AI protein design: which startup is ahead?
In our AI in drug discovery market deck, you will find everything you need to understand the market
SUMMARY
Earendil Labs is ahead in AI protein design today because it has taken an AI-supported antibody pipeline into human trials, including a Phase IIa program, while its closest private rivals remain at the design, laboratory or preclinical stage.
Chai Discovery is still the strongest reusable therapeutic protein-design platform. Its reported hit rates, rapid laboratory workflow and adoption by Pfizer, Novartis, Lilly and argenx make it the clearest platform challenger to Earendil.
The comparison is really between several different businesses. Earendil is a drug developer, Chai and Nabla design therapeutic proteins for pharmaceutical teams, Profluent builds gene editors, Cradle sells engineering software, and EvolutionaryScale develops broad protein foundation models.
Clinical progress changes the ranking more than funding or model size. Earendil has already dealt with manufacturing, regulatory preparation, human safety testing and patient dosing, which is a much harder test than generating a strong binder in the laboratory.
Cradle has the broadest visible pattern of regular customer use. More than 50 reported projects, work across six of the 25 largest pharmaceutical companies and several expanded or multiyear relationships make it the closest thing to an everyday enterprise tool for protein engineers.
Profluent has the strongest specialist lead. OpenCRISPR-1 is the field’s clearest peer-reviewed example of an AI-generated protein performing a consequential biological task, and no direct competitor has matched its combination of published evidence and gene-editing partnerships.
Nabla Bio is the capital-efficiency standout. It has raised far less than the leaders but produced unusually strong antibodies against difficult GPCR targets and turned an initial Takeda relationship into a broader multiyear agreement.
EvolutionaryScale probably has the broadest general protein model, but proprietary data may become just as important as model architecture. Basecamp Research’s biological collection is harder to reproduce than another large public-sequence training run, although it still needs stronger functional and commercial proof.
Earendil’s lead is real but fragile. A weak Phase IIa result for HXN-1001 would quickly reduce the value of its clinical advantage, while one named Chai-designed candidate entering formal development could narrow the gap sharply.
The current order is Earendil first, Chai second and Profluent third, followed by Cradle and Nabla. The leaders are ahead for different reasons, so the ranking could change quickly once clinical efficacy data, named development candidates or recurring revenue figures become public.
This market map, featured in our AI in drug discovery market deck, highlights top companies and startups in the AI in drug discovery market
Which startups really belong in the AI protein design race?
The serious private AI protein design field currently contains about ten companies, but only Earendil Labs, Chai Discovery, Profluent, Cradle and Nabla Bio have enough evidence to shape the overall leadership debate.
We include companies that use AI as a central part of designing, generating or improving proteins. That covers therapeutic antibodies, gene editors, enzymes, peptides, miniproteins and general protein foundation models. A company does not qualify simply because it predicts protein structures or uses machine learning somewhere in its research process.
The field also contains several different businesses. Earendil Labs develops its own antibody and biologics pipeline. Chai and Nabla mainly help pharmaceutical teams design therapeutic proteins. Profluent specializes in gene editors and recombinases. Cradle sells software that scientists use to improve proteins themselves. Arzeda and Biomatter focus more heavily on enzymes and industrial applications.
We exclude Generate Biomedicines and Absci because both are now public companies. Generate remains the most useful outside benchmark because its platform has already produced several clinical-stage programs.
Isomorphic Labs is also left out. Its technology supports the wider drug-discovery process, including small molecules, and its Alphabet ownership makes a direct startup comparison less useful.
Other companies sit just outside the boundary. Shiru mainly searches for useful natural proteins rather than designing entirely new ones. Xaira Therapeutics, Iambic and Insilico Medicine use AI across broader drug-discovery workflows where protein generation is only one possible tool.
