This study uses the Speos graph machine-learning framework on multi-modal molecular networks to identify core genes for complex traits, focusing on ulcerative colitis (UC). It shows tissue-specific core-gene expression, coordinated regulation of core genes after perturbation, and frequent non-linear interactions in co-perturbations.
0:15Midnight in the lab, the screens are cold and bright. Welcome to Base by Base, the papercast that brings genomics to you wherever you are. Thanks for listening, and don't forget to follow and rate us in your podcast app.
0:27So what really happens when a disease isn't caused by a single, you know, broken gene, but by this quiet conspiracy of 1000s of seemingly unrelated genes scattered across your entire DNA. Yeah, that is the $1000000 question in genetics right now.
0:43Right. And how could this change your understanding of diseases that run in families, but somehow defy simple genetic tests? Think of your genome like this massive corporate hierarchy? Okay, I like that analogy.
0:56So what if the actual executives, like the genes directly causing the disease, are totally hidden behind closed doors? And all we can see from the outside are the 1000s of middle managers constantly sending them emails, shifting paper around, making all the noise.
1:11Well, the real decisions are being made in completely secret meetings, which is, honestly, a frustrating reality for geneticists, we spend an enormous amount of time and resources tracking those noisy middle managers.
1:22Because they're the ones we can actually see. Exactly. We can see them working overtime, but because we can't see who they are reporting to, we don't actually understand how the catastrophic cellular decisions are being finalized.
1:33And unmasking those hidden executives is the core mission of our deep dive today. We are exploring this groundbreaking research that uses artificial intelligence to map out the hidden power structures inside ourselves.
1:47It kind of challenges everything we thought we knew about how genetic diseases actually work. It really does. It's a complete paradigm shift. Yeah. So today we celebrate the work of the research team at the Helmholtz Center, Munich and Ludwig Maximilian's Universe of it mentioned, who have fundamentally advanced her understanding of the genetic architecture of complex traits.
2:04And to really, um, appreciate how massive this advancement is we have to look at the brick wall genetics hit over the last decade. We've relied heavily on genome wide association studies, which are widely known as GWA.
2:17Right, GWAs. It's everywhere in the literature. It is. And the basic premise is incredibly powerful. You basically scan the DNA of 10s of 1000s of people, compare the sick to the healthy, and identify the tiny genetic variations.
2:33Those are the single nucleotide polymorphisms, right? The SMPs. Exactly, the S&Ps. You look for the ones that show up more frequently in people with a specific disease. I mean, it sounds like a perfect system for finding the culprit.
2:45On paper, yes. And GUS has successfully identified 1000s of these genetic variants linked to various diseases. But it revealed this glaring flaw, something geneticists call the missing heritability problem.
2:59Right, because when you add up all those 1000s of variants, G-Lewis finds, they don't actually account for the full genetic risk of the disease. Far from it, actually. For a lot of complex traits, those 1000s of associated genetic variants cumulatively explain only about 30% of the disease's heritability.
3:15Wait, 30%? That's it. That's it. That means 70% of the genetic influence on a trait remains completely unexplained. We know from family histories that the disease is in the genes, but the math just doesn't add up.
3:30We can't figure out the mechanistic link between the genetic code and the actual biological triggering of the disease. So we have a massive pile of puzzle pieces, but literally no picture on the box to show us how they connect.
3:43That's a great way to put it. Which brings us to a concept that attempts to draw that picture, which is the omnogenic model. And the researchers in this deep dive use this model as their foundation. Yeah, the omnogenic model offers a radical way to look at the genome.
3:57Instead of looking for, you know, one or two broken genes, it divides the genome into two functional camps. First, you have the core genes. Going back to the corporate hierarchy, these are our executives.
4:07Right, the executives, they are deeply tissue specific, and they have a direct, immediate biological effect on the disease trait, and then you have the peripheral genes, the vast army of middle managers.
4:18The ones making all the noise? Exactly. These peripheral genes don't directly cause the disease. Their job is simply to modulate, tweak, and regulate the expression of the core genes. Okay, but let me ask you this.
