Briefing Chat: The Bunsen burner myth that turns out to be just hot air

Nick Howe (0:04)

Welcome to The Briefing Chat Podcast, the Friday show where we talk about a couple of papers that have been highlighted in The Nature Briefing, Nature's daily email roundup of the latest science news. And this week I have with me the wonderful Charmonee Bundell. Charmonee, what have you been reading about this week?

Shamini Bundell (0:21)

So Nick, Nick, you know how genetics is really complicated?

Nick Howe (0:26)

I am aware of this, yes.

Shamini Bundell (0:28)

Like really complicated. We got all these genes in our body, and some of them make proteins, and some of them influence the making of the proteins, and some of them are like, what are you doing, stretch of DNA? Why are you there? We don't know. They're always doing these different things. Well, Google Deepmind have just released an Atlas, a sort of database of predicted effects of 9 billion genetic mutations.

So 9 billion different mutations and the possible predicted effects that each of them could maybe have, with the caveat that genetics is very complicated.

Nick Howe (1:06)

I was going to say, as you mentioned, genetics really complicate. A lot of things interact with other things. Sometimes genes, you know, go in different ways. They get spliced differently. There's all sorts of ways. Genetics is really complicated. So how have they, I guess, done this with 9 billion mutations?

Shamini Bundell (1:25)

So I was reading this article in Nature, and this work is building on previous Deepmind work, which is an alpha genome AI model. And that's the thing that's actually doing the predictions. So what this alpha genome Atlas is doing is basically they've pre-computed the 9 billion answers so that researchers have it available immediately rather than having to go and ask the AI, hey, can you calculate this and that? Which can be a bit complicated, can be a bit slower.

So this AI model has basically taken loads and loads of data from publicly available databases of experiments. This is both human and mouse cell experiments. And then using what the humans have already been doing, it makes predictions based on things that haven't been done yet. And again, there's lots of sort of caveats to this, but what it essentially does is for any given DNA base pair, you know, you've got all your letters, you've got your GTCA, take one, swap it out for each of the three alternatives that it could be, and then the AI model sort of looks at all of the potential impacts of that, and then predicts what the actual outcome would be in terms of function.

Nick Howe (2:43)

But you know, I think a lot of people would be aware that AIs can sometimes be wrong. So how accurate are these predictions? Do we have any sense?

Shamini Bundell (2:51)

They're not, in a way, they're not intended to be accurate. So the idea here isn't like, oh, I wonder what this genetic change will do. I know, I'll just ask the AI.

The idea is that, and this is particularly for, sort of think about rare diseases, that's the sort of case study that they've given it so far, is, you know, very occasionally, you might get a really obscure disease that has arisen from a random mutation in one person. And sometimes that could be a single base pair mutation, although it's often not, but it could be. And what they've done in their early access trial is given some rare disease researchers access to this atlas and all the predictions that it has. And they've used it to hunt down a particular mutation that they think could be responsible for the symptoms that a particular patient was experiencing.

Nick Howe (3:47)

Okay, so the AI is like guiding them towards the right answer, I guess.

Shamini Bundell (3:51)

Yeah, so it's kind of like the AI is saying like, hey, have you thought of this? Which they hadn't, no. It's, you know, it's some obscure little letter in some obscure little gene. But then they can go off and start experimenting with that, which they did. And they looked at what it did and said, yeah, actually, I think this is probably a pathogenic mutation that we didn't know about before that can cause problems. But it was only in the brain and therefore it wasn't being detected in the blood, so they had struggled to find it.

Nick Howe (4:19)

So is this tool now out and people can use it in the labs?

Shamini Bundell (4:22)

It is. It's free for non-commercial use. If you're commercial, you're going to have to wait for the licensing conditions, which they haven't determined yet. But yeah, basically, any researcher can go and access this in a browser and query, you know, yeah, just this huge amount of information to guide you towards maybe where you should be looking for your next piece of research. So one of the things that it does is it gives you a number, which is called the AVI score, the Alpha Genome Variant Impact Score. And it's basically like a really rough summary of how, almost how important a change is.