Private funding figures are not perfectly consistent across databases, so we use the best available cumulative figure and round it when necessary.
| Startup | What it does | Approximate disclosed funding |
|---|---|---|
| Earendil Labs | Uses AI and high-throughput biology to develop antibodies and other protein therapeutics | About $787 million |
| Chai Discovery | Designs antibodies, miniproteins and other therapeutic molecules through AI models | About $630 million |
| Profluent | Generates functional proteins, especially gene editors and recombinases | About $150 million |
| EvolutionaryScale | Builds broad foundation models for understanding and generating proteins | $142 million |
| Cradle | Sells protein-engineering software to pharmaceutical and industrial scientists | About $103 million |
| Arzeda | Designs and commercializes enzymes and industrial proteins | About $95 million |
| Basecamp Research | Builds proprietary biological datasets and models for protein and genetic design | About $85 million |
| Latent Labs | Generates peptides, miniproteins and antibody formats through accessible AI tools | $50 million |
| Nabla Bio | Designs de novo antibodies and protein therapeutics against difficult targets | About $37 million |
| Biomatter | Designs enzymes for industrial and life-science customers | About €7 million |
Is Earendil Labs now ahead of Chai Discovery in AI protein design?
Earendil Labs is ahead overall because it has moved an AI-supported protein pipeline into human trials, including a Phase IIa program. Chai Discovery has not publicly shown a named drug candidate entering clinical development.
Earendil’s lead program, HXN-1001, is a long-acting antibody targeting TL1A for inflammatory bowel disease. The company completed enrollment in a Phase I healthy-volunteer study and began dosing patients with ulcerative colitis in Phase IIa. Earendil reported that the antibody was well tolerated and had a half-life of roughly eight weeks at the projected therapeutic dose.
Chai can generate promising molecules and place its software inside large pharmaceutical companies, but its disclosed candidates remain before clinical testing. Earendil has already taken a molecule through manufacturing, regulatory preparation, human safety testing and into an early efficacy study.
Earendil also says its platform has produced more than 40 internal and partnered programs across conventional antibodies, bispecifics, trispecifics, antibody-drug conjugates and fusion proteins. Several have reached candidate selection or IND-enabling work.
Sanofi licensed two Earendil bispecific antibodies and later expanded the relationship into a broader discovery collaboration. Earendil therefore has both internal clinical progress and external pharmaceutical validation.
Chai remains easier to judge as a pure design platform because it publishes clearer hit rates, provides direct software access and names several major pharmaceutical users. Earendil reveals less about how much each final drug sequence came directly from generative AI rather than conventional screening and optimization.
Still, clinical execution carries more weight in the overall ranking. Earendil has crossed the hardest threshold reached by any private company in this group, while Chai remains the strongest reusable platform challenger.
If you want more recent data on this point, please see our latest AI in drug discovery market report.
As this chart shows, and as featured in our AI in drug discovery market deck, search interest in AI drug discovery has grown rapidly
Which AI protein design startup has the strongest real customer demand?
Cradle shows the broadest day-to-day customer usage, Chai Discovery has the strongest group of major pharmaceutical platform users, and Earendil Labs has the deepest asset-level relationship through Sanofi.
Cradle reported more than 50 protein-engineering projects across six of the 25 largest pharmaceutical companies. Its named users have included Johnson & Johnson, Novo Nordisk and Lundbeck. It also signed a three-year collaboration with Bayer to place its software inside antibody-discovery and optimization workflows.
Several early Cradle projects reportedly expanded into multiyear commitments, while existing customers added more proteins and research teams. That is better evidence of recurring use than a one-off pilot.
Chai’s customer base is narrower but strategically important. Pfizer licensed the platform and gained access to Chai-3 with software adapted to its research. Novartis expanded more than a year of technical work into a multi-target antibody collaboration. Chai has also worked with Eli Lilly, argenx and Lilly’s TuneLab initiative.
Chai has not disclosed annual revenue, contract values or active-user numbers, so the commercial scale remains unclear even though the customer quality is exceptional.
Earendil’s Sanofi relationship goes further into drug ownership. Sanofi obtained worldwide rights to two Earendil antibodies and later agreed to use Earendil’s platform across additional autoimmune programs.
Nabla has a smaller customer footprint but one of the clearest examples of repeat demand. Takeda began working with Nabla in 2022 and returned for a broader multiyear agreement with double-digit millions in upfront and research payments.
Profluent’s Lilly collaboration carries up to $2.25 billion in possible payments, but most of that value depends on future scientific, clinical and commercial milestones rather than current revenue.
Cradle leads on regular usage, Chai on external platform adoption by major pharmaceutical companies, and Earendil on licensed drug assets.
