Cyrus Biotechnology began with an unglamorous observation: one of academic science's most capable protein-design toolkits was a chore to use. Rosetta could predict structures, stabilize proteins and redesign molecular interfaces, but it asked industry scientists to master specialist software and serious computing. In 2014, three researchers from David Baker's University of Washington orbit - Lucas Nivon, Yifan Song and Javier Castellanos - decided the bottleneck was no longer only the algorithm. It was the doorway.
Their first product, Cyrus Bench, put Rosetta behind a browser interface and cloud infrastructure. Nivon once described the early suite as Microsoft Office with Word and Excel but without PowerPoint or spellcheck. The joke neatly captured both its utility and incompleteness. Cyrus was not inventing protein modeling from zero. It was packaging a deep, unruly body of academic work for scientists who had deadlines.
The paid apprenticeship
The software found buyers. In 2017, when Cyrus raised an $8 million Series A led by Trinity Ventures, it reported more than two dozen customers, including seven of the world's 20 largest pharmaceutical companies. Over time, Cyrus Bench reached more than 120 pharmaceutical firms. The company also sold support, training, customization and hands-on discovery work. Publicly named collaborators or customers include Genentech, Janssen, Nimbus, Selecta Biosciences and the Broad Institute.
That customer work mattered beyond revenue. Each assignment exposed Cyrus to the parts of protein engineering that refuse to behave like a tidy software demo: immunogenicity, solubility, stability, affinity, manufacturability and the uncomfortable distance between a computational score and a useful molecule. More than 30 discovery collaborations became a paid apprenticeship in what drug developers actually needed.
The product was software. The durable asset was judgment about proteins.
What failed first: the clean software story
There is no public record of a Cyrus clinical failure, because its disclosed internal programs remain preclinical. The first thing that clearly broke was the idea that a static software catalog could keep pace with the field. Protein AI was moving too quickly, customer datasets were too specific, and a predicted design still needed experimental proof. Levitate Bio CEO Sam DeLuca later summarized the problem plainly: static products did not fit pharma's dynamic needs.
Cyrus changed its mind in stages. First, it layered services around the software. Then it added high-throughput laboratory feedback. In December 2021, the company raised an $18 million Series B and acquired Orthogonal Biologics, a University of Illinois spinout founded by professor Erik Procko. The acquisition brought deep mutational scanning - the ability to build and measure large libraries of protein variants - under the same roof as Rosetta and AI design. The purchase price was not disclosed. The round brought Cyrus's announced funding at the time to roughly $28.9 million.
This hybrid is the company's real differentiation. AI can explore a huge sequence space and suggest structures. Rosetta contributes atomistic and statistical scoring. Mutational scans reveal which changes improve binding or stability in the real assay. Cyrus openly argues that machine learning can struggle outside its training domain and that physics still matters when closely related designs must be ranked. It is a less cinematic pitch than “AI discovers drug,” but a more credible operating system.
The molecule that must prove the pivot
Cyrus began its own therapeutics pipeline in 2021. Its lead program is CYR212, a redesigned version of IdeS, an enzyme produced by Streptococcus pyogenes. IdeS cleaves immunoglobulin G, or IgG. That makes it intriguing for diseases in which IgG autoantibodies attack the body's own tissues, including generalized myasthenia gravis and immune thrombocytopenia.
Wild-type IdeS presents two obvious problems for chronic use: it leaves the bloodstream quickly and, as a foreign bacterial protein, can provoke antibodies against itself. Cyrus says it engineered CYR212 for a longer half-life and reduced T-cell and B-cell immunogenicity while preserving enzymatic activity and manufacturability. In rabbit studies reported by the company, the candidate produced sustained, dose-dependent IgG depletion and no detected anti-drug antibody response after first or second dosing during the stated observation window.
In January 2025, Cyrus selected CYR212 as its clinical development candidate. It argues that an enzyme could lower IgG faster and at much smaller doses than FcRn-blocking antibodies, a class that reduces antibody recycling. That is the promise. The caveat is larger: rabbit immunogenicity is not human immunogenicity, projected dosing is not observed dosing, and a preclinical candidate has a long road through manufacturing, toxicology, regulatory review and trials.
The rest of the pipeline shows the same re-engineering instinct. ACE2.v2 began as a receptor decoy aimed at SARS-CoV-2 variants; peer-reviewed animal work supported the underlying engineered ACE2 approach. Cyrus now lists the asset for acute respiratory distress syndrome, where natural ACE2's short half-life has limited its usefulness. Discovery-stage cytokines include an IL-22 designed to resist its binding protein and last longer, plus an IL-15 with a slow-release modification intended to widen its oncology safety window. A human cytomegalovirus receptor decoy and partnered programs round out the list.
Where Cyrus sits in the protein gold rush
Protein design has become crowded in a way it was not when Cyrus left the Baker lab. Generate Biomedicines trains generative models to create protein drugs. Absci combines generative AI with high-throughput wet-lab systems. Isomorphic Labs attacks drug discovery with machine learning descended from DeepMind. EvolutionaryScale builds large language models for proteins. Schrödinger and Chemical Computing Group remain familiar software alternatives inside pharmaceutical research departments. Then there is the most stubborn competitor of all: a large drugmaker's internal computational biology team.
