The problem with suggesting AI could cure all disease within a decade is that, sooner rather than later, people will want to see signs of that future arriving.

Isomorphic Labs has partnerships with pharmaceutical giants, a promising drug-design system in IsoDDE and plenty to say about how much time its AI is already saving. But outsiders still have little basis for judging how well its drug candidates actually work. Its latest website update leaves that question largely unanswered, while its first human trials are now expected by the end of 2026, after slipping from last year. Trust in the mission takes more than another account of what the technology can do.

Speaking of trust: a researcher says his team had already studied the enzyme system Anthropic recently presented as a discovery by Claude. He had also shared unpublished details with the chatbot. Anthropic denies training on those conversations. But scientists should not have to wonder whether asking an AI for help could mean losing control of their own findings. The companies asking researchers to entrust them with unpublished work need to do more to explain, and demonstrate, how that work is protected.

In this issue we also have DeepMind’s watermarks for AI-designed proteins, a $7.5 million raise for Aleph Surgical’s attempt to build an autonomous robot surgeon, and a tool for modelling how hearts beat and lungs move from medical scans.

Let’s dive in.

To make an AI-designed protein, a lab can order DNA containing its instructions. Suppliers screen those orders for dangerous sequences. But AI can produce harmful designs that look unlike known threats.

Google DeepMind's SynthID Bio aims to add another security check: a watermark that identifies the AI tool behind a design. BAIO covered the proposal in July. Now a Nature paper reports laboratory results.

The method steers the model's choice of amino acids, the building blocks of proteins, to embed a detectable pattern. In tests against three targets, watermarked proteins attached about as strongly as unmarked versions.

The watermark is no safety certificate. The researchers also showed that redesigning a sequence could remove it.

The authors suggest that labs could get their DNA orders cleared faster and more cheaply if they design proteins using AI models with verified safety protections. Watermarks would let DNA suppliers recognize designs from those models. That potential saving could attract more labs to use them, giving AI developers a commercial reason to add safeguards and watermarking.

A week ago, we covered Anthropic’s claim that Claude had identified an enzyme system with echoes of CRISPR, the bacterial defense system adapted for gene editing. Now a researcher says his team had already studied that system and shared unpublished details with Claude.

The system occurs in viruses that infect bacteria. Its enzyme was already known; Anthropic’s claim was that Claude spotted its association with a partner gene and repeating DNA sequences resembling a CRISPR array. Human experiments showed that those repeats produce short RNA molecules.

In CRISPR, RNA guides the machinery to specific DNA targets, making it programmable. Anthropic suggested something similar might be happening here - but has not demonstrated that the system is programmable or can edit genes. Its natural function also remains unknown.

Mario Rodríguez Mestre, a computational biologist at the University of Copenhagen, told The New York Times that his team had identified the associated components more than a year earlier. The newspaper reviewed records of his research and conversations with Claude.

“My concern is that this information was used to train future versions of the models,” he said.

Anthropic says Claude was not trained on user conversations and its biology team cannot access them. It also says it knows of no published work describing the system. Mestre acknowledges he cannot establish whether his research influenced Claude.

It is past time for frontier labs to make clear how they use the data people entrust to them. Scientists should be able to work with AI without worrying that unpublished findings might later emerge as someone else’s discovery. Claude may well have rediscovered this system independently. But Anthropic, OpenAI and their peers have more work to do to earn that trust. Clear rules on training, access and reuse, backed by evidence that outsiders can scrutinize, would help. Leaving users to wonder where their work might end up will not.

Alphabet’s Isomorphic Labs was built on the idea that AI could design useful drug molecules with far less laboratory trial and error. In a new post on the company’s website, president Max Jaderberg says that “hypothesis has been proven.”

Not that the post gives outsiders enough detail to assess that claim. It includes two videos illustrating the company’s design process. The workflow starts with Isomorphic’s drug designers specifying what a molecule should act on and which properties it needs, such as dissolving readily or entering cells. AI agents then generate successive batches of molecular designs and evaluate them using predictive models.

The aim is to balance those requirements: a molecule that acts strongly on its target may fall short on other properties. Jaderberg describes the search starting broadly, then concentrating on promising designs while continuing to explore alternatives. Improving the best trade-offs found so far is what he calls pushing the “Pareto frontier.”

Jaderberg says these searches run autonomously on computers over 2–4 days. He compares that with conventional cycles taking 1–3 months per step to design, make and test molecules. Selected AI designs still have to be made and tested in the lab.

The post includes one lab result: increasing the concentration of an AI-designed molecule reduced the activity of its target, which remains unnamed. A separate chart compares its properties with those of a molecule from earlier research, using coloured ratings rather than numerical results. Jaderberg estimates that the earlier molecule took 3–5 years of work, but provides no like-for-like test of the claimed time savings.

Earlier this year Jaderberg told BAIO that first-round results had surpassed 10–20 years of work on some targets. He now says preclinical data are helping prepare for human testing. The latest reported target for first trials is the end of 2026, postponed from 2025, according to Reuters. The new post gives no revised date or animal or human results.

Isomorphic is asking us to accept a sweeping conclusion from a selective disclosure. A paper with methods and data would let outsiders assess the work. Detailed results from its drug programmes would go further: showing how the molecules perform in living systems, and how close they are to testing in people.

Aleph Surgical has raised $7.5 million in pre-seed funding led by Andreessen Horowitz to build robots that can perform surgery autonomously.

“We’re building the first surgeon that scales: one system that can acquire, refine, and transfer surgical skills across millions of procedures,” co-founder Ryan McGuire wrote on X. He says its early system is learning surgical tasks from demonstrations and improving through its own experience.

Aleph is developing both the AI and its own robot, Mark 0. In a March research preview, the company described a model that breaks complex procedures into smaller tasks and keeps the robot moving smoothly while it plans its next actions. The preview included a demonstration of vascular anastomosis: joining blood vessels together.

Aleph’s approach could broaden where a surgical robot learns its skills. The company is designing Mark 0 to draw on training data from humans and other robots, potentially reducing the need to teach every movement through surgical demonstrations.

MONAI Physio turns medical scans into 3D organ models and uses AI to estimate how hearts beat and lungs move. Built for imaging researchers, it includes a demo that predicts lung motion from a single chest CT scan. Start with Tutorial 00, which downloads sample data and a pretrained model. The toolkit is for research and visualization and is not validated for clinical use.

☑️ Basecamp Research raises $140 million to develop AI-designed therapies.

☑️ Red Queen Bio wants AI-designed antibodies to help counter future pandemics.

☑️ HealthFlux combines medical histories and biological measurements to predict future disease.

☑️ MindRank signs a China commercialization deal for its AI-designed obesity pill, still in Phase 3 testing.

☑️ Roche outlines plans for autonomous AI laboratories.

☑️ Synthetic Hospital tests medical AI on fictional patients with complicated histories.

☑️ Biossil reaches unicorn status with its AI approach to rescuing shelved drugs.

☑️ Annovis enlists Weave Bio’s AI to prepare an Alzheimer’s drug submission to the FDA.

☑️ NIH launches Linked Discoveries to help readers find connected research papers.