For decades, finding a new therapeutic antibody has looked less like engineering and more like an enormous search problem.
Scientists start with a target linked to a disease. They create, collect, or screen huge numbers of possible antibodies. They test which ones bind. Then they spend months or years trying to improve the winners.
The problem is that binding is only the beginning.
An antibody can bind strongly and still make a poor medicine. It might stick to the wrong proteins. It might aggregate. It might break down too quickly. It might be difficult to manufacture. It might fail to reach the tissue where it is needed. It might produce an unwanted immune response. Or it may simply fail to change the disease in the way scientists expected.
That is why a growing group of Boston-area biotech companies is trying something different.
Instead of searching through what biology happens to give them, they are trying to design antibodies around a desired job.
Generative AI is becoming part of that process.
Models can propose new protein sequences, predict structures, target specific binding sites, optimize multiple properties and increasingly help scientists decide which molecules should be built and tested next. Companies including Generate, Nabla Bio and Manifold Bio have built large experimental systems around this idea, while groups such as Seismic Therapeutic, Metaphore Biotechnologies, Moderna and others are bringing machine learning deeper into biologics engineering.
The important story is not simply that Boston has companies using AI.
Boston is beginning to build something more valuable: closed learning systems where AI designs molecules, experiments test them, the resulting biological data returns to the model, and the next generation of molecules gets better.
That could change how antibody companies compete.
And Boston may be unusually well positioned for that change.
The Short Version: Antibody Discovery Is Starting to Become Antibody Design
Traditional antibody discovery starts with a large pool of possible molecules and searches for useful ones.
Generative antibody design tries to reverse the process.
Instead of asking:
“Which molecule in this enormous library happens to work?”
the new question becomes:
“What molecule should we create to produce the biological behavior we want?”
That sounds like a small change in wording. It is actually a major change in the economics and structure of drug discovery.
Companies can increasingly specify the part of a target they want an antibody to bind. They can generate molecules computationally. They can filter them for binding, selectivity, stability and other properties before expensive experiments begin. The best systems then create real molecules, measure what happens and feed those results back into the next model.
Nabla Bio describes its JAM system as a multimodal generative system capable of designing biologics from molecular context such as a target or epitope. Its experimental systems then measure binding, developability, cellular function and in-vivo performance.
Generate uses machine learning trained on protein sequence, structure and proprietary experimental data to generate new protein medicines, including antibodies. The company says it has generated, built and tested more than 42,000 proteins and now operates more than 140,000 square feet across its Boynton Yards and Andover locations.
Manifold Bio is pushing a different part of the loop. Its technology is built around testing designed biologics directly in living systems at enormous scale. Its latest platform says it can measure the tissue distribution of more than 500,000 binders in a single in-vivo experiment.
These are different technical strategies.
But economically, they point toward the same idea.
The model alone may not be the moat. The learning loop may be the moat.
Why Antibodies Are Such an Important Test for Generative AI
Antibodies are attractive drugs because nature already created them to recognize very specific molecular features.
That does not mean making a great antibody medicine is easy.
An antibody is a large protein with a complex three-dimensional shape. Small changes to its amino-acid sequence can change how it binds, folds, behaves in the bloodstream, interacts with the immune system or performs during manufacturing.
That creates a huge design space.
Binding Is Only One Requirement
Imagine a company wants an antibody against a protein involved in cancer.
Finding something that attaches to that protein is useful, but it does not answer the bigger questions.
Where does the antibody bind?
Does that binding site change the target’s function?
Does it bind similar proteins that should be left alone?
Can the antibody survive long enough in the body?
Will it aggregate during production?
Can manufacturers make it at high concentration?
Does it trigger the right immune behavior?
Will it actually reach a tumor?
A molecule can look impressive on one measure and fail because of another.
That is one reason AI could become valuable.
Computational systems can potentially search many requirements at once instead of optimizing one property and discovering much later that another property became worse.
Nabla, for example, describes measuring properties including affinity, cross-reactivity, stability, polyspecificity, aggregation and hydrophobicity as part of its antibody developability work.
The Real Goal Is Not More Candidates
Generative AI can produce an enormous number of possible protein sequences.
That is not automatically useful.
Biotech does not suffer from a shortage of theoretical molecules. It suffers from a shortage of molecules that work inside complicated biological systems and can become safe, manufacturable drugs.
This distinction matters enormously.
The winners in AI antibody design may not be the companies capable of generating the most sequences.
They may be the companies that can learn fastest which sequences are worth making.
What Generative Antibody Design Actually Does
The phrase “generative AI” can make this field sound more mysterious than it is.
At the simplest level, a generative model learns patterns from existing proteins and experimental data. Scientists then ask it to propose new proteins that meet certain conditions.
A model might be given the structure of a disease target.
Scientists can then ask for antibody candidates that bind a particular part of that target.
The computer produces possible structures and sequences.
Other models or computational filters may rank those designs.
A smaller set is physically made.
Those molecules are tested.
