Artificial intelligence is starting to change biology in a much deeper way than simply helping scientists search papers or predict one protein structure.
A new group of companies is trying to build AI models that learn the underlying rules of biology itself. Instead of training a system to solve one narrow problem, they want models that can learn patterns across millions of proteins, cells, experiments, patients, diseases, and biological signals. Those models can then be adapted to many different scientific problems.
This is broadly what people mean when they talk about foundation models for biology.
Boston has become one of the most important places to watch this shift because the region has something few technology centers can easily copy. It has world-class AI researchers, large hospitals, major biotech companies, academic laboratories, pharmaceutical partners, and thousands of scientists working within a relatively small geographic area.
More importantly, Boston companies are beginning to connect AI models directly to physical experiments.
Generate is creating protein medicines through a continuous design-build-measure-learn loop. Cellular Intelligence is building models designed to understand how cells change state. Manifold Bio is connecting AI protein design with huge in-vivo experiments. BostonGene is developing multimodal models of tumors and immune systems. Lila Sciences is connecting scientific AI with autonomous laboratories. Asimov is combining biological engineering, lab automation, and predictive models. Dyno Therapeutics has used machine learning to search the enormous design space of gene-therapy vectors. Even a very young Boston startup, AbInitio Bio, is taking foundation-model ideas into biologics manufacturing.
Taken together, these companies point toward something bigger than another biotech trend.
Boston may be building the infrastructure for a future in which scientists do not use separate AI tools for proteins, cells, disease, manufacturing, and experiments. Instead, increasingly general biological models could sit inside continuous learning systems that move between computation and the laboratory.
That could change how medicines are discovered, designed, tested, manufactured, and improved.
The Short Version: Boston Is Building Models of Biology, Not Just AI Tools for Biologists
The easiest way to understand this market is to separate traditional biotech AI from biological foundation models.
Traditional biotech AI usually solves one defined task. A model might predict whether a molecule binds to a protein, classify a pathology image, identify a patient who qualifies for a clinical trial, or estimate whether a manufactured protein will aggregate.
Those tools can be very useful. But each one is generally designed around a particular job.
A foundation model tries to learn something broader.
It may learn the relationships between protein sequence, structure, and function. It may learn how cells respond when genes, chemicals, or environmental signals change. It may learn patterns across genetics, pathology, immune activity, treatment history, and patient outcomes.
Once the underlying representation is learned, the same system can potentially support many downstream tasks.
That distinction matters because biological R&D contains thousands of related problems. If knowledge learned from one experiment can transfer into another problem, companies may no longer have to start from zero every time they design a new drug.
Boston companies are attacking this opportunity from several directions.
| Boston-area company | Main biological layer | Core model or platform idea | Physical-data loop |
|---|---|---|---|
| Generate | Proteins | Generative protein models | Very strong |
| Cellular Intelligence | Cells | Virtual cell-signaling and cell-state models | Very strong |
| Lila Sciences | Cross-domain science | Scientific AI and autonomous experimentation | Very strong |
| Manifold Bio | Proteins + in-vivo biology | AI-guided protein design | Very strong |
| BostonGene | Disease + patients | Multimodal tumor and immune models | Strong |
| Asimov | Cells + manufacturing | Predictive synthetic-biology models | Strong |
| Dyno Therapeutics | Gene delivery | Machine learning for AAV capsids | Strong |
| AbInitio Bio | Biomanufacturing | Foundation models for manufacturing | Emerging |
This table reveals the most important point.
Boston does not have one biological foundation-model market. It has several connected markets forming around different levels of biology.
Proteins sit near the molecular layer. Cells sit one level higher. Tumor and immune models attempt to describe disease systems. Manufacturing models learn what happens when biological products move from design into industrial production. Scientific-agent platforms may eventually coordinate several of these layers.
That creates the possibility of an AI stack for biology.
What Exactly Is a Foundation Model for Biology?
The term can become confusing because companies use words such as “foundation model,” “generative biology,” “virtual cell,” “scientific superintelligence,” “protein language model,” and “multimodal AI” differently.
The underlying idea is simpler.
A foundation model learns useful patterns from a very large and broad dataset. Instead of being built for only one narrow prediction, it creates an internal representation that can be reused across many related tasks.
Large language models provide an easy comparison.
A language model does not learn only how to write emails. It learns broad patterns in language. That same model can then summarize documents, answer questions, translate text, write software, classify information, and perform many other tasks.
Biological foundation models attempt something similar with biological information.
Protein models try to learn the language of amino acids
Proteins are long chains built from amino acids. Their sequences affect how they fold, what they bind to, where they travel, how stable they are, and what they do inside the body.
Traditional drug discovery searches nature and existing protein libraries for useful starting points.
Generative models offer another possibility.
If a model learns the patterns connecting sequence, structure, and function, it may be able to design entirely new proteins with properties chosen by scientists.
Generate is one of Boston’s clearest examples. The company says its platform learns from hundreds of millions of known proteins and combines public biological information with proprietary experimental data. Its platform then generates proteins, physically builds them, measures their properties, and feeds those results back into its models.
That last step is critical.
The model is not simply trained once and then used forever. New physical experiments create new information that can make future designs better.
Cell models face an even harder problem
Cells are far more complicated.
A cell contains genes, proteins, signaling networks, metabolism, physical structures, and thousands of interacting processes. The same cell may behave differently depending on its environment.
This creates a huge modeling challenge.
Scientists increasingly want systems that can answer questions such as:
What happens when a particular gene is turned off?
How will a cell respond to a drug?
What signals cause a stem cell to become a neuron instead of another cell type?
Why does a diseased cell enter one state while a healthy cell enters another?
Cellular Intelligence, formerly Somite, is pursuing this problem from Boston. The company originally described DeltaStem as a foundation model for human stem cells. In January 2026, it repositioned itself around the broader goal of creating a universal virtual cell-signaling model designed to understand and predict how cells change.
The company pairs its models with a proprietary experimental system designed to generate cell-state data efficiently. That relationship between AI and data generation could become more important than model architecture alone.
Disease models connect biology to patients
Disease represents another level of complexity because real patients contain enormous biological variation.
BostonGene is approaching the problem through multimodal oncology.
Its platform combines multiple forms of biological and clinical information to model tumor and immune biology. In January 2026, the company announced an AstraZeneca collaboration involving its foundation model to help predict patient safety and efficacy profiles during oncology development. A separate Johnson & Johnson collaboration involves applying its pretrained cancer and immune-system model to generate patient digital twins intended to support treatment-sequencing research in multiple myeloma.
That represents a different type of foundation-model opportunity.
Instead of asking, “What protein should we design?” the model asks, “How does this disease system behave, and how might a specific patient respond?”
Why Boston Has an Unusual Advantage
AI research is not limited to Boston. The San Francisco Bay Area remains an enormous center of AI talent, compute, venture capital, and startup creation.
Biology changes the equation.
Building useful biological AI requires much more than GPUs and software engineers.
Companies need biological samples. They need experimental systems. They need wet labs. They need people who understand drug development. They need clinical experts. They need manufacturing knowledge. They need partnerships with pharmaceutical companies. They may eventually need patients and clinical trials.
Boston already has many of those pieces.
The result is an unusually dense feedback environment between computation and physical biology.
Models need experiments
Software companies can generate enormous amounts of digital training data cheaply.
Biotechnology companies cannot.
Someone must run the experiment.
Cells must be cultured. Proteins must be expressed. Animals may need to be studied. Assays must be performed. Samples must be sequenced. Instruments must be operated.
That makes proprietary experimental capacity strategically important.
Generate says its platform has generated, built, and tested more than 42,000 proteins. The company describes its system as a continuous loop in which machine learning generates candidates, biological infrastructure builds them, measurement systems test them, and experimental results improve future designs.
Manifold Bio is pushing the same principle in another direction.
In September 2026, the company said its platform could measure tissue distribution for more than 500,000 binders in a single in-vivo experiment. Its goal is not simply to generate more protein designs. It is to generate extremely large amounts of information about how those designs behave inside living systems.
That difference may become one of the defining competitive advantages in biological AI.
Original BostonBizScene Analysis: The Biological Foundation Model Stack
To understand Boston’s position more clearly, BostonBizScene reviewed a focused sample of eight Greater Boston companies whose public descriptions show at least one of the following characteristics: broad pretrained biological models, generative biological platforms, virtual-cell systems, transferable machine-learning models, or closed-loop AI platforms intended to learn across multiple biological problems.
The analysis uses information publicly available through October 4, 2026.
We scored each company from zero to five across five dimensions.
Model breadth measures whether the platform appears useful across many related problems rather than one narrow task.
Proprietary data generation measures whether the company can create unique biological training data itself.
Experiment integration measures how directly computation connects with physical biological testing.
Commercial validation considers partnerships, products, clinical programs, or other evidence that outside organizations are using or validating the platform.
Translation depth measures how closely the platform connects AI outputs with actual therapeutic development, clinical decisions, or manufacturing.
The score is an editorial analytical framework, not a scientific ranking. It is designed to compare business models consistently using publicly disclosed information.
Boston Biological AI Platform Score
| Company | Model breadth /5 | Proprietary data /5 | Experiment integration /5 | Commercial validation /5 | Translation /5 | Total /25 |
|---|---|---|---|---|---|---|
| Generate | 5 | 5 | 5 | 5 | 5 | 25 |
| Lila Sciences | 5 | 5 | 5 | 4 | 4 | 23 |
| Cellular Intelligence | 5 | 5 | 5 | 4 | 4 | 23 |
| Manifold Bio | 4 | 5 | 5 | 5 | 4 | 23 |
| BostonGene | 5 | 4 | 4 | 5 | 5 | 23 |
| Asimov | 4 | 4 | 5 | 4 | 4 | 21 |
| Dyno Therapeutics | 4 | 4 | 4 | 5 | 4 | 21 |
| AbInitio Bio | 4 | 2 | 3 | 1 | 3 | 13 |
The important finding is not which company has the highest number.
The important finding is that data generation and physical experimentation are unusually prominent across the Boston group.
That suggests the region is developing a different type of AI moat.
Chart: Where Boston’s strongest advantage appears
Physical experiment integration ████████████████████ Very high
Proprietary biological data ███████████████████ Very high
Model breadth ██████████████████ High
Drug-development integration ██████████████████ High
Pure software distribution ███████████ ModerateOur analysis suggests that Boston’s strongest advantage is not building the world’s largest general AI model.
It is building AI systems that can continually interact with physical biology.
That is a much harder capability to reproduce.
Original Finding #1: Boston’s Real Moat May Be Experimental Data, Not Model Size
Much of the AI industry talks about parameters, GPUs, compute clusters, and model scale.
Biology introduces another variable: the cost of discovering the correct answer.
A language model can learn from text that already exists.
A biological model often needs information that nobody has generated yet.
Suppose an AI system designs 10,000 proteins.
Which proteins actually express correctly?
Which remain stable?
Which reach the right tissue?
Which trigger unwanted immune responses?
Which work inside a living organism?
Those answers require experiments.
This means a biotech company with slightly less compute but dramatically better experimental infrastructure may eventually build a stronger model.
Manifold Bio offers an unusually clear example.
The company’s September 2026 mBER-2 announcement describes a system in which enormous numbers of protein binders can be tested directly in living animals. According to Manifold, its barcoding technology can measure tissue distribution across more than 500,000 binders in one experiment.
That turns biology itself into a data engine.
Instead of training AI only on what scientists already know, the company can deliberately ask new biological questions at scale.
Generate has built a similar flywheel around protein therapeutics. Its public platform description explicitly centers on generating proteins, building them, measuring their characteristics, and using that information to improve the model.
Cellular Intelligence uses the same logic at the cellular level. The company says its capsule-based approach produces dense cell-state data for training models, while its earlier Somite platform claimed roughly 1,000-fold greater efficiency than conventional approaches to generating certain differentiation data.
The competitive question therefore changes.
The question is not simply:
Who has the biggest model?
It becomes:
Who can generate the most useful new biological truth per dollar and feed it back into the model fastest?
That is a very different race.
Generate Is Building a Protein Learning Machine
Generate may be Boston’s most mature example of foundation-model thinking becoming a full drug-development company.
The Somerville company describes its platform as a therapeutic-area and protein-modality-agnostic system combining computation with scalable biological hardware. Its 2026 SEC filing emphasizes a tightly integrated design-build-test-learn loop designed to create proprietary therapeutic data.
Generate’s system learns from protein sequences and structures, proposes new molecules, physically creates those molecules, measures them, and uses the results to make future predictions better.
The commercial scale of the company also matters.
Generate raised $370 million in 2021 and another $273 million Series C in 2023. The company said the Series C brought equity financing since 2020 to nearly $700 million. In February 2026, it priced an IPO of 25 million shares at $16 each, representing $400 million in gross offering value.
That transition is significant for Boston.
Foundation-model biology is beginning to move from venture-funded experimentation toward public-market biotechnology.
Why Generate matters beyond Generate
Generate’s importance is not limited to whether individual drugs succeed.
It is testing a bigger business hypothesis.
Can a reusable biological learning engine improve enough over time that every therapeutic program benefits from previous experiments?
Traditional biotech development is often project based. A company develops Drug A, learns a great deal, and then develops Drug B.
Some knowledge transfers, but much of the work remains program specific.
A foundation platform could make learning cumulative.
If every molecule provides useful training data for future molecules, the economics of biotech R&D could eventually change.
Cellular Intelligence Is Trying to Build a Model of Cell Behavior
Proteins are complicated.
Cells are another order of difficulty.
Cellular Intelligence is attempting to build models capable of predicting how cells move between states.
The company started as Somite and focused strongly on stem-cell differentiation. Its DeltaStem model was designed to help predict how stem cells could be directed into specific cell types. In January 2026, Somite became Cellular Intelligence and described its broader target as a universal virtual cell-signaling model.
This shift matters.
A useful virtual cell could become one of the most powerful tools in biotechnology.
Instead of testing every biological intervention physically, scientists could first simulate possibilities computationally.
Such systems will not replace experiments anytime soon. Biology is too complicated and current models remain incomplete.
But even partial predictive capability could reduce the search space dramatically.
The Novo Nordisk deal adds a real therapeutic test
In May 2026, Cellular Intelligence acquired global rights to Novo Nordisk’s clinical-stage stem-cell-derived Parkinson’s therapy. Novo Nordisk also became an equity investor.
That gives Cellular Intelligence something especially valuable.
It now has an opportunity to connect an AI cell platform with a real clinical-stage cell therapy.
This creates a much stronger test than publishing model benchmarks.
If virtual-cell technology can improve development, manufacturing, quality, or future cell programs around actual therapies, the commercial case becomes much more powerful.
Lila Sciences Is Expanding the Foundation-Model Idea Beyond Biology
Lila Sciences represents an even broader vision.
The Cambridge company emerged publicly in March 2025 with $200 million in seed financing and a plan to combine AI with autonomous laboratories across life science, chemistry, and materials science.
By October 2025, Lila said it had completed a $350 million Series A, bringing total capital raised to $550 million.
The interesting idea is not merely autonomous lab equipment.
Lila describes its facilities as “AI Science Factories.”
The objective is to allow AI systems to propose experiments, run them through automated infrastructure, observe the results, and use those results to generate more scientific knowledge.
If that works at scale, scientific data becomes something a company can manufacture deliberately.
Scientific tokens could become a new type of training data
Traditional AI companies acquire text, images, video, and human feedback.
Scientific AI companies may build experimental factories designed specifically to produce high-value training data.
This could become one of Boston’s largest opportunities.
The region already contains thousands of laboratories. The next step is making those laboratories increasingly machine-readable and machine-operable.
The winner may not be the company with the most publications.
It may be the company whose AI can run the largest number of valuable learning cycles.
Manifold Bio Shows Why In-Vivo Data Could Become Extremely Valuable
Protein models have become increasingly good at proposing molecules.
That creates a new bottleneck.
Scientists still need to know what happens after the molecule enters a living body.
Manifold Bio is focusing heavily on that gap.
Its platform combines computational design with massively parallel in-vivo measurement. In 2025, Roche agreed to pay Manifold $55 million upfront in a collaboration involving brain-targeting shuttles, with potential research, development, clinical, and commercial milestones exceeding $2 billion.
That deal shows why proprietary experimental platforms matter commercially.
A model that predicts binding is useful.
A system that designs molecules and then generates enormous amounts of proprietary information about where those molecules travel inside a living organism could be much more difficult to copy.
Manifold’s September 2026 mBER-2 announcement reinforces that strategy. Its AI platform is being scaled alongside its capacity to measure hundreds of thousands of binders in living systems.
The model and the experimental platform reinforce each other.
That is exactly the pattern appearing across Boston.
BostonGene Is Building Foundation Models Around Disease
BostonGene approaches biological AI from another direction.
Instead of primarily generating new molecules, the Waltham company attempts to model cancer, the immune system, and patient response.
Its platform combines genomic, transcriptomic, immune, pathology, and clinical information into broader representations of disease.
This matters because diseases are not controlled by one gene or protein.
Cancer can involve tumor genetics, immune activity, cell composition, treatment history, tissue structure, and many other variables.
A useful model therefore needs to combine many forms of information.
BostonGene calls its approach omnimodal.
In early 2026, the company announced major collaborations with AstraZeneca and Johnson & Johnson around foundation-model-driven oncology research.
This shows another likely direction for the foundation-model market.
Models will increasingly be judged not only by laboratory benchmarks but by whether they can influence development decisions.
Can a model help choose patients?
Can it identify who may benefit from a therapy?
Can it predict safety problems earlier?
Can it suggest better treatment sequences?
Those are questions pharmaceutical companies will pay to answer.
Asimov Shows Where Biological Models Meet Manufacturing
Drug discovery gets most of the attention in biotech AI.
Manufacturing could become nearly as important.
A medicine is not useful if companies cannot produce it reliably, at sufficient scale, and at an acceptable cost.
Boston-based Asimov is building an AI-native synthetic-biology platform around the engineering and production of biological medicines.
In July 2026, the company announced that it would work with Boston University’s NSF-backed programmable cloud laboratory to generate experimental data and develop models that predict protein manufacturability directly from DNA sequence.
That is an important direction.
Future biological models may need to optimize two things simultaneously:
Does this medicine work?
and
Can this medicine actually be manufactured?
Those questions are usually addressed at different stages.
AI could bring them closer together.
A design system might eventually reject molecules that look therapeutically exciting but are likely to create expensive manufacturing problems.
That would save enormous amounts of time.
Dyno Therapeutics Demonstrates the Power of Searching Huge Biological Design Spaces
Gene therapy creates another foundation-model opportunity.
Adeno-associated viruses, commonly called AAVs, can act as delivery vehicles that carry genetic material into cells.
The problem is that natural AAVs were not designed by humans to become perfect medicines.
Dyno Therapeutics uses machine learning to search and design the much larger space of possible capsids.
The Cambridge company raised a $100 million Series A in 2021 to expand its CapsidMap platform and partnerships, including work with major pharmaceutical companies such as Novartis, Sarepta, and Roche.
Dyno illustrates an important economic feature of biological AI.
The theoretical design space can be enormous.
Humans cannot manually inspect every possible protein or viral capsid.
Machine learning allows companies to navigate spaces far larger than traditional experimental screening could practically cover.
Foundation models make that search more powerful because they learn general rules rather than treating every candidate independently.
AbInitio Bio Shows Foundation Models Moving Into Biomanufacturing
Some of the most interesting signals come from very young companies.
AbInitio Bio entered Y Combinator’s Spring 2026 batch in Boston with a plan to build foundation models for biologics manufacturing.
Its first model, Echo, is designed to predict manufacturing outcomes and transfer learning across different products and production settings. The company says it intends to expand the architecture into cell-line engineering, multispecific-antibody development, CMC risk, and other manufacturing problems.
The company remains extremely early, so it should not be compared directly with a public biotech such as Generate.
But strategically, its existence matters.
It suggests foundation-model thinking is spreading throughout the biotech value chain.
The first generation focused largely on molecule discovery.
The next generation may include models for:
protein design,
cell engineering,
disease biology,
clinical development,
process development,
cell-line optimization,
manufacturing,
and eventually regulatory operations.
If that happens, the biological AI stack could become far larger than today’s AI drug-discovery category.
Original Finding #2: The Boston Market Is Splitting Into Four Model Layers
Our review of the Greater Boston ecosystem suggests biological foundation-model companies can be grouped into four main layers.
| Layer | Core question | Boston examples |
|---|---|---|
| Molecular models | What molecule should we build? | Generate, Manifold Bio, Dyno |
| Cellular models | How will cells behave? | Cellular Intelligence, Asimov |
| Disease models | How will disease or patients respond? | BostonGene |
| Scientific systems | What experiment should we run next? | Lila Sciences |
A fifth category is now emerging around manufacturing, represented by companies such as AbInitio Bio and parts of Asimov’s platform.
These layers will probably not remain separate.
The long-term opportunity is connecting them.
Imagine a system that begins with a disease model identifying a biological mechanism.
A protein model designs a therapeutic candidate.
A cell model predicts how relevant human cells will respond.
An autonomous laboratory tests those predictions.
An in-vivo platform measures distribution and activity.
A manufacturing model determines whether the candidate can be produced reliably.
Results from every stage return to the models.
That would create something much closer to a learning operating system for drug development.
Original Finding #3: The Closed Loop Is Becoming More Important Than the Foundation Model
One of the strongest conclusions from our analysis is that the term “foundation model” may eventually matter less than the system surrounding it.
Models are becoming easier to build.
Open-source architectures are improving. Compute remains expensive but increasingly accessible. Academic research spreads quickly. Employees move between companies.
Proprietary biological feedback loops are harder to reproduce.
Generate has protein design, protein construction, measurement, and learning.
Manifold has protein generation tied to high-throughput in-vivo experiments.
Cellular Intelligence combines cell models with proprietary cell-state data generation.
Lila is building autonomous laboratories specifically to produce scientific data.
Asimov is combining models with biological engineering and production systems.
These companies are all moving toward the same strategic shape.
MODEL
↓
PREDICTION
↓
EXPERIMENT
↓
PHYSICAL RESULT
↓
PROPRIETARY DATA
↓
BETTER MODEL
↺That loop may become one of the strongest competitive moats in biotechnology.
A competitor can sometimes reproduce a model architecture.
It cannot instantly reproduce five years of proprietary experiments.
Original Finding #4: Boston’s Biological AI Companies Are Already Moving Toward Commercial Validation
A common criticism of AI biotech has been that impressive models do not necessarily produce successful medicines.
That criticism remains valid.
Drug development is difficult, and no AI system removes clinical risk.
But Boston’s ecosystem is moving beyond purely computational demonstrations.
Generate has clinical-stage programs and became a Nasdaq-listed company in 2026.
Cellular Intelligence now controls a clinical-stage Parkinson’s cell-therapy program acquired from Novo Nordisk.
Manifold Bio signed a Roche collaboration carrying $55 million upfront and potential milestones exceeding $2 billion.
BostonGene has announced collaborations involving its foundation model with AstraZeneca and Johnson & Johnson.
Dyno has worked with major gene-therapy companies including Roche, Sarepta, and Novartis.
The important metric is changing from model accuracy to economic usefulness.
What Investors Should Measure Instead of AI Hype
Investors evaluating this sector should be careful with words such as “foundation model.”
The label alone tells you almost nothing about whether a company has a durable business.
A more useful question is whether the platform becomes stronger as the company operates.
Does every experiment improve the platform?
A powerful biological AI company should ideally create a compounding data advantage.
If each drug program generates new information that improves all future programs, the company can become more efficient over time.
If every new project requires almost entirely new work, the platform may be less scalable than marketing suggests.
Is the data actually proprietary?
Training on public protein databases may be necessary, but many competitors can access those databases.
The strongest moat may come from information nobody else possesses.
That is why experimental infrastructure matters.
Can predictions survive contact with biology?
AI benchmarks are useful.
Real experiments matter more.
The important progression is:
computer prediction,
laboratory validation,
animal validation,
clinical validation.
A company moving through those stages is producing much stronger evidence than one producing computational results alone.
Is there transfer between programs?
Foundation models become economically powerful when knowledge transfers.
Investors should ask what the model learns from Drug A that materially improves Drug B.
That question gets much closer to the real platform value.
What Boston Biotech Leaders Should Do Now
Large biotech companies do not necessarily need to build giant foundation models themselves.
They do need a strategy for using them.
The first task is identifying proprietary datasets that could become strategically valuable.
Many pharmaceutical and biotechnology companies have years of experimental information scattered across spreadsheets, laboratory systems, PDFs, instrument files, electronic notebooks, manufacturing records, and individual teams.
That information may eventually be more valuable than management realizes.
Organizing it now creates optionality.
Build systems around learning loops
Companies should stop viewing AI only as software layered onto scientists.
The bigger opportunity is redesigning workflows around faster learning.
Ask:
How quickly can we move from prediction to experiment?
How quickly does experimental information return to the model?
How much does each experiment teach us?
Can one program improve another?
Those questions matter more than how many AI applications employees use.
Connect computational and laboratory teams
The companies discussed in this article repeatedly remove the boundary between computation and wet-lab biology.
Traditional biotech organizations often separate those functions.
The AI group produces predictions.
The lab group tests them.
The results are eventually transferred back.
That structure may become too slow.
Future organizations will increasingly design computational and experimental systems together.
The Talent Market Will Become More Hybrid
Foundation models for biology require unusual teams.
Pure software skills are not enough.
Pure biology skills are not enough.
Companies need people who can operate at the intersection.
Machine-learning researchers need enough biology to understand why experimental data is messy.
Biologists need enough computation to understand what models can and cannot infer.
Automation engineers need to connect laboratory instruments with software.
Drug developers need to decide whether model outputs matter clinically.
This hybrid talent requirement is another reason Boston is well positioned.
The region contains MIT, Harvard, the Broad Institute, major teaching hospitals, biotechnology companies, pharmaceutical research groups, and a growing AI startup ecosystem.
That does not guarantee Boston will dominate the category.
But it creates unusually fertile ground for teams that need several scientific disciplines at once.
What Could Still Go Wrong
Biological foundation models remain early.
Anyone claiming biology is about to become as predictable as software is getting ahead of the evidence.
Biology contains noise, hidden variables, measurement errors, rare events, and enormous differences between laboratory systems and real humans.
Even extremely accurate models can fail when moved into new biological contexts.
Foundation models can also inherit weaknesses from their training data.
If important populations, conditions, cell states, or biological mechanisms are poorly represented, predictions may fail exactly where companies need them most.
Another risk is economics.
Running massive experiments is expensive.
A closed-loop platform may create better models, but it also requires laboratories, equipment, scientists, automation, consumables, and operational complexity.
This is not SaaS.
The strongest companies will need to prove that improved learning rates justify the physical infrastructure required to achieve them.
What Boston Business Leaders Should Watch Through 2027
Three signals will tell us whether foundation models for biology are becoming a durable industry rather than another AI cycle.
The first is clinical proof.
Watch whether AI-designed or AI-optimized medicines produce convincing patient data.
The second is platform reuse.
Watch whether companies repeatedly produce programs rather than celebrating one successful molecule.
The third is experimental scaling.
Watch how quickly companies can create new, useful biological data.
That third metric may be especially important.
The next major breakthrough may not come from doubling model parameters.
It may come from increasing experimental throughput by 10, 100, or 1,000 times.
Original Finding #5: Experimental Throughput Could Become Biology’s Equivalent of Compute
The modern AI industry became more powerful partly because companies could scale computation.
More compute allowed larger models to process more data and run more experiments.
Biology may develop a parallel scaling law.
But instead of only scaling GPUs, biotech companies may scale experiments.
Consider the pattern.
Lila wants autonomous science factories capable of producing scientific data continuously.
Manifold says it can measure more than 500,000 binders within a single in-vivo experiment.
Cellular Intelligence has built technology specifically intended to generate cell-state information efficiently for AI training.
Generate has built a large-scale protein design-build-measure-learn system and reports tens of thousands of proteins already generated, built, and tested.
Asimov is connecting AI with programmable cloud-lab infrastructure to create experimental datasets around protein production.
These are not isolated examples.
They point toward a new scaling model.
OLD BIOTECH SCALING
More scientists
↓
More experiments
↓
More discoveries
AI-NATIVE BIOLOGY SCALING
Better models
↓
Better experiment selection
↓
More automated experiments
↓
More proprietary data
↓
Better models
↺The company that closes this loop fastest may gain a compounding advantage.
Original Finding #6: Boston Could Become Strongest Where AI Meets the Physical World
Boston probably will not beat Silicon Valley by copying Silicon Valley.
It does not need to.
The region has a different advantage.
AI is moving out of purely digital environments.
It is entering laboratories, robotics, medicine, manufacturing, materials, and biology.
Those fields require physical infrastructure and scientific expertise.
Boston already has both.
Foundation models for biology may therefore become one of the best examples of a broader Boston technology strategy.
Build companies where advanced AI must interact with difficult physical systems.
That includes drug discovery.
It includes autonomous laboratories.
It includes robotics.
It includes manufacturing.
It includes healthcare.
It includes advanced materials.
The more AI moves into the physical world, the more valuable deep scientific ecosystems become.
The Bigger Business Opportunity: Biological Intelligence as Infrastructure
The foundation-model market could eventually split into several business models.
Some companies will develop their own medicines.
Others will license models or platforms.
Some will partner with pharmaceutical companies.
Some will sell infrastructure.
Others may provide laboratory capacity.
Several will probably combine all of these approaches.
This makes biological AI look less like a normal software category and more like infrastructure.
Companies that control important models, datasets, laboratories, and workflows could become foundational suppliers to the entire biotech industry.
A protein foundation model could support dozens of therapeutic companies.
A virtual-cell model could help researchers understand drug mechanisms.
A disease model could help pharmaceutical companies design smarter clinical trials.
A manufacturing model could reduce development failures.
A scientific-agent platform could coordinate experiments across all of them.
The value would not come from replacing scientists.
It would come from giving every scientist a much larger search space and a faster feedback loop.
The Next Competitive Battle Will Be Over Who Owns the Feedback Loop
The next phase of biological AI will probably not be defined by who releases the most impressive model demo. It will be defined by who controls the best feedback loop between prediction, experiment, and real-world evidence.
That distinction matters because biological models can become outdated very quickly if they are separated from new experimental data. A company may begin with an excellent model trained on public protein, cell, or disease datasets, but competitors can often access much of the same starting information. What becomes harder to copy is the flow of proprietary information produced after the company starts operating.
Every failed protein can teach a model something.
Every unexpected cell response can reveal a weakness in an earlier prediction.
Every manufacturing problem can become training data for better process models.
Every clinical response can improve the company’s understanding of disease.
The strongest companies will therefore treat failures as data rather than simply setbacks.
Failed Experiments Could Become Valuable Assets
This is one of the least discussed advantages of AI-native biotech.
Traditional scientific organizations often celebrate successful experiments and move past unsuccessful ones. In reality, failed experiments can contain some of the most useful information in the entire development process.
A protein that refuses to express properly tells researchers something about manufacturability.
A cell that behaves differently than predicted tells them where the model is incomplete.
A molecule that works in a dish but fails inside an animal reveals something about the gap between simplified laboratory systems and real biology.
AI systems can potentially learn from all of this.
That means the value of a biotech company may increasingly depend on how well it captures negative results, unexpected outcomes, borderline results, and failed designs.
The organization that records only successes throws away part of its learning advantage.
The organization that turns every experiment into structured training data gets stronger even when individual projects fail.
Data Quality Will Matter More Than Raw Data Volume
The AI industry often talks about having more data.
Biology requires a more careful approach.
One million poorly controlled experiments may be less useful than 50,000 highly consistent experiments designed specifically to answer important questions.
Scientific data can be noisy. Different instruments produce slightly different readings. Different laboratories may perform the same experiment in different ways. Samples can vary. Cell cultures change. Human patients are even more complicated.
This means biological foundation-model companies need more than huge datasets.
They need clean experimental systems.
They need standardized measurements.
They need metadata explaining exactly how every result was produced.
They need strong quality controls.
They need systems that make results comparable across time.
This is another area where automation could matter enormously.
The more standardized an automated laboratory becomes, the easier it may be to generate machine-readable data at scale. That could make automated labs valuable not only because they reduce labor, but because they create more consistent training data for AI.
Boston Could Build a Shared Infrastructure Layer Around This Market
There is also a larger regional opportunity.
Not every Boston biotech startup will be able to build its own automated laboratory, large-scale sequencing operation, protein-expression system, animal-testing infrastructure, and machine-learning platform.
That creates room for shared infrastructure companies.
Boston could support specialized businesses that provide automated experimental capacity, high-throughput assays, biological data pipelines, AI-ready laboratory software, robotics, model evaluation, secure compute, and scientific data management.
This would allow smaller companies to participate in the foundation-model economy without recreating every layer internally.
The effect could look similar to what cloud computing did for software startups.
A small software company no longer needs to build a data center before launching a product.
In the future, a small biotech company may not need to own every experimental system before building an AI-native discovery platform.
If Boston develops this infrastructure early, the region could benefit even when individual drug programs fail.
The ecosystem itself becomes more productive.
The Winning Companies May Look Less Like Software Businesses Over Time
There is another important implication for investors.
Some of the most valuable biological AI companies may become more capital intensive as they improve.
That sounds counterintuitive because software usually becomes more efficient as it scales.
Biological AI can scale differently.
A better model may generate more useful hypotheses.
More useful hypotheses create demand for more experiments.
More experiments require more automation, instruments, laboratory capacity, manufacturing infrastructure, and biological materials.
The result is a company that may combine the economics of software, biotechnology, and industrial infrastructure.
That makes traditional SaaS metrics less useful.
Gross margins matter, but so does experimental throughput.
Software revenue matters, but so does the value of the proprietary dataset being created.
Headcount matters, but so does the number of useful experiments each scientist can supervise.
Drug pipelines matter, but so does how much every program improves the underlying platform.
Boston investors will need new ways to evaluate these businesses.
The best question may eventually be simple:
How much smarter does this company become every time it spends another dollar on science?
Final Takeaway: Boston Is Building the Learning Systems Behind the Next Generation of Biotech
Foundation models for biology are sometimes described as biotechnology’s version of large language models.
That comparison is useful, but incomplete.
Biology has something language does not.
The truth often has to be discovered experimentally.
You cannot scrape the internet to learn how a completely new protein behaves inside the human body.
Someone must build it and test it.
You cannot download a perfect dataset showing every possible way a human cell responds to every possible intervention.
Those experiments have not happened yet.
This is why Boston’s emerging model is so interesting.
Generate is connecting generative protein models with large-scale physical testing.
Cellular Intelligence is connecting virtual-cell models with proprietary cell experiments.
Manifold Bio is linking AI design with massive in-vivo measurements.
BostonGene is modeling disease and patient response.
Lila Sciences is trying to automate the scientific learning cycle itself.
Asimov is bringing AI into biological engineering and manufacturing.
Dyno is using machine learning to explore gene-delivery designs that humans could never search manually.
AbInitio Bio is attempting to bring transferable foundation models into biologics manufacturing.
The common theme is not simply AI.
It is learning.
The most valuable platforms may become systems that learn faster every time scientists run an experiment, design a molecule, test a cell, study a patient, or manufacture a biological product.
That changes what competitive advantage means.
The biggest model may not win.
The company with the most GPUs may not win.
The company with the largest public dataset may not win.
The strongest businesses may be the ones that build the fastest loop between prediction and reality.
Boston already has many of the ingredients needed to build those loops: software talent, laboratories, hospitals, pharmaceutical partners, biological infrastructure, research institutions, capital, and people who understand how difficult science actually works.
That is why foundation models for biology matter far beyond another wave of AI startups.
They could become the intelligence layer sitting underneath a new biotechnology industry.
And if the next generation of biological AI depends as much on experiments as it does on algorithms, Boston has a genuine chance to become one of the places where that future is built.
