Anna Greka

Whether you, you know, the terminology, whether you call it a virtual cell, I like the idea of sort of saying that this, to state it as a problem, you know, the way that people who think about it from a mathematics perspective, for example, would think about it.

I think stating it as the cell prediction problem appeals to me because it actually forces us biologists to think about

Setting up the way that we would do these cell perturbation data sets, the way we would generate them to set them up to serve predictions.

So, for example, you know, the way that I would think about this would be, you know, can I in the future have so much information about how cell perturbations work that I can train a model so that it can predict when I show it a picture of another cell?

under different conditions that it hasn't seen before, but it can still tell me, ah, this is a neuron in which, you know, you, you know, perturbed the mitochondria, for example, and now this is sort of the outcome that we, that you would expect to see, right?

And so to be able to have this ability to have

a model that can have the ability to predict in silico what cells would look like after a perturbation.

I think that's sort of the way that I think about this problem.

It is very far away from anything that exists today.

But I think that the beginning starts, and this is one of the unique things about my institute, if I can say, you know,

we have um you know a place where you know cell biologists uh geneticists mathematicians machine learning experts um we all come together you know in the same place to really think and grapple with these problems and of course we're very outward facing interacting with scientists all across the world as well but there's this sort of idea of bringing people into one institute where

We can just think creatively about these big aspirational problems that we want to solve.

I think this is one of the unique things about the ecosystem at the Broad Institute, which I'm proud to be a part of.

And it is this kind of out of the box thinking that will hopefully get us to generate the kinds of data sets that will serve the needs of building these kinds of models with predictive capabilities down the road.

You know, as you astutely said, you know, AlphaFold, of course, was based on the protein database existing, right?

And that was...

a wealth of available information on which one could train models that would ultimately be predictive, as we have seen this miracle that Demis Hassabis and John Jumper have given to humanity, if you will.

But, you know, as Demis and John will also, you know, would also say, I believe, is, as I have discussed with them, in fact, the cell prediction problem is really

a bigger problem because we do not have a protein data bank to go to right now but we need to create it to generate these data and so my ladder secures accelerator is here to basically provide some part of the answer to that problem create this kind of well-controlled database that we need for cell perturbations while at the same time maximizing our learnings

about these fully penetrant coding mutations and what their downstream sequelae would be in many different human cells.