The results become new training data.
Then the process begins again.
It Is Moving Beyond Predicting Protein Structure
Protein AI first became widely known for prediction.
Give a system a protein sequence, and it tries to predict what shape the protein will take.
Design asks the reverse question.
Scientists have a desired structure, interaction or function. The system needs to propose a sequence likely to produce it.
That is harder.
Antibody design goes further because the molecule must interact correctly with another molecule.
A useful generative system therefore needs to understand not just an antibody, but the interface between an antibody and its target.
Epitope Control Could Be a Major Advantage
The exact place where an antibody binds is called its epitope.
Two antibodies can bind the same disease target but behave completely differently because they touch different parts of that target.
This is where computational design becomes particularly interesting.
Nabla says its approach can specify the binding epitope before generating an antibody. That allows the company to explore different binding sites systematically instead of accepting whichever epitopes conventional discovery happens to produce.
Absci, an important non-Boston benchmark in this market, similarly describes using generative models to design antibodies toward specified epitopes and then optimize attributes including potency, selectivity, half-life and manufacturability. Its Origin-1 research reported experimentally validated antibodies against four of ten difficult “zero-prior” targets using fewer than 100 design attempts per target.
That matters because future competition may move from:
Who can find an antibody against this protein?
to:
Who can program exactly how the antibody engages this protein?
BostonBizScene Original Research: Mapping Boston’s AI Antibody Design Cluster
To understand how far Boston has actually moved toward this future, BostonBizScene built a company-level dataset using publicly available information.
The goal was not to count every biotech that occasionally uses machine learning.
We used a stricter screen.
Our Methodology
The analysis was completed using publicly accessible company materials, technical publications, job postings, financing announcements and disclosed partnerships available through October 6, 2026.
A company was included in our broad Boston AI-antibody screen when it met three conditions.
First, it had significant headquarters, laboratory or technical operations in Greater Boston.
Second, it publicly described the use of machine learning, AI, generative design or computational learning in antibody or biologic discovery.
Third, the technology was connected directly to molecule design, optimization, experimental selection or biologic development rather than general business analytics.
We then coded publicly verifiable signals across five categories:
| Signal | What We Looked For |
|---|---|
| Generative or de novo design | Explicit evidence that models create new protein or antibody designs |
| Experimental learning loop | AI output is built and physically tested |
| Antibody-specific work | Public antibody programs, platforms or development activity |
| Human-relevant or in-vivo testing | Cellular, mammalian, in-vivo or clinical validation |
| External validation | Disclosed pharma collaboration, substantial financing or clinical advancement |
Missing public information was treated as unknown, not as proof that a capability did not exist.
This matters because many private biotech companies reveal only a fraction of their technical systems.
The Boston Cluster We Identified
Our broad screen found at least nine notable organizations fitting most or all of those criteria.
| Organization | Boston Presence | Public AI/ML Antibody or Biologics Signal | Main Angle |
|---|---|---|---|
| Generate | Somerville | Yes | Generative protein therapeutics |
| Nabla Bio | Cambridge | Yes | De novo antibody design |
| Manifold Bio | Boston | Yes | AI design plus massive in-vivo testing |
| Seismic Therapeutic | Cambridge | Yes | ML-designed immunology biologics |
| Metaphore Biotechnologies | Cambridge | Yes | Functional-data-driven antibody engineering |
| Moderna | Cambridge | Yes | AI-enabled de novo antibodies and binders |
| Converge Bio | Boston presence | Yes | Generative antibody optimization |
| Nona Biosciences | Cambridge presence | Yes | Antibody language models |
| AstralBio | Boston | Yes | ML-enabled antibody therapeutics |
Sources supporting this screen include company platforms, technical pages, recent hiring descriptions and public announcements.
This is not meant to be a complete list of every Boston company touching antibodies.
It is a view of the most visible part of the emerging AI-design cluster.
Original Finding #1: Boston Is Building Experimental AI Companies, Not Just Software Companies
Our most important finding came from looking at what these businesses actually built around their algorithms.
The Boston cluster is heavily oriented toward physical experimentation.
That may prove more important than which company currently has the best model.
Why Experimental Data Matters So Much
Public protein databases contain enormous amounts of sequence information.
But drug development requires information that is much harder to find.
How does a candidate behave in a human cell?
Does it aggregate?
Does it reach a particular tissue?
Does it activate the right pathway?
How selective is it?
What happens inside an animal?
Those measurements are expensive.
That makes proprietary experimental data far more defensible than simply training another model on widely available protein sequences.
Boston Is Closing the Loop
Nabla describes its AI and human-relevant experimental systems as one integrated engine rather than separate departments. Its lab generates measurements of binding, developability, cellular function and in-vivo performance that can strengthen its JAM system.
Generate describes a similar feedback loop in which machine learning drives protein design while experiments continuously provide new data.
Manifold takes the idea further toward whole-organism measurement. Its platform combines computational design with highly multiplexed testing of biologics inside living systems.
