OncoGAN is a multimodel generative AI pipeline that simulates realistic, privacy-preserving cancer genomes (VCFs, CNAs, SVs) across eight tumor types to support benchmarking and training of genome-analysis tools.
0:19Welcome 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 you have the blueprint to cure a rare cancer.
0:31You have the advanced artificial intelligence ready to find that cure. But the data you need to turn that AI on is legally locked in a vault. And, you know, even looking at it is a privacy violation. So how do you train a life-saving AI on data you aren't allowed to see?
0:47Right, it sounds like an unsolvable riddle. You need 1000s of real cancer genomes to teach the algorithm what to look for. But human DNA is, well, it's the ultimate private identifier. Exactly. The moment you share it, you expose a patient's most intimate biological secrets.
1:01Exactly. So instead of trying to pick the lock on that legal vault, What if we just built a new reality? Like, what if we use AI to hallucinate completely fake yet biologically flawless cancer patients?
1:13Okay, let's unpack this. Today we celebrate the work of Androdes and Navarro, Lincoln Stein, and their team at the Ontario Institute for Cancer Research, the University of Toronto and the Vector Institute, who have advanced our understanding of synthetic cancer genomics.
1:26Yeah, we are taking a deep dive into their groundbreaking 2025 paper published in cell Genomics. It details a highly sophisticated computational pipeline they created called on Kogan. So the overarching mission of this deep dive is to understand how this team figured out a way to bypass a massive privacy bottleneck by, you know, literally inventing genomic data for cancer patients who do not actually exist.
1:51But before we get into the fake data. If you are listening to this, you might be wondering why researchers desperately need this information in the 1st place. We really need to look at the daily reality of precision oncology.
2:01When a tumor develops in the human body, its genome becomes heavily damaged. Over time, it accumulates thousands, sometimes 10s of thousands of genetic alterations. It is absolute biological chaos inside those cells.
2:14Total Chaos. But not all of those alterations are actually responsible for making the cancer grow and spread. Far from it. Out of all that chaos, you might only have a handful of what we call driver mutations.
2:27Ah, driver mutations. These are the actual culprits driving the cellular proliferation. Precisely. And everything else is just a passenger mutation. Passengers are harmless background noise. It's genetic damage that just happens to be along for the ride as the cell divides.
2:43So if I'm trying to picture this, It's like building an incredibly advanced spell checker to find a very specific, dangerous typo in a 1000000000 word document. That's great analogy. And the spell checker has to do this while completely ignoring all the other weird, harmless formatting quirks scattered across the pages.
3:00Yeah, that captures the scale of the challenge perfectly. The research community builds complex algorithms called variant callers. These act as those advanced spell checkers, hunting through the genome for those elusive driver mutations.
3:13But here is where I get stuck. If the real patient data is locked away for privacy reasons, how are scientists testing these variant callers right now? I mean, how do you know your spell checker works if you aren't allowed to read the documents?
3:25Currently computational biologists are forced to use some pretty severe workarounds. Often, they benchmark their mutation calling tools against a tiny legacy set of about 50 genomes derived from decades old cancer cell lines.
3:40That's it. Yeah, just 50. And if they don't use those, they take a normal, healthy genome and use a computer script to artificially spike a series of randomly generated mutations into it. Wow. That seems like it would completely miss the reality of a human body.
3:54A real tumor has been exposed to specific tissue conditions, complex immune system attacks, maybe chemotherapy or, you know, environmental toxins. The scientific community fully recognizes that limitation.
4:06Those rudimentary methods fail miserably at capturing the true, messy complexity of an actual patient derived cancer genome. We do have massive trolls of real data, to be fair. The pan cancer analysis of whole genomes, or PC guy at G, sequenced over 2600 whole cancer genome.
4:25Right. But because of strict privacy laws globally, you can't just email that treasure trove to a startup or university lab so they can benchmark their new algorithms. Which brings us to the core methodology of this deep dive.
4:38Here's where it gets really interesting. If we can't share real genomes, how exactly do we build synthetic ones that an advanced algorithm will actually believe? Because we aren't just faking a basic spreadsheet here.
4:48A human genome is 3000000000 base pairs of DNA. To pull this off, the team built the Ankajan pipeline using a multimodel ensemble approach. The heavy lifting is done by generative adversarial networks, or jams.