Does Chai Discovery still have the strongest lab-proven antibody platform?
Chai Discovery has the strongest broad antibody-design platform in the laboratory, although Profluent offers the clearest peer-reviewed functional result and Nabla Bio may be better on the hardest antibody targets.
Chai reports experimental success rates above 10% for de novo antibodies and above 50% for miniproteins. Its platform can design against selected regions of a target, work with several antibody formats and move chosen designs into laboratory characterization in less than two weeks.
A double-digit hit rate can sharply reduce the number of candidates a laboratory needs to test. The limitation is that Chai’s results remain mainly company-reported, and performance can vary greatly by target.
Profluent’s OpenCRISPR-1 provides stronger independent proof. The AI-generated editor successfully modified the human genome, showed activity comparable with SpCas9 and achieved higher specificity in reported tests. The work later appeared in Nature.
Nabla has produced some of the most impressive results against difficult therapeutic targets. Its JAM system generated hundreds of single-domain antibodies against the GPCRs CXCR4 and CXCR7, with the strongest designs reaching picomolar to low-nanomolar binding affinities.
Latent Labs reports the highest raw hit rates. Latent-X achieved 91% to 100% for macrocyclic peptides and 10% to 64% for protein mini-binders across seven targets. Several reached picomolar affinity, although smaller binders are easier to generate than complete therapeutic antibodies.
Chai’s pharmaceutical reach strengthens its technical case. Pfizer, Lilly, Novartis and argenx have used or evaluated the platform across their own targets. Nabla has deeper public evidence on difficult membrane proteins, while Earendil has done more to turn antibody discovery into a clinical pipeline.
Chai still leads the external antibody-platform race because it combines broad design performance, target flexibility and major pharmaceutical adoption.
| Startup | Strongest reported laboratory result | How much confidence should we place in it? | Main limitation |
|---|---|---|---|
| Chai Discovery | More than 10% success for antibodies and more than 50% for miniproteins | Strong multi-target wet-lab evidence, although mainly company-reported | Full target-level data and independent replication remain limited |
| Profluent | OpenCRISPR-1 edited the human genome with activity similar to SpCas9 and higher reported specificity | Highest confidence because the work was peer-reviewed in Nature | Evidence remains concentrated in gene-editing proteins |
| Nabla Bio | Hundreds of GPCR antibodies, with leading designs reaching picomolar to low-nanomolar affinity | Detailed experimental preprints on commercially difficult targets | No disclosed clinical candidate yet |
| Latent Labs | 91% to 100% macrocycle hit rates and 10% to 64% mini-binder hit rates | Extensive company-run testing across seven targets | Smaller binders are easier to generate than complete therapeutic antibodies |
| EvolutionaryScale | A distant but functional fluorescent protein generated by ESM3 | Strong scientific demonstration | One flagship protein does not prove repeatable therapeutic performance |
| Cradle | Repeated improvements in affinity, activity, expression and stability during customer projects | Real laboratory feedback from customer programs | Cradle mainly optimizes existing proteins rather than designing complete drugs from scratch |
If you want more recent data on this point, please see our latest AI in drug discovery market report.
This chart, featured in our AI in drug discovery market deck, shows annual venture capital investment in AI drug discovery startups
Which AI protein design startup is closest to helping real patients?
Earendil Labs is already closest to patients because HXN-1001 has entered a Phase IIa ulcerative colitis trial, several development stages ahead of the other private protein-design platforms.
HXN-1001 has moved beyond binding tests and animal models. Earendil completed its Phase I enrollment in healthy volunteers and began testing the antibody in people with ulcerative colitis.
Earendil has also started a Phase I study of HXN-1011, a biparatopic antibody targeting TSLP for asthma and chronic obstructive pulmonary disease. A second clinical program makes its progress harder to explain as one unusually advanced asset.
The main caveat concerns the phrase “AI-designed.” Earendil describes an AI-native platform combining foundation models, antibody libraries and high-throughput laboratory work, but it has not published a full design history for HXN-1001.
Chai’s pharmaceutical partners may already have undisclosed candidates moving toward development, but we cannot give clinical credit to programs that have not been named.
Nabla’s Takeda programs remain earlier, while Profluent’s gene editors still need to solve delivery, off-target effects, immune responses and long-term safety before reaching patients.