Seismic integrates structural biology, machine learning, protein engineering and biological testing to design immunology medicines.
BostonBizScene Interpretation
This suggests Boston’s AI-biotech advantage may be fundamentally different from the advantage associated with consumer AI.
In consumer software, having a better foundation model can create enormous leverage.
In biotech, a model eventually has to confront biology.
A model can predict that a protein should work.
The cell, animal or patient gets the final vote.
That makes experimental infrastructure part of the AI stack.
Original Finding #2: The Real Bottleneck Is Moving From Generating Molecules to Measuring Them
The cost of creating computational designs is falling quickly.
The cost of obtaining high-quality biological truth is not falling nearly as fast.
That creates a new bottleneck.
The Design Funnel Is Becoming Extremely Wide
Manifold provides one of the clearest demonstrations.
In March 2026, the company reported a joint study with NVIDIA in which it experimentally tested one million binder designs against 127 targets and measured more than 100 million protein-protein interactions. Specific binders were found for 68% of the tested targets.
Its own mBER antibody work previously reported testing more than one million designs across 145 targets and measuring more than 100 million interactions.
Those numbers point to a profound change.
If a model can suggest one million molecules, a company does not need another way to make one million suggestions.
It needs an efficient way to learn which suggestions matter.
Chart: The Emerging AI-Biologics Funnel
| Stage | Traditional Constraint | AI-Era Constraint |
|---|---|---|
| Generate candidates | Expensive | Becoming cheap |
| Rank candidates | Moderate | Model quality matters |
| Build proteins | Expensive | Automation reducing cost |
| Measure properties | Very expensive | Increasingly the bottleneck |
| Test function | Very expensive | Critical differentiator |
| Test in vivo | Extremely expensive | Potential strategic moat |
| Reach patients | Slow and risky | Ultimate proof |
This changes where value may accumulate.
The companies with the best experimental engines can create datasets competitors cannot simply download.
That makes the wet lab a data factory.
Manifold Bio: Turning Living Systems Into a Training Dataset
Manifold Bio is especially interesting because it attacks a problem that AI protein design cannot solve through computation alone.
A molecule can look excellent in a test tube and still behave badly inside an organism.
Its distribution may be wrong.
Its lifetime may be wrong.
It may fail to cross into the correct tissue.
From In-Vitro Design to In-Vivo Design
Manifold’s argument is simple: biologics should be measured in the environment where they eventually need to work.
The company uses molecular barcoding to track large numbers of protein designs simultaneously inside living animals.
Its September 2026 mBER-2 announcement said its platform can now measure tissue distribution for more than 500,000 binders in one in-vivo experiment.
That is a remarkable shift in experimental scale.
Traditional drug development often treats animal testing as a late filtering step.
Manifold is trying to make living biology part of the design loop itself.
Why This Could Matter for Antibodies
An antibody intended for the brain, tumor or other difficult tissue has a delivery problem.
Binding the target is useless if the molecule never reaches it in sufficient concentration.
An in-vivo-first system can potentially optimize both molecular recognition and physical distribution.
That expands the design target.
The model is no longer merely asking:
“Does this antibody bind?”
It can increasingly ask:
“Does this biologic reach the correct tissue, persist there and behave as needed inside a living organism?”
That is a much more valuable problem.
Nabla Bio: Trying to Turn Antibody Discovery Into Engineering
Among Boston startups, Nabla Bio may be the clearest pure example of AI-first antibody design.
The company was built around the idea that antibodies can be designed directly rather than discovered through large blind searches.
Its JAM system uses sequence, structure and experimental information to generate new biologics.
Designing Against Hard Targets
Traditional antibody discovery performs well against many accessible proteins.
It becomes harder when targets sit inside cell membranes or expose very small binding areas.
GPCRs, ion channels and transporters are important examples.
Nabla has specifically focused on difficult membrane proteins and has reported de novo antibody work against targets including CLDN4 and CXCR7.
This matters commercially.
If generative design only makes antibody discovery slightly faster against easy targets, its value is helpful but limited.
If it makes previously difficult target classes practical, the economic impact becomes much bigger.
Developability Is Built Into the Problem
Nabla also highlights something the industry increasingly understands: a computational binder is not automatically a drug.
Its platform combines generative design with measurements of drug-like behavior, including manufacturing-related properties and functional testing.
That moves AI away from novelty generation and toward therapeutic engineering.
Pharma Has Already Paid Attention
Nabla announced a $26 million Series A in 2024 and collaborations with AstraZeneca, Bristol Myers Squibb and Takeda carrying more than $550 million in potential upfront and milestone economics plus royalties. Current company information says it has raised about $37 million.
Takeda later signed a second collaboration with the company.
The financial numbers should not be interpreted as guaranteed future payments.
Milestone-heavy biotech partnerships can be far larger on paper than the money eventually received.
But they still show something important.
Large pharmaceutical companies are willing to place real programs on these platforms.
