So this AI thing I keep hearing about, it’s getting pretty good, you say? So good it makes mathematicians cry?

Alright then. Welcome to biology.

It so happens that, as I was almost finished with this issue, I realized we’d accidentally turned it into a stress test for AI in biology.

Like the new paper in Cell that lays out 15 challenges for generative AI in cell biology, from restoring immune cells to predicting drug toxicity. Or another one asking what a biological world model would actually need to do before the name means much.

But why stop there?

FutureHouse and Edison Scientific have now proposed 12 “Millennium Problems for Biology.” Regenerate an amputated mouse limb. Revive a mouse after cryopreservation. Create self-replicating cells from non-living chemistry. And importantly, success has to show up in an experiment.

Navier-Stokes was child’s play. Now it’s time for expert mode.

Let’s dive in.

China already runs more clinical drug trials than the US, according to Reuters. At the end of 2025 it also had 4,751 innovative drugs (the Chinese term for novel medicines rather than generics) in development - about a third of the global pipeline and the largest number in the world. The remaining question, perhaps, is whether that speed and scale translate into original innovation.

Well, Beijing has now put a number on the ambition: by 2030, it wants at least 25% of the world’s first-in-class drugs - the first treatments of their kind - to be developed in China.

Its new five-year pharma plan makes AI an integral part of the route. It calls for high-quality pharmaceutical datasets, trusted spaces for sharing health data, pharma-specific large models and agents. Named applications range from target screening, molecule design and virtual screening to drug-property prediction, mining animal-model data and AI-assisted review of clinical-trial protocols. It even proposes virtual-patient and disease-progression models.

AI is also supposed to enter the regulator itself. The plan calls for AI models and agents to help review and inspect drugs, medical devices and cosmetics - handling tasks such as classification, document review, knowledge retrieval, problem identification and report generation.

The clinical-development machinery gets attention too: optimize trial protocols and ethics review, shorten trial startup and further digitize trials. Other 2030 targets include at least 20% annual growth for the innovative-drug industry and ranking among the world leaders in new-drug programs entering clinical trials and innovative drugs reaching market.

America, your turn.

MIT researchers led by Robert Langer and Ana Jaklenec have used AI-guided experiments to make mRNA vaccines far more heat-stable. After two months at 37°C, their dried formulations still generated antibody responses in mice comparable to fresh vaccine. Their new preprint also reports what the authors say is the first mRNA vaccination in monkeys using dissolving microneedle patches.

mRNA temporarily instructs cells to make a protein. Covid vaccines showed how powerful that can be, and the same platform is being explored for cancer and other diseases. But mRNA is fragile. It is usually packaged in tiny fat particles called lipid nanoparticles, or LNPs, which protect it and carry it into cells. These products can require refrigeration or freezing, raising costs and making distribution harder where reliable cold storage is scarce. Spoiled or expired doses are thrown away. The paper cites estimates that thermostable vaccines could cut storage costs by 71–86% and wastage costs by at least half.

The MIT team’s solution is to turn normally liquid mRNA-LNPs into a dry, solid form. They first screened 40 stabilizing ingredients already used in pharmaceuticals, including sugars, amino acids and polymers, then selected five with complementary properties for the optimization step. Finding the right ingredients and concentrations for each LNP is not trivial. The authors illustrate how quickly the possibilities multiply: even a fairly restricted search would already involve nearly 350,000 different formulations.

To search that space, the researchers built AGENT (which, despite the name, isn't an AI agent, but short for “Algorithm-Guided Experimental design for lipid Nanoparticle Thermostabilization”). It uses Bayesian optimization, a machine-learning method designed for problems where experiments are expensive and data is scarce. AGENT starts with results from a small number of laboratory tests, estimates which recipes are likely to work and where it is still uncertain, then chooses the next recipes worth testing. Those results feed back into the next round.

The researchers tested the optimized formulations with SARS-CoV-2 vaccine mRNA. They also put the dried vaccine into patches covered with microscopic needles that dissolve in the skin. In mice and monkeys, the patches produced immune responses similar to injections.

If this holds up, the payoff is unusually tangible: fewer freezers, lower storage costs, less wasted vaccine and easier access to mRNA medicines in places where reliable refrigerated storage and transport are scarce. And it may not stop with mRNA. The authors say the same approach could be used to stabilize proteins, antibodies and other drug-delivery particles. Cool stuff, hot delivery.

The idea behind world models goes back to psychologist Kenneth Craik, who proposed that organisms could carry an internal model of reality, allowing them to try out possible actions before taking them.

Over the following decades, versions of that idea entered AI, robotics, video games and autonomous driving. The shared ambition is to anticipate what might happen under different actions. But the world being modeled, and what the model can actually do, varies considerably.

Now that “world model” is becoming a favorite phrase in AI × biology, some definitions would be helpful. Researchers at Harvard Medical School, Oxford and the Broad Institute have taken on that task in a Perspective in Cell’s AI in biology special issue.

We took a look. Here are four things to take home.

The authors propose a functional definition: give the model a biological starting state and an action, such as administering a drug. It should predict possible subsequent states, then let you continue simulating from those predictions.

Do note that last part. A cancer treatment changes the tumour, including which cells survive. The next treatment encounters a different biological situation. A world model should help explore how those sequences might unfold.

If you read our GenBio feature and interview with Eric Xing and our recent special report, parts of the paper might feel familiar. GenBio similarly envisions models that connect different levels of biology, allowing researchers to simulate how a drug or gene edit affects cells, tissues and eventually whole organisms over time.

