Science often takes a leap forward when the invisible becomes visible. When we can see it with our own eyes.

Robert Hooke looked through a microscope and gave us the word “cell.” Galileo turned a telescope to the sky and changed humanity’s place in the universe. The DNA double helix became biology’s most famous image because it made heredity physically imaginable.

Biohub’s laser phase plate might belong in that tradition, if it works at the scale its builders hope: a tool for seeing inside the cell in a way biology has wanted for decades.

In this issue, we also follow Claude Fable’s sudden trip from biology controversy to export-control story, Terray’s attempt to give AI-designed molecules a recipe, TITO’s faster molecular simulations, and Lilly’s TuneLab expansion with Tamarind.

Let’s dive in.

Biohub and UC Berkeley have made a laser phase plate work in a modern cryo-EM microscope - a contrast technology that could let scientists see many proteins inside intact cells. Biohub says fewer than 1% of proteins can be imaged in their native cellular environment today. With laser phase plates, scientists believe more than half of functional proteins could become visible.

Cryo-EM (cryogenic electron microscopy) creates images by sending a beam of electrons through frozen biological samples. That works for many purified proteins. But inside cells, many proteins are small and crowded among thousands of other molecules. They do affect the electron beam, but often only faintly - in ways that do not show up as a strong brightness difference on the detector.

A phase plate is a contrast tool: it turns more of that weak, hidden information into visible image contrast.

Physicists have known this basic trick since 1942, when physicist Frits Zernike described phase contrast for light microscopes. That work later earned him the 1953 Nobel Prize in Physics.

But for electron microscopes, the problem has been much harder. Material-based phase plates, including thin carbon films, can sit in the electron path and change how the beam forms an image. But electrons can damage or charge that material, which makes the image unstable or blurry.

More than 15 years ago, UC Berkeley physicist Holger Müller and microscopist Robert Glaeser proposed a cleaner version: use a laser instead of a physical film.

To many, that idea sounded impossible.

Light barely interacts with electrons. To make the effect strong enough, the laser had to be focused to a micron-scale spot inside the microscope, held stable for hours, and bounced between mirrors about 10,000 times until it reached roughly 100 million times the intensity of the Sun’s surface.

At a 2019 microscopy workshop, Müller pitched the idea again. Biohub then spent seven years funding, building and extending it, including a dual-laser version developed at its own institute.

There are now two working phase-plate designs. Berkeley’s single-laser phase plate improved resolution by up to 44% on small purified proteins. Biohub’s crossed laser phase plate uses two beams in an X-shaped setup. It imaged apoferritin at 1.8 angstroms, boosted contrast in frozen E. coli, and is designed to reduce ghost images - faint copies that can obscure the weak biological signal researchers actually want.

“When you can see inside a cell with phase contrast - when you can actually watch proteins interact with other proteins in context - that’s when biology starts to come alive,” says David Agard, founding scientific director of imaging at Biohub, in a blog on the company’s website.

This connects back to two Biohub stories BAIO has covered recently. In late April, we reported on Biohub’s Virtual Biology Initiative - committing $500 million to generate open cellular datasets for predictive models of the cell. Then, a month later, Biohub released its protein world-model work.

Why it matters: Biohub is explicit about the AI link. Better contrast could open a new frontier in structural biology by showing how molecular machines assemble and interact inside cells. And Biohub says the laser phase plate will be critical to generating the detailed images needed for its Virtual Biology Initiative - training data spanning molecules to organisms, in health and disease.

Did you know? Biohub shares tomography data through the CryoET Data Portal.

Claude Fable has been pulled offline only days after Anthropic released it. Anthropic says the US government issued an export-control directive requiring it to suspend access to Fable 5 and Mythos 5 for any foreign national, including foreign-national Anthropic employees. To comply, Anthropic disabled both models for all customers. Other Claude models are not affected.

That sharply escalates the story BAIO covered in our last issue, and in a separate column. Then, the fight was over what Fable would answer: Anthropic had released it publicly, but most biology and chemistry questions were routed away from the strongest model, while Mythos was reserved for vetted users. Now Fable and Mythos are offline, and the decision came from the US government, not Anthropic.

The trigger appears to be cybersecurity, not biology. Anthropic says the government letter gave no specific details, but that its understanding is that officials had seen a possible Fable jailbreak. Axios reports Amazon shared research with the White House showing how it could jailbreak Fable and access parts of Mythos. Anthropic disputes the severity, saying the technique found only a small number of previously known, minor vulnerabilities that other public models could also find.

Semafor reports the White House move was partly linked to concern that a China-linked group may have accessed Mythos, but that has not been established. Anthropic told Semafor that Chinese access was not raised in its export-control discussions, and that access from China is prohibited.

Why it matters: Fable-gate started as a heated debate about biosecurity safeguards and scientific access. It is now also a test case for whether frontier AI models become export-controlled strategic assets. As models become more useful, access may be decided by labs, governments and security agencies before researchers ever reach the prompt box. We are in uncharted territory here.

Did you know? Anthropic says it is working to restore access as soon as possible. Whether that includes biology and chemistry queries remains to be seen.

TerraSynth is a model that designs small molecules together with a recipe for making them, developed by Terray, a California biotech focused on AI-driven drug discovery.

Note the recipe part: in drug discovery, a molecule is not useful just because an AI model can draw it on a screen. Chemists also need to know whether it can be made from available building blocks, using reactions likely to work in the lab.

In the standard DMTA loop - design, make, test, analyze - the make step is usually the slowest. Terray says complex AI-generated molecules can still take weeks to months to synthesize.