Generate:Biomedicines: The Biggest Boston Proof Point
If investors want evidence that generative protein design can move beyond research demonstrations, Generate is becoming one of Boston’s most important examples.
The company went public in February 2026 at $16 per share, selling 25 million shares for gross IPO proceeds of $400 million before underwriting discounts and expenses.
By June 30, 2026, Generate reported approximately $457.4 million in cash, cash equivalents and marketable securities.
But the more important story is the pipeline.
An AI-Engineered Antibody Is in Phase 3
Generate’s GB-0895 is a long-acting antibody targeting TSLP for severe asthma.
The company’s SOLAIRIA-1 and SOLAIRIA-2 Phase 3 studies were recruiting across six global regions as of its August 2026 update.
That matters because the AI-biotech industry has spent years talking about what models might eventually do.
Clinical development creates a much harsher test.
Can the molecule be manufactured?
Can regulators review it?
Can physicians administer it?
Can it produce meaningful results in patients?
Moving an AI-engineered antibody into Phase 3 does not prove that generative AI will transform drug discovery.
Clinical trials can still fail.
But it moves the argument from software capability toward drug-development evidence.
Generate Has Also Become a Partnership Platform
Amgen signed an initial collaboration with Generate involving five programs and a $50 million upfront payment, with the original agreement carrying up to $1.9 billion in potential transaction value plus royalties. Amgen subsequently opted into a sixth program.
Novartis later signed a multi-target collaboration providing $65 million in total upfront consideration, including $15 million in equity, with more than $1 billion in possible milestone payments plus royalties.
Again, potential milestones are not the same as cash already received.
Still, the pattern is strategically important.
Pharma companies are not merely buying access to software.
They are outsourcing part of molecule creation.
Original Finding #3: Boston’s AI-Antibody Cluster Has Already Attracted Roughly $1.4 Billion in Selected Disclosed Financing
To measure commercial commitment, BostonBizScene analyzed clearly disclosed financing from four major companies in the local AI-biologics ecosystem.
We deliberately separated equity financing from partnership milestone values.
That prevents a common mistake in biotech reporting: treating theoretical future milestone payments as if they were money already invested.
Selected Financing Dataset
| Company | Financing Included in Analysis | Amount |
|---|---|---|
| Generate | Private financing reported through 2024 | ~ $700M |
| Generate | 2026 IPO gross proceeds | $400M |
| Seismic Therapeutic | Total capital raised reported after Series B | $222M |
| Nabla Bio | Total financing reported | ~ $37M |
| Manifold Bio | 2022 Series A | $40M |
| Total selected disclosed financing | ~$1.399B |
Generate’s roughly $700 million private financing figure was reported after its $273 million Series C, before the company later raised $400 million gross in its IPO. Seismic said its $121 million Series B took total capital raised to $222 million. Nabla reports roughly $37 million raised, while Manifold announced a $40 million Series A.
This $1.399 billion figure should be read as a conservative selected financing total, not a complete estimate of every dollar invested in Boston AI antibody companies.
Some companies are omitted because comparable financing information is not public.
Some earlier Manifold financing is not included.
Large partnership economics are also excluded.
Chart: Selected Disclosed Financing
| Company / Financing Base | Approx. Capital Included |
|---|---|
| Generate | $1.10B ████████████████████████████ |
| Seismic | $222M ██████ |
| Manifold Series A | $40M █ |
| Nabla | $37M █ |
| Combined | $1.399B |
The concentration is obvious.
Generate dominates the capital picture.
But smaller companies are also attracting meaningful institutional investment before reaching late clinical development.
That suggests investors are funding multiple approaches rather than one single platform architecture.
Why Seismic Therapeutic Matters
Seismic Therapeutic is not simply trying to generate conventional antibodies faster.
The company is using machine learning to engineer more complex immunology biologics.
Its IMPACT platform combines machine learning, structural biology, protein engineering and immunology.
Machine Learning Can Modify How an Antibody Behaves
Seismic describes using ML to design selective Fc domains, predict antibody epitopes, improve developability and identify immunogenic regions that should be removed.
This shows why the phrase “AI antibody design” is broader than generating a new binding sequence.
The technology can potentially affect:
which immune receptor is engaged,
how strongly it is engaged,
which cell types interact,
how immunogenic the protein is,
and whether the overall molecule behaves like a practical drug.
That makes AI a protein-engineering system rather than merely a search engine.
Capital Has Followed the Model
Seismic raised a $121 million Series B in 2023, bringing reported total capital raised to $222 million. The company said the financing would support its immunology pipeline and continued expansion of its ML-enabled platform.
Its current positioning describes Seismic as a clinical-stage immunology company developing biologic medicines for autoimmune and inflammatory diseases.
That puts another Boston AI-enabled biologics company into the translational test that ultimately matters.
Metaphore Is Expanding What an Antibody Is Supposed to Do
Most people think about antibodies as blockers.
An antibody recognizes something harmful or disease-related and prevents it from working.
Metaphore Biotechnologies is pursuing a broader idea.
