visible in hindsight and correctly flagged. But, as she puts it, retrospective studies “do not tell us whether radiologists will trust it, ignore it, over-rely on it or adapt their behavior in unexpected ways.”
So MASAI used the AI to reorganise the work rather than to do it. The system produced a risk score from 1 to 10, and that score was shown to radiologists before they read the examination: high-risk cases were double-read, lower-risk cases single-read. The reasoning is unusually candid about human factors. Radiologists already carry a bias — they know cancer is rare in a screening population, and rare targets get missed more often — so rather than trying to eliminate that effect, Lång tried to aim it, raising suspicion where disease probability was high and lowering it where probability was low. A second choice went the same way: the AI's detection marks were deliberately withheld until the end of case reading, because eye-tracking work shows that visible marks pull readers toward highlighted regions instead of a holistic search. “This was a behavioral decision rather than a technical one.”
The endpoint choice mattered just as much. Detection rate is the obvious metric and a misleading one: an intervention that preferentially finds indolent lesions scores well while helping nobody, and earlier detection can inflate survival through lead-time bias without moving mortality. MASAI used interval cancer rate instead — cancers appearing after a negative screen but before the next round, which carry worse prognosis and serve as a recognised proxy for screening efficacy. The trial ran as a non-inferiority design with a pre-specified safety analysis at roughly 80,000 women, which found detection had not declined despite a 44% cut in screen-reading workload. The completed trial then reported a 29% increase in cancer detection, driven predominantly by small, node-negative invasive cancers including a higher share of aggressive subtypes, and 12% fewer interval cancers — with recall rate, false positive rate and specificity essentially unchanged.
The Beijing deployment supplies the unglamorous counterpart. AI-TEC is not one model but a coordinated set of agents spanning the whole ophthalmic pathway — pre-consultation, triage, diagnosis, clinician decision support, patient follow-up — deployed into the hospital information system at Beijing Tsinghua Changgung Hospital in November 2025. Its authors argue that bolting AI onto an existing workflow tends to create “an additional information layer that clinicians must now manually review and integrate, creating duplicated effort rather than reducing workload.” Then their own usage numbers made the point for them.
Usage peaked at 25.7% of fundus examinations in December 2025, slid to 3.8% by April 2026, and came back to 23.0% in May after the team sped up loading, trimmed the interaction to three clicks and cut manual input. The model did not change. Their data story runs the same way: RETFound fine-tuned on 26,839 routinely collected images did poorly on the first prospective cohort, and what fixed it was four ophthalmologists hand-selecting 1,426 high-quality images with verified labels, which lifted macro-AUROC from 0.7996 to 0.9383; a further prospective feedback round reached 0.9482. They also name four traps in how this work is normally evaluated: the “silent trial paradox” — trials that prevent clinicians from acting on the AI can only measure detection, never care; the assumption that a hospital diagnosis is ground truth when specialists routinely disagree; the mismatch between disease-centric models and symptom-centric medicine, where the clinician's question is why this person came in; and the “referral fallacy,” that detection is one step in a pathway whose later stages decide whether anything improves.
The week's fourth item is the control case. ProteinTalks, a virtual-cell model built on proteomics rather than single-cell sequencing, was trained on more than 38 million protein measurements from 18 breast-cancer cell lines treated with 63 approved drugs and 59 combinations, sampled before treatment and at 6, 24 and 48 hours. It predicts cellular response to 81 held-out drugs with 88% accuracy, and in biopsies from 501 patients it predicted clinical outcomes of the treatments actually given. For three patients it screened 3,000 compounds and recommended the drugs they had received — plus three with better efficacy in the lab. Impressive numbers, and co-author Tiannan Guo is the one insisting it remains a prototype pending prospective trials. Which is precisely the step the other three papers are about. Meanwhile the Validation Accords is recruiting collaborators for a consensus framework on validating generative AI in medicine, on the observation that real-world validation is frequently skipped altogether, with external validation treated as the ultimate test.
Across all three deployments the binding constraint sits downstream of the model: endpoint choice in MASAI, click count in Beijing, site-to-site generalisation in the lung-cancer study. That is the same lesson the atlas and the benchmark arrive at from the other end — a score is not a finding, and a finding is not a changed decision. Lång's version is the most quotable: the future of medical AI “may depend less on building smarter algorithms than on designing better collaborations between humans and machines.”
