Four leading blastoid methods match real human embryos unevenly on a 14,000-cell reference map

Blastoids are wanted because human embryos are hard to study. They are grown from stem cells. That offers a route into questions about fertility and pregnancy success that technical and ethical constraints have long blocked, without relying on donated embryos ^[9]. Their value depends on how closely they behave like the real thing. "Human embryo models have enormous potential for studying the earliest days of an embryo's development, but there has been no consistent way to assess how accurately these reflect real human development," said Pengyi Yang, an associate professor in the University of Sydney's School of Mathematics and Statistics ^[10][12].

The paper, in Cell Systems, starts by building the yardstick ^[1]. The team pooled and harmonised more than 14,000 single-cell transcriptomes, the RNA each cell is making, from human embryos in the days just before and after implantation. They arranged them as a map of how cells normally differentiate ^[3]. Each blastoid cell is then placed on that map. It either sits close to a state that real embryonic cells occupy or it cannot be confidently matched to any ^[5].

The differences between protocols were substantial, and no model perfectly replicated a natural blastocyst ^[2][6]. The unmatched cells are the result I would most want a number for. A protocol can produce something that looks like a blastocyst while a share of its cells correspond to no known embryonic state, and any measurement pooled across the whole structure then mixes the two ^[4][5]. The university's account does not name which protocol fell where, or give the unmatched share for each.

The thing this benchmark doesn't tell you is whether a model that matches on RNA also behaves like an embryo. The reference is built from transcriptomes ^[3]. A high score is therefore evidence that a blastoid's cells express the right programs at about the right time. Whether the same model does what an embryo does around implantation is a separate experiment. The map is also bounded by the embryos that went into it, although the university describes it as one of the most comprehensive reference maps of early human development ^[15].

Yang put the result both ways. "The encouraging finding is that some models capture important aspects of early embryonic development relatively well, although each model has limitations," he said ^[13]. "But our study also shows that current models are not biologically equivalent to real human embryos, and researchers need to be careful about the conclusions they draw from them" ^[11].

In my view, a blastoid finding about any one lineage is provisional until that model's cells for the lineage have been run through this comparison. That holds where the reference samples the relevant lineage and stage well. Where its coverage is thin, a good match means less.