The Cambridge company says its platform combines a proprietary experimental functional-data engine with a mechanism-aware computational system to create antibodies able to activate, tune or engage biological pathways in new ways.
This is important for the next stage of AI biologics.
Function May Become More Important Than Binding
A weak version of AI antibody discovery asks:
Can we create a binder?
A stronger version asks:
Can we create a molecule that changes biology in exactly the desired way?
Those are very different goals.
The second requires functional data.
That reinforces the broader Boston pattern.
As models get better at generating plausible proteins, the competitive frontier shifts toward learning what those proteins actually do.
Moderna Is Quietly Bringing Generative Antibody Design Into a Much Larger Platform
Moderna is primarily associated with messenger RNA.
That could make its antibody-design activity easy to overlook.
But a recent Cambridge research hiring description provides an unusually clear view of where the company is moving.
Moderna is hiring for AI-enabled design of de novo antibodies, binders and antigens, with work spanning generative design, structure prediction, sequence design, docking, scoring and experimental testing.
More interestingly, the company explicitly links those capabilities to mRNA therapeutics.
AI-Designed Proteins Could Meet Programmable Delivery
Moderna says such binders could function as standalone biologics, targeting modules or components of more complex protein systems.
This points toward an important convergence.
The future may not be divided neatly into:
antibody companies,
mRNA companies,
AI companies,
and protein-design companies.
The same computational design systems could produce targeting elements for many therapeutic formats.
For Boston, that makes the ecosystem more interesting than a collection of isolated antibody startups.
The technologies may reinforce one another.
Original Finding #4: The Boston Moat May Be Experimental Density, Not Model Size
There is a popular assumption that AI competition eventually becomes a contest over the largest model.
That may be wrong in biologics.
Protein design does benefit from stronger algorithms and more compute.
But biology creates a second source of scale that cannot be replaced by GPUs.
Experimental throughput.
Three Boston Examples Show the Pattern
Generate says it has already generated, built and tested more than 42,000 proteins.
Nabla integrates high-throughput human-relevant experimental measurements directly into its generative system.
Manifold has demonstrated experiments involving one million computationally designed binders and more than 100 million measured interactions.
These numbers cannot be compared directly because the platforms measure different things.
But the direction is clear.
AI-biotech companies are increasingly competing on the number and quality of useful biological observations they can generate.
Why Boston Has an Structural Advantage
Boston already has unusual density across:
protein science,
immunology,
biologics engineering,
academic laboratories,
hospitals,
pharma R&D,
venture capital,
automation,
clinical development,
and experienced biotech operators.
A generative model can theoretically be trained anywhere.
Building an organization that knows which experiment should be run next, can execute it correctly, can interpret the result and can translate the winning molecule into a clinical program is far harder.
That institutional knowledge is difficult to copy.
Chart: How the Competitive Moat Is Shifting
| Earlier Biotech Advantage | Emerging AI-Biotech Advantage |
|---|---|
| Large screening library | Large proprietary learning dataset |
| Strong assay | Automated high-throughput assay system |
| Protein engineering expertise | Protein engineering + model engineering |
| Single-property optimization | Multi-property optimization |
| Discovery followed by testing | Continuous design-build-test-learn loop |
| Animal testing late | In-vivo learning earlier |
| Data as experimental record | Data as training asset |
| Best individual molecule | Fastest improving discovery system |
This is one of the biggest changes executives and investors should understand.
The value may move away from a single successful screening campaign.
It may move toward an engine that gets systematically better every time it runs.
Original Finding #5: Pharma Validation Is Concentrating Around Platforms That Combine Models With Experiments
Partnership activity provides another useful signal.
Large pharmaceutical companies have internal AI teams.
They can license software.
They can train models.
So when they commit drug programs to an outside biotech, they are often buying something more difficult to recreate.
Generate
Amgen’s original Generate collaboration involved $50 million upfront and up to $1.9 billion in potential transaction value for the original programs.
Novartis later committed $65 million in total upfront consideration to another Generate collaboration.
Nabla
Nabla disclosed collaborations with AstraZeneca, Bristol Myers Squibb and Takeda potentially worth more than $550 million in upfront and milestone payments plus royalties.
Takeda subsequently expanded the relationship with a second collaboration.
Manifold Bio
Manifold’s public news archive reports a 2025 Roche collaboration involving $55 million upfront and potential payments of up to $2 billion to develop next-generation brain shuttles.
This does not prove these platforms will produce successful medicines.
Biotech partnerships frequently end without an approved drug.
But the structure of these deals tells us what pharmaceutical companies value.
They are increasingly willing to pay for integrated capability spanning design, proprietary data and biological validation.
Converge Bio Shows That Generative Antibody Tools Are Becoming Easier to Deploy
Not every company in this market is trying to become a full-stack drug developer.
Some are building generative systems that other biotech companies can use.
Boston-connected Converge Bio provides a good example.
In 2026, the company announced work with Purple Biotech using generative AI to optimize a tri-specific antibody platform.