Two independent lines of work this week ask what cancer therapy leaves behind in tissue that was never cancerous. Duplex sequencing of normal tissues from childhood-cancer patients shows platinum chemotherapy drives paediatric mutation burdens up to adult levels and leaves a liver-specific mutational signature not seen in other tissues, generating nonsynonymous variants — some potentially leukaemogenic — that may help explain late effects in survivors; an accompanying Perspective reads the damage as accelerated ageing of the organ. From the other direction, a Nature Genetics study of normal oesophagus finds that short-course chemo and radiotherapy before surgery selects pre-existing mutant clones — in DNA damage response, HLA and kinase genes — without leaving any detectable mutational signature at all, meaning the usual signature-based forensics would miss it entirely. The Research Briefing notes that some of the selected mutants are druggable and confer treatment resilience on normal cells. Sequencing healthy tissue from treated patients, in other words, may reveal resistance genes that sequencing the tumour does not.
Four papers, four different hiding places. PLA2G2D-high macrophages in tumour-draining lymph nodes track with poor prognosis and inhibiting the enzyme improves anti-tumour immunity in mice — the draining node, not the tumour, as the site of suppression. A serpin–myeloid axis in pancreatic cancer builds fibrin-rich niches that reprogram macrophages locally and physically exclude T cells, producing spatially organised pockets of evasion rather than a uniformly cold tumour. Chronic type II interferon resolves a long-standing paradox: sustained exposure releases double-stranded mitochondrial RNA into the cytoplasm, triggering a type I interferon response that synergises with IFN-II through prostaglandin synthesis to accelerate tumour growth — the same cytokine, opposite sign, depending on duration. And proximity-guided graph learning defines “proximity antigens” — pairs of surface proteins that sit close together on tumour cells but not healthy ones — identifying EGFR–CDCP1 as a co-target pair for multispecific therapeutics, which turns spatial arrangement rather than expression level into the selectivity mechanism. Separately, a Perspective on PD-1 argues the checkpoint works by dissolving the biomolecular condensates that organise T cell receptor signalling, a mechanism distinct from the phosphatase-recruitment model underpinning current checkpoint-inhibitor thinking. The clinical counterpart to all of this is to intervene before the niche exists: in the randomised phase 2 ImmunoPRISM trial, the BCMA-directed bispecific teclistamab produced higher complete-response rates than lenalidomide–dexamethasone in high-risk smoldering myeloma — treating patients who do not yet have overt disease — though follow-up is not yet long enough to show durable prevention of progression.
Nature ran a drug-delivery Outlook this week, and the collection has a clear direction of travel. The framing piece argues that targeting precision now limits therapeutic benefit more than potency does — a good drug that misses is not a good drug. One feature covers the competing nanoparticle and exosome approaches to the blood–brain barrier, the obstacle that has stalled brain-cancer and neurodegeneration treatment for decades; another goes a level finer, to therapies aimed at mitochondria and other organelles rather than merely the right cell, several of which are now in clinical testing. The collection's least glamorous entry may be the most consequential: a Q&A with health psychologist Rob Horne on medication non-adherence — the largest single source of lost therapeutic benefit in medicine, and one that no amount of delivery engineering addresses.
Complementing the prediction story, several papers this week are simply about reading sequence more completely. A telomere-to-telomere mouse Y chromosome closes all 25 gaps in the GRCm39 reference, corrects structural errors, adds 8.7 Mb of previously unassembled sequence and 142 new genes, maps pseudoautosomal recombination loci and turns up a Y-chromosome dosage-compensation mechanism from X–Y homologous gene expression. On the human side, the T2T-HG002v1.1 benchmark extends variant-calling ground truth into regions that were previously unassessable. And two population-genetics papers push on what ancestry actually explains: a within-family design in a Mexico City cohort — which controls for the population stratification that confounds standard ancestry comparisons — still finds significant ancestry-associated differences in height and type 2 diabetes, while a schizophrenia CNV study in East Asian populations identifies rare copy-number variants and, meta-analysed with European-ancestry data, additional risk loci enriched in loss-of-function-intolerant genes.