MIT's IRIS model infers a cell's signaling history from its gene activity

The test crosses a species line. Training was on human embryonic stem cells, and the test was on single cells taken from mouse embryos ^[10]. A model that had only memorized its training conditions would have been caught out there.

Pulin Li, who led the study with graduate student Nicholas Hutchins, explained the cross-cell-type result as transfer learning ^[2]. "Think of voice recognition systems like Siri, which are trained mainly in English, but then use that training to help them recognize other languages," Li said. "This is called transfer learning, and this is why IRIS can work across many different cell types." ^[9] Li's analogy covers cell types, and the jump from human to mouse is a further stretch on top of it.

MIT's description calls the mouse predictions accurate ^[7]. It does not include an accuracy figure or a count of the cells tested. The model's output is an estimate that a pathway was likely on, read from overall gene activity ^[5]. Nor does the description say how often that estimate is wrong, or whether some pathways separate more cleanly than others.

Traditional methods require testing every pathway in every cell type, MIT says ^[12]. IRIS still needs its atlas first, and the atlas covers six pathways ^[6]. In our view the experimental cost moves into training, and for now the cross-cell-type result is a result about those six.

The authors say the work could speed stem cell engineering and organoid creation ^[11]. MIT says it opens up the possibility of mapping the signaling histories of every cell in a mouse or human embryo ^[12]. The paper was published open access in Nature Methods on Sept. 8 ^[8], so the accuracy question can be checked against it.