5:01I get how a JAN generates a fake face on social media. It's just arranging pixels based on patterns, but how does it generate DNA? Are we just making up A's, C's, T's and G's at random? Not quite at random.
5:12Think of a JAN as a highly sophisticated forgery game. You have 2 neural networks locked in a battle. The 1st network is the generator. Its job is to create the fake genomic data. The 2nd network is the discriminator.
5:26Acting as the detective. Look at the fake data, compares it to the real locked away PCCWG data, and tries to catch the forgery. Wow. And they just train against each other. Exactly. They train against each other until the detective network simply cannot tell the difference anymore.
5:40But genomic data is incredibly complex. You have point mutations, areas where whole chunks of DNA are duplicated, structural changes. Which is why the JAN alone isn't enough. They specifically used a variant called CTaveJam Plus to handle the complex cabular data, the overarching statistics, like the total number and types of mutations per donor.
6:01Okay, I see. So then they brought in a 2nd technology called tabular variational auto encoders or TBAEs. Wait, tabular variational auto encoders? Let's translate that for me and the listener, what does that actually do in the pipeline?
6:13Let's strip away the jargon. A TVAE is essentially a tool that learns the hidden underlying structure of a massive data set. In this case, it helps handle the staggering scale of the genomic positions.
6:25So it looks at the structure. Right. It learns the statistical probabilities of where mutations are most likely to occur across those 3 billion base pairs. Here's where it gets really interesting to me, just from a computing standpoint.
6:37How does an AI simulate exact locations across 3 billion base pairs of DNA without the computational weight just crashing the entire system? The researchers introduce a brilliant mathematical step called discretization.
6:51To handle the sheer size of the genome, the AI temporarily compresses it down from 3 billion base pairs to 30 million base pairs. So they were shrinking it down by a massive ratio of one to 100. Yes, exactly.
7:03If I'm trying to visualize this, it's like trying to memorize a map of the world. You don't start by looking at every single house on Google Street view. You zoom out to see where the continents and cities are first, learn the general population density, and then later you zoom back in to place the actual houses.
7:19Is that what the AI is doing with DNA? That is a phenomenal way to look at it. They divide this zoomed out compressed genome into over 100,000 bins. By doing this, the generator can learn the overall mutation density.
7:34Meaning where the genetic typos tend to cluster. Right, without getting bogged down by single letter resolution. Once the AI understands the continents in cities, the TVAE model learns the specific positions within those bins.
7:47Finally, the expand the simulated positions back out to the actual 3000000000 base pair map. That is a staggering technical workaround, but I have to push back here because if you are a researcher listening to this, alarms might be going off.
8:00If this AI is learning the exact patterns from real, highly sensitive patients, couldn't it accidentally spit out a real patient's DNA sequence? If the forgery is too perfect, doesn't that violate the very privacy we are trying to protect?
8:16That is the absolute most critical question in synthetic data generation, if you memorize the training data and just regurgitate it, you have failed. This is where the Ankogan pipeline implements a fascinating privacy safeguard.
8:29During the simulation phase, the model doesn't actually have access to the human reference genome at all. It doesn't. No, it simulates the mutations location and its trinucleotide context, meaning the specific DNA letters immediately surrounding the mutation completely independently of each other.
8:46So the AI guesses a location on, say, chromosome 5. And then entirely separately in a different part of the system, it guesses that the mutation should be surrounded by the letters A, C, and G. You've got it.
8:56It only brings those 2 separate guesses together at the very end of the pipeline. The system checks a small 100 nucleotide window around the singulated position on the real reference genome. Just to verify it.
9:06Exactly. It searches to see if that independently generated ACG context actually exists there in real human biology. If it doesn't find a match, It discards it and resimulates the position until it finds one.
9:18That is so clever. Because they're generated in total isolation from one another, the chance of perfectly reconstructing a real patient's exact profile is incredibly low. The data bears that out beautifully.
9:31When they analyze the final synthetic genomes, only 0.021% of the simulated mutations accidentally match the exact mutations in the training set. Wait, let me make sure I'm hearing this right. 0.021%. To put that microscopic number in perspective, The natural rate of identical mutations happening by pure random chance between 2 totally unrelated real human beings in the data set is actually higher.
