For most of modern drug discovery, scientists have worked by searching nature for useful proteins and then improving what they found. They might begin with an antibody created by the immune system, an enzyme found in a living organism, or a protein identified after years of laboratory screening. That process has produced many important medicines, but it also creates a basic limit because researchers are usually starting with something biology already made and trying to modify it rather than designing exactly what they want from the beginning.
Artificial intelligence is starting to change that model in a very important way. Instead of searching through existing proteins and hoping to find a useful starting point, a new generation of companies is trying to design therapeutic proteins much more deliberately. Their systems can suggest new amino-acid sequences, estimate how those sequences may fold, predict how they could interact with disease targets, and then learn from laboratory experiments that show which ideas actually worked. The result is a new kind of drug-discovery process in which computation and biology are becoming much more tightly connected.
Boston has become one of the most important places to watch this shift because the region already has the ingredients needed to turn protein-design software into real medicines. Generate in Somerville is using generative AI to create and optimize proteins across several therapeutic formats. Nabla Bio in Cambridge is designing antibodies against targets that have traditionally been difficult to reach. Manifold Bio in Boston is trying to solve another important part of the problem by testing very large numbers of protein designs directly in living systems. Seismic Therapeutic has used machine learning to optimize biologic drugs for autoimmune disease, while Converge Bio is building generative systems for antibody design and protein manufacturing optimization.
The important point is that this is no longer simply a group of startups attaching the word “AI” to traditional biotechnology. Several of these companies are beginning to show what a much more advanced version of computational biology could look like. Their models generate possible medicines, automated experiments test those medicines, the resulting data are fed back into the models, and another design cycle begins with better information. That feedback loop may eventually matter far more than the ability to generate impressive protein structures on a computer screen because therapeutic proteins have to survive the real world of biology, manufacturing, animal testing, clinical development, and human medicine.
The Short Version: Boston Is Moving From Predicting Proteins to Engineering Medicines
The easiest way to understand the change happening in protein design is to separate two different questions. The first question asks what a protein looks like, while the second asks what protein should be built to perform a specific job.
The first problem has received enormous attention because systems such as AlphaFold showed how powerful machine learning could become when predicting protein structure.
AlphaFold helped demonstrate that AI could estimate three-dimensional protein shapes at a level that would have seemed extremely difficult only a few years earlier, and AlphaFold 3 expanded that work by modeling interactions involving proteins, DNA, RNA, ligands, ions, and other molecules.
Protein design moves in the opposite direction. Instead of beginning with an existing amino-acid sequence and predicting its structure, a protein-design system may begin with a desired function and search for a sequence capable of producing it.
A scientist might want a protein that binds a very specific location on a disease target, avoids another protein, remains stable in the bloodstream, can be manufactured efficiently, reaches the right tissue, and does not trigger an unwanted immune reaction. The design system is then asked to search for sequences that may satisfy several of those requirements at once.
That sounds similar to generative AI systems that create text or images, but the problem is far more demanding because a protein must function in physical reality.
A computer can generate a sequence that looks plausible, but that does not mean the molecule will fold correctly, bind strongly, remain stable, avoid unintended targets, manufacture well, survive in the body, or prove safe in patients.
This is why Boston’s emerging protein-design ecosystem is becoming less focused on one impressive model and more focused on the complete design-build-test-learn process.
Our BostonBizScene analysis suggests that this distinction helps explain where companies are putting their money and scientific effort. The most advanced Boston companies are building not only software but also laboratories, experimental systems, clinical pipelines, and proprietary datasets.
That is a much harder strategy to execute, but it may also create a much stronger competitive advantage if the models become better every time another molecule is designed and tested.
What AI Protein Design Actually Means
Proteins are built from chains of amino acids, and the order of those amino acids plays a major role in determining how the protein folds into a three-dimensional shape. That shape influences what the protein can bind to, where it can travel, how stable it is, and what biological function it performs. Because even a relatively short protein can have an enormous number of possible sequences, scientists cannot simply manufacture every possible version and test them one by one.
Traditional protein engineering therefore depends heavily on careful searching. Researchers often begin with something that already works reasonably well, make mutations, test variants, identify the strongest candidates, and then modify those candidates again. This process can be very effective, but it is often slow because scientists are exploring only a tiny fraction of the possible sequence space.
AI changes the search process by learning patterns from large collections of protein sequences, structures, and experimental results. Generative systems can then propose sequences that do not simply copy existing proteins. Some systems can work backward from a desired shape or biological function and suggest molecules that may satisfy those requirements. In practical terms, this allows researchers to ask more ambitious questions than they could ask with traditional screening alone.
Generate researchers, for example, developed Chroma, a diffusion-based model designed to create proteins while considering both structure and sequence. The research demonstrated that protein generation could be controlled using constraints related to shape, symmetry, and function. The important idea is not simply that the system can create unusual proteins. The more meaningful point is that researchers are trying to make protein generation programmable, which could eventually turn some parts of therapeutic discovery into a much more deliberate engineering process.
If scientists can reliably specify the properties they need and then generate molecules that satisfy those properties, protein development begins to look very different from traditional discovery. Instead of asking what biology happens to contain, researchers can increasingly ask what biology could be made to contain. That shift may become one of the most important changes in biotechnology over the next decade.
