The Problem With Treating Biology Like Software
The AI industry is right to be excited about biology. The scientists are right to be impatient with the hype. The investable truth sits in the gap between those two moods.
A few weeks ago, I wrote that a bigger AI war is starting, and that the next battlefield for the frontier labs may sit under a microscope rather than inside a chat window.
That view still looks right. Google DeepMind has moved from AlphaFold 2 to AlphaFold 3, AlphaProteo, AlphaGenome, and Isomorphic Labs. OpenAI has made its own moves through Eli Lilly and Retro Biosciences. Microsoft is building BioEmu for protein dynamics. NVIDIA is turning digital biology into an infrastructure market. Anthropic is trying to make Claude useful inside the daily workflow of scientists.
The pattern is too broad to dismiss as branding. The leading AI companies are looking at biology because biology has started to look, at least from a distance, like an information problem. DNA has sequence. Proteins have sequence and structure. Cells have states. Disease perturbs a system. Drug discovery searches through an enormous space of possible interventions. Experiments generate feedback.
Once you see biology that way, it becomes tempting to believe the same forces that scaled language models can scale scientific discovery.
That optimism has real evidence behind it. But the best version of the thesis needs a stricter sentence attached to it:
AI may make biology more programmable, but biology will punish anyone who confuses programmability with control.
The progress is real
The strongest case for optimism begins with AlphaFold.
Protein folding was never a toy benchmark. For decades, figuring out the three-dimensional structure of a protein could require years of experimental work. Then AlphaFold changed the default expectation. DeepMind later released predicted structures at enormous scale, turning one of biology’s central bottlenecks into a shared scientific resource.
That gave the AI-for-science thesis something rare in technology investing: a visible scientific win.
AlphaFold 3 pushed the story further. In 2024, Google DeepMind and Isomorphic Labs said the model could predict the structure and interactions of proteins, DNA, RNA, ligands, antibodies, and other molecules. Their own description was careful but still striking: for interactions between proteins and other molecule types, AlphaFold 3 showed at least a 50% improvement over existing methods, with some categories doubling prediction accuracy. The model was also made available through AlphaFold Server for non-commercial research.
That does not make drug discovery easy. It does shift the frontier from isolated protein shapes toward molecular relationships, which is closer to where drug discovery actually lives. A drug works because a molecule binds, a pathway changes, an antibody recognizes a target, or a mutation alters expression. If AI can reduce the cost of exploring those relationships, it can raise the return on experimental capital.
DeepMind’s newer systems point in the same direction. AlphaProteo designs protein binders and has shown wet-lab success across several targets, while also failing on at least one difficult target. AlphaGenome moves into regulatory genomics, trying to predict how DNA variants affect gene regulation across long sequence contexts. The company is no longer making a single protein-folding claim. It is building layers of a larger biological model.
This is why Demis Hassabis matters in the story.
In my earlier essay, Demis Hassabis’s 37 ideas about AI, science, and the next human era, I argued that his deepest theme is the move from chatbots to discovery engines. Hassabis keeps returning to the same idea: AI should become the “ultimate tool for advancing human knowledge.” In his worldview, AlphaFold is a prototype for a broader scientific machine. Protein structure was one layer. The next layers are interactions, regulation, cell state, disease mechanisms, and drug response.
On CBS’s 60 Minutes, he pushed the optimism much further. Drug development, he said, takes roughly ten years and billions of dollars for one drug; AI could perhaps reduce parts of that process to months or weeks. He then made the quote that now hangs over every AI-biology discussion: “one day maybe we can cure all disease with the help of AI,” perhaps “within the next decade or so.”
That sounds extreme. It is easier to understand if you see where Hassabis is coming from. He is not thinking mainly about a chatbot answering medical questions. He is thinking about search, simulation, hypothesis generation, and experimental loops. His career runs from games to reinforcement learning, from AlphaGo to AlphaZero, from protein structure to scientific reasoning. In games, the system can search a defined space, evaluate outcomes, and improve through iteration. Hassabis’s biology thesis is that parts of science can gradually acquire that same loop: a model proposes, an experiment tests, the data returns, and the next proposal improves.
That is the dream. In narrow domains, the dream is already becoming practical.
