Somewhat, we are assuming that a model trained on human data entirely is able to 'project' mouse data into a human transcriptomic space. It feels like something that should obviously fail (isn't it out of distribution?), but it works surprisingly well according to the perturbation controls we had! Morphology of tissue may simply be a rather universal substrate.
And yes, it is trained on 18,963-plex spatial transcriptomics :)
Thanks for posting this here! And surprising its attracting attention, I have to imagine that the TAM for 2-hour-long biologics-manufacturing podcasts is small :)
If you're more interested in this person's work, his website is here: https://www.iku.bio/
I completely agree, but I also think there is some truth to the related statement: 'cancer research often isn't conducted in a way that is actually useful'!
For example, in-vivo tumor experiments in mice can yield completely different results depending on exactly where the tumor was implanted. E.g. a 'lung cancer mouse model' may have the lung cancer injected just under the skin, also known as subcutaneous tumor models, instead of in the lung! Entirely because it's a lot more efficient + yields more trustable data, but the results are often deeply disconnected from how the tumor would naturally grow + respond to drugs within its host organ.
I think it has very limited therapeutic applications with what we know about RNA structure today! But there's a great deal of completely unknown RNA biology (some of which I touch on in the essay) that may greatly benefit from RNA structure. The bit I mention about Arrakis Therapeutics preclinical work in drugging the (structured) RNA version of the MYC protein points to that being a very real possibility. All interesting biotech startups are built on bets on where the future is going, and I'm very happy that someone (AtomicAI and others) is betting on this, because clearly the answer of 'is RNA structure useful' isn't super open-and-shut