Last week ended on old biology getting retooled into something new — a paralytic neurotoxin re-engineered to kill cancer cells, a bacterial pore protein turned into a four-in-one molecular sensor. This week sits a step earlier in that same pipeline: not repurposing something already understood, but producing the raw material — the lead, the map, the biomarker — that a future tool or trial will eventually be built from.
Anthropic's new AI-driven biology lab turned roughly 950 autonomous AI agents loose on a database of viral genomes and surfaced a repeated DNA pattern that structurally resembles a CRISPR array — intriguing, and honestly described by the company's own head of life sciences as "a promising lead," not a discovery. A consortium nearly a decade in the making built something closer to infrastructure than a single finding: the largest gene-activity map of the human prefrontal cortex yet assembled, spanning donors from infancy to age 108 and eight separate brain disorders, meant as a resource other studies compare against rather than a standalone result. And a phase 2 trial of the CD47-blocking drug evorpacept in HER2-positive gastric cancer didn't stop at a response-rate win — it built the biomarker case for exactly which patients that win applies to, the groundwork a phase 3 trial needs before it can even start.
Below the three focus stories, Also This Week collects the rest: two new catalogs of the regulatory DNA behind human-specific anatomy, a second AI system this week finding structure in genomic regions nobody had gotten around to labeling — and, strikingly, also centered on a reverse-transcriptase gene of unknown function, the same raw material as this week's lead focus story — three more angles on cancer biology, and the search for the next targets in metabolic and cardiovascular disease.
Anthropic has launched a biology "wet lab" — a research group where human scientists and AI agents work side by side to design and run experiments. This week the company announced one of the team's first finds: a peculiar, repeated pattern of DNA in the genomes of several giant viruses, structurally reminiscent of the arrays that power microbial CRISPR immune systems. It hasn't determined what the pattern actually does.
The method is as much the story as the finding. Anthropic's researchers set roughly 950 autonomous AI agents — systems that often rely on large language models — loose on a database of DNA sequences encoding billions of proteins, directed to search for proteins that work alongside reverse-transcriptase enzymes. The agents explored self-directedly for more than 21 hours, discussing preliminary results and deciding on next steps among themselves. While examining the DNA surrounding a reverse-transcriptase gene in one giant virus, they spotted the same short sequence of DNA letters, repeated over and over.
Those repeats resemble what's inside a bacterial CRISPR array: multiple repeated chunks of DNA, separated by segments copied from viruses or other invaders, which RNA later uses to guide a partner enzyme to cut an intruder's genetic material. But that's where the resemblance currently ends — there's no evidence yet that this newly spotted repeat does anything similar, and no known DNA-slicing partner enzyme has been identified alongside it. "This is a promising lead," said Eric Kauderer-Abrams, Anthropic's head of life sciences. "What's hard is to then go and completely characterize it, understand its function, and develop it into an interesting tool."
Kauderer-Abrams also pushed back on reading this as AI replacing biologists the way AI-generated proofs have unsettled parts of mathematics. Biology, he said, "requires actually running experiments in the physical world" — the agents can search and flag, but a human still has to go characterize what's actually there. The result was posted as a preprint on alphaXiv, not yet peer reviewed.
The headline isn't that this specific repeat will become the next gene-editing tool — it may not. It's that a swarm of AI agents, working self-directedly across billions of sequences and comparing notes with each other, surfaced a genuinely novel structural lead in genomic territory that's been public for years without anyone noticing it. That's a new way of searching, whatever this particular lead turns out to be.
Researchers have produced the largest map to date of gene activity in the human prefrontal cortex — the region supporting planning, decision-making, and behavioral and emotional regulation. The work, published as a collection of eight studies including three in Nature, draws on single-cell RNA sequencing across more than 6.3 million individual brain cells from tissue donated by nearly 1,500 people after death. "The scale of this project and the amount of work required to assemble it are genuinely impressive," says Zhichao Miao, a computational biologist at the Guangzhou National Laboratory who wasn't involved in the work.