It separately reported an internal experiment using its ConvergeAB system to generate a modified version of cetuximab with higher measured binding affinity, according to the company’s own announcement.
These company-reported experiments still require independent validation and much more development before they can say anything about clinical performance.
But they show another important direction.
AI antibody design could become a service layer used by companies that do not want to build large internal model teams.
Nona Biosciences Shows How Specialized Antibody Foundation Models Are Emerging
General protein models are powerful.
But antibody companies increasingly want models trained on data specific to the kinds of molecules they care about.
Nona Biosciences announced HCAbLM in September 2026, which it describes as a language model trained specifically on fully human heavy-chain-only antibodies. The company says the training set included 31.8 million antibody sequences from 73 independently immunized transgenic mice.
This illustrates an important trend.
The industry may not converge around one universal protein model.
It may instead develop many specialized systems trained on different molecule types, biological contexts and experimental objectives.
That could create opportunities for companies with unusual proprietary datasets.
Better AI Will Not Remove the Need for Better Biology
It is easy to imagine a future where computers simply produce finished drugs.
That is unlikely anytime soon.
Biology remains too complicated.
Prediction Error Compounds Quickly
Suppose a model correctly predicts that an antibody should bind a target.
It might still be wrong about selectivity.
Even if selectivity is good, the antibody may fail functionally.
Even if it changes the desired pathway, it may have poor pharmacokinetics.
Even if animal performance looks excellent, human biology may differ.
Every stage introduces another source of uncertainty.
AI can improve the odds.
It cannot eliminate biology.
Experimental Validation Is Becoming More Important, Not Less
A 2026 review of AI-driven protein design emphasized the full pipeline from data curation and candidate generation through experimental validation.
That is the correct way to think about this market.
The important unit is not the model.
It is the whole system.
Data → model → molecule → experiment → biological result → new data.
The shorter and richer that loop becomes, the faster a company can learn.
What Founders Should Learn From Boston’s AI-Antibody Companies
The Boston market provides several practical lessons for biotech founders.
Do Not Build a Model Without a Data Strategy
A model trained only on public information can be useful.
It is also vulnerable.
Competitors can access much of the same data.
A stronger company asks what proprietary biological measurements will become more valuable as the model improves.
That could be binding data.
Developability data.
Cellular function.
Human tissue data.
In-vivo distribution.
Manufacturing performance.
The specific answer depends on the product.
But there needs to be an answer.
Optimize the Drug, Not the Benchmark
AI teams naturally want clean metrics.
Protein design is full of them.
Prediction accuracy.
Structure confidence.
Binding score.
Hit rate.
These numbers matter.
But the ultimate objective is not to win a computational benchmark.
The objective is to create a molecule that can become a useful medicine.
That means teams should connect model objectives to therapeutic objectives as early as possible.
Build Experimental Throughput Where It Changes Decisions
More experiments are not automatically better.
A company can generate enormous amounts of low-value data.
The important question is whether an experiment helps choose between competing designs.
A good experimental system should reduce uncertainty.
If the experiment changes what the model designs next, it has much more strategic value than a measurement that simply confirms what scientists already know.
What Pharma Teams Should Ask Before Partnering With an AI-Antibody Company
Drug companies now face a difficult vendor and partnership landscape.
Almost everyone claims to use AI.
That makes technical diligence critical.
The first question should be:
What Is Actually Generated?
Does the system create complete antibodies?
Only CDR loops?
New binding domains?
Protein backbones?
Optimized variants of existing antibodies?
The phrase “AI design” can describe very different technical capabilities.
What Is Experimentally Proven?
A beautiful computer structure is not enough.
Teams should ask how many generated molecules were physically created.
What fraction bound?
What were their affinities?
Were the results reproduced?
Was function measured?
Was developability tested?
Was structural accuracy experimentally confirmed?
What Happens After the First Hit?
A platform should not end when it finds a binder.
Drug development has only started.
The company needs a path toward potency, selectivity, stability, manufacturability, pharmacokinetics and safety.
Does Every Campaign Improve the Platform?
This is one of the most strategically important questions.
If the company performs 20 discovery programs, does program number 20 benefit from what happened during programs one through 19?
If yes, the business may have a compounding data advantage.
If no, it may simply be a sophisticated contract research operation.
What Investors Should Measure Differently
AI-biotech investors also need better metrics.
The normal software metrics are not enough.
Model Performance Alone Can Be Misleading
A company can publish a strong model while lacking the ability to turn predictions into drugs.
Investors should examine the conversion funnel.
How many computational designs are generated?
How many are selected?
How many are built?
How many express correctly?
How many bind?
How many have the desired function?
How many survive developability screens?
How many progress into animals?
How many become development candidates?
That funnel tells a more important story than a model benchmark.
Data Efficiency May Matter More Than Model Size
The strongest platform may not need millions of new experiments for every target.
It may learn from previous campaigns and require fewer physical tests over time.
That creates an important potential KPI:
experiments required per viable drug candidate.
If that number falls steadily, the learning system may genuinely be improving.
A Better KPI Dashboard for AI Antibody Companies
BostonBizScene believes investors and operators should watch a set of metrics that connect computational performance to drug-development performance.
| KPI | Why It Matters |
|---|---|
| Designed molecules per campaign | Measures computational exploration |
| Molecules physically tested | Shows real experimental capacity |
| Experimental hit rate | Connects model output to reality |
| Desired-epitope hit rate | Measures control, not just binding |
| Functional hit rate | Shows biological relevance |
| Developable hit rate | Tests drug quality |
| Experiments per development candidate | Measures learning efficiency |
| Time from target to validated lead | Measures practical speed |
| Percentage of data that is proprietary | Indicates defensibility |
| Share of new programs benefiting from prior data | Measures compounding |
| In-vivo validation throughput | Shows ability to learn beyond test tubes |
| Clinical candidates generated by platform | Provides ultimate translational evidence |
No single number proves a platform works.
The direction across several numbers is what matters.
The Biggest Risk: AI Could Make Mediocre Molecules Faster
Every major technological improvement can accelerate bad decisions as well as good ones.
Generative models are no exception.
If a company optimizes the wrong objective, AI can produce thousands of highly optimized answers to the wrong question.
Binding Is the Classic Example
Imagine that the model is rewarded primarily for affinity.
It learns to create molecules that bind very tightly.
That sounds good.
But extremely strong binding might not produce the correct function.
The molecule might also become less selective or harder to manufacture.
A platform needs several objectives at once.
That means companies need excellent assays and excellent judgment about which properties matter.
Bad Training Data Creates False Confidence
Biological data is noisy.
Different laboratories can measure the same property differently.
Assays can contain hidden biases.
A model trained on those measurements can learn the bias instead of the biology.
This is another reason the best AI-biotech companies increasingly control both computation and experimentation.
They can control how training data is produced.
Why Manufacturing Needs to Enter the Design Loop Earlier
A therapeutic antibody ultimately has to be manufactured repeatedly, consistently and at scale.
This is not an operational detail.
It is part of molecule quality.
Poorly behaved proteins can create problems during expression, purification, storage and dosing.
AI companies increasingly try to identify these problems before selecting the final lead.
Nabla’s public hiring material, for example, describes measuring properties such as stability, aggregation, hydrophobicity, cross-reactivity and expression as part of candidate assessment.
This is exactly the direction the industry should move.
Instead of discovering a powerful molecule and then asking engineers to fix it, platforms can attempt to design manufacturing quality into the molecule from the beginning.
Why Multispecific Antibodies Make AI More Valuable
A traditional monoclonal antibody usually recognizes one main target.
Multispecific antibodies can interact with multiple biological targets.
That can produce powerful effects.
It also makes engineering much harder.
Every additional binding domain creates more opportunities for structural problems, unwanted interactions and manufacturing difficulty.
This is a natural use case for computational design.
Nabla describes using generative modeling to explore multispecific architectures while attempting to preserve target affinity and overall developability.
Seismic is developing multi-functional antibodies intended to influence multiple immune pathways and cell types.
As antibody drugs become more complex, the number of possible designs grows dramatically.
Human intuition alone becomes less effective at exploring that space.
AI becomes more valuable precisely because the molecule becomes harder to reason about manually.
Original Finding #6: Boston’s Opportunity Is Moving From Designing Binders to Designing Biological Behavior
Looking across the local companies reveals another shift.
The earliest AI protein-design story centered heavily on structure.
Can a computer design a protein that folds correctly?
Then the industry moved toward binding.
Can it design a protein that recognizes a chosen target?
Boston companies are increasingly pushing toward the next question.
Can we design biological behavior?
Consider the progression.
Nabla wants epitope-specific antibodies with functional and developability properties.
Seismic is designing biologics that regulate immune pathways.
Metaphore is trying to create antibodies with new functional mechanisms.
Manifold is optimizing behavior inside living organisms.
Moderna is examining how de novo binders can operate as parts of larger therapeutic systems.
That is a much bigger market than faster antibody discovery.
If successful, AI becomes a way to program what biologic medicines do.
The Next Competitive Frontier: Choosing the Right Experiment
As generative models improve, the limiting question changes.
Imagine that a company can propose ten million plausible antibody designs.
Which 1,000 should it make?
After those results return, which 100 follow-up designs should it test?
Which animal experiment will eliminate the most uncertainty?
Which assay is actually predictive of clinical success?
These are experiment-selection problems.
AI will increasingly need to solve them too.
Active Learning Could Become Central
The most useful model may not merely generate a drug candidate.
It may identify what information it is missing.
Then it can recommend an experiment likely to teach the system the most.
The experiment produces new evidence.
The model updates.
A second experiment is selected.
That creates a much more intelligent loop than brute-force screening.
The long-term objective becomes maximum biological learning per experiment.