Why Protein Design Has Suddenly Become Much More Important
AI protein design is accelerating because several technologies are improving at the same time. Protein structure prediction has become much more accurate, protein language models can learn from enormous databases of amino-acid sequences, generative systems can propose new molecular structures, DNA synthesis has become easier, laboratory automation allows more experiments to be conducted, and high-throughput screening can generate much larger amounts of biological data than traditional manual workflows.
These improvements reinforce each other. Better models generate more promising candidates, better experiments reveal which candidates actually work, those experimental results become additional training data, and the next generation of models becomes better informed. The process becomes especially powerful when the same company controls both the computational layer and the experimental layer because it can repeatedly produce proprietary information that competitors do not have.
This creates an important strategic change. The strongest AI-biotech company may not be the company with the largest model or the most impressive benchmark result. The more durable advantage may belong to the company that can create high-quality experimental data faster than competitors and then use that data to improve the next round of designs.
Boston has unusual strengths in this area because the region already has a dense network of biotechnology laboratories, hospitals, research universities, pharmaceutical companies, investors, and scientists with experience developing biologic medicines. That gives local companies the opportunity to connect advanced computation with sophisticated biological experiments much more quickly than a pure software cluster might be able to do.
BostonBizScene Original Research: Mapping Boston’s AI Protein-Design Economy
To understand the Boston market more clearly, BostonBizScene created a focused original dataset of Boston-area companies where machine learning or generative AI is a central part of protein or antibody design. The goal was not to count every company that uses computational biology in some way. Instead, we wanted to identify businesses where AI is directly connected to designing, generating, optimizing, or experimentally improving protein therapeutics.
The dataset includes Generate, Seismic Therapeutic, Manifold Bio, Nabla Bio, and Converge Bio. Each company has a clear Boston-area presence, and each is using machine learning in a way that directly affects protein or antibody discovery rather than simply using AI for general business analytics, clinical data processing, or small-molecule research.
| Company | Boston-area base | Core protein-design angle | Publicly disclosed capital used in our analysis |
|---|---|---|---|
| Generate | Somerville | Generative protein therapeutics across multiple formats | Nearly $700M private equity disclosed by 2023 |
| Seismic Therapeutic | Cambridge/Watertown | ML optimization of antibodies, enzymes and immune therapeutics | $222M |
| Manifold Bio | Boston | AI-guided protein design plus high-throughput in-vivo testing | $40M Series A |
| Nabla Bio | Cambridge | De novo antibody and biologic design | $37M conservatively counted |
| Converge Bio | Boston | Generative antibody design and protein optimization | $30M |
The financing figures were intentionally handled conservatively. Generate stated that it had raised nearly $700 million in equity financing before its public offering. Seismic’s $121 million Series B brought its total capital raised to $222 million. Manifold disclosed a $40 million Series A. Nabla had publicly stated $37 million raised before additional later financing estimates began appearing in private-market databases, so we used the lower directly supported number in the core calculation. Converge Bio has said its 2026 Series A brought total funding to $30 million.
Our Methodology
This dataset is intentionally narrow because the goal is to understand the protein-design category rather than inflate Boston’s AI-biotech market by including loosely related companies. AI or machine learning had to be directly involved in designing, generating, or optimizing a protein therapeutic or antibody. Companies using AI mainly for clinical data analysis, target discovery, general biotech software, or small-molecule drug discovery were not included in the core group.
We also avoided adding undisclosed financing estimates or mixing different types of capital together. Where a company had private financing followed by a public offering, we treated those as separate financial events. That makes the analysis more conservative, but it also prevents the market from appearing larger than the publicly available evidence supports.
Original Finding #1: At Least $1.03 Billion of Private Capital Has Already Backed Boston’s Core AI Protein-Design Companies
The five companies in our conservative dataset account for approximately $1.029 billion in publicly disclosed private financing. That is a significant amount of capital for a category that only recently began moving from academic research into large-scale commercial development.
| Company | Capital included | Share of tracked private capital |
|---|---|---|
| Generate | ~$700M | 68.0% |
| Seismic Therapeutic | $222M | 21.6% |
| Manifold Bio | $40M | 3.9% |
| Nabla Bio | $37M | 3.6% |
| Converge Bio | $30M | 2.9% |
| Total | ~$1.029B | 100% |
The distribution is extremely uneven because Generate alone represents roughly 68% of the capital in the sample. That concentration tells us that investors are not spreading money evenly across the category. Much of the capital is moving toward companies attempting to build large, integrated platforms capable of both designing molecules and advancing them through development.
Generate then entered the public markets in February 2026 by pricing 25 million shares at $16 each, creating expected gross proceeds of $400 million before underwriting discounts and other expenses. If that IPO is added as a separate capital-market event rather than mixed into the private financing numbers, the combined capital access represented by this small Boston cohort rises to approximately $1.43 billion.
That figure should not be confused with current cash on hand or company valuation. It is simply a measure of how much publicly disclosed financing these companies have been able to access through major financing events. Even with that caution, the number shows that AI protein design has moved far beyond a collection of small research experiments. Serious capital is now being used to build laboratories, generate proprietary data, run drug programs, form pharmaceutical partnerships, and conduct clinical trials.
Generate Is the Biggest Test of the Entire Idea
Generate is important because it is attempting something much larger than using AI as a research assistant. The company wants to make protein medicines more programmable by combining machine learning, large-scale protein generation, laboratory validation, and clinical development inside one organization.