Cyrus occupies a narrower patch. It does not claim that one giant model can invent every medicine. Its public pipeline emphasizes naturally occurring proteins with known biological activity that can be made more drug-like: keep the useful function, repair liabilities such as a short half-life, off-target binding or immunogenicity, and verify the repairs experimentally. That is less open-ended than de novo generation. It can also be easier to explain to a development team because the starting biology already has a record.
The competitive test changes with each program. Against modeling vendors, ease, customization and Rosetta expertise once mattered most. Against contract engineering shops, Cyrus could offer an integrated computational and screening loop plus shared economics. Against other drug companies, only the asset profile counts. CYR212 does not win because its designers know Rosetta. It wins only if rapid and deep IgG depletion produces a clinically useful balance of efficacy, convenience, safety and repeat dosing.
Who pays, and for what?
The original buyer was a pharmaceutical scientist or research group paying for enterprise access, cloud computation, implementation and expertise. The exact price of Cyrus Bench was not publicly posted. Partnered discovery deals shift payment toward research funding, milestones and royalties. The 2021 Selecta collaboration, reportedly carrying up to $1.5 billion in potential development and sales milestones across programs, illustrates the headline-rich structure of biotech partnerships: most of that number is contingent, not cash in the bank. The acquisition price for Orthogonal Biologics was also undisclosed.
Owned drugs reverse the arrangement. Instead of customers funding defined work, investors finance years of experiments in exchange for equity and the chance of a much larger asset payoff. Cyrus's $18 million Series B supported the acquisition and independent discovery push. That is modest beside the hundreds of millions raised by some AI-biotech companies, but “capital efficient” remains a hypothesis until a program reaches the clinic. Preclinical thrift can be erased by manufacturing studies, toxicology, trial enrollment and a single disappointing readout.
In biotech, a billion-dollar partnership headline is a map of possible payments, not a bank statement.
The culture hidden in the architecture
Cyrus's culture is easiest to see in the institutions it keeps near. It grew out of an academic lab, commercialized open scientific infrastructure, co-founded the OpenFold consortium and later placed its software team under the Rosetta Commons Foundation. This is not pure open science: Cyrus still builds proprietary algorithms, datasets and drug intellectual property. It is a deliberate boundary. Shared infrastructure can improve the field; differentiated molecules can finance a company.
The leadership mixes computational specialists with people who have carried drugs toward investigational filings and negotiated biotech deals. Nivon is trained in computational biophysics. Song helped develop core Rosetta functionality. Procko brought large-scale mutagenesis and receptor engineering from Illinois. Development executive Eric Tarcha arrived with experience moving early therapeutics toward clinical work, while Mark Benjamin brought business-development history across biotech partnerships and transactions. The organizational design mirrors the platform: algorithms beside assays, discovery beside development.
There is also a useful restraint in the company's technical writing. Cyrus has argued publicly that AI models can fail outside their training data and that traditional scoring remains important. That view does not immunize the company from AI marketing, but it tells employees what kind of skepticism is permitted. The lab is not a ceremonial validation step added to a machine-learning pitch. It is where attractive predictions are allowed to lose.
A company sheds its first skin
In June 2024, Cyrus spun out the software and services team as Levitate Bio, owned by the nonprofit Rosetta Commons Foundation. Levitate inherited graphical tools, APIs and custom software work; Cyrus retained access while concentrating capital and attention on drugs. It was an unusually explicit admission that the original wedge and the eventual business belonged in different homes.
The split also clarifies Cyrus's customers. Historically, they were scientists buying tools and protein-engineering help. Today, Cyrus's commercial audience includes drug-company collaborators willing to share milestones and royalties, investors financing owned assets, and eventually clinical partners or acquirers. Its economic model has moved from recurring software and services revenue toward the lumpy, delayed value of therapeutic intellectual property.
Why the model can work
Paid partner projects create real constraints, experimental data and pattern recognition before the company risks its own pipeline capital.
Why it can fail
Good assays can still model the wrong biology. A protein can work in animals and fail on human safety, dosing, manufacturing or efficacy.
What a reader can copy
The portable lesson is not “start a drug company.” It is to sequence risk. Cyrus first solved an access problem around proven research software. It sold into demanding organizations. It used services to discover repeated pain. It added an experimental capability where customer work exposed a blind spot. Only then did it move upstream into assets it owned.
The Cyrus playbook, minus the lab coat
- Package a powerful but inconvenient technology around an existing workflow.
- Use services as paid research, but record which problems repeat.
- Build a feedback loop that measures reality, not vanity metrics.
- Acquire the missing capability when integration creates compounding learning.
- Separate the wedge from the endgame when they need different teams and capital.
This pattern does not work under every condition. It fails when service projects are too bespoke to teach a shared system, when customers will not permit learning across engagements, when measurement is slow or misleading, or when the move into owned products requires capital the core business cannot support. In therapeutics, it also fails if the biology simply refuses the design.
Cyrus now sits in a less comfortable but more consequential market. Its software history supplies credibility, its Rosetta lineage supplies methods, and its partner work supplies judgment. None supplies clinical proof. CYR212 will have to do that. The company's most interesting choice was not adopting AI. It was deciding that predictions were valuable enough to test, and not valuable enough to trust alone.