DARPA's biotechnology director acknowledges that "virtual cell models" and de novo protein design do not work yet
DARPA's biotechnology director acknowledges that "virtual cell models" and de novo protein design do not work yet
Mike Koeris heads DARPA's Biological Technologies Office (BTO), the arm of the U.S. defense agency that allocates a share of its multibillion-dollar biotechnology budget. At SynBioBeta 2026 panels in May he openly criticized two of the most heavily promoted AI directions in biology. On September 28, conference founder John Cumbers published exact quotes, and the following day Koeris confirmed them himself on LinkedIn.
Koeris has led BTO since April 2024. Before that, he trained in the lab of Jim Collins, one of the founders of synthetic biology (who built the first genetic toggle switch in E. coli in 2000), and founded and sold two biotech startups.
At the panels, Koeris spoke without the usual caution of a government official:
Virtual cell models don't work right now. That'll be a surprise, yeah. De novo protein design doesn't work either.
Both statements target the two most heavily promoted AI applications in biology. A virtual cell is a computational model meant to predict a cell's response to a drug without a laboratory experiment; this concept is championed by Mark Zuckerberg's Chan Zuckerberg Initiative and the biomedical Arc Institute, with support from Nvidia. De novo protein design means creating protein structures that do not exist in nature, with a specified function; David Baker received half of the 2024 Nobel Prize in Chemistry for a method of such design. The same generative approach recently failed against tau fibrils in Alzheimer's disease, although it had worked on other targets. According to Koeris, current algorithms do not build protein structure from true scratch: they take an existing natural scaffold, adjust it slightly, and call that design. BTO itself officially lists whole-cell modeling among its goals, which means Koeris is criticizing a direction his own office funds.
Alongside these two, Koeris raised a third challenge from the same field and assessed it differently. He called protein sequencing (deciphering an unknown amino acid sequence without a reference for comparison) a technology that will be fully solved within ten years and will cost as little as DNA sequencing. The difference lies in the nature of the task: sequencing reads a sequence already present in nature, while de novo design must produce a structure that has never existed. Behind his optimism stands BTO's own PROSE program, launched in 2026 with a measurable target of reading proteins of 300 amino acids or longer at above 99% per-residue accuracy.
This distinction rests on a philosophy of risk that Koeris laid out in 2024:
About half the projects we fund may fail, but that gives us enormous learning. And the other half is a resounding success: GPS, the internet, mRNA vaccines. I don't think that's a bad trade.
DARPA landed in that successful half in 2011, when it funded early mRNA vaccine technology through the ADEPT program; Moderna, then a little-known startup, took part in that work.
When Cumbers published his post, Koeris responded under it on LinkedIn without softening his tone:
I stand by these quotes. Onward. We do this because it's hard. We do this for the mission!
Confident promises about AI in biology typically come from people selling these technologies to investors. Koeris controls part of DARPA's budget and decides which efforts receive funding. He confirmed his words the following day, without a single qualifier.