Generate Biomedicines shows how much further the category can go. The public company now reports five clinical-stage programs enabled by its platform, including a pivotal asthma program.
Earendil has crossed the private field’s clearest clinical threshold. The next test is whether HXN-1001 produces meaningful efficacy in patients.
Can any startup catch Profluent in AI-designed gene editors?
No other startup is close to Profluent in AI-designed gene editors today.
OpenCRISPR-1 showed that a protein language model could create a gene-editing system far from known natural sequences and still edit the human genome. Profluent also made the editor openly available for ethical research and commercial use.
The Nature publication gave researchers access to the sequences, experimental setup, activity measurements and specificity results rather than just a company summary.
Profluent has since expanded into CRISPR proteins with different targeting requirements, base editors and site-specific recombinases capable of moving much larger pieces of DNA.
The Lilly partnership focuses on these recombinases. Standard CRISPR tools are better suited to relatively small cuts or corrections, while recombinases could insert, remove or rearrange larger DNA segments.
Profluent has also created several routes toward deployment. Ensoma is working with it on genetic medicines for blood stem cells, Corteva is exploring agricultural proteins, and an ARPA-H-backed program is applying Profluent technology to a rare liver disease.
Basecamp Research is the closest long-term threat. Its proprietary biological data covers mobile genetic elements and large serine recombinases, but it still lacks Profluent’s published functional proof and partnership depth.
Profluent’s main risk comes from biology rather than a direct competitor. Its editors still need to reach the right cells, avoid unwanted edits and control immune reactions.
If you want more recent data on this point, please see our latest AI in drug discovery market report.
This chart, featured in our AI in drug discovery market deck, shows how Shrödinger is positioned in AI drug discovery
Does EvolutionaryScale have the best protein model, or does proprietary data matter more?
EvolutionaryScale still leads the general protein-foundation-model race, while Basecamp Research owns the raw biological dataset that competitors will find hardest to reproduce.
ESM3 works across protein sequence, three-dimensional structure and biological function. Researchers can provide part of one or more of these elements and ask the model to complete the rest.
The esmGFP experiment showed that ESM3 could generate a functional fluorescent protein far from known natural fluorescent proteins in sequence space. EvolutionaryScale also develops ESM Cambrian models for prediction and classification tasks.
Profluent has built the strongest direct model alternative. Its ProGen3 family was trained on more than 3.4 billion protein sequences and connects broad protein generation with experimentally tested gene-editing applications.
Basecamp is taking a different route by collecting biological data missing from standard public databases. Its network spans 31 countries across six continents, and the Trillion Gene Atlas aims to gather genomic information from more than 100 million species.
That dataset could expose natural protein functions and genetic systems that competing models have never encountered. The difficult part will be turning its scale into reliable designs and useful products.
Earendil and Cradle have more immediate experimental feedback loops. Earendil combines model predictions with antibody expression, purification and functional testing. Cradle learns from laboratory results generated inside customer projects, although confidentiality limits how freely that data can be pooled.
EvolutionaryScale remains first in general model breadth. Basecamp leads on unique biological inputs, while Earendil has the strongest closed drug-development loop.
Is Cradle becoming the default AI tool for protein engineers?
Cradle is the closest thing to a default enterprise AI tool for protein engineering.
Scientists upload sequences and experimental results, use the software to propose improved variants and feed new laboratory data back into the system. Customers keep control of the molecule, experiments and intellectual property.
That approach fits normal protein-engineering work. Many teams already have an antibody or enzyme that functions but needs better affinity, stability, activity, expression or manufacturability.
Cradle reported more than 50 projects across six top-25 pharmaceutical companies. Some proof-of-concept projects have expanded into multiyear relationships, and Bayer agreed to deploy the platform for three years across parts of its antibody pipeline.
The company says its software can speed projects by as much as twelve times and reduce costs by up to 90%. Its case studies are more useful than those headline figures. One enzyme customer had already tested roughly 1,200 variants through ten rounds before Cradle helped nearly triple activity within three additional rounds.
The missing number is revenue. Cradle has disclosed projects, customers and expansion, but not annual recurring revenue or retention rates.
Even with that gap, Cradle has made protein AI feel more like a normal working product than most competitors.
If you want more recent data on this point, please see our latest AI in drug discovery market report.
This chart, featured in our AI in drug discovery market deck, shows annual funding in AI drug discovery startups
Which AI protein design startup is moving fastest right now?
Earendil Labs is moving fastest, Chai Discovery follows closely, and Latent Labs has the quickest pace of visible product releases.
Earendil raised about $787 million, broadened its Sanofi relationship, added development and manufacturing collaborations, and advanced another respiratory antibody into human testing. Its recent progress covers financing, clinical execution and pipeline expansion.
Chai raised $400 million at a reported $3.8 billion valuation after raising $130 million only months earlier. Its recent Pfizer, Novartis, argenx and Lilly-related agreements show that pharmaceutical adoption is keeping pace with the funding.
Chai is also widening distribution through Lilly’s TuneLab, which allows selected biotechnology companies to evaluate its miniprotein-design tools.
Profluent’s Lilly deal gives it a well-funded route to develop AI-designed recombinases for several genetic diseases. Cradle’s three-year Bayer agreement shows a similar move from experimentation into longer deployment.
Latent Labs has released products faster than it has signed major pharmaceutical deals. It moved from peptides and miniproteins into antibody formats and then introduced Latent-Y, which automates more of the design process.
Nabla remains the capital-efficiency standout. It raised roughly $37 million before securing double-digit millions in upfront and research payments from its second Takeda agreement.
Earendil leads the current acceleration because its financing, clinical programs and partnerships are advancing together.
What could knock Earendil Labs or Chai Discovery off the top?
Weak clinical results from HXN-1001 would quickly damage Earendil Labs’ lead, while Chai Discovery needs a named development candidate to close the largest gap in its own case.
Earendil’s position now depends heavily on clinical execution. HXN-1001 could prove safe without delivering enough benefit to patients, or it could struggle to stand out from other TL1A antibodies already in development.
Its pipeline of more than 40 programs reduces dependence on one molecule, but repeated setbacks would raise questions about the wider platform.
Chai faces the opposite problem. Its technology and customer adoption look increasingly credible, but the clinical evidence has not caught up. A Chai-generated antibody entering formal preclinical development would help; a clinical trial could put Chai back in first place.
Profluent has another path to the top. A generated recombinase or gene editor entering human testing would validate a new type of therapeutic machinery rather than another antibody platform.
Cradle could climb through financial evidence. Strong recurring revenue, high retention and wider use across customer portfolios would prove that protein AI can support a large software business without taking drug-development risk.
Nabla needs one of its Takeda programs to become a named development candidate. EvolutionaryScale needs a flagship commercial use for ESM3.
Large outsiders are also getting stronger. Generate Biomedicines now has five clinical-stage programs, while Isomorphic Labs can draw on DeepMind research, Alphabet infrastructure and billions of dollars in funding.
Earendil remains closest to proving the full path from generated sequence to effective medicine, but the decisive clinical data has not arrived.
This chart, featured in our AI in drug discovery market deck, compares the main business model options for AI drug discovery biotech companies
Which AI protein design startups are actually ahead?
Earendil Labs is ahead overall, Chai Discovery leads the reusable protein-design platform race, and Profluent remains the strongest specialist.
Earendil is the only private company in our comparison with an AI-supported protein pipeline reaching Phase IIa. Its lead also extends across several antibody formats, more than 40 programs and a major Sanofi relationship.
Chai ranks a close second. Its models have clearer public design-performance evidence, and its platform is spreading across leading pharmaceutical research groups. A named development candidate could move Chai back into first place.
Profluent sits third because it has the strongest peer-reviewed functional protein result and a commanding lead in AI-designed gene editors.
Cradle comes fourth after building the broadest visible base of regular enterprise users. Nabla takes fifth and could overtake Cradle if Takeda advances one of its designed antibodies.
EvolutionaryScale owns the broad foundation-model lead without matching the commercial or clinical progress of the companies above it. Arzeda follows as the industrial-protein leader.
Basecamp has the most differentiated data strategy. Latent Labs is moving quickly at the product level but still needs deeper customer proof. Biomatter has credible enzyme-design capabilities on a much smaller scale.