The physical scale of the company helps explain why this matters. Generate says it has generated, built, and tested more than 42,000 proteins, while its operations cover more than 140,000 square feet across facilities in Boynton Yards and Andover. Those numbers make one point very clear: advanced AI drug discovery is not only a software problem. Models may be able to generate enormous numbers of theoretical proteins, but those proteins still have to be manufactured, tested, compared, optimized, and eventually developed into medicines.
Chroma Showed What Generative Protein Design Could Become
Generate researchers developed Chroma as a programmable model for protein generation. Instead of simply predicting the shape of an existing amino-acid sequence, Chroma can generate new proteins while considering constraints related to shape, symmetry, and biological requirements. That represents an important shift because it allows scientists to think about protein design in much the same way engineers think about designing other complex systems.
The long-term goal is not simply to create unusual protein structures. The larger goal is to make biological design more intentional. If researchers can define useful properties in advance and then generate molecules that have those properties, discovery becomes less dependent on finding a fortunate starting point in nature.
Generate Is Now Testing Whether AI-Engineered Molecules Can Survive Clinical Development
The real test of this idea does not happen on a computer. It happens when a molecule enters drug development and eventually reaches patients. Generate’s GB-0895, a long-acting antibody targeting TSLP, is one of the most closely watched examples because the company announced global Phase 3 studies involving approximately 1,600 adults and adolescents with severe asthma.
The significance of that program is broader than one drug. AI-designed therapeutics have often been discussed using preclinical benchmarks, laboratory results, or model performance. A late-stage clinical program moves the discussion into a much harder environment. Researchers can now begin asking whether computational design created properties that matter in real patients, whether it reduced the amount of optimization required, and whether the same design principles can be repeated across additional medicines.
Generate also reported other clinical programs, including GB-4362 and GB-5267, which shows that the company is not betting everything on a single molecule. That matters because one successful drug can happen for many reasons. A platform becomes much more convincing when it repeatedly produces molecules that survive development.
Nabla Bio Is Attacking One of the Hardest Antibody Problems
Nabla Bio represents another important Boston strategy. The Cambridge company was founded by researchers with strong academic backgrounds and is focused on designing antibodies and other biologics against targets that have traditionally been difficult to reach.
Multipass membrane proteins are a useful example. These proteins move through cell membranes several times, which gives them complex shapes and makes them difficult to isolate, present, and screen against using conventional approaches. Many important drug targets belong to this category, yet scientists have often struggled to create suitable antibodies against them.
Nabla’s Bet Is That Antibodies Can Be Designed More Directly
Nabla is working on the idea that antibodies can be designed more deliberately instead of being found mainly through large-scale screening. Its JAM system is intended to generate antibodies for difficult targets and specific binding locations while combining computational design with laboratory testing of selectivity and biological function.
This approach could change the economics of antibody discovery if it works consistently. Traditional discovery may require researchers to screen very large collections of naturally derived or engineered antibodies and then spend significant time improving the best candidates. A strong generative system could narrow the search much earlier by creating molecules aimed at a specific location and biological function.
Nabla has reported research involving the de novo design of hundreds of functional antibodies against G protein-coupled receptors, which are among the most important classes of membrane proteins. The company has also described internal laboratory systems capable of generating and testing antibody candidates at very large scale. Those claims should still be evaluated carefully because testing methods and definitions vary across companies, but the general direction is important. AI protein design is moving toward much larger experimental cycles than traditional protein engineering teams could manage manually.
Original Finding #2: Pharma Is Paying for Access to Protein-Design Engines, Not Only Individual Drugs
The structure of pharmaceutical partnerships may be one of the clearest signals that the category is becoming strategically important. Traditional biotechnology licensing often focuses on a single molecule. By contrast, many AI protein-design agreements cover multiple targets, programs, or modalities, which suggests that large pharmaceutical companies are paying for repeated access to a platform rather than simply purchasing one drug candidate.
Generate’s collaboration with Amgen originally covered five programs, with $50 million paid upfront and up to $1.9 billion in potential future value plus royalties. Amgen later exercised an option for a sixth program that could add hundreds of millions of dollars in additional milestone value. Generate also signed a collaboration with Novartis that included $65 million in upfront consideration and more than $1 billion in potential milestone payments.
Nabla has also built a partnership-heavy strategy. The company announced collaborations with AstraZeneca, Bristol Myers Squibb, and Takeda carrying more than $550 million in combined upfront and potential milestone payments, and a later second Takeda collaboration added another agreement with more than $1 billion in possible success-based payments.
Manifold Bio followed a similar pattern when it signed a collaboration with Roche involving $55 million upfront and more than $2 billion in potential research, preclinical, clinical, and commercial milestones.
Taken together, these agreements show that large pharmaceutical companies are willing to commit meaningful upfront capital and much larger contingent payments to platforms capable of producing multiple therapeutic programs.
The Headline Opportunity Is More Than $6 Billion, but the Number Needs Context
When the maximum potential values of the largest disclosed collaborations are added together, Boston-area AI protein companies in our dataset are connected to more than $6 billion in headline partnership value. That figure sounds enormous, but it must be interpreted carefully because much of the money is not guaranteed.
Most pharmaceutical partnerships are structured so that only a relatively small portion is paid at signing. The much larger numbers depend on molecules reaching specific research, clinical, regulatory, or commercial milestones. Many drug programs never achieve those milestones, so headline deal values should never be treated as actual revenue.
Even with that warning, the structure matters. Large pharmaceutical companies are putting multiple programs behind AI protein-design platforms and returning for additional collaborations. That is stronger evidence of strategic interest than a demonstration, benchmark, or promotional claim.