9:57It's 0.28%. Two real strangers walking down the street are more likely to have identical mutations by pure chance than this AI is to accidentally copy a real patient. That is wild. The fake data is statistically more unique than real human DNA.
10:13So they built the pipeline, they solved the computational scale problem with the map zooming technique, and they mathematically proved the privacy. But does the forgery actually hold up under a microscope?
10:24The team successfully generated 800 synthetic genomes across 8 distinct cancer types. We are talking breast cancer, liver cancer, kidney cancer, esophageal cancer, and the biological fidelity is astounding.
10:38Meaning it's not just faking random typos. It's flawlessly replicating the specific handwriting of the person making the typos. In clinical genomics, we refer to that handwriting as mutational signatures.
10:49Different biological mechanisms or environmental exposures leave highly specific patterns of damage on the DNA. For example, Ankogorgian successfully simulated a signature called SPS4, which is the specific pattern of genetic damage associated with tobacco smoking.
11:03And it reproduced that smoking signature, specifically in the tumor types where you would expect to see it like lung or esophageal cancer. Yes, maintaining the exact frequency and density. They also accurately generated a signature called SBS 9.
11:17This is a highly specific pattern linked to somatic hypermutation driven by an enzyme in the body, and it is a hallmark of a type of leukemia called mutated lymph CLL. Wow. When they ran standard signature extraction tools on the completely synthetic genomes, the tools found the exact same mutational distribution profiles you observe in real human donors.
11:40It even fooled the driver detection algorithms, right? The spell checkers we talked about earlier. They ran a widely used algorithm called Active Driver WGS on the synthetic data. In the real locked away data set, this tool detects about 89% of the coding driver genes.
11:55And on the fake data. In the entirely fake OnCogin data set, it detected 87%. The synthetic genomes preserve the complex biological rules of disease, like which mutations tend to occur together and which ones never do.
12:06Okay, the statistical checks are passing with flying colors. But what happens when you put this fake data to the ultimate test? When you feed it to an AI diagnostic tool that is designed to look at a genome and diagnose a cancer.
12:20They tested the synthetic genomes using deep tumor. This is an advanced deep learning system built to predict a cancer's tissue of origin based entirely on these passenger mutation patterns. How did it handle the synthetic data?
12:33When deep tumor analyze the fake onkagan donors? It achieved nearly 100% prediction accuracy for most of the tumor times. The diagnostic AI looked at the fake genomes and accurately classified the fake cancers?
12:47But the paper doesn't stop there. This is the moment that genuinely blew my mind. They didn't just test deep tumor with the fake data to see if it would be fooled. Right. They took it a step further. They actually took the fake data and fed it back into deep tumors training textbook.
13:01This is where the true practical value of synthetic data shines. They took 100 of these completely artificial genomes for each tumor type and mixed them into deep tumors original training data. They padded the textbook.
13:13And the diagnostic AI got smarter. Its overall accuracy improved by 0.9%. But the most stunning improvements were seen in the rare cancers, where real world data is extremely limited. Let's humanize this for the listener, because it is incredibly hard to train an AI to spot a rare disease if you only have like a handful of real patients to show it.
13:36Take that leukemia we mentioned earlier, lymph MCOL. In the real world, the original training data set only had 35 real donors for that subtype. Deep 2 are really struggled to diagnose it. Only 35. That's barely anything.
13:50Right. It frequently misclassified it as a completely different disease B cell non-Hodgkins lymphoma because their mutational patterns are almost identical. With only 35 examples to study, the algorithm just couldn't learn the subtle differences in the signatures.
14:03But when they've padded the AI's textbook with these synthetic hallucinated patients, deep tumor's accuracy for that rare cancer jumped dramatically. Its F1 score, which measures overall precision and recall, went from 75% up to 84%.
14:18It learned how to spot a real disease by studying fake people. They used completely artificial data to fix a real world-class imbalance problem. The faked data filled in the gaps, and it made the algorithm measurably better at its job.
14:32Exactly. If we connect this to the bigger picture. It completely changes the accessibility of cancer research globally. Because there is 0 privacy risk. No real patient identifiers. These synthetic genomes are entirely open access.
14:47So anyone listening to this could go get them right now. Anyone. All 800 synthetic genomes, complete with their matched, complex structural variants, are available right now on public repositories like Hugging Phase and Zenoto.
