This paper introduces scPrediXcan, which combines a deep-learning model (ctPred) built on Enformer-derived features with single-cell RNA-seq to perform cell-type-specific TWAS via a linearized SNP predictor (ℓ-ctPred), improving gene discovery for T2D and SLE.
0:00Welcome to Base by Base, the paper cast that brings genomics to you wherever you are. Thanks for listening, and don't forget to follow and rate us in your podcast app. Imagine for a second. that you are tasked with finding a single malfunctioning instrument in a massive symphony orchestra.
0:17Oh, that sounds like a nightmare. Right. But there's just one catch. You're entirely locked outside the concert hall. Oh, wow. Yeah. So you can hear the timpani booming and, you know, the violins soaring through these thick brick walls, but your boss tells you, hey, one of the 100 instruments in there is playing a B flat instead of a B.
0:36Go fix it. Good luck with that. Exactly. I mean, you press your ear to the cold brick. You might be able to tell that something is generally wrong with the brass section. But trying to pinpoint whether it is the 3rd trumpet or like the 2nd French horn, it's physically impossible.
0:52The music all just blends together into one loud, overlapping soundscape by the time it reaches you. That is, honestly, a remarkably accurate way to describe the exact frustrating problem that geneticists have faced for years.
1:05I mean, when studying complex diseases, we have basically been stuck outside the concert hall. When we look at genetics and human health, historically, we've had to look at whole organs. Which means, uh, we are forced to listen to the entire orchestra at once.
1:21Right, and that brings us to the core question of this deep dive. What happens when we can finally bypass those thick walls? What if we could zoom right in and listen to the specific individual cellular instrument that is playing out of tune and causing a disease?
1:35It's the holy grail, really. It is. And, you know, how does artificial intelligence finally give us the key to those concert hall doors? Well, it completely reframes our understanding of human biology.
1:46Because for a long time, the tools we had to map disease genetics were they were inherently limited by that outside the hall perspective. Because of the noise. Exactly. If you were averaging out the noise of 1000000s of different cells, you are also averaging out the critical signals from the few cells that are actually sick.
2:03Today, we celebrate the work of Yakazu, Manji Chen, Hai Kyung Yim, and their collaborative teams at the University of Chicago, Harvard Medical School, UNSW, and Argonne National Laboratory, who have advanced our understanding of cellular level genetics.
2:19A powerhouse team. Right off the bat, looking at that list of institutions, I mean, I have to ask, why did it require such a massive multi-institutional effort to tackle this? We map genes all the time.
2:31Yeah, we do, but mapping disease at the cellular level requires fusing entirely different scientific universe. Oh, so? Well, you need cutting edge deep learning, massive supercomputing infrastructure, and highly advanced single cell biology.
2:44Okay yeah. There's a combination that's simply no single discipline, and really no single laboratory, could achieve entirely on its own. They had to build a functional bridge between massive computational power and, you know, microscopic biological reality.
2:58To understand the bridge they built, though, we 1st need to understand the tool they realized was fundamentally broken. In the genetics world, there is a workhorse method called TWS. That stands for transcript on wide association study.
3:12Yes, a very classic tool. And at a basic level, TWS helps researchers find candidate genes that cause diseases by linking predicted gene expression to physical traits. Precisely. You look at variations in a person's DNA, you predict how those variations affect the way certain genes are turned on or off.
3:30Their expression. Rather expression. And then you see if those changes associate with a trait or disease. But tell US has a massive blind spot. It is what geneticists call the bulk problem. Yes. Canonical T West, the traditional way we have been doing this for years, relies on bulk tissue.
3:48To get enough statistical power to make confident predictions, researchers have to use whole tissues. Like a whole sample of blood. Exactly, or an entire chunk of liver or pancreas. Grinding up the whole orchestra into a single smoothie.
4:01That's exactly it. You are taking 1000000s of cells of all different types, mashing them together and looking at the average gene expression. Which dilutes everything. Exactly. The problem is that many complex diseases, like type 2 diabetes or lupus, they do not start everywhere at once.
4:17They often originate in highly specific, sometimes very rare cell types. In a bulk sample, the loud, common cells completely drown out the faint critical signals from those rare cells. Okay, I have to push back on this premise, though.