Manifold Bio Is Solving the Problem That Comes After Generation
As AI makes it easier to generate huge numbers of candidate proteins, another bottleneck becomes more obvious. Someone still has to determine which molecules actually work inside a living biological system.
A protein may look excellent in a computer model or perform well in a simplified laboratory assay and still behave poorly in the body. It may disappear too quickly from circulation, accumulate in the wrong tissue, fail to reach the intended organ, or interact with other biological systems in ways that were not obvious during early testing.
Manifold Bio is focused on this downstream problem. The Boston company combines AI-guided protein design with high-throughput in-vivo measurement. Its technology is designed to track large numbers of biologics inside living systems so researchers can understand properties such as biodistribution, pharmacokinetics, and tissue targeting at much greater scale.
One Experiment Tested One Million Binder Designs
In March 2026, Manifold announced work with NVIDIA involving the experimental validation of a generative protein-binder model. The scale was unusually large. Manifold reported testing one million binder designs against 127 targets in a single multiplexed experiment and measuring more than 100 million protein-protein interactions.
The company reported finding specific binders for 68% of the tested targets. Those numbers should not be compared directly with success rates reported by other companies because experimental systems, definitions, and target difficulty can vary widely. However, the broader point is important because the cost of generating candidate proteins is falling much faster than the cost of physically testing them.
This creates a new economic opportunity. If generative models can produce millions of designs, technologies that efficiently identify which designs work may become just as valuable as the models themselves.
Original Finding #3: Boston’s Competitive Advantage May Be the Feedback Layer, Not the Foundation Model
Much of the public conversation around AI biology focuses on models. People ask which company has the largest model, which architecture is newest, or which system was trained on the most biological data. Those questions are understandable, but they may become less important as strong models become more widely available.
Our analysis of Boston companies suggests that the deeper competitive advantage may come from proprietary experimental feedback. Generate connects machine learning to large-scale wet-lab validation. Nabla combines generative design with proprietary antibody testing. Manifold puts in-vivo measurement at the center of its platform. Seismic combines machine learning with protein engineering, structural biology, and immunology. Converge Bio combines generative systems with laboratory validation and customer-facing development work.
The common pattern is a repeated loop in which AI generates ideas, physical experiments test those ideas, proprietary data reveal what worked and what failed, and the next generation of designs becomes better informed. That process may become much more defensible than simply owning one model.
Why Proprietary Experimental Data Could Become the Real Moat
Many protein sequences and structures are publicly available, and new AI research spreads quickly through academic papers, open-source repositories, and public model releases. This makes it difficult for any company to maintain a permanent advantage based only on architecture.
Experimental data are different. If a company designs thousands of proteins and measures binding strength, stability, expression, immune response, pharmacokinetics, and biological function, it creates a dataset that competitors cannot easily reproduce. The failed molecules are valuable too because they reveal which design choices did not work and under what biological conditions.
Over time, that information can become one of the strongest assets inside the company. Every experiment improves the dataset, the dataset improves the model, and the model improves the next set of experiments. Boston’s concentration of wet-lab talent and infrastructure makes this type of feedback loop particularly important.
Seismic Therapeutic Shows Why Designing One Property Is Not Enough
A therapeutic protein almost never needs only one strong characteristic. An antibody may bind its target extremely well but become unstable during manufacturing. An enzyme may perform its intended biological function but trigger an immune reaction. A molecule may have good laboratory activity yet bind to unintended targets or disappear from the bloodstream too quickly.
Seismic Therapeutic has built its IMPACT platform around this multi-property challenge. The company combines machine learning, structural biology, protein engineering, and immunology with the goal of optimizing several therapeutic properties at the same time rather than solving them one after another.
This approach is important because drug development is ultimately an optimization problem with many competing requirements. The strongest molecule on one laboratory measurement may be a poor medicine overall if it performs badly on other characteristics that matter later in development.
An Effective Protein Can Still Be a Bad Medicine
One particularly important issue is immunogenicity. Therapeutic proteins are introduced into the human body, and the immune system may recognize them as foreign. If patients develop antibodies against the therapy itself, the treatment can become less effective or cause unwanted reactions.
Seismic has used machine learning and wet-lab validation to redesign therapeutic proteins with the goal of reducing potentially immunogenic regions while keeping the desired biological activity. Its work on S-1117, a pan-IgG protease, illustrates this approach because the company has described using computational methods to reduce predicted immune epitopes while maintaining important functional properties.
Seismic Has Also Crossed Into Human Trials
Seismic’s programs have now moved beyond laboratory research. S-1117 entered Phase 1 testing, and S-4321, a bifunctional antibody designed for autoimmune disease, also moved into clinical development.
The importance of this transition is simple. AI protein design cannot be judged only by computational results. Eventually, molecules must be tested in humans. Clinical entry does not prove that a platform is successful, but it represents a much more meaningful level of evidence than early discovery benchmarks.