The project, run by the NIH-funded PsychAD Consortium, began in 2019. Panos Roussos, director of the Center for Disease Neurogenomics at Mount Sinai's Icahn School of Medicine and a co-author on all eight papers, says the group deliberately focused on one region studied consistently rather than spreading thin across many: donors ranged from infants to a 108-year-old, spanned diverse ancestries, and included both neurotypical controls and people diagnosed with one of eight brain disorders — Alzheimer's disease, dementia with Lewy bodies, Parkinson's disease, vascular dementia, tauopathy, frontotemporal dementia, schizophrenia, and bipolar disorder. "Studying the same region under different conditions enables more consistent comparisons and helps connect our results to existing genetic and molecular studies," Roussos says.
A foundational study within the collection tracked how the dorsolateral prefrontal cortex changes across a lifespan, using healthy brains from people under one year old to age 97. It identified three distinct phases: rapid remodeling during early life, stability from roughly age 24 onward, and a second remodeling phase beginning around age 65 — with different cell types showing distinct gene-activity patterns in each: developmental signals early on, and shifts tied to immune activity, stress responses and circadian rhythms late in life. "This is an important window into brain disease," Roussos says, "but additional regions will be needed to understand the full picture."
This atlas doesn't itself explain what causes Alzheimer's or schizophrenia — that's not what it's for. What it provides is a shared reference: the same brain region, the same techniques, the same consortium, spanning normal aging and eight separate disorders at once. Future studies chasing a specific disease mechanism now have a real baseline to compare against, instead of stitching one together from incompatible smaller datasets.
CD47 is a "don't eat me" signal — a checkpoint protein overexpressed on many tumor cells that blocks macrophages from attacking them, but also expressed on healthy cells throughout the body, which is exactly why earlier anti-CD47 drugs ran into trouble: block CD47 broadly enough to help against a tumor, and you also expose healthy blood cells to the same attack. Evorpacept was engineered around that trade-off. It fuses the D1 domain of SIRPα (CD47's natural binding partner) to a deliberately inactive Fc domain, stripped of the antibody machinery that caused off-tumor toxicity in earlier drugs like magrolimab. The cost of that fix: evorpacept has no independent tumor-killing activity of its own. It only works paired with an active, tumor-targeting antibody — here, trastuzumab.
The phase 2 portion of the ASPEN-06 trial, an international, randomized phase 2/3 study, tested that pairing in 127 patients with HER2-positive advanced or metastatic gastric or gastroesophageal-junction cancer that had progressed on prior anti-HER2 treatment — a second- or third-line setting, across 49 institutions in ten countries. Patients received either evorpacept added to trastuzumab, ramucirumab and paclitaxel (evo+TRP, n=63) or that same three-drug regimen alone (TRP, n=64).
In the intent-to-treat population, the objective response rate was 40.3% with evorpacept added, versus 26.6% for TRP alone. The real story, though, is what a follow-up biomarker analysis found underneath that topline number: restricting to patients with both confirmed retained HER2-positive disease and high tumor CD47 expression (≥5% of cells staining strongly positive, n=48), the response rate climbed to 63.6% versus 23.1%, median duration of response nearly tripled (25.5 vs. 8.4 months), and the risk of disease progression roughly halved (hazard ratio 0.38). Patients who had lost HER2 expression saw no benefit from evorpacept at all, regardless of their CD47 status — the biomarker wasn't incidental, it was load-bearing.
Safety was consistent with what a four-drug chemo-antibody regimen would predict: treatment-emergent adverse events of any grade occurred in 100% of patients in both arms, most commonly neutropenia (69.8% vs. 55.6%), anemia (58.7% vs. 38.1%) and diarrhea (41.3% vs. 36.5%), each modestly more frequent with evorpacept added. Discontinuation rates and the incidence of febrile neutropenia were similar between arms. The authors describe this as the first randomized study to demonstrate both real clinical activity and a manageable safety profile for CD47–SIRPα inhibition — a combination that has tripped up other drugs in this class before.
