Drug discovery has always involved a painful amount of searching.
Scientists identify something in the body that may be linked to disease. They search for a molecule that might affect it. They test many possible molecules. Most fail. The promising ones are changed, tested again, and slowly pushed through years of laboratory work before anyone knows whether they could become useful medicines.
Generative biology is starting to change that process.
Instead of asking artificial intelligence to simply predict whether an existing molecule might work, scientists are increasingly asking AI systems to design new biological molecules with specific properties from the start.
That difference is important.
A traditional computational system might search through a library of known proteins and rank the most promising options. A generative system can attempt to create a protein sequence that did not previously exist, based on the job scientists want the molecule to perform.
Boston has become one of the most interesting places to watch this shift because the region already has nearly every ingredient needed to turn computational designs into real experimental programs: machine-learning researchers, protein engineers, major research hospitals, pharmaceutical companies, wet labs, venture investors, and a very large biotechnology workforce.
The result is a growing group of companies that are not merely putting AI software on top of traditional biotechnology.
They are rebuilding the discovery process itself.
Generate Biomedicines is designing new protein medicines and has already moved an AI-engineered antibody into global Phase 3 trials. Manifold Bio is connecting generative protein design with highly scaled testing inside living systems. Nabla Bio is using computational and experimental tools to design antibodies against difficult drug targets. Dyno Therapeutics is applying machine learning to the protein shells used to deliver gene therapies. Neoclease is applying generative AI to the design of custom gene-editing enzymes.
These companies are approaching different parts of medicine, but they point toward the same larger idea.
Biology is becoming something scientists can increasingly design rather than simply search.
That does not mean drug discovery has suddenly become easy. Biology remains messy, clinical trials remain risky, and even an excellent AI-designed molecule can fail once it meets the complexity of a living human body.
But Boston’s generative-biology companies are showing what the next drug-development system could look like.
The Short Version: Generative Biology Turns Drug Discovery Into a Design Problem
The easiest way to understand generative biology is to compare it with generative AI systems people already know.
A text-generation system learns patterns from enormous amounts of language. Give it instructions, and it can create a new sequence of words that fits those instructions.
A generative protein system works with a very different kind of sequence.
Proteins are built from chains of amino acids. The sequence influences how a protein folds, what shape it takes, what other molecules it can bind to, how stable it is, and ultimately what biological job it can perform.
Generative models can learn patterns connecting protein sequence, structure, and function. Scientists can then use those models to propose new sequences intended to meet particular biological requirements.
Generate Biomedicines describes its own platform as a continuous cycle of generating protein sequences, building them as physical proteins, measuring their properties, and using the resulting data to improve future designs. The company says it has generated, built, and tested more than 42,000 proteins.
That cycle explains more about generative biology than the phrase “AI drug discovery” ever could.
The important part is not simply generation.
The important part is generation followed by physical testing followed by learning.
Boston appears unusually well suited to this model because computational biology and experimental biology can exist within the same company, the same neighborhood, and sometimes the same building.
That shortens the distance between an AI prediction and a laboratory result.
And that distance may become one of the most important competitive advantages in biotechnology.
What Generative Biology Actually Means
The term can sound more complicated than the underlying idea.
Traditional drug discovery often begins with what already exists.
Researchers may examine natural proteins, antibodies, chemical libraries, previously discovered molecules, or biological samples. They search for candidates that have useful characteristics and then improve them through repeated rounds of experimentation.
Generative biology changes the starting point.
From Finding Molecules to Specifying Molecules
Imagine that researchers need a protein that can:
bind very strongly to a disease-related target,
avoid binding to similar healthy proteins,
remain stable long enough to function as a medicine,
avoid creating an unwanted immune response,
and be manufactured efficiently.
Historically, scientists might begin with an existing molecule that does one or two of those things reasonably well and spend years improving it.
A generative approach attempts to work in the opposite direction.
Researchers describe the properties they want, and computational systems search the enormous space of possible biological sequences for designs likely to satisfy those requirements.
The final molecule still has to be built and tested.
But the search process becomes far more directed.
Protein Space Is Far Bigger Than Nature Has Explored
This matters because the number of proteins that could theoretically exist is enormous.
Nature has explored only a tiny fraction of the possible sequence combinations.
Evolution did not spend billions of years trying to create the perfect asthma medicine, tumor-targeting antibody, gene-therapy delivery vehicle, or gene-editing enzyme.
It produced proteins that helped organisms survive and reproduce.
Generative biology asks a new question:
What useful proteins could exist that evolution simply never needed to create?
Generate’s Chroma model, for example, was developed to create protein structures and sequences while allowing scientists to steer designs toward desired geometric and functional properties. Generate says its wider platform combines data from natural proteins with proprietary experimental information to learn relationships between protein sequence, structure, and function.
That opens a very different search space from conventional drug discovery.
Why Boston Is Particularly Important to This Market
Artificial intelligence can be developed anywhere computers and skilled researchers exist.
Generative biology is harder.
It needs computation, but it also needs laboratories.
It needs molecular biology.
It needs automated experiments.
It needs manufacturing knowledge.
It needs people who understand disease.
Eventually, it needs clinical-development expertise.
Boston already has unusual density across all of those layers.
That is why the region’s competitive advantage should not be described simply as “Boston has good AI talent.”
San Francisco has outstanding AI talent too.
Boston’s more distinctive advantage is that advanced computation sits beside one of the world’s deepest concentrations of biological research.
Harvard, MIT, the Broad Institute, major hospitals, biotech companies, pharmaceutical research operations, incubators, venture firms, and specialized scientific service companies create an environment in which an idea can move between software and physical biology unusually quickly.
The origin stories of several companies make this visible.
Dyno Therapeutics grew from work connected to George Church’s laboratory at Harvard Medical School and the Wyss Institute. Manifold Bio also emerged from work connected to the Church lab and Wyss Institute. Nabla Bio’s founders similarly came through the Harvard ecosystem.
This is not simply an academic story.
Those companies have turned research into commercial platforms used by major pharmaceutical businesses.
That transition from university research to industrial-scale biological design is becoming one of Boston’s most valuable technology pipelines.
BostonBizScene Original Research: How We Analyzed the Local Generative-Biology Market
To understand what is actually happening in Boston, BostonBizScene analyzed publicly available information for a focused group of Boston-area companies that meet a strict definition of generative or AI-directed biological design.
The analysis was conducted using public company materials, scientific descriptions, announced partnerships, financing announcements, investor materials, SEC filings where available, and independent institutional sources.
The cutoff for this analysis is October 3, 2026.
Our Inclusion Rules
We included a company when it met three conditions.
First, the company needed a meaningful Boston-area operating presence.
Second, AI or machine learning needed to participate directly in the design or engineering of a biological molecule, protein, gene-editing system, or therapeutic delivery component.
Third, the resulting designs needed to connect to experimental biological validation rather than remaining a pure software product.
Using this method, five companies emerged as especially useful examples for comparison:
| Company | Boston-area base | Primary design problem | What AI helps design |
|---|---|---|---|
| Generate Biomedicines | Somerville | Therapeutic proteins | Antibodies and other protein medicines |
| Manifold Bio | Boston | Tissue-targeted biologics | Protein binders, antibodies and shuttles |
| Nabla Bio | Cambridge | Difficult protein drug targets | Antibodies and custom biologics |
| Dyno Therapeutics | Greater Boston | Gene-therapy delivery | AAV capsids |
| Neoclease | Boston | Precision gene editing | Custom nucleases |
Sources describing each company’s current approach support these classifications.
This is intentionally not a ranking of every AI-biotech company in Massachusetts.
The goal is narrower.
We want to understand how Boston companies are turning generative models into experimentally validated medicine-design systems.
Original Finding #1: Boston’s Real Strength Is the Closed Biological Learning Loop
One of the strongest patterns in our analysis is that the leading companies are not treating AI as an isolated prediction tool.
They are building feedback systems.
The basic architecture looks like this:
Design → Build → Test → Measure → Learn → Design again
Generate makes this loop explicit in its platform description. Manifold similarly connects AI-driven protein generation with large-scale experimental testing. Neoclease describes a closed-loop workflow where design, simulation, validation, testing, and new structured training data feed one another.
We coded the five companies for whether public evidence showed both computational design and a meaningful experimental feedback layer.
Boston Generative-Biology Feedback Loop Analysis
| Company | AI-guided design | Internal or integrated experimental validation | Feedback into future design | Closed-loop evidence |
|---|---|---|---|---|
| Generate Biomedicines | Yes | Yes | Yes | Strong |
| Manifold Bio | Yes | Yes | Yes | Strong |
| Nabla Bio | Yes | Yes | Yes | Strong |
| Dyno Therapeutics | Yes | Yes | Yes | Strong |
| Neoclease | Yes | Yes | Yes | Strong |
The important result is not that five out of five companies use artificial intelligence.
That was part of the inclusion rule.
The important result is that all five connect computational design to physical biological data.
This suggests that Boston’s generative-biology ecosystem is developing around a different competitive moat from general-purpose AI.
The model itself matters.
But proprietary experimental data may matter just as much.
Why Proprietary Experimental Data Could Become the Bigger Moat
Public protein databases give every company a powerful starting point.
The problem is that competitors can often access much of the same public information.
An experiment run inside a company’s laboratory is different.
If a company designs 100,000 molecules, physically produces them, measures what they do, and records the failures as carefully as the successes, it begins building a dataset competitors cannot easily copy.
That dataset can then improve the next generation of models.
The improved models create better designs.
Better designs can produce better experiments.
Better experiments create even more valuable data.
The advantage can compound.
This is one reason Boston’s wet-lab infrastructure matters so much.
The most valuable generative-biology company may not eventually be the business with the largest AI model.
It may be the company that can run the most informative biological learning loop.
Generate Biomedicines: The Clearest Test of the Entire Generative-Biology Thesis
Generate Biomedicines is probably the most direct example of what the phrase “generative biology” means commercially.
The company grew out of Flagship Pioneering and is based in Somerville.
Its platform is designed around a continuous process of generating potential protein medicines, physically building those proteins, measuring their properties, and learning from the resulting data.
This is important because Generate is no longer only an interesting research platform.
It has become a clinical-stage biotechnology company.
The Number That Matters: 42,000 Proteins
Generate says it has generated, built, and tested more than 42,000 proteins.
That number is more meaningful than the size of a model by itself.
Every experimentally tested design can potentially teach the platform something about the connection between sequence, structure, biological function, manufacturability, stability, or other properties.
This is the central economic idea behind closed-loop generative biology.
The company is not merely buying compute to create predictions.
It is investing in a machine that converts physical biological experiments into proprietary training information.
GB-0895 Moves the Debate From Models to Medicines
The most important proof point may be GB-0895.
GB-0895 is an AI-engineered antibody targeting TSLP, a protein involved in airway inflammation. Generate began global Phase 3 development for severe asthma, with two studies planned to evaluate roughly 1,600 patients.
By mid-2026, recruitment was underway across six global regions. Generate also reported Phase 1 activity for other programs, including GB-4362, while continuing additional preclinical work.
This matters far beyond one company.
Generative drug design has often been discussed using model benchmarks, laboratory performance, or the speed at which molecules can be proposed.
Clinical development creates a much harder test.
The question becomes:
Can a molecule partly shaped by generative design survive the same safety, efficacy, manufacturing, regulatory, and commercial filters faced by every other medicine?
That is the point where generative biology stops being mainly an AI story.
It becomes a pharmaceutical-development story.
The Capital Required Is Also Becoming Visible
Generate completed an initial public offering in February 2026, pricing 25 million shares at $16 and targeting approximately $400 million in gross proceeds before fees and expenses.
The company’s second-quarter update reported approximately $457.4 million of cash, cash equivalents, and marketable securities as of June 30, 2026.
This highlights an important truth that gets lost in discussions about AI reducing drug-development costs.
Generative AI may improve discovery efficiency.
It does not make clinical development cheap.
Moving drugs through human trials still requires enormous amounts of capital.
Manifold Bio: The Bottleneck May Be Testing What AI Designs
If Generate illustrates the complete design-to-clinic vision, Manifold Bio highlights a different problem.
AI systems are becoming very good at producing molecular ideas.
What happens if models eventually generate useful designs faster than laboratories can test them?
That creates a new bottleneck.
Manifold is attacking that bottleneck.
Designing Millions of Proteins Is Not Useful If You Cannot Learn From Them
Manifold’s platform combines AI-guided protein design with massively multiplexed experimental measurement.
Its strategy is unusual because the company wants to move biological testing deeper into living systems rather than relying only on simplified laboratory experiments.
The company’s reasoning is straightforward.
A molecule can look excellent in a test tube and still behave badly inside an organism.
Distribution matters.
Metabolism matters.
Immune response matters.
The ability to cross tissue barriers matters.
The location where a drug accumulates matters.
Manifold therefore describes its system as a direct-to-vivo discovery engine designed to learn properties such as pharmacokinetics and biodistribution inside living systems.
One Million Designs Creates a Very Different Type of Dataset
In March 2026, Manifold reported work with NVIDIA in which its platform tested one million protein-binder designs against 127 targets in a multiplexed experiment measuring more than 100 million protein-protein interactions.
The company reported finding specific binders for 68% of the targets tested.
Those numbers illustrate why experimentation is becoming part of the AI infrastructure itself.
Chart: Scale of Manifold’s Reported 2026 Protein-Binder Validation Experiment
| Measure | Reported scale |
|---|---|
| Binder designs tested | 1,000,000 |
| Targets | 127 |
| Protein-protein interactions measured | 100,000,000+ |
| Targets with specific binders identified | 68% |
Source: Manifold Bio’s March 2026 announcement.
A model can create enormous numbers of sequences cheaply.
But without high-quality labels showing what those sequences actually do, model improvement eventually stalls.
Manifold is effectively trying to industrialize the creation of those labels.
Roche Gives the Strategy Commercial Validation
In November 2025, Roche agreed to work with Manifold on next-generation brain shuttles designed to help therapeutic payloads reach the brain.
Manifold disclosed a $55 million upfront payment and potential research, development, and commercial milestone payments exceeding $2 billion, plus royalties.
Again, the most interesting point is not the headline milestone number because those payments depend on future success.
The important signal is that a major pharmaceutical company is paying meaningful upfront money to access a discovery system built around AI plus in-vivo biological measurement.
That suggests the experimental layer itself has commercial value.
Nabla Bio: Generative Biology Is Moving Toward Difficult Targets
Nabla Bio provides another important piece of the Boston story.
Its focus is antibody and protein design, particularly around difficult targets that conventional discovery methods have struggled to address.
The Cambridge company combines AI with wet-lab experimentation and has announced collaborations involving AstraZeneca, Bristol Myers Squibb, and Takeda.
Why Difficult Targets Matter
Drug discovery has never been evenly distributed across biology.
Some disease targets are relatively easy for conventional drug technologies to reach.
Others are extremely difficult because of their structure, location, flexibility, or other biological features.
If generative design simply creates slightly better versions of molecules against targets that pharmaceutical companies can already drug easily, the economic impact will still be useful.
But the more important opportunity is different.
Generative models may allow companies to design molecules for biological targets that traditional discovery systems frequently miss.
Nabla says its platform is being applied to antibodies and other biologics and is designed to give scientists control over important drug-like properties from the beginning.
That points toward one of the largest long-term opportunities in generative biology.
The technology may not simply make existing discovery faster.
It may expand the set of biological problems that drug companies can realistically attempt to solve.
Pharma Partnerships Are Becoming a Business Model
Nabla is also useful because it illustrates the platform-partnership business model.
Not every generative-biology company has to finance dozens of drugs through clinical development itself.
A company can develop a powerful biological-design engine and allow larger pharmaceutical companies to apply that engine across their own pipelines.
Nabla announced a $26 million Series A in 2024 and said at the time that collaborations with AstraZeneca, Bristol Myers Squibb, and Takeda represented more than $550 million in potential upfront and milestone payments plus royalties.
The company later announced another Takeda collaboration in October 2025.
The distinction between guaranteed payments and possible future milestones is important.
Milestone-heavy deal values should never be treated like money already received.
But repeated deals remain useful evidence that major drug companies are willing to put generative protein design into real discovery programs.
Dyno Therapeutics: Generative Biology Is Not Only About the Medicine
Dyno Therapeutics broadens the definition of what AI can design.
Instead of focusing mainly on the therapeutic payload, Dyno applies artificial intelligence to the delivery vehicle used in gene therapy.
The company designs AAV capsids.
A capsid is the protein shell around the genetic material delivered by an adeno-associated virus-based gene therapy.
That shell plays a major role in determining where the therapy goes and how effectively it reaches particular tissues.
Delivery Is One of the Hardest Problems in Genetic Medicine
Gene therapy can only work if the therapeutic genetic instructions reach the right cells.
Naturally occurring viral vectors were not designed by evolution to function as perfect medical delivery systems.
They come with limitations involving tissue targeting, immune response, packaging capacity, manufacturing, and other characteristics.
Dyno’s CapsidMap platform uses machine learning with experimental data to search for new capsids with improved properties.
This represents another form of generative biology.
The AI system is not necessarily designing the active drug itself.
It is redesigning the biological infrastructure required to deliver treatment.
That distinction matters because delivery remains a major bottleneck across gene therapy, RNA medicine, gene editing, and other advanced therapeutic areas.
Dyno Shows the Longevity of Boston’s AI-Biology Thesis
Dyno was applying machine learning to biological design before the current generative-AI boom.
The company announced collaborations with Novartis and Sarepta in 2020 and later expanded work with Roche. Its 2021 Series A raised $100 million.
A second Roche partnership announced in 2024 included $50 million upfront and potential milestones exceeding $1 billion.
The lesson is useful.
Boston’s current generative-biology market did not appear because ChatGPT made generative AI fashionable.
Researchers in the region had already been using machine learning to navigate biological sequence space.
What changed is the scale and capability of the models.
Neoclease: Designing the Gene Editor Itself
Neoclease pushes the design concept into another important area: gene editing.
Many gene-editing systems rely on enzymes originally found in nature.
Neoclease is trying to use generative AI to design custom nucleases for particular genetic targets instead.
The company says its NeoMini editors are about half the size of existing standard editors and are designed for gene-specific editing.
Its first publicly highlighted program targets LRRK2 in Parkinson’s disease.
Why Smaller Editors Could Matter
Delivery is one of the major limitations of gene editing.
A powerful editor is not useful if it cannot efficiently reach the cells where treatment is needed.
Smaller molecular systems may create more delivery options.
That means AI-designed gene editors could potentially be optimized around several requirements simultaneously:
editing activity,
target specificity,
molecular size,
delivery constraints,
and unwanted biological effects.
That multi-property design problem is exactly where generative approaches become interesting.
Nature offers researchers a collection of gene-editing tools.
Generative biology asks whether scientists can deliberately create better ones.
Original Finding #2: Boston Generative Biology Is Much More Protein-Heavy Than the AI Hype Suggests
When people hear “AI drug discovery,” they often imagine one giant software system searching all of chemistry.
The Boston market looks more specialized.
Four of the five companies in our core sample directly design or engineer proteins as a central part of their platform.
The fifth, Dyno Therapeutics, designs AAV capsids, which are themselves protein structures.
In other words, our sample is effectively 5 out of 5 protein-centric at the physical design layer.
Chart: Biological Objects Being Designed in Our Boston Sample
| Design category | Companies in sample | Share |
|---|---|---|
| Therapeutic proteins / antibodies | 3 | 60% |
| Protein-based gene-editing enzymes | 1 | 20% |
| Protein-based gene-delivery capsids | 1 | 20% |
| Total involving protein engineering | 5 | 100% |
This does not mean Boston lacks AI companies working on small molecules, cells, RNA, clinical development, genomics, or other forms of biology.
It means that the generative-design companies in our tightly defined sample show a striking concentration around proteins.
Why?
Proteins are particularly suitable for this type of computational approach because sequence, structure, and function are deeply connected.
They are also extraordinarily flexible therapeutic tools.
Antibodies can bind targets.
Enzymes can catalyze reactions.
Cytokines can change immune behavior.
Protein shells can deliver gene therapies.
Engineered proteins can alter cellular pathways.
The protein itself can therefore become programmable therapeutic infrastructure.
That appears to be one of Boston’s strongest emerging specialties.
Original Finding #3: The Market Is Splitting Into Three Different Business Models
Another pattern appears when we look at how these companies expect to make money.
They are not all building the same type of biotech company.
We see three business models developing simultaneously.
Model One: Build Your Own Medicines
Generate provides the clearest example.
It owns internal drug programs and is financing clinical development.
This model offers the greatest possible long-term upside if a drug reaches the market.
It is also extremely expensive.
Generate reported $57.8 million of research and development expense in the first quarter of 2026 alone, partly reflecting investment in its Phase 3 asthma program.
That number shows why “AI makes discovery cheaper” cannot be confused with “AI makes biotech cheap.”
Clinical trials remain capital intensive.
Model Two: Sell Access to the Design Engine Through Partnerships
Nabla and Dyno illustrate this model particularly clearly.
Their technology becomes part of larger pharmaceutical discovery programs.
The platform company can receive upfront payments, research funding, milestones, and royalties without carrying every resulting medicine all the way through development.
The tradeoff is economic.
The partner keeps a large share of the eventual drug economics.
But the platform company can potentially spread its technology across many programs and several partners.
Model Three: Combine Partnerships With an Internal Pipeline
Manifold is pursuing a hybrid structure.
Its Roche collaboration validates the platform commercially while the company also retains rights and develops its own programs.
This may become increasingly common.
The partnership generates cash, outside validation, and development experience.
The internal pipeline preserves greater upside.
Boston’s Emerging Generative-Biology Business Model Map
| Company | Internal therapeutics | Pharma partnerships | Platform economics | Hybrid model |
|---|---|---|---|---|
| Generate | Strong | Strong | Yes | Yes |
| Manifold | Yes | Strong | Yes | Strong |
| Nabla | Developing/licensing | Strong | Strong | Yes |
| Dyno | Partner-led focus | Strong | Strong | Yes |
| Neoclease | Yes | Earlier stage | Developing | Developing |
The implication for founders is important.
Building a generative-biology company does not automatically require becoming a fully integrated pharmaceutical company.
But the business model needs to match the technology.
Original Finding #4: Experimental Throughput May Become More Valuable Than Model Size
The AI industry has trained people to look at model size, compute budgets, and benchmark scores.
Those measures may be less useful in biology.
A drug-design model eventually needs biological truth.
That comes from experiments.
Consider Manifold’s million-design experiment.
Its strategic value is not only that one million proteins could be generated.
Generative models can increasingly produce sequences at enormous scale.
The harder capability is connecting those sequences to more than 100 million measured interactions.
Generate’s disclosure that more than 42,000 proteins have been generated, built, and tested tells a similar story.
The words built and tested are doing enormous work there.
Anyone can create millions of computer files containing theoretical protein sequences.
A biotechnology company creates value when it can determine which sequences become useful physical molecules.
A Better KPI for Generative Biology
This suggests that investors and operators should stop asking only:
“How powerful is your AI model?”
A better set of questions would be:
How many designs can you physically build?
How quickly can you test them?
How biologically relevant are those tests?
How much useful information does each experiment produce?
How rapidly does that information return to the model?
How much does performance improve between design cycles?
Those questions measure the learning machine rather than merely the model.
Original Finding #5: Clinical Validation Is Becoming the New Benchmark
For several years, AI-biotech companies competed largely through computational benchmarks and preclinical experiments.
That phase is changing.
Generate’s GB-0895 program has entered Phase 3 development for severe asthma.
This creates a much harder standard for the field.
The ultimate test of generative biology is not whether AI can produce a protein that looks convincing on a computer.
It is whether the resulting medicines can help patients safely and meaningfully.
The Generative-Biology Evidence Ladder
BostonBizScene groups evidence into five levels:
| Evidence level | What it proves |
|---|---|
| Computational generation | A model can produce plausible designs |
| In-vitro testing | The physical molecule can perform in controlled experiments |
| In-vivo testing | The molecule can function in a living biological system |
| Human clinical testing | The candidate can be evaluated in patients |
| Late-stage clinical validation | The technology has produced a candidate capable of reaching major pivotal trials |
Different companies sit at different points on this ladder.
That is expected.
A platform may be scientifically powerful years before its own drug enters the clinic.
But the distinction helps prevent investors, journalists, and customers from treating every AI-biotech claim as equivalent.
A beautiful model demonstration is not the same thing as clinical evidence.
The Biggest Strategic Shift: AI Is Moving From Prediction to Creation
The first generation of AI in biotechnology was heavily focused on prediction.
Predict protein structures.
Predict which molecules will bind.
Predict toxicity.
Predict clinical outcomes.
Prediction remains enormously useful.
But generative biology changes the economic role of AI.
The system becomes an inventor of candidates.
It proposes the molecular object that will be physically made.
That is a much deeper place in the biotechnology value chain.
Prediction Helps You Choose
A predictive model might examine 10,000 existing molecules and estimate which one is most promising.
Generation Helps You Create
A generative system asks whether the best molecule may not exist yet.
It can create candidates specifically shaped around desired characteristics.
That does not eliminate selection.
Generative systems may produce thousands or millions of candidates that still need ranking and testing.
But the starting universe becomes far larger.
What Pharmaceutical Companies Should Do Differently
The rise of generative biology changes how pharmaceutical companies should evaluate AI partnerships.
The obvious mistake is choosing vendors mainly because their models appear impressive.
A strong demonstration is not enough.
Pharma companies should examine the complete learning system.
Ask Where the Training Data Comes From
Public data can help start a platform.
Proprietary experimental data can make it increasingly differentiated.
A useful question is therefore not simply how much data a company has.
It is whether the company can continuously generate new, relevant biological data that competitors cannot easily reproduce.
Ask What Happens After the Model Produces a Sequence
A design is not a medicine.
Who manufactures the molecule?
How quickly is it tested?
Which properties are measured?
Does testing happen only in vitro?
Does the company have access to in-vivo validation?
How does the result change the next model?
A platform without a strong answer to those questions may simply move the bottleneck downstream.
Judge Platforms on Useful Molecules, Not Generated Molecules
“One million molecules generated” sounds impressive.
The economically important number is how many useful molecules emerge from the million.
Even better questions involve improvement.
How many experiments are required to reach an acceptable molecule?
How quickly does the company move from design requirements to a development candidate?
What percentage of designs survive each stage?
These metrics would tell customers much more about productivity.
What Boston Startups Can Learn From This Market
Generative biology looks attractive because enormous pharmaceutical markets are involved.
But it is a difficult startup category.
A new company should avoid building “AI for biology” without choosing a very specific bottleneck.
The Boston companies that stand out are solving sharply defined problems.
Generate designs therapeutic proteins.
Dyno designs gene-delivery capsids.
Nabla tackles difficult antibody targets.
Manifold focuses on connecting protein design with biologically relevant in-vivo data.
Neoclease designs custom gene-editing enzymes.
The technology may be broad.
The first customer problem should not be.
Own a Feedback Loop, Not Just a Model
A model can become commoditized.
Open-source biological AI is improving.
Academic research moves quickly.
Large technology companies can spend far more on compute than most biotech startups.
That makes a defensible feedback loop particularly valuable.
A startup that controls proprietary experimental systems may have something much harder to copy.
Solve an Expensive Bottleneck
The strongest applications of generative biology are likely to sit where conventional biotechnology already spends large amounts of time and money.
Examples include:
finding antibodies against hard targets,
improving tissue-specific delivery,
optimizing many molecular properties at once,
designing enzymes for precise biological functions,
and testing large numbers of molecules in relevant biological systems.
The bigger the existing bottleneck, the more economic value there is in reducing it.
What Investors Should Watch Instead of AI Hype
Generative biology creates a difficult diligence problem.
Nearly every biotech company can now describe itself using AI language.
Investors need stronger filters.
The first question should be whether the AI meaningfully changes the biology.
If the same company could exist almost unchanged without machine learning, AI may be a feature rather than the core platform.
The second question should be whether the company produces proprietary biological data.
The third should be whether experimental throughput increases as computational throughput increases.
The fourth should be whether external organizations are paying real money to use the platform.
The fifth should be whether molecules are advancing toward increasingly difficult forms of validation.
These measures provide a much clearer picture than the number of model parameters.
The Funding Data Needs Careful Interpretation
Funding headlines can make generative biology appear simpler than it is.
Generate’s $400 million IPO is not comparable with Dyno’s $100 million Series A, Manifold’s earlier $40 million Series A, or Nabla’s $26 million Series A.
They occurred in different years and at very different company stages.
BostonBizScene therefore does not add these rounds together and call the result the “size of the Boston generative-biology market.”
Doing so would produce a neat number but weak analysis.
The more useful conclusion is structural.
Capital is available across several stages: early venture funding, large growth rounds, pharmaceutical collaborations, and now the public markets.
Generate’s 2026 IPO is particularly important because it shows that a generative-biology company can progress far enough for public investors to evaluate the platform alongside clinical assets.
That does not prove the business model will succeed.
It proves the category has moved well beyond seed-stage experimentation.
Why Pharma Partnerships Matter Almost as Much as Venture Funding
Partnership economics provide another useful signal.
Across the Boston companies examined here, large pharmaceutical companies including Roche, Novartis, AstraZeneca, Bristol Myers Squibb, Takeda, Sarepta, and others have entered collaborations involving AI-directed biological design.
Those partnerships matter because pharmaceutical companies already have large internal research organizations.
If they choose an external platform, they are effectively saying that the platform offers a capability worth accessing rather than rebuilding immediately inside the company.
The details still matter.
Potential milestone values can stretch over many years and depend on successful development.
A “$2 billion deal” does not mean a startup received $2 billion.
Manifold’s Roche agreement, for example, included a disclosed $55 million upfront payment while the much larger number represents possible future milestones. Dyno’s 2024 Roche deal similarly included $50 million upfront alongside more than $1 billion in potential milestones.
Upfront cash is therefore a stronger near-term validation signal than the maximum headline value.
Original Finding #6: Boston Is Building an AI-to-Experiment-to-Clinic Stack
Looking across the companies together reveals something larger than individual startups.
Boston is developing multiple layers of a biological-design stack.
At one layer are generative models.
At another are automated or highly scaled wet-lab systems.
At another are in-vivo measurement systems.
At another are delivery technologies.
At another are therapeutic-development teams.
At another are hospitals and clinical investigators capable of running human studies.
This means the region is not merely creating AI software for biotechnology.
It is creating infrastructure that connects computation all the way to medicine.
Boston’s Emerging Generative-Biology Stack
| Layer | Example capability |
|---|---|
| Biological foundation models | Learning sequence, structure and function |
| Generative design | Creating new protein or enzyme sequences |
| Computational filtering | Removing poor candidates before testing |
| Automated build systems | Producing designed molecules |
| High-throughput testing | Measuring binding, function and stability |
| In-vivo measurement | Learning behavior inside living systems |
| Delivery engineering | Getting therapies into correct tissues |
| Clinical development | Testing AI-designed candidates in humans |
Different Boston companies control different combinations of these layers.
The ecosystem advantage comes from the fact that all of them can exist within the same regional network.
The Next Bottleneck Will Be Biological Measurement
Generative models are likely to keep getting cheaper and more powerful.
That means candidate generation could become abundant.
If ten companies can each generate one million plausible protein designs, the ability to generate the millionth sequence provides less differentiation.
The scarce resource moves elsewhere.
The bottleneck becomes reliable biological measurement.
Which protein really binds?
Which one stays stable?
Which one reaches the right tissue?
Which one avoids unwanted immune activity?
Which one works in animals?
Which one behaves correctly in humans?
This may explain why companies such as Manifold are emphasizing experimental infrastructure so heavily.
As generation becomes easier, truth becomes expensive.
Experiments provide that truth.
The Next Competitive Battle May Be Over Data Quality, Not Data Quantity
Biotech AI companies often talk about enormous datasets.
Size matters, but data quality matters more.
A million low-quality observations may provide less value than 10,000 highly controlled experiments measuring exactly the property a model needs to learn.
Generative-biology companies therefore need to think like information factories.
Every experiment should not merely answer whether one molecule worked.
Ideally, it should generate structured information capable of improving thousands of future designs.
That changes laboratory design.
Experiments become part of the machine-learning architecture.
This is one of the most important ways generative biology differs from conventional biotech.
The wet lab is no longer simply where computational predictions are checked.
It becomes part of the training system.
Could Generative Biology Really Make Drug Development Faster?
Potentially.
But people need to be precise about where the speed improvement occurs.
Generative biology may dramatically reduce the time needed to identify and optimize promising molecules.
It may decrease the number of design cycles required.
It may help scientists optimize several properties simultaneously.
It may make difficult targets accessible.
None of those improvements automatically shortens every clinical trial.
Patients still need to be enrolled.
Safety still needs to be monitored.
Regulators still require evidence.
Manufacturing still needs validation.
Long-term outcomes may still require time.
The biggest near-term gains are therefore likely to appear in discovery and preclinical optimization, followed gradually by improvements downstream.
Could AI Design Better Medicines, Not Merely Faster Medicines?
This is the more interesting question.
Speed is useful.
Quality is transformational.
Suppose generative systems eventually allow researchers to optimize potency, selectivity, stability, dosing frequency, manufacturing characteristics, tissue targeting, and immune response together.
Drug discovery would stop being only a search for something that works.
It would become a multi-variable engineering problem.
Generate’s GB-0895 provides an early example of the idea. The antibody was engineered with the goal of strong TSLP binding and extended half-life, with twice-yearly dosing being evaluated in severe asthma.
Whether the program ultimately succeeds clinically remains a question for the trials.
But the design objective shows where the field is heading.
AI is increasingly being used not simply to discover a biological effect.
It is being used to shape the complete product profile.
Why Boston Could Build a Long-Term Advantage
AI models themselves spread quickly.
Research papers are published.
Engineers move between companies.
Open-source models improve.
Cloud compute is available globally.
Boston’s harder-to-copy advantage is the surrounding biological system.
The region has decades of institutional knowledge around drug development.
It has specialized laboratories.
It has people who have taken drugs from discovery into clinical trials.
It has venture investors who understand the unusual economics of biotechnology.
It has major pharmaceutical companies.
It has hospitals capable of translating scientific ideas into human research.
And it has universities producing both biological research and machine-learning talent.
Those ingredients reinforce one another.
Generative biology turns that existing ecosystem into something even more useful because the new discovery process requires computation and experiments to interact continuously.
Boston already has both sides of that equation.
What Could Go Wrong
Generative biology deserves serious attention, but not blind optimism.
Several major risks remain.
Models Can Learn the Wrong Things
Biological datasets contain biases.
Experiments can be noisy.
A model may become very good at optimizing laboratory measurements without producing a medicine that works in humans.
That is why increasingly realistic testing matters.
A Beautiful Protein Can Still Fail as a Drug
Binding the intended target is only one requirement.
A molecule may be unstable.
It may create immune reactions.
It may be difficult to manufacture.
It may disappear from the bloodstream too quickly.
It may accumulate in the wrong tissue.
It may show toxicity.
Generative systems therefore need to optimize several characteristics together.
Clinical Biology Remains Extremely Difficult
Humans are much more complicated than experimental systems.
The ultimate success rate of generative medicines will only become clear after substantially more candidates pass through clinical development.
Generate’s late-stage programs will therefore be watched closely across the industry.
Capital Requirements Remain High
Even excellent computational platforms eventually meet expensive biological development.
Generate reported a net loss attributable to common stockholders of $203.2 million for 2025 before completing its 2026 IPO.
That does not mean the model is failing.
Clinical-stage biotechnology regularly requires large investments long before commercial revenue exists.
It simply demonstrates that generative biology does not remove biotech’s capital intensity.
The KPI Dashboard Boston Generative-Biology Companies Should Be Measured Against
A better industry dashboard would move beyond funding and model announcements.
BostonBizScene would track eight operating metrics.
| KPI | Why it matters |
|---|---|
| Designs generated | Measures computational throughput |
| Designs physically built | Shows connection between software and reality |
| Experimental success rate | Measures design quality |
| Design-to-test cycle time | Shows speed of the learning loop |
| Proprietary experiments produced | Measures data moat |
| In-vivo validation rate | Tests biological relevance |
| Development candidates nominated | Measures translation into real drug programs |
| Clinical-stage programs | Measures ultimate platform productivity |
No company currently publishes all of these metrics in a standardized form.
That should change.
As generative biology matures, investors and pharmaceutical partners will need better ways to distinguish genuine platform productivity from impressive demonstrations.
What to Watch During the Rest of 2026 and Into 2027
The next phase of Boston’s generative-biology market will be determined by execution.
For Generate, Phase 3 progress for GB-0895 will remain particularly important because it moves the generative-biology thesis into late-stage clinical testing.
For Manifold, the key question is whether massively scaled in-vivo measurement can repeatedly produce differentiated therapeutic molecules and useful pharma programs.
For Nabla, watch whether partnerships continue expanding from technical validation into nominated development candidates.
For Dyno, the important question is whether AI-designed capsids can demonstrate increasingly strong translational performance across partnered gene-therapy programs.
For Neoclease, watch how rapidly custom AI-designed editors move from platform demonstrations into convincing preclinical therapeutic evidence.
Across all five companies, one metric matters most:
Does each design cycle produce better biology?
If the answer keeps improving, the Boston generative-biology ecosystem could become far more important than the current AI hype cycle.
Final Takeaway: Boston Is Trying to Turn Biology Into an Engineering Discipline
Generative biology is often described as AI being used to discover drugs faster.
That description is too small.
The deeper change is that scientists are beginning to move from searching biological space toward deliberately designing within it.
Boston companies are designing proteins, antibodies, delivery vectors, and gene-editing enzymes. They are building systems that connect computational generation to real experiments. Some are now connecting those experiments to pharmaceutical partnerships and human clinical development.
Our analysis of five Boston-area generative-biology companies found that all five depend on some form of integrated computational and experimental learning. All five are centered on protein engineering at the physical design layer, even though they apply that capability to very different therapeutic problems. And the companies are developing several business models at the same time, from fully owned clinical pipelines to platform partnerships and hybrid structures.
Those findings suggest that Boston’s advantage is not simply access to powerful AI.
The region’s real advantage is the ability to close the loop between code and biology.
A model creates a molecule.
A laboratory builds it.
An experiment measures it.
The result becomes new data.
The system learns.
Then it designs again.
When that cycle operates fast enough and produces high-quality biological information, drug discovery begins to look less like a long search through nature and more like an engineering process.
That transition will take years.
Many molecules will still fail.
Some companies will disappear.
Clinical trials will continue to deliver surprises.
But the direction is becoming clearer.
The next major biotechnology platform may not be defined by how many molecules it can screen.
It may be defined by how quickly it can imagine, build, test, and learn from molecules that never existed before.
Boston is becoming one of the most important places where that idea is being tested in the real world.