Original Finding #4: Boston’s Protein-Design Market Is Splitting Into Several Business Models
One of the most interesting findings from our dataset is that Boston companies are not all trying to build the same business. At least four different commercial models are beginning to appear, and each creates a different balance between risk, capital needs, and long-term upside.
| Model | Boston example | What is being sold |
|---|---|---|
| Integrated platform plus internal pipeline | Generate | Proprietary medicines and partnerships |
| Design platform plus pharmaceutical collaborations | Nabla | Designed biologics and partner programs |
| Experimental infrastructure plus therapeutics | Manifold | In-vivo discovery capability and tissue-targeting biologics |
| Disease-focused machine-learning biotech | Seismic | Internal immunology medicines |
| AI software and service layer | Converge Bio | Protein-design systems used by other drug developers |
Some companies want to own the final medicines. This creates the greatest possible upside because a successful drug can become extremely valuable, but it also requires enormous amounts of capital and patience because clinical development takes years and failure rates remain high.
Other companies are using partnerships to share development risk with large pharmaceutical organizations. Nabla is a good example because its technology can be applied across multiple partner programs without requiring the company to independently finance every downstream clinical trial.
Manifold points toward another possibility in which experimental infrastructure becomes valuable on its own. If protein generation becomes widely available, the ability to test huge numbers of molecules in living systems could become a scarce and valuable service.
Converge Bio represents yet another model in which AI protein design becomes more productized and accessible to companies that do not want to build large internal machine-learning teams. This could create an important software-and-services layer underneath larger biotechnology companies.
Converge Bio Shows Another Direction: Productizing Protein AI
Converge Bio approaches the market more like an AI platform provider than a traditional vertically integrated drug company. The company says it trains models on DNA, RNA, and protein sequences and builds systems covering target discovery, antibody design, manufacturing optimization, and biomarker discovery.
Its ConvergeAB product is designed to generate and score antibody candidates across several characteristics including affinity, stability, specificity, expression, immunogenicity, and manufacturability.
The goal is not simply to create a binder but to help customers identify molecules that have a better chance of becoming practical therapeutic candidates.
This business model could become particularly useful for smaller biotechnology companies. Not every startup can build a large internal AI organization and automated laboratory infrastructure. Many companies may instead use specialized platforms for certain parts of the design process, just as software companies rely on outside cloud infrastructure rather than building their own data centers.
Original Finding #5: The Market Is Moving From Single-Objective AI to Multi-Property Drug Design
The earliest demonstrations of protein AI often focused on one headline result. Researchers asked whether a model could predict a structure, generate a binder, or improve affinity. Those were important scientific achievements, but they represent only part of the problem faced by real drug developers.
Our review of current Boston platforms found repeated emphasis on optimizing several characteristics together. Nabla discusses binding, selectivity, developability, cellular function, and in-vivo behavior. Seismic focuses on therapeutic function, immunogenicity, and developability. Converge discusses affinity, stability, solubility, expression, immunogenicity, and structural fit. Manifold brings biodistribution, pharmacokinetics, and tissue targeting into the design process, while Generate is developing several therapeutic formats rather than treating protein generation as the final objective.
This suggests that the market is moving toward a much more practical definition of success. The best molecule is not necessarily the molecule with the highest score on one laboratory measurement. It is the molecule that performs well enough across the complete set of properties needed to become a medicine.
The Best Binder Does Not Automatically Become the Best Drug
Consider two hypothetical proteins. Protein A binds a disease target extremely strongly, but it is difficult to manufacture, unstable in storage, and likely to trigger an immune response. Protein B binds slightly less strongly but remains stable, expresses well, can be manufactured efficiently, reaches the right tissue, and lasts longer in the bloodstream.
Protein B may be the far better therapeutic candidate even though Protein A wins on the single measurement of binding strength. This is why AI protein design is increasingly becoming a multi-variable engineering problem rather than a simple search for the maximum score on one assay.
Why Boston Has an Advantage Here That Silicon Valley Cannot Easily Copy
Silicon Valley has extraordinary AI talent and enormous amounts of capital, but Boston does not necessarily need to compete by building the largest general-purpose AI models. Its advantage is the ability to connect computation to complex biology unusually quickly.
MIT, Harvard, Harvard Medical School, the Broad Institute, major hospitals, pharmaceutical companies, and hundreds of biotechnology firms sit within a relatively compact region. Scientists can move between universities and startups. Biotech companies can recruit researchers who already understand protein engineering, translational medicine, clinical development, regulatory work, and manufacturing.
That creates a different type of ecosystem. Boston is not simply an AI cluster. It is an AI-plus-biotechnology cluster, and that distinction becomes extremely important once a protein moves beyond the model and into physical experiments.
The hardest parts of therapeutic development begin after the computer generates a sequence. The molecule must be built, purified, tested, optimized, manufactured, and eventually evaluated in patients. Boston already has deep infrastructure around those later stages.
The Wet Lab Is Becoming Part of the AI Product
For many years, software companies could improve products almost entirely through digital feedback. Search engines learned from search behavior, recommendation systems learned from clicks, and online platforms could collect enormous amounts of data without conducting physical experiments.
Biology does not work that way. The correct answer often requires a real-world experiment. A protein must be synthesized, cells may need to be grown, binding must be measured, biological function must be tested, and preclinical studies may be required. Eventually, clinical trials are needed to understand what happens in people.
This means AI-biotech companies cannot separate their models from their laboratory operations as easily as normal software companies can. The wet lab is increasingly becoming part of the AI system itself because every experiment creates information that can shape the next computational decision.
Laboratory Throughput Could Become an Important AI Metric
Investors often evaluate software companies using metrics such as users, revenue growth, retention, or computing performance. AI biology may require a different set of measures.
The important questions may include how many designs the company can physically build, how quickly those molecules can be tested, how many biological properties can be measured at once, how expensive each design-test cycle is, and how much proprietary information each cycle produces. Investors should also ask whether later generations of designs are actually improving because of the experimental data being collected.