Earlier anti-CD47 drugs got stuck on a hard trade-off: block CD47 well enough to matter, and you also hit healthy cells hard enough to matter. Evorpacept's inactive-Fc design sidesteps that by giving up independent potency — and this trial's real contribution is showing that trade works when you also select for the right patients by biomarker. That's not a finished drug yet. It's the case a phase 3 trial can now actually be built on.
Two papers this week catalog the regulatory DNA behind traits that make humans look like humans. A study on the gene-regulatory evolution of the human skeleton finds that variants in genes producing extracellular matrix molecules — not the structural genes themselves — could have enabled the skeletal adaptations that distinguish humans from other great apes, using massively parallel reporter assays and human–ape hybrid skeletal cells to test candidate regulatory elements directly. A separate atlas of cell types and regulatory elements underlying human facial diversity combines single-cell transcriptomics, spatial transcriptomics and chromatin accessibility maps of the developing human craniofacial region, aiming to explain both normal facial variation and craniofacial pathology from the same regulatory map. Neither paper claims to have found "the" gene for either trait — both are building the reference catalog a mechanistic answer would eventually be found in.
Two more AI systems this week went looking in genomic territory that's easy to skip past. AlphaFold's database expanded to include protein complexes from common viruses, predictions the researchers behind it say could aid pandemic preparedness — though, like any AlphaFold prediction, they'll need experimental confirmation before anyone treats them as ground truth. More strikingly, this week's BioByte digest covers MINERVA, a tool built by researchers at Arc Institute and Stanford that predicts which positions in a genome structurally interact directly from sequence — without needing a known related gene family to align against, the blind spot that alignment-based methods hit hardest in non-coding DNA and RNA. Scanning 150 bacterial genomes in about 100 minutes, MINERVA flagged 57,787 likely RNA-hairpin-forming regions, nearly 70% of them outside any existing annotation. One signal, inside "UG27" reverse-transcriptase systems embedded in dormant bacterial-virus genomes, was validated in the lab: expressed in E. coli, five UG27 systems each produced short single-stranded DNA molecules mapping precisely to the predicted RNA hairpins, and disabling the reverse transcriptase eliminated the product. What those DNA molecules are actually for is still unknown — a second, independently run AI system this week surfacing an uncharacterized nucleic-acid structure next to a reverse-transcriptase gene, the same raw shape as this issue's lead story.
Pancreatic cancer that's metastasized to the liver was found to reprogram the liver's own hepatocytes into supplying serine, an amino acid the tumor cells can't synthesize on their own once they lack the necessary enzyme — a host-tissue hijack that promotes the metastasis's growth and survival in foreign territory. A separate mechanistic study identifies ZMYND8 as a brake on CD8+ T cell activity: it suppresses IL-2 receptor signalling by blocking a transcriptional activator needed to switch the receptor gene on, enforcing the "terminal exhaustion" state that lets tumors and chronic infections evade a sustained immune response — deleting ZMYND8 pushed T cells toward a more effector-like state and markedly improved both antiviral and antitumor immunity in the studied models, making it a new checkpoint-adjacent target. And Nature Cancer's new review series on next-generation cancer therapeutics takes stock of where the field's newer therapeutic approaches actually stand against the existing clinical landscape, rather than against each other in isolation.
Two pieces this week look past today's best available drugs toward what comes next. Some patients see no meaningful benefit from potent GLP-1 anti-obesity drugs at all, and researchers are now specifically chasing why — understanding non-response could point toward combination therapies or genuinely personalized weight-loss regimens, rather than one drug fitting everyone. And in cardiac disease, a newly characterized atypical G protein–coupled receptor signaling pathway is proposed as a fresh therapeutic target for cardiac fibrosis — the scarring process that stiffens heart tissue after injury and drives progression toward heart failure, and one current drugs don't treat directly.