Ты зачем у меня идею из головы украл

I don't know much enough multi-modal models though, to claim that it'd work.

I do think that representing scRNAseq data in this rank-order way may be constrained. (As it gets rid of a lot of information). It would be interesting to train a multi-modal model with inputs like:
- DNA/RNA seq
- Protein seq
- scRNAseq/HiSeq data
- structural data
then we are def talking

I know wanna see whether I can replicate the our Interdependence Coefficient/Modes paper findings, e.g. if I feed it a bunch of neurons scRNAseq, will it be able to detect Protocadherins as exclusively expressing. Here goes my Sunday I guess.

How does it do with genes of naturally low expression (also what units are used for ordering of genes? some normalisation techniques will result in different orders)

Does it only support only one gene naming convention?

Does it model "on/off" gene state well enough?

It's also not always true that if exp(A) > exp(B) and A is downregulated -> B is downregulated.

Basically counts mean different things for different genes.

It's literally a text generation model, which was fine tuned on "cell sentences" -- space-separated gene names ordered by descending expression. Claims to be able to do quite a few things at inference (see the diagram)

I wonder about a few things though:

There are still more shows here in Manchester for a week or so, and then she's going for a tour.

This is exactly how Marina Abramovich's ‘Balcan Erotic Epic' felt like. At first, I tried to perceive it as an exhibition, but it seemed cheesy. Looking at it as a very abstract theatrical performance on the other hand... Loved it a lot.