The Cell Perspective also emphasizes that a useful world model needn’t reconstruct every molecular detail. It needs to preserve enough information about the biological system and its history to predict how it would respond to subsequent interventions.

Many molecular measurements destroy the cells being studied. Researchers can compare populations before and after treatment, but cannot follow the individual cells between those observations. A larger share of one cell type after treatment could mean that those cells multiplied, that other cells changed into that type, or that other cell types died. The snapshots alone may not reveal which happened.

One way the authors propose addressing this is to combine repeated measurements of living cells with the established technique of lineage tracing - tracking which cells descend from which.

This also helps explain the interest in systems such as Cell Cinema, which we covered in Issue 40: repeatedly observing the same living cell could supply information that destructive measurements miss.

Patient records, on the other hand, follow the same person over time, but they can make treatment effects difficult to untangle. A drug given mainly to very sick patients might appear to produce worse outcomes simply because those patients were sicker to begin with. If the records don’t adequately capture that difference, a model could wrongly attribute the worse outcomes to the drug.

The authors propose training models on both patient records and data from randomized trials, including trials run within routine care.

To simulate further into the future, a model can use each predicted biological state as the starting point for its next prediction. If the first prediction is just slightly wrong, the next step begins from an inaccurate picture of the system. As this repeats, later predictions can drift further from reality, eventually producing biologically impossible outcomes.

The paper proposes a concrete test: predicting 48 hours ahead should give results consistent with predicting 24 hours ahead, then another 24 from that predicted state.

For practical use, the authors suggest repeatedly reconnecting simulation with reality: explore several steps ahead, perform one intervention, measure what actually happened, then revise the plan.

Imagine using a model to decide which drug to test next. It might predict most drugs’ effects very accurately, yet make a small error that puts the wrong candidate at the top of the list. Another model could be less precise about the size of each drug’s effect but correctly identify which drug works best. For deciding what to test next, that second model could be more useful despite its lower overall prediction score.

The authors propose testing whether world models help researchers choose better experiments. For example, let a world model, a simpler method and human experts each select the same number of comparable drug experiments. Which approach finds more drugs that produce the desired effect?

Restoring an immune cell’s ability to fight disease, predicting drug toxicity and identifying who benefits from treatment are among 15 challenges for generative AI in cell biology proposed in the latest issue of Cell. The authors come from Stanford, Columbia, the Chan Zuckerberg Initiative and Biohub, among other institutions.

The list also covers more fundamental questions about the molecular machinery inside cells and how cells interact. Each challenge comes with proposals for how to test progress. The authors acknowledge AI’s successes in predicting protein structures and designing proteins, while arguing that predicting how cells and tissues behave remains much less advanced.

In other domains, giving AI more data and computing power has produced major advances. That history underpins AI researcher Rich Sutton’s “bitter lesson”: general methods for learning and searching have repeatedly overtaken approaches that encode experts’ ideas about how a task should be solved.

The authors question whether the same pattern will hold for these challenges in cell biology. They argue that cells present a special difficulty: many molecules act together, creating an astronomical number of possible combinations, while experiments cover only a tiny fraction. Scaling up today’s models and datasets, they suggest, may therefore be insufficient.

Their proposal is to build known molecular interactions and physical constraints into models, reducing what those models must discover from experimental data alone. They also say those biological assumptions should be reassessed as the models’ predictions are tested. Whether that provides a lasting advantage, or the temporary boost Sutton warns about, remains unsettled.

The biggest barrier they identify will be familiar to many of you: insufficient high-quality training data. A model could have millions of measurements describing untreated cells, yet few examples showing what happens when a drug is added. To learn those responses, it needs data connecting the treatment to the biological changes that follow. The authors want more experiments designed specifically to generate that evidence.

For predicting patient outcomes, they propose a public-private consortium to collect and share clinical-trial data. It would link molecular measurements taken before and during treatment with patients’ outcomes, helping models learn how the biology of people who benefited differed from those who did not. They also want companies to preserve patient samples so missing molecular measurements can be made. The consortium would initially focus on failed trials, where companies have fewer commercial reasons to withhold data, with a goal of assembling data from at least 500 trials within five years.

There may be a shortage of training data, but there is no shortage of ideas for obtaining it. Failed drug programmes and stored patient samples could still yield evidence that helps future models. Part of solving the data problem is therefore organizational: funding the missing measurements, preserving the material and getting companies to share what they learn.

FutureHouse and Edison Scientific have published 12 “Millennium Problems for Biology”: very difficult challenges with one important rule - the answer should be easy to verify in a normal wet lab within days or weeks.

Examples include regenerating an amputated limb in an adult mouse, reviving a whole mouse after cryopreservation, and getting self-replicating cells to emerge from a plausible primordial soup.

Each problem comes with concrete acceptance criteria, so an AI cannot pass by writing a persuasive explanation or gaming a benchmark. FutureHouse co-founder Sam Rodriques calls them a possible “last reasonable eval” for AI in biology: if a system can solve one of these and the result works in the physical world, something pretty meaningful has changed.

The list is still being refined, and the team is inviting corrections and new problems. Browse the challenges and their acceptance criteria.

☑️ Large Drug Discovery Model experimentally validates generative drug design across five targets.

☑️ Mithrl raises $20M to expand its biomedical world model across biopharma.

☑️ Big Picture Bio emerges from stealth to build a world model for cancer drug combinations.

☑️ Sovereign AI fund in talks to back drug discovery start-up Potential Sciences.

☑️ Waypoint Bio argues new medicines are fundamentally an out-of-distribution problem.

☑️ AI finds previously unknown RNA viruses in 14,300-year-old sequencing data.