A protein model can propose a new sequence in seconds. Biology is less cooperative.
Before anyone knows whether that sequence works, it has to become DNA, be expressed as a physical protein, survive quality control and then face an assay that measures what it actually does. In Adaptyv Bio’s current binding workflow, that journey can still take three to four weeks. The gap between machine-speed design and physical validation is becoming one of the harder constraints in AI biology.
Adaptyv has spent the past several years trying to shrink that gap. On August 25, the Lausanne company raised a $40 million Series A led by Highland Europe, with existing investors Ace Ventures, ByFounders and Y Combinator participating again. Over the previous year, Adaptyv says laboratory throughput grew by more than five times, while its customer base passed 100, spanning AI-biology companies such as Chai Discovery and Boltz and pharmaceutical groups including Roche and Novo Nordisk.
The financing is large, but the more interesting story is how the laboratory itself has been changing.
The founders started with the part AI could not do
Julian Englert and Daniel Nakhaee-Zadeh Gutierrez teamed up in early 2021 after becoming frustrated with how difficult biological experimentation remained. They entered Y Combinator that summer and later raised a pre-seed round led by Wingman Ventures. Julian Englert is now CEO and Daniel Nakhaee-Zadeh Gutierrez CTO.
Their backgrounds help explain why they approached protein AI through physical infrastructure. Before Adaptyv, Julian Englert worked in R&D engineering in Lausanne. Daniel Nakhaee-Zadeh Gutierrez’s earlier research included bioengineering work around lung-on-chip systems, biomaterials, drug delivery and immunotherapy, according to the founder-profile material reviewed for this article.
By 2025, Julian Englert was describing the company’s original conviction in unusually simple terms:
“AI models for biology are only as good as the experimental data they’re trained on and the hypotheses they can test in the real world.”
That idea has remained surprisingly consistent. Adaptyv did not begin by building another protein-generation model. It concentrated on the place where computational confidence meets physical evidence.
Its earliest public description in 2023 was a foundry that could connect protein design to experimental validation using synthetic biology, lab automation, robotics and data infrastructure. Two years later, the company opened its cloud lab more broadly after testing well over 10,000 proteins in 2025 and working with more than 30 companies that year. At the same time, it disclosed an $8 million seed round led by Ace Ventures, with ByFounders and Founderful returning and LongGame joining.
Then the wet lab started behaving more like software
A protein submitted to Adaptyv does not remain a digital sequence for long.
The platform converts that sequence into DNA instructions, physically assembles the DNA, uses cell-free synthesis to manufacture the protein and then measures characteristics such as expression, binding or stability. In its binding service, designs move from digital submission through DNA preparation, expression, kinetic measurements and final data review. Adaptyv says it now expresses thousands of proteins every month with expression success rates above 90%.
The more consequential change arrived when that laboratory became callable by software.
Adaptyv’s API allows a researcher, application or AI agent to search available targets, create an experiment, obtain pricing, track its progress and retrieve structured results. Finished experiments can return binding classifications, kinetic measurements, melting temperatures and expression values as machine-readable data. An AI system can then use those results to decide what to design next.
That changes the role of the lab. Instead of being somewhere a scientist sends an order and waits for a report, it starts to resemble a physical backend for computational biology.
Claude sent 1,320 ideas into the physical world
A recent experiment shows what that architecture looks like when the machine begins doing more of the work.
Using Adaptyv’s laboratory, 1,320 proteins designed using Claude were tested across 16 targets. Of those designs, 354 bound their intended targets, producing an average 26.8% hit rate. Adaptyv also reported that 95% expressed successfully. The laboratory received the designs anonymously and measured their binding using its standardized workflow.
Earlier in 2026, Adaptyv ran a different experiment in which 10 human teams and six autonomous AI-agent teams designed proteins against TREM2. The group produced 141 designs, with the top 100 entering wet-lab validation. Humans produced binders at 38.5%, while agents achieved 34.3%. Adaptyv’s statistical analysis found no meaningful difference in hit rates or binding-affinity distributions between the cohorts.
The result was encouraging for autonomous protein design, but it also revealed a limitation. All six agents independently converged on the same primary design tool, while human teams explored a broader mix of approaches. Agents also recorded a higher failure-to-express rate among tested designs.
That matters because autonomous science requires more than generating plausible answers. It needs enough experimental feedback to learn when its reasoning is narrow, repetitive or simply wrong.
Every failed protein may be useful data too
This is where Proteinbase makes Adaptyv’s strategy more interesting than laboratory throughput alone.
The company launched Proteinbase in October 2025 with more than 1,000 experimentally tested novel proteins, linking computational predictions to physical results and the methods used to create each design. Crucially, the dataset also includes negative results that are often missing from published research. All of the initial experimental data came through standardized Adaptyv laboratory protocols.
Proteinbase has since expanded to more than 3,000 validated proteins. Its current database connects individual proteins with design methods, biological targets and experimental measurements, allowing researchers to compare which approaches actually survive contact with the wet lab.
That creates the possibility of a compounding loop: models generate designs, Adaptyv tests them, standardized results accumulate, and those successes and failures can inform the next generation of models.
It is not yet proof of a defensible data moat. It is, however, a much harder asset to reproduce than software output alone because every useful row ultimately requires a physical experiment.
The $40 million is really a bet on closing the loop
Adaptyv now plans to double its Lausanne team and open a new London office and laboratory in Q4 2026. The company describes its longer-term ambition as building high-throughput infrastructure capable of supporting AI systems that design experiments, run them, learn from the resulting data and iterate.
There is still a considerable distance between demonstrating that an AI-designed protein binds a target and developing a medicine. Adaptyv itself makes that distinction. Its Claude experiment tested the first stages of protein function, not finished therapeutics. Later work would still need to address stability, manufacturability, cellular function, realistic biological models, safety and in-vivo evidence.
That limitation is precisely why the Series A matters. AI is making the design side of biology faster. The physical world has not accelerated at the same rate. Adaptyv is betting that the next important infrastructure layer in AI biology will belong to whoever can make experiments fast enough for machines to learn from reality, not merely predict it.