AI Latest dispatch
Five pharmaceutical companies fine-tuned OpenFold3 on 20,167 private protein–ligand structures without pooling the data. Protein–ligand accuracy jumped from 35.6% to 52.1%. Here is what that means for drug discovery, antiviral design and pandemic preparedness.
Everything else from the field.
Anthropic's Dario Amodei asked rivals to slow AI development and let outside evaluators inside frontier labs. OpenAI and Google DeepMind agreed within days. Here's what's actually being tested, bioweapons risk included.
UK Biobank's Research Analysis Platform reopens this September after a data security incident, and a 2026 study shows why AI is quietly eroding old promises of genetic anonymity.
Anthropic's September 2026 threat report reveals Claude was used in five bioweapons-adjacent research attempts, including a chikungunya gain-of-function request. Here's why virologists and AI safety experts are both worried.
UC Berkeley's GPN-Star, a genomic language model published in Nature, predicts harmful DNA variants better than far larger AI models, and trains in hours not months.
An AI-discovered, AI-designed drug lowered predicted biological age across six independent ageing clocks in a phase 2a trial. What it means, and what it doesn't.
For the first time in an Ebola response, WHO and Flowminder are using anonymised mobile-phone data and modelling to predict where Congo's Bundibugyo outbreak will spread next. An early test called all ten hotspots. Here is what it means for pandemic preparedness.
Google DeepMind's AlphaGenome Atlas precomputes the molecular effects of all 9 billion single-letter human DNA changes and offers them free in a browser. What it does, why it matters, and where it falls short.
Stanford's SPHERE turns locked biomedical datasets into shareable 'synthetic twins' for AI, and opens an Alzheimer's cohort to prove it. The promise, and the caveats.
OpenEvidence's Darwin is the first medical AI to score 100% on MedQA, yet the company is withholding it over biosecurity risk. The claim, and its caveats.
AI in 2026 is moving past ageing clocks that spit out a single number towards foundation models that reason about biological age, predict disease and nominate targets. Here is what changed and why it matters.
German scientists used generative AI to design Synthetic Transfer Vehicles, protein shells that carried RNA into cells more efficiently than the lipid nanoparticles they tested and delivered a CRISPR edit in a pig.
BiteNetI, a new deep-learning model, out-predicts AlphaFold 3 at finding protein metal-ion binding sites, a blind spot for today's protein language models.
Artificial intelligence is moving from the lab into the working machinery of pandemic preparedness, from spotting risky viruses in genetic data to scanning the world's news for early signals. Here is what it can and cannot do in 2026.
Scientists mutated every letter of the ΦX174 phage genome and measured the result. Top AI models predicted it only moderately well, a reality check for AI biology.
A Johns Hopkins team built an AI agent that screened 201,064 studies for a suicide-prevention meta-analysis, cutting human workload by 99% at under half a cent per paper. Here is what the preprint actually shows, and where it falls short.
Google Research has released the planetary prediction engine, an AI agent that builds geospatial models from a plain-language question. It was tested on a live Ebola outbreak. Here is what the numbers actually show, and what they do not.
The first blinded, wet-lab-validated benchmark for AI antibody design just landed in Nature Biotechnology. Here is what the models got right, and wrong.
A new bioRxiv preprint, Prot-LAMBDA, teaches protein language models explicit 3D geometry and reportedly matches a model five times its size. Why efficiency, not scale, is the real story.
A Nature Medicine study trained AI on 25,712 human tissue slides and found that organs keep separate biological clocks. Some of that signal shows up in a blood sample. What it means, and what it does not.