These operating metrics could eventually become more meaningful than the raw number of parameters in a biological foundation model.
Original Finding #6: Experimental Scale Is Expanding Faster Than Most Public Discussions Suggest
Several publicly disclosed numbers show how quickly experimental scale is increasing. Generate reports more than 42,000 proteins generated, built, and tested. Manifold reported an experiment involving one million binder designs, 127 targets, and more than 100 million measured protein-protein interactions. Nabla has described internal systems capable of generating very large numbers of antibodies and evaluating how they interact with targets within a matter of weeks.
These figures are not directly comparable because each company is measuring different things. One company’s “tested protein” may involve a very different experimental process from another company’s interaction measurement. Even so, the overall direction is difficult to ignore.
Protein engineering is moving toward experimental datasets of a size that would have been very difficult for traditional teams to manage manually. That matters because larger experimental datasets can improve the quality of future AI models, especially when the data includes failed designs as well as successful ones.
AI Does Not Remove Biology Risk
This is where some discussion of AI drug discovery becomes unrealistic. Generating a molecule more quickly does not eliminate the difficult parts of making a medicine.
Biological systems remain highly complex. A target may turn out to matter less than researchers expected. A molecule may produce unexpected toxicity. A drug can look promising in laboratory and animal studies and fail in people. Clinical endpoints can disappoint, competitors can develop superior treatments, manufacturing can create problems, and regulators still require strong evidence of safety and effectiveness.
AI can improve certain probabilities and timelines, but it does not remove the fundamental uncertainty of biology.
Clinical Proof Matters More Than Computational Proof
The protein-design field will therefore enter a more demanding stage during the next several years. Investors and pharmaceutical companies will care less about whether a company can generate attractive molecular structures and more about whether those molecules repeatedly survive real development.
The best platforms should eventually show measurable advantages such as stronger prospective hit rates, fewer optimization cycles, better developability, shorter time to candidate selection, lower research cost per successful program, or improved access to difficult biological targets.
The strongest evidence will come when several drugs created by the same platform advance successfully. A single strong molecule is encouraging, but repeated success begins to suggest that the underlying system is genuinely better.
Why the Generate Phase 3 Program Matters for the Entire Boston Ecosystem
Generate’s GB-0895 deserves attention beyond the company itself because Phase 3 represents a different level of evidence from the early stages where most AI-designed medicines remain.
The central question is not simply whether the drug succeeds or fails. The more strategic question is what researchers learn from the complete development journey. Did computational design improve the molecule’s properties? Did it reduce the amount of optimization required? Did it enable a longer dosing interval? Did the platform identify useful design features that traditional approaches might have missed? Can those lessons be applied to future molecules?
Those questions determine whether AI protein design becomes a repeatable industrial system rather than a collection of isolated successes.
Protein Design Could Expand the Number of Drug Targets
One of the most ambitious possibilities is that generative protein design could make previously difficult biological targets more accessible.
Many proteins are scientifically interesting but hard to drug. Some have awkward shapes, some lack obvious binding pockets, some sit inside cell membranes, and some require an antibody to bind a very specific location before a useful biological effect appears.
Traditional antibody discovery can struggle in these situations because researchers are often searching through very large libraries and hoping that a suitable molecule exists.
Generative design creates another path. Instead of waiting to find an antibody that happens to recognize a difficult target in the correct way, researchers can attempt to design that interaction from the beginning.
This Is Why Nabla’s Membrane-Protein Work Matters
Nabla’s work on multipass membrane proteins such as GPCRs, ion channels, and transporters is especially important because these proteins are biologically valuable but technically challenging.
If generative systems can repeatedly produce high-quality antibodies against specific locations on these targets, AI protein design will do more than simply make existing discovery faster. It could increase the number of biological mechanisms that can realistically be turned into medicines.
That would expand the market rather than merely improve the efficiency of the current one.
The Next Race Is Designing Where a Drug Goes
Protein design is also moving beyond the question of what a molecule binds. Researchers increasingly want to control where a therapeutic protein travels inside the body.
A drug may need to cross the blood-brain barrier, accumulate in a tumor, avoid healthy tissue, remain in circulation for a certain length of time, or concentrate inside a particular organ. Those distribution properties can determine whether a therapy works.
Manifold is especially interesting in this area because its platform focuses on measuring pharmacokinetics, tissue distribution, and biological behavior inside living systems. Its collaboration with Roche around brain shuttles also shows how protein design could eventually be used to control drug delivery more deliberately.
If this approach works, protein-design systems may eventually optimize both what a molecule binds and where it travels.
Another Race Is Designing Proteins That Can Actually Be Manufactured
A theoretical protein is not a pharmaceutical product. Drug companies need molecules that can be manufactured consistently at industrial scale.
A protein may look excellent in a computer model but express poorly in production cells, aggregate during manufacturing, become unstable during storage, or produce such low yields that the final medicine becomes too expensive.
This is why companies are increasingly including manufacturability inside the design process rather than waiting until late development to discover problems.
Converge Bio explicitly works on protein manufacturing optimization alongside antibody design, while Nabla has recruited scientists focused on expression, purification, stability, aggregation, hydrophobicity, and cross-reactivity.
These are practical signals that the industry is moving beyond impressive models and toward medicines that can survive the full development process.
