Interesting hypotheses but I dont see / missed any preprint or publication of their tools (they show a paper published by a key scientist before he joined their company).
This sounds good but the hard thing for these sweeping experiments is making a discrete, new finding or saying something interesting about an old problem. I’ll be curious to see how they approach that goal at the level of a statement that can withstand peer review
If I understand this: the issue is that mouse physiology doesn't really represent human physiology, so they have a mapping from one to the other used to predict what happens in mouse experiments to know what the effect would be on humans?
But this specifically is an AI model (like a LLM) not trained on text, but on specific medical data: "18,963-plex spatial transcriptomics" [0].
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 :)
Interesting hypotheses but I dont see / missed any preprint or publication of their tools (they show a paper published by a key scientist before he joined their company).
This sounds good but the hard thing for these sweeping experiments is making a discrete, new finding or saying something interesting about an old problem. I’ll be curious to see how they approach that goal at the level of a statement that can withstand peer review
Very fair! We're working on a preprint right now
If I understand this: the issue is that mouse physiology doesn't really represent human physiology, so they have a mapping from one to the other used to predict what happens in mouse experiments to know what the effect would be on humans?
But this specifically is an AI model (like a LLM) not trained on text, but on specific medical data: "18,963-plex spatial transcriptomics" [0].
Which is a really interesting approach!
[0] https://www.noetik.blog/p/tario-2-a-whole-transcriptome-foun...
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 :)
(I work at Noetik and wrote this article)
Not to be confused with illumina's MARS-Seq (Massively Parallel scRNA-seq)