Drug discovery platforms · Protein design
Diffuse Bio
A Y Combinator company founded by one of the researchers who first pointed diffusion models at proteins. It now sells the design software and the laboratory screen that checks whether the designs actually bind.
Four thousand designs, one number that matters
On June 18, 2024, Diffuse Bio published its first results. It had asked a model it calls Diffuse StructGen-1 to invent nanobodies against two target antigens, at a specified patch on each. It made 4,000 designs per target and put them on yeast.
Then it counted. About three percent of the proteins that expressed on the yeast surface also bound the target. Several of the ones it tested individually bound at 10 nanomolar or tighter. Competition experiments showed the designs hit the epitope they had been asked to hit, and did not stick elsewhere.
In protein engineering, a hit rate is the whole argument. The company’s own comparison, in a footnote, puts the historical rate for panning a giant random library at roughly one in a billion.
The post was self published, not peer reviewed, and the targets were never named.
Where the bottleneck moved
Computational protein design is old. What changed in 2022 was the arrival of generative models that build a protein backbone out of noise, the mathematics behind image generators. Namrata Anand and Tudor Achim posted one of the first to arXiv on May 26, 2022. That December the Baker lab at the University of Washington released RFdiffusion, trained at far greater scale, and published it in Nature in 2023. Diffusion became the field’s default tool.
Better generators created a second problem. If a model can write ten thousand candidate binders in an afternoon, the constraint is no longer the model. It is the bench. Diffuse Bio put this in writing when it launched its screening platform in December 2025: “As protein generative methods improve, experimental screening becomes the rate-limiting step.”
That sentence explains the shape of the company. It is not only selling a model.
The founder
Namrata Anand took a bachelor’s in applied mathematics at Harvard in 2014, a master’s in computer science at Stanford in 2016 and a PhD in bioengineering at Stanford in 2021, and was on the founding team at Encoded Therapeutics.
In February 2022, Nature Communications published her paper on designing protein sequences with a learned potential. Of 16 designs tested, 15 expressed in bacteria, 10 looked well folded under circular dichroism, and two high resolution crystal structures matched the computed models. That is experimental validation of a neural network design, done before the company existed.
Three months later came the diffusion preprint with Achim. On the day Chroma and RoseTTAFold Diffusion were announced, she told MIT Technology Review that the significance was scale.
From a paper to a price list
The provisional patent went in on May 19, 2022. Diffuse Bio, Inc. went through Y Combinator in the Winter 2023 batch, then stayed quiet for about eighteen months.
Then it shipped, steadily. DSG-1 results in June 2024. DSG2-mini and a web app called DiffuseSandbox on June 11, 2025, the first public access to its models. RamaX, a new screening assay, on December 16, 2025. On February 2, 2026 the two were wired together, so a user can generate up to 10,000 designs and send them straight to the bench. RamaX Opt for affinity maturation in March 2026, ProxyTm for thermostability in May 2026.
The patent office granted number 12,573,475 on March 10, 2026 to Diffuse Bio, Inc., naming Anand and Achim as inventors. The assignment records show they signed the invention over to the company on December 9 and 10, 2024, more than two years after the priority date.
Prices are published: a RamaX screen or optimization run starts at $30,000, and discovery from a naive or AI designed library at $50,000. Customers who let their data train the models pay less, which is the point of the whole apparatus.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| US patent 12,573,475 B2 | Priority May 19, 2022 from provisional 63/343,789, filed May 18, 2023, granted Mar 10, 2026. Inventors Namrata Anand and Tudor Achim. Assignee Diffuse Bio, Inc., California, by assignment signed Dec 9 to Dec 10, 2024. | Public record |
| Founder’s peer reviewed work | Protein sequence design with a learned potential, Nature Communications, Feb 8, 2022. 15 of 16 designs expressed, 10 appeared well folded by circular dichroism, two high resolution crystal structures agreed with the models. Predates the company. | Public record |
| The diffusion preprint | arXiv 2205.15019, submitted May 26, 2022, by Anand and Achim. No peer reviewed journal version appears in the record. | Public record |
| DSG-1 binder results | Company blog, June 18, 2024: 4,000 designs screened per target on yeast display, about 3 percent of expressing cells bound, several designs at 10 nM or tighter, epitope specific by competition. Two targets, unnamed. Not peer reviewed. | Company-stated |
| DSG2-mini benchmarks | Company reports it beats DSG1, RFdiffusion and RFantibody on interface and packing metrics for nanobody design. Computational only, internal test set. | Company-stated |
| RamaX accuracy | Company case study, July 6, 2026: against TrkA, PD-L1 and CTLA-4 libraries with existing SPR labels, the screen recovered over 75 percent of binders at a false positive rate as low as 15 percent. | Company-stated |
| Partner campaigns | Two anonymized projects: 2,000 designed minibinders yielding single digit nanomolar hits in under two weeks, and 6,000 designs yielding a 150 nM binder. Partners not named. | Company-stated |
| Regulatory and clinical record | No FDA 510(k) or PMA, no registered trials, no NIH or NSF award, no PubMed entry under the company name, as of Sep 24, 2026. | Not found |
| Independent coverage | MIT Technology Review, Dec 1, 2022, and Chemical and Engineering News, 2023, quote Anand on the state of the field. No independent laboratory test of the company’s models was found. | Independent |
Read plainly: the public record establishes who these people are and what they invented, and stops there. The patent is granted, the crystal structures are in Nature Communications, the preprint is a landmark. Everything about the product, the hit rate, the benchmark wins, the partner campaigns, rests on the company’s own posts. Those posts are unusually detailed for a private startup, with target names, sample counts and false positive rates. They are still self reported.