What Pharmaceutical Companies Should Measure Before Partnering With an AI Protein Startup
Pharmaceutical companies evaluating AI protein-design platforms should avoid becoming overly impressed by model architecture or retrospective benchmark results. The strongest evaluation should focus on prospective evidence and real operating performance.
Companies should ask how often the platform produces experimentally useful molecules on targets that were not selected specifically to make the system look good. They should also examine what happens after the first successful binder appears because the number of additional engineering rounds required to create a developable molecule is often more important than the initial hit.
Another important issue is ownership of experimental data. If a pharmaceutical partner pays for thousands of experiments, the contract should clearly define whether the AI company can use those results to improve its general model and whether information learned from one partnership can indirectly benefit another.
Pharma teams should also calculate the total cost of a complete design-test-learn cycle. A platform that produces results quickly but requires extremely expensive experiments may not create a meaningful economic advantage. The important metric is not speed alone but useful progress per dollar and per experimental cycle.
What Boston Biotech Founders Should Learn From the Current Leaders
The strongest lesson from today’s Boston companies is that simply adding AI to an existing biotech workflow will probably not create a durable advantage. The more interesting companies are redesigning the workflow itself.
Generate links machine learning with large-scale experimental validation. Nabla connects generative design to antibody testing. Manifold created technology for measuring many biologics inside living systems. Seismic built a system around optimizing several therapeutic properties in parallel.
These companies are not merely using AI software. They are changing how experiments are selected, executed, and learned from.
Own a Difficult Feedback Loop
A new Boston company does not necessarily need the biggest foundation model. It may be much more defensible to own a unique source of biological feedback.
That could be a specialized cellular assay, a better method for measuring tissue targeting, a high-throughput functional screen, a manufacturing dataset, a human-relevant disease model, or a new way to measure therapeutic safety.
If those experiments create proprietary data that improves the model, the company can build a compounding advantage that is difficult for competitors to reproduce.
Start With a Narrow Biological Problem
“AI for drug discovery” is far too broad to be a strong market position. The most credible companies usually understand one difficult problem very deeply before expanding.
Nabla has concentrated heavily on difficult targets and antibody design. Seismic has focused on immune disease. Manifold has focused on in-vivo behavior and tissue targeting.
A narrow problem allows a company to collect better data, develop stronger experimental systems, and prove value more clearly.
What Investors Should Track Instead of AI Hype
AI biotech requires a different diligence framework from normal software investing. Revenue growth alone is not enough, and model performance alone is not enough.
The central question is whether the platform is becoming better at producing useful medicines.
| Metric | Why it matters |
|---|---|
| Prospective experimental hit rate | Shows whether generated molecules work outside retrospective benchmarks |
| Number of design-test cycles | Measures how much optimization is still required |
| Time to development candidate | Connects AI performance to actual drug-development speed |
| Developability pass rate | Shows whether molecules can become practical medicines |
| Proprietary experiments per quarter | Measures the rate of new learning |
| Cost per validated design | Tests whether the platform improves economics |
| Independent targets validated | Shows whether results generalize |
| Partner repeat rate | Reveals whether customers found the first program useful |
| Clinical-stage molecules | Shows whether computational output survives development |
| Human clinical results | Provides the strongest real-world test |
Several Boston companies already provide some of these signals, but the industry still lacks a standardized way to compare protein-design platforms. That gap will probably narrow as more companies move molecules into clinical development and investors demand clearer evidence of platform performance.
Original Finding #7: Repeat Partnerships May Be a Better Near-Term Signal Than Headline Deal Size
Large partnership numbers attract attention because they create impressive headlines, but most of the value is usually contingent. A more useful near-term signal may be whether pharmaceutical partners return for additional work.
Amgen expanded its Generate collaboration by adding another program, while Takeda entered a second collaboration with Nabla after earlier work between the companies. These repeat engagements matter because the pharmaceutical partner has usually seen much more information than outside investors or journalists.
A partner can evaluate molecule quality, experimental results, platform speed, communication, scientific depth, and the quality of the development team. A second collaboration does not guarantee success, but it may be stronger commercial evidence than a large theoretical milestone total.
Boston’s Biggest Constraint May Eventually Be Physical Infrastructure
AI models can scale rapidly using cloud computing, but biological experiments cannot scale nearly as easily.
Laboratories require specialized equipment, trained staff, controlled environments, sample storage, reagents, biological materials, and careful quality systems. Proteins need to be produced and purified. Cells need to be grown. Assays have to be conducted. Some experiments require animals or specialized disease models.
As design volume increases, physical laboratory throughput becomes a strategic issue.
Generate’s large footprint across Boynton Yards and Andover shows how different a mature AI-biotech company can look from a conventional software startup. The strongest companies may need excellent computational infrastructure and excellent biological operations at the same time.
Boston’s established life-sciences infrastructure is helpful here, although the region’s high real-estate and labor costs mean companies still need to use that infrastructure efficiently.
The Talent Market Is Also Changing
Protein-design companies require teams that are difficult to assemble.
They need machine-learning researchers who understand biological data, protein engineers who can translate computational designs into experiments, structural biologists, synthetic biologists, assay scientists, automation specialists, and computational biologists.
As drug programs advance, they also need the traditional machinery of pharmaceutical development including toxicology, manufacturing, clinical operations, regulatory strategy, and commercialization.
This creates another reason Boston is well positioned. The region already has a deep labor pool across both computation and drug development.