4:31On behalf of everyone trying to wrap their heads around this, if the peripheral genes are just background noise, just regulators, why do they hold that missing 70% of the heritability? Are they acting like a 1000000 tiny drops of water that eventually burst a dam?
4:47That is the central hypothesis of the omnogenic model right there. Individually, a peripheral gene has an infinitesimally small effect, like almost unnoticeable. Just a single drop. Right. But collectively, 1000s of them funnel tiny regulatory changes through the cells network, eventually converging on a core gene and completely overwhelming it.
5:06The problem is, while the omnigenic model is a highly plausible theory, it has remained mostly theoretical. Because we didn't have the tools. Exactly. Until this paper, we didn't actually know which genes were the core executives for most complex diseases.
5:21We also didn't know the specific biochemical paths. The corporate email chains, basically, that the peripheral genes were using to transmit their influence. Well, because tracing 1000000s of interacting variables across human biology is just way beyond human calculation.
5:37which forced the research team to build a completely new kind of tool. They developed a graph machine learning framework called Spios. Which is an incredibly elegant solution. To find the core genes, they couldn't just give the AI a linear list of genes, they had to represent the cell mathematically.
5:55So they built this massive multimodal biological graph involving over 17,000 genes in their corresponding proteins. Okay, let's break down what a graph actually means in this context, because we aren't talking about a bar chart.
6:07We were talking about a network of nodes and edges, right? What actually constitutes an edge in this biological model. Think of the genes and proteins as the nodes, like cities on a map. The edges are the highways connecting them.
6:19But in SpEOs, they layered multiple different types of highways. Oh, so it's not just a flat map. Not at all. They napped out physical protein, protein interactions, literally, which proteins physically bind to each other?
6:30They layered gene regulatory networks showing which transcription factors turn specific genes on or off. They added metabolic pathways detailing how cells share resources and energy. Wow. And to top it all off, they fed the model gene expression data from 72 different human tissues and cell types.
6:48Wait, 72 tissues? That is a staggering amount of context. So it's not just, you know, Gene A is near Gene B. It's more like genies, protein, binds to Gene B's protein, but only when they are in liver tissue and only when a specific metabolic resource is present.
7:04Exactly. That multidimensional context is the secret sauce of Spios. It allows the AI to see the structural hierarchy of the cell. Similar to how early search engines used algorithms like page rank to figure out which websites were the most important hubs based on how many links pointed to them, Spios looks for the topological hubs in the biological network.
7:23But wait, I have to push back here. If we don't know what a core gene looks like for a complex disease, How do we train the AI to find one? You can't just tell it to look for hubs without defining what a bad hub looks like.
7:36You give the algorithm a cheat sheet. The researchers train the Space model using known Mendelian disorder genes. Mandelian disorders. Those are the ones where a single severe mutation in one gene practically guarantees the disease, right?
7:49Like cystic fibrosis or Huntington's. Spot on, the scientific community universally agrees that these are undeniable core genes. So by feeding these known absolute core genes into the network, they taught the AI to recognize the specific mathematical and biochemical neighborhood that a central executive lives in.
8:08I really have to pause here though. Because whenever we talk about feeding training data into an AI to find hitting patterns, there is a massive red flag. Hallucinations. Exactly. We know machine learning models are notorious for hallucinating.
8:22They will find patterns and static on a TV screen if you tell them to. If we train this model on severe single gene diseases, isn't it just going to hallucinate connections to force complex diseases to look the exact same way?
8:36How do we know this math translates to wet, messy human biology? It's a very valid concern, and the researchers were acutely aware of that exact risk. They knew they couldn't just publish a mathematical prediction.
8:48They had to prove it in living systems. To validate SpEOs. They took the AI's top predictions for core genes and various diseases and cross-referenced them with actual mouse knockout studies. Meaning they physically remove those specific genes from living mice to see what would happen.
9:03Correct. And the results were striping. When you knock out the specific genes that SpOs predicted were core, the mice developed the exact physical symptoms of the disease significantly more often than if you just knock out the peripheral genes, identified by standard GWOs.
9:18Oh, wow. So it's not just a mathematical parlor trick. No, it's very real. Furthermore, they looked at real human patient data. The core genes, predicted by the AI, were significantly measurably differentially expressed in the tissues of sick individuals compared to healthy ones.
9:32The biological evidence thoroughly backed up the mathematical graphs. Okay, so the tool definitely works. It finds the executives. And to really put it through its paces. The researchers aims Bios at a notoriously complex condition.
9:45Ulcerative colitis or UC, which is a severe debilitating inflammatory bowel disease. Oh, sort of colitis is a perfect test case. It has a huge GYS data set, meaning we know a lot of the peripheral middle managers, but a very murky understanding of the actual core drivers.
10:03When Spios processed the data, It spit out 693 highly confident core genes for UC. And this is where the paper delivers a massive biological plot twist. As the name implies, ulcerative colitis is a disease of the colon.
10:17It causes severe inflammation and ulcers in the intestinal lining. So logically, you'd think any medical student would expect the core executives driving this disease to be heavily active in gut tissue.
10:27It seems like the most intuitive assumption in the world. But the data revealed something entirely different. The core genes identified by the AI were not primarily expressed in the colon. Wait, really?
10:38Where were they? The defining feature of these UC core genes was extremely high expression in the immune system. Specifically, they were hyperactive in whole blood, the spleen, and especially an Epstein bar virus transformed lymphocytes, which are a type of white blood cell.
10:55So the disease destroys the gut, but the genes coordinated the destruction or operating out of the blood and the spleen. Yes, which forces us to completely rethink the battleground. If we look at the bigger picture.
11:07This finding proves that ulcerative colitis is fundamentally a dysregulated immune system disorder that merely manifests in the gut. The colon is just the collateral damage. The generals directing the war are stationed in the immune system.
11:20That is mind blowing. It actually aligns beautifully with emerging clinical views that the most effective treatments for severe UC involves systemic immune suppression, rather than just topical treatments applied to the gut tissue itself.
11:32That totally reframes the entire clinical approach. But it brings up a really important question about the GWRs data. If the core genes are in the immune system, What are we looking at when we see those massive signals in the GWO studies?
11:46Because if the paper calls these highly visible genes HSPs, or highly significant peripheral genes, why did the AI flag them as core executives, if their genetic signal is screaming so loudly? To understand the distinction the AI is making, We need to compare their job descriptions.
12:03Let's look at a true core gene identified by the model like HLA DRB 5 or JAK3? These genes are incredibly precise specialists. They are highly tissue specific, functioning almost exclusively in those immune cells we just talked about, and they are directly involved in very specific inflammatory signaling pathways.
12:21They are the precision tools. Right. Now contrast that with a highly significant peripheral gene. Let's look at an HSP called Ineva. It has a massive genetic association with UC in the GWOs data. But when you look at where it operates, it isn't expressed in immune cells at all, it's expressed in gut goblet cells.
12:37It's local to the gut. Exactly. It acts as an accessory regulator that modulates local inflammation, but it isn't the primary driver of the immune response itself. Or take another HSP they highlight, PRRK 7.
12:49It also has a huge GW wallet signal for UC. But PRK 7 is expressed in almost every single cell in the human body. Its job is completely generic. It suppresses reactive oxygen species to keep cells from getting stressed.
13:01Which is important, but it lacks the targeted tissue specific precision of a core gene. Inava and Park 7 are generic master regulators. A mutation in them might make the colon slightly more vulnerable to stress, but they aren't the executives ordering the immune attack.
13:17And this perfectly validates the omnogenic theory. Core genes possess extreme tissue specificity. Peripheral genes act as broad, generic modulators. This brings us to the most crucial mechanistic question of the deep dive.
13:29We have the generic middle managers on the periphery, and the specialized executives in the core. How do the middle managers actually tell the executives what to do? You mentioned a funneling effect earlier.
13:40How did the researchers prove that funneling actually happens? Oh, this is arguably the most amazing part of the methodology? The researchers utilize something called the LINCS data set? This is a massive, incredibly valuable library, where scientists have systematically taken human cells, in this particular analysis, HT 29 colon cancer cells, and individually knockdown or overexpressed 1000s of different genes.
14:06They basically flipped 1000s of different genetic light switches one by one to see which rooms in the cell lit up. Exactly. So they can see the exact cause and effect of tweaking a single peripheral gene.
14:17And when the researchers analyze this massive database of percurbations, they discovered something astonishing. When you randomly perturb genes across the entire genome, a staggering one third of those perturbations caused a coordinated, measurable ripple effect specifically on the ulcerative colitis core genes.
14:35Let me make sure I'm hearing this correctly. A 3rd of the entire geno. Yes. One third. Irrespective of where the peripheral gene was located or what its primary job was, flipping its switch reliably altered the expression levels of the core genes in the center of the network.
14:48It sounds like the cells network is structured exactly like a spider web, with the core gene sitting dead center. No matter where a peripheral gene plucks the silk on the outer edge, the vibration travels down the spokes directly to the middle.
15:01That analogy captures the topological reality perfectly. The core genes occupy a uniquely central position in the cell's regulatory architecture. There, the hubs receiving all the downstream traffic. But the researchers didn't stop at single plex of the web.
15:16Right, because in biology, genes don't mutate one at a time in a vacuum. Exactly. So they used another advanced machine learning framework called gears to simulate co-perturbations. Simulating what happens when you plug 2 strings at the exact same time.
15:29Which raises a fundamental mathematical club for geneticists, does one +one equal two? Historically, the onogenic model mostly assumed that genetic risk was additive. You add up all the little peripheral vibrations, and that total sum equals your disease risks.
15:45But the gear simulation proved that when you perturb multiple core genes at once, the math becomes highly nonlinear. Nonlinear, meaning the effects don't just stack neatly on top of each other. Right. They observed profound, complex interactions.
15:59They saw suppression, where 2 genes that both independently cause inflammation, actually cancel each other out when triggered together, and more importantly, they saw a neomorphism. Neomorphism. break that down for us.
16:11It's a technical term for radical synergy. It's when the combined effect of 2 genes is entirely different and vastly more powerful than the sum of their individual parts. Let me give you a specific example from the paper.
16:23They looked at 2 genes, BAX and NFKBIA. Okay, what do they do? BAX is a pro-apoptotic gene. Its job is to tell a highly stressed cell to self-destruct for the greater good of the tissue. NFKBIA, on the other hand, inhibits a major inflammatory pathway called NF Cow B.
16:41If that pathway is inhibited too much, it also acts as a trigger for cell deck. So both of these genes basically have a hand and cell suicide. Right. Now, if you artificially overexpress just one of them in a simulation, the cell gets stressed, but it usually manages to compensate.
16:57But gears show that if you overexpress both BX and NFKBIA at the same time, it doesn't just double the stress, it creates a massive synergistic trigger. Like pushing it over the edge. Exactly. The cell crosses what is called a beastable threshold.
17:12It completely alters its transcript dome, and irrevocably plunges into programmed cell death. Wow, so they don't just add together. They act like a match in a powder keg. Independently, a match burns out and a barrel of powder just sits there.
17:25But together they explode. That is exactly what happens. It proves that core genes aren't just passively receiving signals from the periphery. They are intricately interacting with each other in the center of the web in ways that can suddenly and violently tip a healthy cell into a disease state.
17:39This architecture is obviously a massive breakthrough for understanding ulcerative colitis. But does this specific spider web structure apply to other diseases? Is this a quirk of autoimmune conditions, or is it a fundamental rule of biology?
17:53It appears to be a fundamental rule. To test the universality of their framework. They broadened the scope and applied Space to 2 completely different, highly complex traits, coronary artery disease and schizophrenia.
18:07And they found the exact same architectural pattern. The same core and periphery spider web just with different genes sitting in the center. Exactly. For coronary artery disease, The AI identified core genes that were heavily driven by metabolic networks, specifically pathways involving the transfer of glycogen and lactate, which makes perfect sense for a disease driven by cardiovascular stress and plaque buildup.
18:31And what about schizophrenia? For schizophrenia, the central web was dominated by regulatory networks operating in the adrenal gland, and fascinatingly, at the interfaces with the gut microbiome. Oh, which highlights the growing evidence of the gut brain axis and psychiatric conditions.
18:45Indeed. But across all 3 vastly different diseases, the central principle held true. Thousands of generic peripheral genes converge their signals onto a tight group of highly tissue specific core genes.
18:57So this research fundamentally validates the omnogenic model by proving that funneling effect is real. But it also drastically updates the model by proving that once those signals reach the core, they interact in these highly complex, explosive, nonlinear ways.
19:14It adds a necessary layer of complexity that genetics has struggled to quantify for years. If we truly want to understand complex diseases, we can't just tally up genetic variants. We have to map the math of their interactions.
19:28Which brings us to the clinical reality. If a 3rd of the entire genome can send vibrations down the web to influence these core genes, doesn't that make gene therapy an absolute nightmare, if the target is spread across 1000s of peripheral middle managers, how do we ever fix it?
19:43At first glance, it sounds overwhelming. But clinically, this is actually incredibly hopeful. Think about it this way. Instead of playing whack-a-mole with 1000s of peripheral symptoms or trying to invent therapies to fix 100s of tiny genetic variations scattered across the DNA, we now have a map to the executives, this vastly improves a process called trans EQTL mapping.
20:05Let's define transit QTL mapping for everyone listening. A transit QTL is essentially a genetic variant on one chromosome that regulates the expression of a completely different gene on another chromosome.
20:15It's the long distance signaling we've been talking about. Historically, finding these long distance connections was like looking for a needle in a haystack, but Spios gives us the exact coordinates of the central hubs receiving these signals.
20:28So if drug developers know exactly which core hubs are at the center of the web, they don't have to worry about the thousands of peripheral vibrations. Precisely. Drug development can focus exclusively on how to stabilize that specific network intersection.
20:43We can design small molecule therapies that act as shock absorbers at the center of the web, neutralizing the disruptive signals coming from the periphery before they can trigger that nonlinear explosion.
20:54It shifts our approach from trying to fix the whole city's traffic grid. To simply installing a highly effective stoplight at the main intersection. That's beautifully put. To summarize it all, complex traits are not simply the result of a few isolated broken genes.
21:08They are the result of a vast, interconnected biological web, where thousands of peripheral genes funnel their noise into central, highly tissue specific core genes. Understanding the topology of this network allows us, for the 1st time, to mathematically map how scattered genetic variants collaborate to cause disease.
21:28Which leaves us with a really profound question about the future of healthcare. What does this mean for how we approach personalized medicine in the future? If your unique genetic risks are scattered across 1000s of generic peripheral genes, perhaps the ultimate cure won't require editing your DNA to fix those genes themselves.
21:46Perhaps the future of medicine is simply learning how to dynamically reroute the traffic in your cellular network. It's an incredible thought. This episode was based on an open access article under the CCBY 4.0 license.
21:58You 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. If you'd like to support our work, use the donation link in the description.
22:11Now stay with us for an original track created, especially for this episode, and inspired by the article you've just heard about. Thanks for listening and join us next time as we explore more science base by base.
22:38Midnight in the lap, the screens are cold and bright, tracing tiny typos in the cold of life tonight. A motor in the cell. But you still hold the line now slips between the teeth, where often meets the spine, bine, bine, Breath running higher, but the lights are getting low.
22:58Power without a promise. Where did the current go? If it's built to be a turbine. mine Why is it breaking in the flow, uncoupled, spinning fast? Still coming up short oxygen on the rise. But the spur won't hold the floor, one wrong piece in the ring can bend the whole machine.
23:19But we can see the fault line we can read. What it means. What it means With the contact points, where the subunits align, A dominant shadow leans in Those the rhythm out of time. Those are the complex standing Less of the ATP flame.
23:50Yet the engine keeps on burning like it never learned its name from the gels. Watch the bands, they hear the silence and the spin, measurement branes, slip and feel the leak under the skin, model the breaking, living systems.
24:04Then at the wild tight back, and the noise turns into signal like a path we found in black, blah, blah, blah, blah. Uncouple spinning fast. Still coming up short. we name the hidden damage We met the missing boards from restless consumption to a we can build the screen.
24:22Now the story makes the circuit now the diagnosis. Is clean. But not unbroken. We're closer than we've been.
