We've covered 100+ AI × bio stories - here's what they reveal

After our first 20 issues and more than 100 AI × biology stories, a pattern is starting to emerge that no single paper, funding round, product launch, or benchmark can show on its own.

Yes, BAIO has covered a lot of news. But the point is that the news is beginning to add up. We’re starting to see a shape.

AI × biology is moving from models to systems. The bottlenecks are moving downstream. Virtual biology is climbing from proteins toward cells, patients, and physiology. Big AI companies are treating biology as a platform category.

And AI can produce impressive biological outputs. More important, though, is who validates it, who explains it, who has the right data, and who turns it into something scientists, clinicians, or companies can actually use.

So for this special report, I went back through BAIO’s first 20 issues and distilled eight signals that say where AI × biology stands in the first half of 2026.

Let’s dive in.

The early AI × biology story was often about models: better protein structure prediction, better molecular generation, better cell-state prediction, and so on. That still matters. But across BAIO’s first 20 issues, the more important shift is that models are becoming components inside larger systems.

A lot of work now is happening around the model: lab access, agent workflows, data pipelines, validation loops, deployment environments, regulatory traceability. A model connected to a wet lab, wrapped in a workflow a scientist can actually use, and able to learn from the next experiment is more useful than one that scores well in a notebook.

That is why the field increasingly looks less like a leaderboard race and more like an infrastructure race. The durable advantage may belong to whoever builds the tightest loop between prediction, experiment, feedback, and decision.

Further reading:

For years, the promise of AI in biology was that better prediction would make discovery faster. Now another pattern keeps showing up: once AI makes one step easier, the bottleneck moves to the next one.

If models can generate molecules, someone still has to make them. If agents can design experiments, someone still has to run them. If virtual cells can predict perturbations, someone still has to explain why the prediction is trustworthy. If AI can find more drug candidates, the hard part may shift to patient selection, clinical trial design, manufacturing, regulation, and development.

That is a useful corrective to both hype and dismissal.

Further reading:

In the world of virtual biology predicting protein structures or designing molecules remains foundational. But the ambition keeps moving upward: from molecules to proteins, from proteins to cells, from cells to tissues and organs, and now toward virtual patients.

A protein model asks: what does this molecule look like, bind, or do? A virtual cell asks: what happens if we perturb this gene, pathway, or program? A virtual patient asks something harder: how does biology unfold across time, disease, treatment, and clinical history?

This is still early, uneven, and easy to oversell, not least of all because the concept of a virtual cell mostly remains a vision. But one can see the direction. AI × biology is trying to build computational representations of living systems at increasing levels of abstraction. The closer those representations get to patients, the more useful - and more dangerous to trust too early - they become.

Further reading:

Google has Isomorphic Labs. OpenAI is funding life sciences through its Foundation and has already connected frontier models to biological experiments. Anthropic is building Claude into life sciences workflows, acquiring biology talent, and testing models on real biological design tasks. Meta is releasing brain models. In addition to selling GPUs to biotechs, NVIDIA is building models, infrastructure, datasets, and workflows for them. AWS is packaging AI-driven drug discovery as a service.

Biology is becoming one of the arenas where frontier AI companies compete for developers, enterprise customers, scientific legitimacy, and long-term platform control. The question is which AI company becomes the default operating layer for biological work.

Further reading:

AI × biology is full of impressive outputs: generated molecules, predicted structures, ranked targets, simulated cells, automated analyses, and fluent scientific reasoning.

No matter: you still need validation.

That can take different forms. Wet-lab validation. External cohorts. Prospective trials. FDA review. Randomized clinical data. Independent benchmarks. Sometimes even a failed prediction is useful, if it shows where the system breaks. The important question is “what kind of test did the claim pass?”

This is where the field becomes both more serious and more complicated. Some AI × biology systems are now producing experimentally validated designs, clinically useful predictions, and regulatory-grade evidence. Many others are still preprints, demos, internal benchmarks, or computational claims waiting for the lab.

Further reading:

In mainstream AI, scale often means more parameters, more compute, more tokens. In biology, that story is incomplete. The recurring lesson is that better models need better biological coverage: more cell types, more perturbations, more species, more patient cohorts, more experimental conditions, more modalities.

That matters because biology is not one distribution. A model trained on familiar proteins, common cell lines, clean benchmark datasets, or narrow patient populations may look strong until it meets a new tissue, organism, disease context, or clinical setting. The field needs datasets that expose models to the messy variation of living systems.

This may become one of the deepest moats in AI × biology. Compute can be bought. Model architectures spread. But causally rich, diverse, well-annotated biological data - especially linked to experiments and outcomes - is much harder to build.

Further reading:

One of the strangest shifts in AI × biology is that the barrier to trying things is falling faster than the barrier to knowing whether they worked.

That does not only mean better tools for professionals. It means a non-biologist can use public AI tools to help design a dog cancer vaccine. Someone with no lab training can sequence their own genome at home with a MinION, Claude-written protocols, and a camping table. A small team can use a browser-based protein platform to redesign the binding loops of Keytruda in minutes.

This is not the same as saying the results are reliable. Often they are not proven, not peer-reviewed, not clinically useful, or not validated in the lab. But that is sort of the point. AI is making biological capability more accessible before it has made biological judgment more scalable.

Further reading:

As AI moves deeper into biology, “the model says so” becomes less useful. A drug target, protein function, clinical prediction, or experimental recommendation is only valuable if scientists can inspect why the system reached it - and decide whether the reasoning holds up.

Explainability matters for debugging models, designing follow-up experiments, convincing collaborators, satisfying regulators, and knowing when a system is confidently wrong. In biology, trust is built by showing the evidence chain, not by nicer chat interfaces.

Further reading: