This episode explores M-REGLE, a multimodal deep‑learning pipeline that jointly learns representations from ECG and PPG waveforms to boost GWAS discovery and polygenic risk prediction for cardiovascular traits, including atrial fibrillation, and validates results across multiple biobanks.
0:00Welcome 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. Absolutely, glad to be here. So I want you to take a 2nd right now and just, you know, look down at your wrist.
0:12Assuming they're wearing a watch, right? Yeah, exactly. If you are wearing a modern smartwatch, like an Apple Watch or Garmin, whatever it might be, there is a very, very good chance. It is quietly and constantly measuring 2 specific things.
0:27Right. The uh, the electrical and the optical. Yes. First, there's the electrical activity of your heart, which is your ECG. And second, there's that little blinking green light on the back of the watch.
0:37Which is measuring the optical blood flow. Right, known as a PPG. And usually we just think of these as, you know, fitness trackers. Heart rate monitors. Just basic metrics, yeah. But what if, and this is kind of wild, What if you combine those 2 simple signals, and by doing that you can actually map out your underlying genetic risk for cardiovascular disease?
0:59It sounds like science fiction, honestly. It really does. But it's happening right now. And it's what we're getting into for today's deep dive. We're looking at exactly how a massive new AI framework is doing this.
1:10Fusing those biological data streams together. Exactly. Finding hidden genetic markers predicting disease with this level of accuracy that, well, it genuinely shifts the paradigm of modern genomics. It's huge leap forward.
1:23And before we get too deep into the weeds, today we celebrate the work of Google research and multiple academic institutions who have advanced our understanding of multimodal AI and genetics. Yeah, they put together a really incredible piece of work.
1:36This is a 2025 paper from the American journal of human genetics by Zao and colleagues. And the framework they've developed, the AI itself is called M-R-E-G-L. M. Rayelli, which stands for multimodal representation learning, right?
1:51Exactly. And the core ambition of this paper, the real mission here is to solve a fundamental computational bottleneck in Genome Wide Association Studies, or GWS. Because we have all this data. Right. We have these massive bio banks.
2:06And they're filled with high-dimensional continuous waveform data. Like the ECGs and PPGs we were just talking about. Yeah, and spear grams, all sorts of stuff. But the problem is mapping that raw, noisy physiological data to specific low sci on the human genome.
2:22Which is not easy. No, it requires a mathematical bridge that, honestly, until now has been incredibly inefficient. And correct me if I'm wrong, but that inefficiency basically comes from how we historically process the data, right?
2:35Yeah exactly. The authors call the standard approach U-E-G-L-E. The unimodal approach. And it treats these data streams as like totally isolated silos. That's the perfect way to describe it. Silas. So if I understand the architecture correctly.
2:51You're essentially running 2 completely separate neural networks. That is exactly what happens. Okay, so one model is trained exclusively on the ECG to extract its latent features, and then a completely 2nd model is trained exclusively on the PPG.
3:03Right, and they don't talk to each other at all. Not until the very end, right? Like at the fully connected layer, you finally do this, I don't know, post-hoc statistical mashup of those extracted features.
3:13Yeah, that late stage integration is known in the field as intermediate fusion or late fusion. And it's been the default for a long time, simply because computationally, it is way easier to train isolated models.
3:26mathematically simpler. Right. But biologically, it is entirely counterproductive. Why? Because the heart isn't just doing one thing at a time. Exactly. The electrical pacing of the heart, so the ECG and the mechanical volumetric pumping of the blood, the PPG.
3:43They aren't independent events. Right, they're physically connected. They are inextricably linked by hemodynamics. So if you isolate the data streams during the feature extraction phase. You're kind of forcing the AI to learn the physiology blind.
3:57Blind to the broader context, yes. It makes me think of, like, music. Music. How so? Well, that unimodal approach, Like trying to appreciate the complex rhythm of a song, right? But you isolate the drum track in one room.
4:10Okay, I see where you're going. And then you walk down the hall into a totally different room to listen to the bass, right? Right, right. I mean, sure, you can transcribe the individual notes perfectly, but you completely miss the groove.
4:22You miss how they interact. Exactly. You miss the real-time crosstalk, you know, the way the mechanical output slightly lags behind the electrical signal. Which is where the actual physiological truth lives.
4:33Yes. That groove is exactly what MRGLE is designed to capture. So how does it do it? Instead of late fusion, MRG Eli uses early fusion. Early fusion. So right at the beginning. Exactly. The researchers take the raw waveform data from both the ECG and the PPG, and they literally concatenate them.
4:54They stitch them together before the deep learning analysis even begins. Oh, wow. So it's one big block of data. Yes, the combined signal is fed directly into a joint AI model. This forces the neural network to analyze the electrical and the mechanical data simultaneously.
5:10From the very 1st step. From the very 1st convolutional layer, yeah. You know, I was looking at the canonical correlation analysis they did, the CCA. Oh, yeah, the CCA is crucial. They ran that on these signals prior to feeding them into the model, right?
5:24And they established that ECG lead I and the PPG have a correlation of .43. Which is a very specific, fascinating number. Right. Because it actually proves why early fusion is so necessary. Like a correlation of .43 means there's a significant shared variance.
5:42A joint mathematical space, essentially. Yeah, but it also means that over half the signal is entirely unique to the individual sensors. Exactly. And if you fuse them late. You've already thrown away all those subtle interactions.
5:53Right. The interactions between that shared core and the unique variances are just gone. It just lost data. The CCA provides the whole mathematical justification for their architecture, because these signals share an underlying generative factor, the heartbeat, but they capture different physical manifestations of it.
6:10The electricity versus the actual fluid moving. Right. So one signal can act as a natural regularizer for the other. A regularizer. Meaning like, it keeps it in check. Exactly. When the AI looks at them together, it can cross-reference the electrical spike with the optical pulse wave.
6:26Oh, I get it. So it can tell the difference between a real physiological anomaly. And just random sensor noise. Like if you just bumped your watch on a desk or something. Exactly. Which brings us to the actual engine they use to do this cross-referencing.
6:39The convolutional variational auto encoder or VAE. It's VAE, yeah. Now, when I look at a VAE architecture. My immediate concern, and I think a lot of people's concern is the bottleneck. The compression phase.
6:53Right. Because the network has this encoder that basically forces 10s of 1000s of data points from these massive combined waveforms. Down into a tiny mathematical summary. A latent embedding of just what, 12 to 96 dimensions?
7:071296, yeah. Depending on the setup. But in standard data compression, forcing high dimensional data through a low dimensional bottleneck is lossy. Like a JPEG. Exactly. Think of a highly compressed JPEG image.
7:19You lose the high frequency details, you get this pixelated mess. Right. So how does squashing a full 12 lead ECG or even smartwatch data into just 96 numbers. How does that preserve the hyperspecific subtle variations we need to find genetic markers?
7:36It's a great question. And the instinct to compare it to standard file compression is really common, but a VAE operates on a totally different principle. Okay, how so? It is learning a generative model of the data.
7:48The objective isn't just to make the file size smaller. Then what is it? The objective is to force the neural network to discover the fundamental rules of the system. See, the decoder network has to take those 96 numbers and perfectly reconstruct the original massive waveforms from scratch.
8:04Like reverse engineering it. Exactly. And by artificially restricting the flow of information through that bottleneck, the VAE simply cannot afford to waste any of its 96 dimensions on stochastic noise.
8:15Oh, so it just drops the static. It drops a movement artifacts, the sensor static, all of it. It has to abandon the noise and memorize the actual structural manifold of the cardiovascular system. That's beautifully put, yes.
8:27It prioritizes the data that has causal predictive power rather than just memorizing the superficial shape of the wave. And we can actually see how well it works, right? Like we can quantify how successfully it learned that structure.
8:38We can, by looking at the reconstruction error. Right. The mean squared error. Exactly. When the researchers compared the reconstructed waveforms of the early Fusion M Argale model against the old unimodal uray grail model, the reduction in error was massive.
8:54The numbers on this are insane. Yeah. For the 12 led ECG data, M aregally reduced the mean squared error by 72.5%. That's huge jump. And even for the paired ECG led eye and PPG so, the kind of data you get straight from a smartwatch, the error dropped by 20.2%.
9:11A 72.5% reduction is just that's a completely different tier of mathematical accuracy. It really is. It tells us that because the VAE had access to both modalities simultaneously, it used the physical blood flow data to fill in the gaps of the electrical data.
9:26Creating a reconstructed wave that is remarkably closer to biological ground truth, but, and I want to push back on this a little bit. A lower means squared error doesn't necessarily mean the latent space is biologically meaningful to a human doctor.
9:40The black box problem. Exactly. AI is famously a black box. If I'm a cardiologist, how do I know this 96 dimensional vector isn't just a highly efficient mathematical shortcut? Right. how do we know it has any basis in actual human physiology?
9:56Yeah. Well, the researchers anticipated that exact skepticism, which is why they designed this brilliant interpretability experiment to basically force the black box open. Okay, I love this part of the paper.
