AI-Redesigned Starting Points Made Directed Protein Evolution 79x Better at One Task. Here's Everything Else That Shipped.

Starting laboratory protein evolution from an AI-redesigned protein, instead of a natural one, produced an enzyme 79 times better at its target reaction, a new Broad Institute study finds. Also this week: a real drop in hospital mortality tied to an AI-linked rapid-response system, two new FDA device clearances, a single-cell clustering benchmark that undercuts its own field's favorite deep-learning result, an antimicrobial-peptide benchmark, a Fast Track designation for an AI-designed cancer drug, and a dengue-forecasting transfer-learning model. Plus this week's feature: why an AI can beat clinicians 2.56-to-1 on one task and still score 44.8% on real clinical work. Full breakdown here.

Biotech & structural biology: AI gives directed evolution a better place to start

A team led by David Liu at the Broad Institute found that starting laboratory-directed evolution from a ProteinMPNN-redesigned botulinum-neurotoxin protease, rather than the natural protein, produces an enzyme 79 times better at cleaving ataxin-2, a protein implicated in neurodegeneration (Nature, DOI 10.1038/s41586-026-10820-0, online July 22). The AI-stabilized starting point has "more stability to spare," in senior author Liu's words, so evolution can afford larger functional changes without breaking the protein — and the activity-boosting mutations found along that path didn't transfer back onto the natural protein without destabilizing it, the clearest sign the AI redesign, not just more rounds of evolution, is what unlocked the gain.

HealthTech & MedTech: a real mortality drop, and two new FDA clearances

Across 11 RWJBarnabas Health hospitals, wiring the Epic Deterioration Index — an AI risk score that recalculates every 15 minutes from EHR data — directly into rapid-response-team activation was followed by in-hospital mortality among 23,132 high-risk patients falling from 23.1% to 18.6%, an 18% reduction in risk-adjusted odds of death, as RRT activations rose from 25.3% to 37.5% of high-risk stays (NEJM AI, DOI 10.1056/AIoa2500973, July 29). A companion audit worth flagging for anyone deploying similar tools: a stress-test of standard clinical-fairness-audit pipelines, KAISEN (arXiv 2607.28608, July 30), found the diagnostic step in a typical five-phase audit correctly flagged 144 of 144 controlled bias cases but caught 0 of 48 realistic model-driven cases once proxy variables were misspecified — built and tested on synthetic tasks, not real clinical data, but a caution about auditing fairness by the book.

On devices: the FDA cleared ThinkSono Guidance, AI software letting non-ultrasound-trained clinicians acquire diagnostic-quality compression-ultrasound images for DVT, with 92.9% sensitivity and 97.1% specificity across a 1,691-patient, 24-site study (July 29). Separately, DeepHealth's breast-ultrasound AI became commercially available with a multi-reader study reporting over 98% lesion localization accuracy, an 8-point sensitivity gain for cancer detection, and 37% less radiologist interpretation time (July 30).

Single-cell & proteomics: two benchmarks that complicate their own field's favorite result

A sensitivity-aware benchmark across 9 clustering pipelines and 10 real scRNA-seq datasets (arXiv 2607.25288) found the best-looking deep-learning method — a contrastive autoencoder at a mean ARI of 0.7872 — wasn't statistically confirmed as better than classical baselines, and that learning rate explained more of the score variance (Sobol total effect 0.70) than latent dimensionality did (0.56): the hyperparameter mattered more than the architecture. On the chemistry side, AMPBench-MT(arXiv 2607.25518), a homology-controlled benchmark spanning 161 endpoint-specific model evaluations, found that a model's accuracy at simply recognizing an antimicrobial peptide doesn't predict its accuracy on the assay-level outcomes — potency, hemolysis, selectivity — that actually decide whether a candidate is worth testing further.

Cheminformatics & clinical: a Fast Track designation and a dengue-forecasting model

Insilico Medicine's ISM6331, a pan-TEAD inhibitor designed with its Chemistry42 generative platform, received FDA Fast Track designation for mesothelioma after immunotherapy and chemotherapy (July 29) — three years after nomination and about 18 months after first-patient dosing, with Phase I data now accepted for an ESMO 2026 rapid oral presentation. And a transfer-learning model, TREA-Net (arXiv 2607.26854), improved dengue-outbreak forecasting in newly instrumented Mexico and Malaysia surveillance sites using only 78–104 weeks of local data, beating its baseline in 9 of 10 tested transfer settings and cutting 8-week prediction-interval width by 29.6% in Mexico when paired with conformal prediction.