This is where synthesis planning comes in. A synthesis planner is software that tries to find a workable lab route for making a molecule - often by reasoning backward from the target molecule to starting materials a chemist can buy, and/or forward through reactions likely to work.

TerraSynth changes how that planning happens. Instead of first generating a molecule and then asking whether chemists can make it, it generates molecules together with the synthetic route. The recipe is part of the design, not a rescue operation afterward.

Terray claims TerraSynth reconstructed drug-like molecules from ChEMBL - a public database of bioactive, drug-like molecules - with a 67% higher reconstruction rate than the next-best planner, while running roughly 1,000x faster. The company also says this is already being used on real in-house drug discovery work: across 17 internal projects requiring custom synthesis, TerraSynth-guided designs were synthesized 2-4x faster than earlier unconstrained designs.

It’s worth pointing out that this comes from a Terray blog, not a peer-reviewed paper, and a technical report is said to be forthcoming.

If you have a very good memory, and were here from the start, you may remember that BAIO wrote about Terray in our very first issue. Back in February, the story was TerraBind, its model for predicting whether a small molecule is likely to bind a protein target. That was also about speed: Terray said TerraBind beat open-source Boltz-2 while running 26x faster.

Why it matters: A recurring BAIO theme is that AI x bio is only as strong as the weakest link in the loop. For small molecules, one weak link is still physical chemistry: someone has to make the thing. This is an interesting effort to speed up that process.

Did you know? Terray is hiring.

TITO is an AI model for molecular dynamics that predicts how molecules move over longer jumps in time, instead of simulating every femtosecond twitch. Researchers at Chalmers University of Technology and the University of Gothenburg say it is more than 10,000 times faster than conventional simulations.

Molecular dynamics is how chemists simulate motion at atomic resolution. Drug discovery depends on that motion: binding, folding, shape changes and drug unbinding are not visible in a still image.

Standard simulations move in tiny steps, around a femtosecond, because fast bond vibrations must be resolved for the calculation to stay stable. But many useful molecular events unfold over nanoseconds, microseconds or longer. That leaves researchers doing billions of steps.

TITO, short for Transferable Implicit Transfer Operators, learns from existing simulation data how atomic configurations change over longer time gaps, then samples those jumps directly. Or, as Chalmers puts it, the model can jump between scenes in “molecular movies” instead of watching every frame.

In Science Advances, the team tested TITO on 12,530 small organic molecules and more than 1,000 short peptides.

“As far as we know, this is the first time this has been done in a way that works for many different molecules,” says Simon Olsson, research leader and associate professor in computer science and engineering at Chalmers and the University of Gothenburg, in Chalmers’ article.

The current version is still far from simulating a real drug binding to a protein in a cell-like environment. The paper tested small molecules and short peptides, under simplified conditions, so scaling this to larger systems remains the next problem.

Why it matters: “In the long term, AI models like ours could help to identify promising drug candidates more quickly and improve accuracy in the early stages,” lead author Juan Viguera Diez, an industrial doctoral student at AstraZeneca and Chalmers, said. “This will hopefully pave the way for the development of more general techniques, which may ultimately facilitate the development of new drugs and new treatments, and, in a broader sense, also improve our understanding of diseases.”

Did you know? TITO’s code and model weights are available on GitHub and Zenodo.

Tamarind Bio has been selected as a technology partner for Lilly TuneLab 2.0. Its job: build, host and operate the inference layer - the software that lets participating biotechs run Lilly’s biomolecular models in private workspaces.

TuneLab is Lilly’s collaborative AI/ML drug discovery platform. It gives selected biotechs access to models trained on decades of its proprietary research data, representing hundreds of thousands of molecules. In return, member companies contribute data through privacy-preserving federated learning, so the models can improve without companies directly exposing proprietary data.

Tamarind is not running the federated-learning part. That’s already handled by Rhino Federated Computing. Tamarind is building the layer around the models: workflows, scalable inference, clean APIs, secure access, and private workspaces for participating biotechs. The company says it already provides more than 300 molecular design and prediction tools to scientists across pharma and biotech.

Why it matters: TuneLab is an interesting bargain between big pharma and smaller biotechs. Lilly brings models trained on proprietary drug discovery data. Biotechs get access to those models without receiving Lilly’s raw data, and can contribute their own data through federated learning without handing it over directly. Tamarind’s role is to make that usable in practice: each member company gets its own private tenancy where it can run the biomolecular models securely.

Did you know? Tamarind Bio is hiring.

SCP-Nano is open-source code from Ali Ertürk’s group for mapping where drug-delivery vehicles go across an entire mouse body at single-cell resolution. Think lipid nanoparticles, AAVs, liposomes and DNA origami - carriers used to deliver RNA, genes or other therapeutic cargo. The method, published in Nature Biotechnology, uses cleared whole-body imaging and deep learning to quantify delivery by organ, tissue and cell. It comes from the same whole-body imaging world as MouseMapper - the Ertürk-led tool BAIO covered in Issue 26 - but applies that logic to drug delivery.

☑️ General-purpose clinical LLMs outperform specialized clinical AI tools on medical benchmarks.

☑️ EinsteinArena lets AI agents collaborate on open research problems.

☑️ UK doctors could be liable for AI-tool mistakes, Medical Protection Society warns.

☑️ Google DeepMind asks how AGI becomes ASI.

☑️ Immorta Bio argues AI can help personalize aging therapies.

☑️ Ageing Research Reviews maps AI across the aging continuum.

☑️ No Priors talks Biohub, virtual biology and open-source biology.

☑️ PBCNet2.0 predicts protein-ligand binding at atomic resolution for probe discovery.