4:31Go for it. If you read any biology news today, single cell RNA sequencing is everywhere. It's a huge buzzword. We literally have the technology right now to look at individual cells. We do. So why could a scientist just take that single cell data, which already exists and plug it directly into these traditional TW frameworks?
4:49Well, that is the logical next step, but it hits a brick wall in reality. It comes down to the hard truth of single cell data. It is incredibly sparse. Sparse, meaning like the data is physically missing.
5:00Yes. Think back to your orchestra. Imagine putting a tiny microphone on that one specific trumpet. Okay. But the trumpet is so fragile, that the very act of micing it up causes the musician to drop 90% of their notes.
5:15Oh, wow. That is single cell sequencing. The RNA molecules inside a single cell are incredibly delicate. We literally lose most of the data as we try to capture it. it just breaks apart. Right. The molecules degrade or they fail to amplify during the sequencing process.
5:31So you get a highly pixelated incomplete picture of what that cell is actually doing. Exactly. Furthermore, to run a traditional T-wise, you need genetic data from 100s, if not 1000s of individuals. To prove the link.
5:45Exactly. To prove that a genetic variation actually causes a disease across a population. This is often done using GWYs, or genome wide association studies. And those pull from databases of what, half a 1000000 people?
5:57Easily. And currently gathering high quality single cell data for even a few 100 people is astronomically expensive and computationally exhausting. So doing it for half a 1000000 is just impossible. You simply cannot do it for half a 1000000 people.
6:11not right now So if we can't easily collect single cell data for 100s of 1000s of people, this research team realized they had to do something entirely different. They decided to predict it using artificial intelligence.
6:23That is the core innovation here. If you cannot measure the single cell expression directly across huge populations, you use AI to simulate it. And they created a framework called slick predictskin. The mechanics behind it are just fascinating to me.
6:37They started by building a deep learning model called SkipPred, which stands for cell type specific prediction. But they didn't build it from scratch, did they? They layered it on top of an existing, very powerful pre-trained AI called Informer.
6:49They did? How does informer actually read human DNA? Well, informer is a sequence to epigenomics model. It looks at raw DNA and predicts the epigenomic features. Okay. It takes in a massive chunk of DNA, a sequence of exactly 1906,608 base pairs centered right where a specific gene starts.
7:08That is a huge window. It is. And from that huge sequence in former extracts, 5313 specific epigenomic features. What do those 5,313 features actually represent in physical reality? Because it sounds a bit abstract.
7:23Sure, they represent the complex regulatory grammar of your DNA. Regulatory grammar. Right. It is predicting physical chemical tags, like methylation, which act as stopper go signs for a gene. Ah, okay.
7:36It is predicting the 3D folding of the DNA strand, figuring out if a piece of DNA loops back on itself to turn a gene on. Informer reads that long sequence and outputs a map of all those physical and chemical instructions.
7:48So the researchers took those 5,313 regulatory features extracted by informer and fed them into their new model, CTPread. Exactly. And seekpred's job is to take those general instructions and predict how that gene will actually behave inside a highly specific cell type.
8:04That's the leap they made, yes. But wait, if it is predicting something this complex, isn't this AI model incredibly heavy? I have seen rival AI models in biology that require massive server farms just to run a single test.
8:16You'd think so, but that is what makes CT bread so impressive. It is astonishingly lightweight. It operates on only .4 million parameters. 0.4 million. Yeah. To put that in perspective, there is a rival model trying to do something similar called M to sell, and that model relies on over 60000000 parameters.
8:36Oh, wow. Yet the tiny .40000 parameter CDPread actually outperforms it. But even with a lightweight model, they still hit a computational bottleneck, right? Because, like we said earlier, they need to run this for entire population.
8:48A massive bottleneck. Even using the highly efficient CT bread, running deep learning predictions for every single gene across 100s of 1000s of people in a genome wide database is practically computationally impossible.
9:02The numbers they reported highlight just how impossible it is, just to run this deep learning model for 500 people. It would take around 2700 GPU hours. It staggering. If you were trying to study half a 1000000 people to find a reliable genetic link for diabetes, you would need supercomputers running for lifetimes.
9:20absolutely cannot scale it. So they had to translate this complex deep learning model into something much, much faster. They performed a mathematical process called lineearization, creating a simplified model they named LCTPred.
9:33The L's stands for linear. This makes me think of like creating a cheat sheet. Or compressing a massive 4K video file into a tiny ZIP file so you can actually email it. I like that. You take this heavy complex AI.
9:47You run it intensely on a highly controlled computer simulated reference panel in this case, using genetic data from 462 European individuals. And you extract a simplified set of linear rules. That cheat sheet analogy is perfect.
10:01Instead of running the massive AI every single time for every single person, scientists can now just take basic genotype data, check the L for T-Pread cheat sheet and get the prediction instantly. And that cheat sheet unlocks the final piece of their framework.
10:15You take this extremely fast linear model and plug it straight into standard summary statistics from huge already existing genetic databases. Exactly. You don't even need raw individual level DNA data from half a 1000000 people, which is incredibly difficult to access due to medical privacy laws anyway.
10:31Which is a huge hurdle in genetics. You just use the broad summary statistics, apply the linear rules, and perform an ultra-high resolution, cell type level TWS. You have essentially bypass the thick brick walls of the concert hall using math and simulation.
10:48So a brilliant theoretical tool is great, but it is completely useless unless it actually uncovers hidden biology. When they pointed this tool at real world data, what did the map reveal? The results were striking.
11:01They unleashed Speed Redexkin on 2 incredibly complex stubborn diseases, type 2 diabetes, or Q2D, and systemic lupus erythhematosis SLE. Let's look at the diabetes findings first, where they focused on the eyelet cells inside the pancreas.
11:15The contrast between the old way and the new way is stark here. When researchers previously used the bulk method on the pancreas looking at the tissue as a single smoothie. It found 111 candidate genes across 64 genomic blocks.
11:27s. A genomic block is just a specific neighborhood on a chromosome. Okay, 111 genes. But when they applied ski products can to look at the individual islet cells, it found 222 candidate causal genes across 108 genomic blocks.
11:39Wow, it practically doubled the discovery rate. And it did not just find more genes. It found them in incredibly specific microscopic locations that the bulk method was totally blind to. Exactly. For instance, there is a gene called C-A-S-R, which is heavily linked to obesity induced type 2 diabetes.
11:59ScareProdexkin found that this gene was only driving disease in gamma cells. Which are rare cells in the pancreas that regulate certain hormones. Right. And another gene, MSRA, was only found in activated stellet cells, which are cells that respond to tissue injury and inflammation.
12:15Those specific cellular drivers are completely masked when you grind up the pancreas. The gamma cells and the cell cell cells make up such a tiny fraction of the organ that their signals just vanish. They just get drowned out.
12:26We see the exact same phenomena when they applied the framework to lupus, focusing on the immune system. The lupus results were just as dramatic. The old bulk method found 54 candidate genes. ScarePredict skin found 129.
12:39And again, it is the specificity that matters for medical research. Ski products can found disease causing genes on entirely different chromosomes than the bulk method did, which is wild. It identified a gene called PY Cardi specifically operating inside T cells, which are the immune systems targeted attackers.
13:00Okay. It found the ITGM genes specifically inside monocytes, the 1st responder white blood cells that rush to an infection. One discovery that really stood out to me was a gene called CFB, the bulk whole blood method missed it entirely.
13:12It completely missed it. But ski predicts can found it operating specifically in T cells. Why does identifying that specific gene in that specific cell matter for someone suffering from lupus? Well, CFB is a crucial component of the alternate complement pathway.
13:27That is a system involved in clearing out dead cellular debris from the body. If that clearance process is impaired by a faulty gene, the body's auto antibodies start recognizing that leftover cellular debris as a foreign threat.
13:41Oh, wow. Yeah, that initiates an autoimmune response, leading directly to the aggressive activation of T cells, which is a hallmark of lupus. So missing the CFB gene means missing a fundamental piece of the disease's entire mechanical origin.
13:55Exactly. Okay, I have to play the skeptic for a 2nd because this sounds almost too perfect. Fair enough. When you switch to a more sensitive mathematical tool, and you suddenly find double the number of candidate genes.
14:06How do we know we aren't just capturing statistical noise? It's great question. Like, how do we know these are actually the right genes driving the disease and not just false positives generated by the AI?
14:17That is the most critical question you can ask about any new computational model? To prove it was not just throwing out random noise? The researchers tested, scare a ketic skin against a silver standard list.
14:30What is that? This is a list of highly validated laboratory proven type 2 diabetes genes. If the AI was just guessing, its hit rate on that specific known list would be poor. Makes sense. Instead, scary ketics can show vastly superior recall.
14:46It's successfully identified 24 of those silver standard genes. And what did the old method get? The old bulk method only caught 13? That proved definitively that the new discoveries are highly relevant disease drivers, not math errors.
15:01So, we have this ultra high resolution map now. We can see the specific trumpet playing out of tune, but we have to talk about the blind spots, because every new technology has them. Of course. Let's look at the limitations of ski product skin as it stands right now.
15:16The primary limitation actually stems from the underlying informer model. Informer is brilliant at predicting that a gene is associated with the disease, but it struggles with what we call directionality.
15:28Meaning it knows the gene is involved in the crime, but it doesn't know if the gene is producing way too much of a protein or entirely failing to produce enough of it. Exactly. It can tell you there's a strong correlation, but it sometimes predicts negative correlations inaccurately.
15:43Or four, ski prediction can tell us a gene is highly associated with lupus in a specific cell, but it cannot reliably tell us if that gene's expression is turned up or turned down. That's tricky. It is.
15:55Additionally, the model misses distal enhancers. These are distant genetic switches that regulate genes from very far away on the DNA strand. Informers design just does not capture those long distance relationships well.
16:07And crucially, there is an ancestry limitation. Yes, there is. The in Silico reference panel was built using data from 462 European individuals because it was trained primarily on European ancestry data.
16:21That is a massive historical bias in genetic databases that the scientific community is still working to overcome. The ancestry limitation is significant, as it means the predictions might not be as accurate for non-European populations.
16:35However, the researchers have made their pipelines and methods publicly accessible so that ancestry specific reference data can be generated by other teams in the future. That's huge step. But I want to go back to that 1st limitation about directionality.
16:48If you have a family history of lupus and scientists are trying to build a drug to help you. Don't they need to know whether to design a drug that boosts a gene or suppresses it? They do. eventually. So if this AI model cannot even tell us whether a gene is turned up or turned down.
17:02How is this actually useful for creating real-world drug targets? You have to remember the overwhelming scale of the human genome. Before this tool, researchers were looking for a needle in a haystack.
17:15They knew a disease was linked to the pancreas, but they had no idea which of the 1000000000s of cells to target or which of the 1000s of genes within those cells was responsible. It was just too big. What Skipper Dixon does is narrow that search space exponentially.
17:30It takes a search for a needle in a haystack and turns it into a search for a needle in a teacup. It hands the laboratory scientists a highly specific treasure map. Precisely. It tells a physical lab. Stop looking at the whole pancreas.
17:43Look exactly at this rare gamma cell and look exactly at the CASR gene. Once the lab knows exactly where to look, They can easily do the physical experiments under a microscope to figure out the directionality, whether it is up or down.
17:55It saves years, maybe decades, of blind searching. It is a profound shift in how we navigate biology. Super predict skin bridges the massive gap between huge genetic databases and high-resolution cellular biology by using deep learning to simulate single cell expression.
18:13It really is a game changer. By predicting how our DNA acts inside specific rare cells, it uncovers hidden disease drivers that traditional whole tissue studies completely miss. We are finally stepping inside the concert hall.
18:26And we are finally hearing the individual instruments clearly. without the noise of the crowd. It changes everything about how we understand the origins of complex traits. What does this mean for the future of personalized medicine when a computer can predict your unique cellular vulnerabilities using nothing more than your raw DNA sequence?
18:42This episode was based on an open access article under the CCBY 4.0 license. You can find a direct link to the paper and the license in our episode description. If you enjoyed this, follow or subscribe in your podcast app and leave a 5 star rating.
18:57If you'd like to support our work, use the donation link in the description. Now stay with us for an original track created especially for this episode and inspired by the article you've just heard about.
19:06Thanks for listening and join us next time as we explore more science based by base.