What This Could Do to Drug Discovery Economics
The standard drug-development model contains enormous waste.
Many compounds are generated.
Many fail.
Years can pass before important weaknesses become visible.
Generative design cannot remove the basic uncertainty of biology.
But it could move information earlier.
If companies identify poor developability before animal studies, money is saved.
If they understand target epitopes earlier, better mechanisms may be selected.
If in-vivo distribution can be measured earlier and at much higher throughput, delivery problems may be discovered before a company spends years optimizing the wrong molecule.
The economic benefit does not require AI to invent a perfect drug in one attempt.
It only needs to reduce the amount of expensive work performed on bad candidates.
That is a far more realistic benchmark.
Why Boston Could Become One of the Most Important AI-Biologics Hubs
Boston’s strongest position may come from combining two worlds.
The region has deep software, mathematics and machine-learning talent.
But it also has something Silicon Valley cannot instantly reproduce: decades of accumulated drug-development infrastructure and biological expertise.
Universities produce new methods.
Biotech companies generate experimental data.
Pharmaceutical companies contribute development experience.
Hospitals connect research with disease.
Investors finance highly technical platform companies.
Experienced scientists move between them.
That creates a dense feedback network.
AI makes that density more valuable because every part of the system can potentially become a source of training data and experimental learning.
Boston Versus a Pure Software AI Cluster
A software startup can often relocate relatively easily.
A biotech experimentation system cannot.
High-throughput labs require facilities.
Protein scientists need specialized equipment.
Clinical programs require regulatory teams.
Manufacturing knowledge matters.
Animal and translational work requires specific expertise.
These constraints traditionally made biotech slower than software.
But they also create defensive advantages.
If the future of biological AI depends on integrating models with physical experiments, a place where software and biology already coexist has a structural advantage.
Boston fits that description unusually well.
What Could Go Wrong
AI antibody design is promising.
It is not guaranteed to transform the pharmaceutical industry.
Several risks deserve attention.
Models May Generalize Poorly
A system can perform extremely well on proteins similar to its training data and struggle with truly novel biology.
That is why results against difficult or low-information targets are particularly important.
Better Preclinical Performance May Not Produce Better Clinical Performance
Humans are not cell cultures.
They are not mice.
Many excellent preclinical molecules fail in people.
AI will not remove that problem.
The Economics Could Become Commoditized
If powerful open models become widely available, basic protein generation may become cheap.
That would reduce the value of companies whose primary advantage is software.
The strategic response is clear: proprietary data, differentiated experiments, clinical assets and accumulated biological learning become more important.
Capital Requirements Remain High
A generative AI company building its own therapeutics is still a biotech company.
Manufacturing costs money.
Regulatory work takes time.
A faster discovery engine does not turn drug development into SaaS.
What Success Would Actually Look Like
The industry should be careful about declaring victory when an AI system creates a novel protein.
That is a research milestone.
Drug-development success requires something more demanding.
A compelling AI-antibody platform should eventually demonstrate repeated evidence that:
the model generates novel molecules,
the molecules work experimentally,
they show useful biological functions,
they possess good drug-like properties,
they perform in animals,
they can be manufactured,
they advance through clinical trials,
and the platform becomes more efficient across successive programs.
Only then will we know whether generative AI has changed the economics of biologics discovery.
Generate’s clinical progression provides an important early test.
Other Boston companies will follow.
Final Takeaway: The Winning AI Antibody Company May Own the Best Learning Loop
The first generation of AI-biotech excitement focused on prediction.
Then came generation.
The next stage will be about learning.
The companies most likely to create lasting value will not necessarily have the model capable of producing the largest number of antibody sequences.
They will have a system that learns faster than everyone else.
A model designs a molecule.
Automation builds it.
Experiments measure it.
Cells reveal what it does.
Animals show where it goes.
Eventually, patients reveal whether the original biological idea was correct.
Every result returns to the platform.
Then the next molecule is designed with more knowledge than the one before it.
Boston is beginning to assemble exactly these kinds of systems.
Generate is showing that computationally engineered biologics can move deep into clinical development. Nabla is pushing toward de novo, epitope-controlled antibody design. Manifold is making living organisms part of the high-throughput learning loop. Seismic is using machine learning to engineer complex immune medicines. Metaphore is targeting function. Moderna is connecting generative protein design with a much broader therapeutic technology stack.
Our original analysis also shows that this is no longer a lightly funded research niche. A selected group of major Boston AI-biologics companies already represents roughly $1.4 billion in disclosed financing under our conservative methodology, before counting the full value of major pharma collaborations.
But money is not the most important advantage.
Neither is model size.
The scarce asset is becoming high-quality biological feedback.
That is why Boston matters.
The region already has the laboratories, protein scientists, clinicians, drug developers, computational researchers and biotech infrastructure needed to turn AI predictions into real experiments and real medicines.
Generative AI may make it possible to imagine millions of new antibodies.
The companies that matter will be the ones that can determine, faster than anyone else, which antibodies deserve to become drugs.