What to watch
- DSG2. The flagship model has been described since June 2025 as upcoming, while only the smaller DSG2-mini is public.
- External validation of RamaX. A preprint or a third party benchmark would settle the accuracy claims.
- A named partner or pharma agreement. Every commercial result so far is anonymized.
- A priced financing. No Form D from this company appears in SEC records as of Sep 24, 2026.
- Which half of the business wins. Screening bills today, models pay later, and the plan is to use the first to build the second.
In their words
“Our goal is to make designing a protein (therapeutic, diagnostic, enzyme, molecular machine, etc) as simple as pushing a button on a computer.”
Namrata Anand, founder, personal site · Founder-stated
“It may be fair to say that this is more like DALL-E because of how they’ve scaled things up.”
Namrata Anand, on Chroma and RoseTTAFold Diffusion, MIT Technology Review, Dec 2022 · Independent
“I think, to actually solve downstream problems, you’re going to need a whole bunch of complementary approaches.”
Namrata Anand, Chemical and Engineering News, 2023 · Independent
“We’re still in the regime where physics-based models are better for certain types of problems.”
Namrata Anand, Chemical and Engineering News, 2023 · Independent
“we found that the AI-designed nanobodies bound their respective targets with a ~3% hit rate”
Diffuse Bio, Our Vision and Initial Results, June 18, 2024 · Company blog
“DiffuseSandbox puts the power of DSG2-mini into researchers’ hands, whether they’re at major institutions or working independently”
Namrata Anand, CEO, launch release, June 11, 2025 · Company release
“These works reveal just how powerful diffusion models can be for protein design”
Joseph Watson, Baker lab, on RFdiffusion, Institute for Protein Design, Dec 2022 · Independent
Related companies
Sources
- Public recordUS patent 12,573,475 B2, Protein sequence and structure generation with denoising diffusion probabilistic models
- Public recordProtein sequence design with a learned potential
- Public recordProtein Structure and Sequence Generation with Equivariant Denoising Diffusion Probabilistic Models
- Public recordDe novo design of protein structure and function with RFdiffusion
- Public recordEDGAR Form D search, no Diffuse Bio entity found
- Public recordRecords pull: no FDA, trial, NIH, NSF or PubMed record
- IndependentBiotech labs are using AI inspired by DALL-E to invent new drugs
- IndependentGenerative AI is dreaming up new proteins
- IndependentA diffusion model for protein design
- IndependentDiffuse Bio company profile
- IndependentDiffuse Bio funding and patent record
- IndependentNamrata Anand speaker profile
- CompanyOur Vision and Initial Results, DSG-1 data
- CompanyDiffuse Bio Launches DSG2-mini AI Model for Protein Binder Design
- CompanyRamaX: Fast, Selective, Sensitive Screening Platform
- CompanyDiffuse Bio Announces Push-Button Platform for Ultra-Fast Biologics Discovery at Scale
- CompanyUltra-fast Binder Optimization with RamaX Opt
- CompanyProxyTm: High-Throughput Protein Thermostability Measurement
- CompanyBenchmarking Minibinder Library Screening with RamaX
- CompanyRamaX workflows and published pricing
- CompanyCompany page and launch post, Winter 2023 batch
- FounderNamrata Anand personal site and publication page
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