Why Silicon Valley is so tempted by biology
The tech industry tends to approach a domain by asking whether it can be represented digitally, searched computationally, and improved through feedback. Biology now gives better answers to those questions than it did twenty years ago.
Sequencing costs collapsed. Cryo-EM improved. Single-cell and spatial omics can measure biological state at extraordinary resolution. CRISPR makes perturbation experiments more programmable. Robotic labs can run experiments at scale. Public databases have accumulated protein structures, gene-expression profiles, molecular screens, clinical trial records, and literature.
For a software person, that looks familiar. A messy human activity becomes digitized. Once digitized, it becomes searchable. Once searchable, it becomes optimizable. After that, someone tries to turn it into a platform. Advertising, maps, media, finance, logistics, code, and customer support all followed some version of that path.
So when the AI world looks at biology, it sees a giant, under-optimized search problem with enormous economic value.
A good molecule can be worth billions. A validated target can define a company. A better assay can redirect a research program. A faster way to design proteins can create an entire platform. If AI can make discovery even modestly faster, the market does not need science fiction to justify investment.
This is also why the infrastructure companies care. NVIDIA does not need to know which AI-designed drug wins. If biology becomes a computation-heavy industry, demand can flow into GPUs, BioNeMo, NIM microservices, model serving, simulation, chemistry, genomics, imaging, and private-cloud deployments. Cloud providers get paid before the drug works. Lab-software companies get paid before approval. Data platforms get paid while pharma is still deciding which targets to pursue.
That investment stack is more robust than a simple “AI will cure cancer” trade. The early money may accrue to compute, cloud, lab automation, data infrastructure, workflow software, and companies with proprietary feedback loops. The higher-upside layer is more fragile: AI-native drug discovery companies, protein-design platforms, longevity programs, and full-stack biology companies that must eventually survive clinical reality.
That is why I still like the broad thesis from A Bigger AI War Is Starting: life sciences may become one of the most important credibility tests for frontier AI.
But credibility cuts both ways.
Biology is two orders more complex than the software analogy suggests
Here is where many scientists start to roll their eyes.
Tech people see information. Biologists see context.
That difference explains much of the disagreement. In software, the same input usually produces the same output. If the program fails, you can reproduce the bug, inspect the stack trace, patch the code, and ship the fix. The system is built by humans, and its abstractions were designed to be legible.
Biology did not inherit those conveniences. A gene is not a function call. A protein is not a fixed machine part. A cell is not a container for code. A disease is not a single bug. The same molecule can behave differently across tissue types, developmental stages, immune states, microbiomes, genetic backgrounds, drug combinations, sex, age, and environment. A pathway that matters in a dish may disappear in a mouse. A result in a mouse may fail in a human. A clean target in one cancer subtype may become irrelevant when the tumor evolves around it.
This is why the “two orders of magnitude” complaint has force.
One hidden order is biological organization. Molecules sit inside cells; cells inside tissues; tissues inside organs; organs inside organisms; organisms inside environments. Each level changes the behavior of the level below it.
Another hidden order is measurement. In software, logging a system often means recording the thing you need. In biology, the measurement can distort the system, miss the relevant time point, average away the important subpopulation, or capture a proxy that looks precise while hiding the causal mechanism.
That is why AI can look brilliant in one biological layer and fragile in another. A model can predict a structure and still fail to predict toxicity. It can design a binder and still fail on delivery. It can optimize affinity and still create immunogenicity. It can identify a target and still fail because the disease mechanism is downstream, redundant, compensated, or patient-specific. It can generate a therapy and still run into manufacturing, dosing, biodistribution, reimbursement, trial design, or regulation.
Scientists have learned to respect these traps because biology has spent decades embarrassing clean theories.
Jennifer Doudna’s skepticism fits here.
In a Bloomberg interview, Doudna pushes back against the claim that AI will simply cure everything. Her reply to the idea that ChatGPT might deserve a drug-sales royalty was only two words: “Good luck.” On Larry Ellison’s claim that AI could generate personalized cancer vaccines in 48 hours, her answer was equally grounded: if that were true, everyone would be happy, but she does not see that day yet.
Her most important point is simple: cancer is not one disease. It is hundreds of diseases. Every oncologist knows this, but the phrase “AI will cure cancer” tends to compress that reality into a slogan.
