Founder Stories: Eli Bogart, Deep Genomics and AI for RNA Therapeutics

Pleasure to be here representing Deep Genomics today. I know we're coming into the coffee break, so I'll be brief. I want to take the next 5 minutes to give an introduction very briefly to the company and then sort of an advertisement for some of the machine learning work that we've recently been doing that I think will be of interest to this community. So, as a quick introduction, we are a drug discovery startup. Our mission is to revolutionize drug development by applying artificial intelligence methods to RNA biology. Uh, we have a fantastic interdisciplinary research team that combines strengths in computational science, especially machine learning, uh, with experimental biology. Uh, we believe that our approach is applicable to a wide range of diseases and different therapeutic modalities. Uh, our main focus in the past couple of years has been developing oligonucleotides, uh, designed to increase the expression of particular target genes to treat genetic diseases. So what I want to highlight here is a foundation model for RNA biology that a large group of my colleagues published a preprint on back in September. This is a model called BidRNA. So BidRNA takes in a wide one-hot encoded DNA sequence as input and fits a number of different genomic data tracks at 128 base pair resolution as output. In this audience, I don't— for this audience, I don't need to spend a lot of time walking through that setup. So I'll highlight some of the distinctive features of this model. We've done a lot of work to incorporate individual whole genome sequencing data on the input side and individual RNA-Seq datasets from the GTEx database on the output side. So we have data from 70 individuals across up to 51 tissues for each individual that go into the training set, leading to 2,956 individual RNA-seq output tracts. We supplement that with another 700 tracts capturing RNA binding, protein binding data, as well as microRNA binding. So we subscribe to this school of thought that says that the key thing about a foundation model is that it can be applied to a range of downstream tasks. And if you take a look at the preprint, we go through multiple tests in a dozen different application domains, but I want to just briefly touch on two that I think will be of interest given discussions yesterday and today. So the first is that we find that the model does very well predicting the effects of variants in non-coding regions. So I've highlighted here a case study where we looked at a known pathogenic variant in the gene NAA10. This is a mutation in the 3' untranslated region of the gene, which is known to decrease expression by disrupting a polyadenylation site. And this comes out very clearly in the model's predictions. I've shown in blue the predicted RNA-seq coverage track for the wild type sequence, in red the same for the variant sequence. And you can see the decreased expression looking at the reduced coverage in the exons, as well as this extension of the high coverage region of the 3' UTR due to the disruption of poly-A site. As you'd hope based off of that anecdote, when we systematically evaluate the model's performance on discriminating between pathogenic and benign variants in 5' and 3' untranslated regions in the ClinGen database, we get quite good classifiers which match or exceed the results from alternative approaches. Finally, coming back around to drug development, what we found is that the model is very good at predicting the effects of steric blocking oligonucleotides in a zero-shot framework. So for those not familiar with the approach, a sterate-blocking oligonucleotide basically is a way of saying, all right, I have a functional region or suspected functional region in RNA transcripts. I'm going to synthesize a complementary molecule to bind to it and prevent it from interacting with its natural interaction partners and suppress its function. The model has never seen data from cells treated with oligonucleotides, but what we find is that you can very reliably get good results by taking the binding region of your oligonucleotide and just zeroing out the one-hot encoding of the sequence. So for instance, you can use these molecules to modulate splicing, for instance, skipping exons. And when we go through and compare bidRNA's predictions to internal experimental data from 620 oligonucleotides skipping a range of exons, We generally see quite good correlations between the model's predictions and the exon skipping effects across those oligos at a consistent dose. We can compare this to doing a similar prediction approach using SpliceAI, which is already a pretty sophisticated splicing model. And we see that in this case, going the extra mile or many miles to build a foundation model has gotten us improved performance on this task, is really relevant to what we want to do. I'll wrap up there. I'm very happy that the company can support this community. Great to be here as a sponsor. If you have questions about this work or deep genomics, I'm happy to talk. I'll be around all day today. My colleague Andrew is in the back and can also take questions. We are currently hiring for a director of machine learning. If you know anyone who might be interested, that posting is on our website.

Thanks, Eli. So with that, that's the last talk before the break. So let's take a 10-minute break for coffee and then come back and we'll have another 3 talks.

Yeah.

Okay.

Just logging this soon and share it.

Yeah, we should, you should probably check that right now. Are you the next? I'm the one, the second one.

Okay.

Yeah, I would check it now because I'm not sure if you have the permissions to share. I think that's just email that I can log in. Oh, okay.

I'm gonna try without preview.

Can you just click on share and just see if it lets you? Okay, good.

Yeah. Okay.

Yeah, so when you, um, basically when you bring your laptop up, we have to put these two in and then I have to switch a bunch of settings on this.

Okay. But it should be okay. So, uh, do we need to open the audio or something else?

Yeah, we'll have to do that when we connect in. I see.

So I also need to join the audio.

Yeah.

Okay.

So do you, do you pronounce your name Gang or Hong?

Gang. Gang. Okay.

Thank you. Ah, this one's not.

I think that one's dead. What are we— have batteries? Oh yeah, there's— I don't know if those are functioning. Yeah, so, okay, the two things I mean, why is this difficult? Start with everybody. The other one's like, how old is that community? How old is that community? Kind of reflecting stuff, like philosophy, because I know there's everything on the FBI website. Yeah, we did that. So we tried that, like you said, now with more. So we did a soft launch. I think something that you could tell the game— it's really good at these tasks, but I think it's also caught on by the fact that like people name them because they are something like that. Like, you know, like CD8 and CD24, embedded into any large web because they share the structure that is the seeding family. The developer fires seeding template with the widget of the content. Or they have that like, like, example, all this really pretty. And then I just know that the chunk should be— what's the phrase? It's not about you, it's about who gets it first. But it's about what The bad thing is it all makes it up. It wants to give you an answer, so it will give you more prompts. I mean, nobody's gonna be able to do that, right? Like, there's just no skip pattern.