10:08It's so cool. They extracted the AI's learned latent embedding for an average healthy cardiovascular profile. Like a baseline normal person. Right. And then they map those dimensions against known clinical phenotypes.
10:22And they found something really specific. They did. They identified that one specific coordinate in the latent space dimension 4 strongly correlated with patients diagnosed with atrial fibrillation. And this was the part that really showcased the power of the model, because they essentially created a virtual slider for dimension 4. They took that healthy baseline embedding, held all other 95 dimensions, completely constant.
10:46Didn't touch them. Right. And they just dragged the value of dimension 4 into the unhealthy range. And as they fed that tweaked vector back into the decoder, the resulting reconstructed waveform visibly morphed in real time.
10:58So crazy. The decoded ECG suddenly developed a prolonged QT interval. Which is a classic sign of heart issues. And simultaneously, the decoded PPG waveform lost its dichrotic notch. That is just, that is the crucial validation, isn't it?
11:13It is, because remember, this is an entirely unsupervised model. The VAE was never fed clinical labels. No one programmed it to look for prolonged QT intervals. Right. Or missing dichrotic notches. It just figured it out.
11:25It simply ingested raw waveforms and tried to compress and reconstruct them as efficiently as possible. And in doing so, the math independently reverse engineered, like centuries of clinical cardiology.
11:37It proves that these abstract latent dimensions are not just mathematical artifacts. They are direct, causal representations of human hemodynamics. That is a profound validation of the unsupervised approach.
11:50It really is. It proves the math is inextricably tied to the biology. And this is exactly where the paper jumps from computer science into genomics. The main event. Because now we have this clean, highly compressed. Biologically validated latent embedding of a patient's cardiovascular system.
12:09Right. But how does this actually map to their DNA? I mean, running a genome wide association study on a 96 dimensional vector across 100s of 1000s of patients. It sounds like a nightmare. It sounds like an incredibly complex multivariate regression problem.
12:23It is a massive computational undertaking, no doubt, but the core mechanism is straightforward. In a GWS, researchers are basically scanning 1000000s of single nucleotide polymorphisms or SNPs. There's little variations in our DNA.
12:36Exactly. Across the genome of each patient. And they run statistical regressions to see if variations at specific genetic lowsi reliably correlate with variations in the patient's physical traits. Which, in this case, are the 96 dimensions of our latent embedding.
12:51Right. And because M-R-E-G-L-E filtered out the noise and isolated the true physiological signal so effectively. The statistical power of the G LEA just exploded. Let's talk about the specific metric they use to measure that statistical power.
13:07Because they reported a 22% increase in the expected Chai squared statistic. For the 12 lead data, yeah. compared to the unimodal approach. Now, in the GLOS literature, researchers are constantly fighting population stratification, right?
13:20Oh, constantly. That statistical noise that makes it look like there's a genetic link, when the correlation is actually just driven by demographic or environmental confounding. It's a huge headache in genetics.
13:29So why is the expected Chi squared statistic, the gold standard for proving this early fusion model is actually better? Because the expected chi squared statistic isolates the true polygenic signal from that exact confounding noise you just mentioned?
13:43Oh I see. When you run a GWS, you generate a massive distribution of association scores, if your phenotype data is noisy. Like with the old unimodal method. Exactly. Your association scores will be weak and the genuine genetic signals will just fail to rise above the statistical background noise.
14:01They get buried. Right. So an increase of 22% in the expected chi squared statistic. That means the signal to noise ratio has vastly improved. It's like wiping the static off a TV screen. It indicates that the genetic architecture driving the MRIG embeddings is significantly more robust and way less polluted by genomic inflation.
14:20And the clarity of that signal translates directly into concrete biological discoveries. Yes, it does. Because when they analyze the 12 lead ECG data, the M Regi Aguias identified 142 distinct genetic loci associated with cardiovascular function.
14:36Which is huge. That is 19.3% more low side than the unimodal method found on the exact same patient cohort. Just by processing the data better. Exactly. And crucially, 20 of those genetic loci were completely novel.
14:49Brand new. Never been previously associated with these cardiovascular traits in the scientific literature before. Finding 20 novel loci is a massive leap forward. It really is. When we map these new low side.
15:01They often point to specific genes involved in, you know, ion channel regulation, myocardial structure, or autonomic nervous system pathway. Exactly. These aren't just statistical trivia. They are highly specific biological coordinates.
15:14Each new locus represents a potential target for pharmacological intervention, or maybe a previously unknown regulatory network that contributes to heart disease. And the success wasn't even limited to the clinical grade 12 lead ECGs, right?