Doudna is not anti-technology. She uses AI tools. She sees value in data organization, reporting, guide-RNA design, DNA-sequence modeling, and running fewer, better experiments. In Scientific American, she described AI as transforming parts of CRISPR science, from guide-RNA design to DNA-sequence modeling, and said the convergence of AI and gene editing can help scientists run “fewer experiments but the right ones.”
That phrase is the sober version of the AI-biology thesis: fewer experiments, better chosen.
It is a long way from that to “all disease cured in ten years.”
The data bottleneck is real
The biggest misunderstanding in AI biology may be the word “data.”
Tech people hear “biology has lots of data” and imagine the internet. Biologists hear the same phrase and ask what kind of data: which assay, tissue, species, cell state, protocol, time point, batch, negative control, patient population, and failed experiment?
Biology has a lot of data, but much of it is the wrong shape for AI.
Public biological datasets were usually created for human scientific questions rather than machine-learning consumption. They are uneven, biased toward successful experiments, biased toward fashionable targets, biased toward easier measurements, and often difficult to compare across labs.
The negative results are especially valuable and especially missing. Failed compounds, failed screens, failed protocols, and failed clinical hypotheses often remain inside companies or disappear into the file drawer.
This matters because models need to learn what fails. A drug-discovery model trained mostly on published success is like an investor trained only on winning pitch decks. It may learn the language of confidence without learning the base rate of failure.
The experimental substrate also matters. A large share of biological data still comes from simplified systems: 2D cell cultures, immortalized cell lines, animal models, narrow assays, and measurements that capture a slice of the disease rather than the disease itself. These systems are useful. They can also mislead. A tumor in the body is a spatial ecosystem, with nutrient gradients, hypoxia, immune interactions, stromal cells, clonal diversity, and resistant subpopulations. A flat plate of cells can remove the very context that determines whether the drug works.
That is why organoids, organ-on-chip systems, spatial biology, perturbation atlases, patient-derived models, and automated labs matter so much. They are data infrastructure. They are attempts to generate the kind of biological evidence AI can actually learn from.
Doudna’s example about the genome is powerful for the same reason. The human genome was sequenced around the turn of the century. Yet more than two decades later, scientists still do not understand the function of a large share of genes even in much simpler organisms. The instruction-book metaphor is useful, but it can mislead. We have the letters. We do not have the operating semantics.
AlphaGenome itself admits some of this. DeepMind says the model can analyze up to one million DNA letters and predict thousands of molecular properties, but it also says the model is not designed or validated for personal genome prediction or direct clinical use. It can predict molecular outcomes, while complex traits and diseases involve developmental and environmental factors outside its direct scope.
That caveat should be printed on the wall of every AI-biology investor. Prediction at one layer is progress. Translation across layers is the hard part.
Doudna’s CRISPR lesson: the bottleneck moves to delivery and access
CRISPR has already produced real medical miracles. Victoria Gray became the first patient treated with CRISPR for sickle cell disease in 2019, and Casgevy later became an approved therapy. The case of baby KJ Muldoon is even more dramatic: a fully personalized in-vivo CRISPR therapy for a rare metabolic disease, designed and delivered on an emergency timeline.
Those stories prove that programmable biology is no longer only a metaphor. They also reveal the next bottleneck.
Casgevy costs around $2.2 million per patient. KJ’s personalized therapy cost roughly $800,000 and depended on an unusual coalition of academic, public, and philanthropic support. Doudna’s question is the right one: how do you save more children like KJ without requiring a heroic one-off operation every time?
AI can help, but the job is more prosaic than the hype. It can help design guides, screen off-target risks, interpret sequence effects, support delivery design, automate documentation, triage actionable rare variants, and help researchers choose experiments more efficiently. The hard parts of CRISPR therapeutics still include delivery, immune response, tissue targeting, dosing, manufacturing, safety monitoring, reimbursement, and ethics.
Those are system problems, not chatbot problems.
Doudna’s caution about designer babies belongs in the same frame. Editing a single disease-causing mutation can be rational when the disease mechanism is clear and the suffering is severe. Editing traits like intelligence or height is a different category. Those traits are polygenic, context-dependent, developmentally mediated, and socially dangerous. Changing a few genes does not guarantee a predictable long-term outcome.
