11.1 Points With a Certified AI Diagnostic Copilot. 11.7 Points With Nothing At All. Here's Everything Else That Shipped.

The ALLIANCE trial randomized 82 physicians to Prof. Valmed — the first LLM-based clinical decision support system in Europe certified as a medical device — or to conventional resources alone. The certified-AI arm improved diagnostic accuracy 11.1 percentage points. The nothing-at-all arm improved 11.7. Below: what that trial actually moved, plus a deployed-label audit, a silent bug in a widely used genomic model, a microbial symbiosis catalog, and the largest public structure-affinity dataset for drug discovery.

Clinical & translational AI: certification didn't move accuracy, it moved speed

The full result is today's feature: 11.1 Points With a Certified AI Diagnostic Copilot. 11.7 Points With Nothing At All. The short version: in a randomized, blinded-assessment trial of Prof. Valmed — Europe's only CE-certified LLM diagnostic copilot — against 82 rheumatology and internal-medicine physicians, top-1 diagnostic accuracy rose from 22.2% to 33.3% with the tool and from 23.3% to 35.0% without it (adjusted OR 0.99, p=0.979). What the tool did move: case-processing time, cut from 206 seconds to 94. And in exploratory analyses, when the tool's top suggestion was wrong, physicians followed it into the wrong diagnosis 95% of the time.

HealthTech: 40.5% of a training label came from accounts that never once flagged impairment

A Peking University Sixth Hospital team audited the routine cognitive-status label behind a deployed community screening programme before training anything on it. 181 data-entry accounts, each with at least 100 diagnoses on file, had recorded zero impairments between them — accounting for 40.5% of the outcome column, with recorded-impairment rate falling monotonically by account volume (15.7% for low-volume accounts to 0.7% for the highest). Against a titled-physician reference standard, no locally fine-tuned language model beat a 21-variable incumbent logistic regression (AUROC 0.926); reinforcement-tuned variants scored worse (GRPO 0.789) than zero-shot. The one recipe that won: distilling a frontier model's labels into a 4B local model reached 0.940 — above both the incumbent and its own teacher — using as few as 50 teacher-labelled cases.

Genomics: a genomic foundation model's fast path was silently wrong

Evo 2, the largest openly available genomic foundation model, can't fit its 40-billion-parameter configuration on a single 80GB accelerator. A new toolkit, TurboQuant-Bio, compresses its weights to 4 bits without calibration data — near-lossless across perplexity, genomic classification, splice-site prediction, and variant-effect prediction — bringing the full model onto one GPU and the 7B variant to its full million-token context, an eightfold reachable-context gain. Along the way, the team found the model's released chunked-prefill inference path — its officially supported fast route through long sequences — was silently returning plausible but uncorrelated likelihoods, not scoring the sequence at all. Their fix brought scoring a complete 580-kilobase bacterial genome from 13.7 hours down to 22 minutes, correctly.

Genomics: machine learning flags 15–23% of uncultured microbes as likely symbionts

Most microbial life resists culturing, which has kept large swaths of it functionally invisible. A Nature Biotechnology team built symclatron, a machine-learning classifier, and ran it across more than 100,000 uncultivated bacterial and archaeal genomes from environmental metagenomes and reference collections. Result: 15–23% show genomic signatures consistent with host-associated or obligate intracellular symbiosis — present across half of all known bacterial and archaeal phyla — including recurring patterns of lost metabolic function that track with host dependence. The resulting Symbiont Genomes catalog is a resource, not a proof of any specific symbiosis; the paper flags loss-of-function signatures as correlational until followed by direct study.

Structural biology & drug discovery: a fine-tune on real crystal structures beats the benchmark

An Oxford-led consortium spanning the Structural Genomics Consortium, XChem, OpenFold, and MSKCC released the first public tranche of OpenBind, what the team describes as the largest public single-target experimental structure-affinity dataset: 925 crystallographic binding events across 699 compounds and affinity measurements for 601, all against one antiviral target, enteroviral 2A protease. Benchmarking docking, cofolding, and affinity prediction against it surfaced a familiar set of failure modes — pose ranking is unreliable, docking is pocket-conformation-dependent, affinity prediction remains hard — but fine-tuning OpenFold3-p2 on the fragment-screen structures substantially improved pose prediction and virtual screening for related follow-on compounds, a working demonstration of target-specific adaptation from a small, real experimental set.