15:28Not at all. When they restricted the analysis to just the smartwatch style data. So just the single ECG lot I and the optical PPG, the early fusion model, still uncovered 13.0% more genetic low side than the old method.
15:41With a 16.4% increase in the expected G squared statistic. This means that right now, the consumer wearables that 1000000s of people have on their wrists, they are generating pair data streams that, when processed through an early fusion VAE, yield research grade genomic insights.
16:00It's wild. The democratization of that data collection is precisely why this framework is so valuable. But, and this is a big but, the theoretical discovery of new genetic pathways, is really only half the battle.
16:12Right, because you have to actually help patients. Exactly. The ultimate metric for any genomic research is clinical utility. Does a cleaner phenotype and a more powerful G wear off actually translate into better preventative medicine for a living breathing patient.
16:27Which brings us to the construction of the polygenic risk score. PRS. Yeah. Because if we've mapped all these new loci, we have to figure out how to aggregate them into a single number. right? A number that tells a clinician whether a patient is actually going to develop atrial fibrillation in the next 10 years.
16:43Right. A polygenic risk score is essentially a weighted sum. You look at a patient's genome and you literally count up how many of these risk alleals they carry. But not all low sci are created equal. Exactly.
16:55So you weight each allel based on the effect size that was calculated during the G-Deli. And because the Mriculi lees had a much higher statistical power. And a cleaner signal to noise ratio. The effect size estimates for every single SNP were significantly more accurate.
17:09You've got it. So the foundation of the math is just tighter. Yeah. The model isn't being misled by the noise of late fusion feature extraction. Which means the resulting polygenic risk score should have a much stronger predictive power.
17:22And the paper demonstrates exactly that. The PRS, built from the MRDL data, significantly outperformed the unimodal PRS in predicting incident atrial fibrillation. Wow. The performance improvement is critical, but the validation methodology they used.
17:39That is what cements the paper's importance. Because of the overfitting risk. Yes. If you train a neural network and calculate a PRS entirely within, say, the UK bio bank. There's always the risk that it just memorized the UK biobank.
17:53Exactly. The model might just be memorizing the specific genetic and environmental idiosyncrasies of the British population. To prove the validity of these genetic associations, you have to test the PRS on completely unseen populations.
18:08And they took this model global. They really did. They applied the M-R-E-G-L-E derived polygenic risk score to entirely independent patient databases, like the Indiana Biobank in the United States. And the Ipic Norfolk data set.
18:23And the British woman's Heart and Health Study. And the fact that the predictive superiority held up across all those distinct populations, it proves that the latent space didn't just overfit. It didn't.
18:34The VAE successfully mapped universal fundamental laws of human cardiovascular genetics. The ability to generalize across biobanks is the definitive proof that early multimodal fusion captures a more accurate representation of human biology.
18:49really is. We are officially moving past the era of single modality diagnostics. By forcing artificial intelligence to reckon with the interconnectedness of human physiology at the very 1st layer of computation.
19:00We can dramatically accelerate our ability to read the genetic code. Exactly. The elegance of this paper really just comes down to acknowledging that human biology doesn't operate in silos. We are complex systems.
19:14Your electrical pulse and your physical blood flow, they're just 2 different echoes of the exact same underlying biological truth. Well said. And the researchers actually hinted at the future of this approach, right?
19:27They ran a secondary test where they added a 3rd signal to the early fusion mix. Oh, right, the spirogram. Yeah, the sparogram, which measures continuous breathing and lung capacity. And when they integrated that respiratory data into the VAE alongside the ECG and PPG, the model's genetic predictive power and the number of discovered loresi increased yet again.
19:47It's amazing. Every complementary physiological signal acts as a new regularizer. Stripping away more noise. And revealing a sharper picture of the genome. It really sets up an incredible trajectory for the future of personalized medicine.
19:59It does. I mean, if simply combining your pulse, your heart rhythm, and your breathing unlocks this much hidden genetic data. The implications for ambient health monitoring are just staggering. Totally. Think about it.
20:15What happens over the next 5 years when these early fusion AI models start simultaneously processing the continuous data from your sleep cycles? Or the voice biomarkers collected by your phone. Right. Or the micro movements of your gate is you walk.
20:29It's gonna be everywhere. Our bodies are constantly broadcasting this massive multimodal symphony of physiological data. We finally have the mathematical tools to listen to the whole song. And it turns out the lyrics are written directly into our DNA.
20:41This 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.
20:55If 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.
21:05Thanks for listening and join us next time as we explore more science base by base.