Her answer to the Gattaca question is the best kind of scientific realism: the future is probably somewhere in the middle, hopefully closer to heaven.
Hassabis and Doudna are less opposed than they look
The easy version of this debate puts Hassabis on one side and Doudna on the other. That is too simple.
Hassabis is the ambitious system-builder. Doudna is the experimental scientist who has watched a breakthrough technology fight its way toward real patients. One speaks from the frontier-lab view of AI as a general discovery engine. The other speaks from the translational reality of making biology safe, deliverable, affordable, and ethically usable.
Both are describing real parts of the same machine.
Hassabis is right that AI can change the search process. AlphaFold already did. AlphaFold 3, AlphaProteo, AlphaGenome, BioEmu, ESM3, and the next generation of closed-loop lab systems will probably change it again. If AI can propose better hypotheses, prioritize better experiments, and help scientists explore a larger design space, it can reshape the economics of research.
Doudna is right that summarization is not discovery, that cancer is not one disease, that simulation cannot replace every kind of testing, and that the field needs better data rather than louder promises.
The most investable synthesis is this: AI will not abolish biology’s complexity. It will make parts of that complexity searchable.
That is still a huge statement. Searchability is how new industries begin. Once a domain becomes searchable, companies can build tools, platforms, workflows, and feedback loops around it. The winners may be the companies that turn biological complexity into a compounding data asset rather than a one-time prediction demo.
What investors should watch
The next phase of AI biology should be judged less by press releases and more by feedback loops.
The key question is whether an organization can test its ideas, learn from the results, and improve the next round faster than competitors. A model that generates a plausible molecule is useful. A company that repeatedly turns model output into experimental learning has a different kind of asset.
That shifts the investor’s attention toward proprietary experimental data, especially failed assays and negative results that never reach the public literature. It also makes wet-lab integration harder to fake. If the loop between model and experiment is slow, brittle, or outsourced in fragments, the moat may be weaker than the pitch deck suggests.
Biological fidelity will matter more as the field matures. Data from patient-derived organoids, spatial biology, perturbation screens, and clinically relevant models may beat larger but weaker datasets. Translational depth will matter too, because delivery, toxicity, manufacturability, dosing, regulatory strategy, and trial design separate scientific demos from medicines.
Workflow adoption deserves attention for a less glamorous reason: it may commercialize first. Tools that help scientists read, design, document, analyze, and coordinate research can create value before any AI-designed blockbuster reaches the market.
This is why the investment map needs layers. NVIDIA and the cloud platforms benefit if biology becomes more computational. Alphabet has a unique full-stack position through DeepMind and Isomorphic Labs. Microsoft, AWS, and Anthropic can win through scientific workflow and infrastructure. Lab-automation, data, and omics platforms can become the rails for the field. AI-native drug companies carry the most dramatic upside, but they also carry the full burden of clinical failure.
The real prize is not a single AI miracle drug. It is a new research operating system where models, labs, data, and scientists form a faster learning loop.
That connects back to The Next Architecture of Intelligence. The transformer era proved that scale can create surprising intelligence from statistical learning. Biology may demand the next layer: models that combine scale with search, memory, simulation, causal experimentation, and real-world feedback.
It also connects to The Billionaire Bet on Reversing Aging. Longevity investing is really a bet that cell state can become more programmable. AI may accelerate that bet, but the same warning applies: resetting a cell marker is not the same as safely resetting an organism.
The right level of optimism
The AI industry is right to be excited. Biology is becoming more measurable, more programmable, and more computational. The progress from AlphaFold to AlphaFold 3 to protein design and genome modeling is meaningful. The movement of talent and capital into AI life sciences is a real signal. The potential market is enormous.
The scientists are right to be skeptical. Human biology is not software. Disease is layered, adaptive, heterogeneous, and contextual. Better predictions do not automatically become approved therapies. Data quality, experimental design, delivery, toxicity, manufacturing, regulation, and access remain hard.
The right forecast should hold both ideas at once.
AI will probably not cure all disease in ten years. It may still make the next ten years one of the most important decades in the history of biological discovery.
That is the investment story worth watching: not a straight line from chatbot to cure, but a slow conversion of biology from an artisanal search process into a machine-assisted learning system.
The prize is not certainty.
The prize is better experiments.
And in biology, better experiments are how the future starts arriving.