15:00Furthermore, the pipeline provides compatibility with sequencing read simulators. Let's clarify that. You mentioned sequencing read simulators. What does that mean in plain English for a lab technician or a student?
15:11It means researchers don't just get a summarized spreadsheet of mutations. They can generate custom FASA and synthetic BAM files. Oh, the actual files from a sequencer? Yes. The literal raw data files that a massive DNA sequencing machine prints out.
15:25The pipeline can simulate the raw sequencing reads, complete with specific variant allele frequencies that mimic the physical reality of a tumor cell. So a small startup in another country, building a new genomic tool, doesn't have to wait 2 years to get legal approval for access to sensitive patient databases.
15:42They can download this today, spin up their tools, and start benchmarking. It democratizes precision oncology in a huge way. But I have to ask for a reality check here. This sounds like magic, but biology is infinitely messy.
15:55What is the AI still struggling with? Where does the forgery breakdown? The study is highly transparent about its boundaries, which is crucial for moving the field forward. For one, Onkogan cannot yet simulate subclonal reconstructions.
16:09Okay, subclonal reconstructions. If I'm visualizing a tumor, it isn't just one solid lump of completely identical cells, it's more like a branching family tree, right? Exactly. As a tumor grows over months or years, different branches of cells evolve their own unique sets of mutations, they compete with each other.
16:26Currently, Ankogand effectively simulates the overall blended mutational profile of a tumor, but it struggles to accurately represent those complex evolving subpopulations within that single synthetic patient.
16:38It can paint the whole forest, but it can't quite capture the distinct families of trees growing inside it. What else does it miss? It also struggles with massive catastrophic cellular events. There is a phenomenon in cancer biology called chromothripsis.
16:53Imagine a chromosome literally shattering into pieces, like glass. Oh, wow. And then the cells panicked repair mechanisms, stitch those pieces back together in a random chaotic order. That sounds like a biological nightmare to try and simulate mathematically.
17:07It is incredibly complex. On Kugan's current architecture doesn't support the simulation of wholesale complex events of that magnitude. Right. Are there any other limitations? It also missed a couple of very rare mutational signatures, simply because of data sparsity.
17:21For example, a signature and esophageal cancer was only present in about 5% of the real donors. So the model just didn't have enough real-world examples to learn how to forge it accurately. But as more whole genome data sets become available globally, the training data will enrich, and these models will only become more comprehensive.
17:40So what does this all mean? Let's bring it all together for you. The Unkadan pipeline successfully leverages advanced generative AI to create highly realistic privacy preserving synthetic cancer genomes.
17:52By faithfully replicating complex biological patterns without exposing actual patient DNA, it provides an unlimited open access resource that is already proving it can train and improve the real-world algorithms used to detect cancer.
18:05This raises an important question for you to consider. We started by looking at a vault, where the most vital human data was locked away. Now, the vault is an empty, but the data inside is entirely synthetic.
18:16What does this mean for the future of medicine when the most valuable life-saving patient data used to train the algorithms diagnosing your illness might belong to a patient who never actually existed?
18:27This 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.
18:42If 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 just heard about.
18:51Thanks for listening and join us next time as we explore more science base by base. Midnight Loud Glow The moon is brief We trace a storm where the old data leaves. Not for a name Not for a face. Just patterns in motion a safer place.
19:34Copy the rhythm, not the soul. Keep the truth but lose control. Of who it was, let it dissolve. So every tool can learn the song. Ghost genomes on the screen tonight. Real enough to teach the lie To my echo signature lights, shed in the clear no privacy cry.
20:16We build what we can freely hold. A future trained on Stories told. Point, point, sparks of line, dells, like, stutters in a hidden line, numbers rise, then break, then fall, copy, count, shadows on the wall, and structural scenes where the big brakes crawl.
21:02Some chapters won't appear no subclones. Whispering in the real, no shadow crown chaos, no chromo sky, the gate world's mapped and we can try again and again without goodbye. Ghost Gnomes on the screen tonight.
21:34Real enough to teach the light To miracle signature lights, shared in clear, no privacy crime. Quiet code to open road. We send 800 Down the road And every new run sings. Go, go, go, go, go, go, go, go, go, go, go, go, go, go, go,