The emerging companies are showing that the most valuable employees may be people who can work across both worlds. Computational scientists increasingly need to understand wet-lab data, while experimental scientists need to understand how their measurements will improve the next generation of models.
The Bigger Story: Protein Design Is Becoming a Closed-Loop Engineering System
The most important change happening in Boston is not any individual model. It is the creation of a new industrial process.
Traditional drug discovery often follows a sequence in which researchers discover a starting point, screen candidates, optimize the best ones, test them, and eventually move them into development.
The emerging model is more circular. Researchers design molecules, build them, measure what happens, learn from the results, and redesign the next generation.
That difference becomes extremely important when the loop runs quickly.
Imagine that each experimental cycle gives an AI system thousands of new examples showing how sequence changes affected binding, stability, expression, cellular activity, immune response, and tissue distribution. The next generation of designs can be based on better information, and each following cycle becomes more informed than the last.
At that point, the company is no longer simply searching a static biological database. It is actively creating its own training data.
That is why automated laboratories, high-throughput experiments, and sophisticated in-vivo measurement may eventually become just as important as the AI models themselves.
What Could Happen Over the Next Five Years
The first major change is likely to be a much larger number of AI-designed protein therapeutics entering human trials. That will produce the type of evidence the field currently lacks. Some programs will fail, and those failures will be valuable because they will show where computational predictions failed to translate into real human biology.
Successful programs will be even more informative because they will show which design choices created real development advantages.
The second major change will be broader multi-property optimization. Models will increasingly be asked to design molecules that balance potency, selectivity, stability, half-life, immune response, tissue distribution, expression, and manufacturability at the same time.
The third change will be deeper integration with automated experimentation. Software systems will increasingly help decide not only which molecules to design but which experiment should be conducted next because that experiment will produce the most valuable new information.
The fourth change will be growing competition around proprietary datasets. As public models become stronger, experimental data that cannot easily be reproduced may become more valuable.
The fifth change will probably be consolidation. Pharmaceutical companies may acquire some protein-design platforms, build internal capabilities, or enter much deeper partnerships with the strongest independent companies.
What Boston Should Watch Closely
For Boston, the next stage of this market can be understood through a few important questions.
Generate must show whether its clinical pipeline can produce repeated evidence that computationally designed medicines offer real development advantages. Nabla needs to show that de novo antibody design can consistently produce development-ready molecules against difficult targets. Manifold needs to prove that large-scale in-vivo measurement can make tissue targeting and pharmacokinetics more predictable earlier in development.
Seismic’s clinical programs will provide evidence about whether multi-property machine-learning optimization can produce better immune therapeutics, while Converge Bio will test whether protein-design systems can become widely used infrastructure for companies that do not want to build large internal AI teams.
These questions matter far more than which company announces the newest model.
What Pharmaceutical Executives Should Do Now
Pharmaceutical companies do not need to rebuild their entire discovery organizations around AI immediately, but they should begin creating their own evidence now.
A useful approach is to choose one difficult but measurable protein-design problem and compare conventional discovery with AI-driven design under similar conditions. Both approaches should work against the same target and use the same success criteria.
Executives should then measure not only whether each method finds a useful molecule but how many experiments, how much money, and how much optimization were required.
The comparison should continue downstream. Researchers should track expression, stability, specificity, pharmacokinetics, manufacturing performance, and other properties that determine whether the molecule can actually become a drug.
This creates evidence that can support real investment decisions.
AI protein design should be treated as an engineering hypothesis that can be measured rather than a branding exercise that has to be accepted on faith.
Final Takeaway: Boston’s Protein-Design Race Is Really a Race to Build Better Learning Systems
The first major wave of excitement around AI and proteins was about prediction. The next wave is about generation, but the most important wave may ultimately be about learning.
A computer can generate an enormous number of proteins, but that capability has limited value on its own. Someone still has to build those proteins, test them, learn why some fail, understand how the best ones behave inside biological systems, and turn the strongest candidates into medicines.
Boston is assembling companies around every part of that process. Generate has built a large generative biology operation and advanced an AI-engineered antibody into late-stage clinical development. Nabla is trying to generate antibodies for difficult targets with greater control over where and how those antibodies bind. Manifold is scaling the measurement of designed proteins inside living systems. Seismic is using machine learning to optimize several therapeutic properties at once. Converge Bio is turning protein-design technology into a platform that other drug developers can use.
Our analysis of five core Boston-area companies found at least $1.03 billion in disclosed private financing before Generate’s 2026 public offering. Their major pharmaceutical collaborations carry more than $6 billion in potential headline value, although most of that money depends on future milestones and should not be confused with guaranteed revenue.
Those numbers matter, but they are not the most important finding.
The deeper finding is that Boston’s strongest AI protein companies are converging on the same fundamental architecture. They combine computation, experiments, proprietary data, and repeated learning in one continuous system.
That structure fits Boston unusually well because the region already knows how to build difficult biological experiments, develop biologic medicines, run clinical programs, and attract both technical and scientific talent.
AI is adding a new design engine to that existing infrastructure.
If the technology works at scale, the effect will be much larger than simply making today’s drug-discovery process faster. Scientists could gradually move from searching biology for useful molecules toward specifying the properties they need and engineering proteins that meet those requirements.
That is the real race now underway across Boston’s biotechnology ecosystem.
The company that generates the largest number of proteins may not become the most important company in the market. The stronger advantage may belong to the company that learns the most from every molecule it designs, builds, tests, and develops.