Earendil holds first place because clinical maturity carries the most weight in a healthcare market. Chai is close enough that one successful development program could reverse the ranking.
| Rank | Startup | Where it leads | Why it holds this position |
|---|---|---|---|
| 1 | Earendil Labs | Overall development and clinical execution | The only private company here with an AI-supported protein pipeline reaching Phase IIa, backed by more than 40 programs and major pharmaceutical relationships |
| 2 | Chai Discovery | External therapeutic protein-design platform | Strong reported antibody and miniprotein performance, rapid model development and adoption by several major pharmaceutical companies |
| 3 | Profluent | AI-designed gene editors and recombinases | The strongest peer-reviewed functional result, a clear specialist lead and several routes into genetic medicine |
| 4 | Cradle | Enterprise protein-engineering software | The broadest disclosed regular usage, with more than 50 projects and several expanding or multiyear customer relationships |
| 5 | Nabla Bio | Difficult-target antibody design | Exceptional GPCR antibody results, repeat Takeda demand and unusually efficient use of capital |
| 6 | EvolutionaryScale | General protein foundation models | ESM3 remains the broadest protein-generation model, although commercial conversion still trails narrower platforms |
| 7 | Arzeda | Industrial protein commercialization | More experience bringing designed enzymes and proteins into scaled commercial products |
| 8 | Basecamp Research | Proprietary evolutionary data | A uniquely difficult dataset to reproduce and a promising route into programmable genetic medicines |
| 9 | Latent Labs | Accessible generative protein-design tools | Fast product releases and impressive reported binder hit rates, with customer adoption still developing |
| 10 | Biomatter | Emerging de novo enzyme design | Credible technical work and industrial applications, but far less capital and operating scale than the leading group |
If you want more recent data on this point, please see our latest AI in drug discovery market report.
OUR METHODOLOGY
This analysis compares the private startups that use AI as a central part of designing, generating or improving proteins. We include therapeutic antibodies, gene editors, enzymes, peptides, miniproteins, protein-engineering software and general protein foundation models, while excluding companies where protein design is only a minor part of a wider drug-discovery platform.
We assessed leadership across clinical progress, laboratory performance, pharmaceutical adoption, recurring customer demand, model capabilities, specialist advantages and recent execution. The companies operate different business models, so no single metric can rank all of them fairly.
Clinical milestones received the greatest weight in the overall ranking because they test far more than design quality. A molecule entering human trials has already passed through candidate selection, manufacturing, regulatory preparation and safety work that most laboratory-stage designs never reach.
For laboratory performance, we looked for experimentally validated proteins, disclosed hit rates, binding affinity, functional activity, target difficulty and independent replication. Peer-reviewed work carried more weight than a company summary, although detailed company-run experiments remained useful where independent data were not yet available.
Commercial demand was judged through named customers, repeat collaborations, multiyear agreements, project expansion and evidence that software or models were being used inside active research programs. Large milestone totals were not treated as current revenue when most of the value depended on future scientific or commercial success.
Foundation-model companies were assessed through model breadth, the ability to work across sequence, structure and function, and evidence that generated proteins were functional. We also considered proprietary biological data because an unusual training dataset may become a more durable advantage than another increase in model size.
Funding figures are approximate cumulative disclosures. Private-company databases often combine equity, grants, debt and round extensions differently, so the figures are used as context rather than as a direct measure of scientific or commercial leadership.
Key sources include Earendil Labs on its platform, Earendil’s clinical and partnership announcements, the published HXN-1001 results, Sanofi’s regulatory disclosure of the Earendil agreement, Chai Discovery’s platform and research materials, and the Chai and Pfizer licensing announcement.
We also relied on the peer-reviewed OpenCRISPR-1 research in Nature, Profluent’s OpenCRISPR materials, Cradle’s enterprise-adoption figures, Nabla Bio’s second Takeda collaboration, EvolutionaryScale’s ESM3 and esmGFP report, and Latent Labs’ experimental results.
The final ranking aggregates the strength, recency and relevance of this evidence. It separates overall leadership from narrower wins in antibody design, gene editing, enterprise software, industrial proteins, proprietary data and foundation models.
This chart, featured in our AI in drug discovery market deck, shows revenue breakdown by customer segment in the AI in drug discovery market
