This episode summarizes Jassim et al.'s introduction of RECODR, a graph-embedding pipeline that measures gene co-expression context drift from sc/snRNA-seq to reveal drivers of tumorigenesis and treatment resistance and to nominate combination therapies validated in mouse models and predicted for human tumors.
0:18Welcome 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. Yeah, thanks for joining us everyone.
0:29So, um, today we're diving into a statistic that, honestly, just it stops you in your tracks, 30 to 50%. That is the staggering reality for oncology patients right now. 30 to 50% of them will actually see their 1st line therapy fail.
0:46It's a devastating hurdle in the clinic, yeah. It really is. And, you know, we often like to think of treating cancer as this straightforward battle, like you identify the bad guy, you deploy the drug, and you walk away with a win.
0:57But tumors aren't just sitting ducks. They are these brilliant, highly adaptable shapeshifters. They fight back. Exactly. When you hit them with a treatment, they develop these incredibly complex resistance mechanisms that eventually lead to relapse.
1:14And the way we've historically tried to understand that resistance has been, well, somewhat one dimensional. So, for decades, genomics has relied really heavily on measuring gene expression levels. Essentially, we take a cancer cell, and we just look at how loud a particular genus turned up.
1:31We kind of just assume that a louder volume means it's the thing driving the resistance. Which, I mean, that makes sense on a basic level, but to put that in perspective for you listening, imagine trying to understand a person just by listening to how loudly they speak without ever looking at who they are talking to.
1:49Oh, that's a good way to put it. Right. Like, are they yelling because they're at a rock concert or are they aggressively whispering in a library? The volume alone doesn't actually tell you anything about their intent in that moment.
2:01It misses all the context. Exactly. So what if we could map a gene's social network instead? What really happens when we track a gene's changing neighborhood during cancer treatment, and how could this change the way we stop tumors from fighting back?
2:15This raises an important question. And, you know, to answer it, we really have to rethink the very fabric of how genes communicate under the extreme stress of these treatments. We have to move away from just measuring volume and start measuring context.
2:29Well, today we celebrate the work of Amir Jessine, Bridget Nimmerville, Richard Gilbertson, and a massive collaborative team, spanning the Cancer Research, UK Cambridge Institute, Boston Children's Hospital, and several other international institutions, who have advanced our understanding of cancer treatment resistance.
2:47The scale of this research is truly impressive, and it tackles a problem that, quite frankly, oncologists grapple with every single day. Because cancers are so heterogeneous, meaning a single tumor is actually made up of this wildly diverse population of cells with different genetic quirks doctors have had to evolve to use combination therapies.
3:07Right. The mix different drugs together. Exactly. The logic is sound, right? You hit the tumor from multiple angles and hopefully prevent overlapping resistance. You try to box it in, you block the main highway with one drug, and you set up roadblocks on the side streets with another.
3:21That is the goal. Yet, these tumors still find a way around the roadblocks. They still relapse. What's fascinating here is that our current computational models, the tools we use to predict this resistance, are inherently flawed.
3:35Flawed how? Like, they just don't work. Well, they rely heavily on comparing a patient's tumor gene expression to static, predefined pathways. So we look at a gene and a tumor, and we check a biological textbook to see what that gene does in other entirely different contexts or, you know, in healthy tissue.
3:53Oh, I see. By doing that, we miss the real-time, highly specific biology happening inside that unique tumor as it actively battles the drugs. Okay, let's unpack this. Are you saying that even if a gene doesn't mutate or change its overall volume, its entire function can shift just because it starts interacting with a completely different group of genes?
4:12Yes, absolutely. It happens all the time under stress. The function of a gene is defined by its cellular neighborhood. If a tumor cell rewires its internal networks to survive a highly toxic drug, a gene that was previously just doing, um, routine structural maintenance might suddenly be recruited into a network that's actively pumping the drug out of the cell.
4:32So it's the exact same gene, expressing it the exact same level. But its job has entirely changed because its context has drifted. That is wild. Which means, if we are only looking for mutations or volume changes, we are completely blind to this rewiring, the tumor is basically plotting an escape route right under our noses.
4:52And we're using the wrong tools to look for it, yeah. So how do we actually see this invisible network? It sounds like try and map the social dynamics of a city of 1000000s while blindfolded. It's a monumental data challenge, for sure.
5:04And that brings us to the technological breakthrough of this team's work. They developed a novel computational pipeline called Rico DR. R-E-C-O-D-R, right? Yes, which sends for resistance through context drift.
5:16I love a good functional science acronym. Okay, let's get onto the hood here. How does Rico ER actually build that map? It starts with single cell RNA sequencing data. And let's clarify what that means.
5:27Instead of taking a biopsy, grinding it all up and getting an average reading of the entire tumor. Like tasting a blended smoothie. Exactly, like a smoothie. Single cell sequencing lets us look at the transcript tome of every individual cell.
5:41We can see exactly which genes are active in cell A versus cell B. So we aren't tasting this movie. We are picking out the individual strawberries, bananas, and spinach leaves to see exactly what they're doing.
5:54A great way to visualize it. Rico ER takes this incredibly high resolution data and builds what we call co-expression graph networks. Okay, graph network. Think of it as a massive multidimensional map where every gene is a node.
6:07And the connections between them represent how closely their activity is linked. So if Gene A and Gene B always turn on and off together across like 1000s of individual cells, they get drawn closer together on this map. Precisely.
6:21But just having a map with 1000000s of lines connecting dots isn't an insight by itself, you have to somehow interpret the traffic on those roads. Right, otherwise, it's just a bunch of noise. And that is where the team brilliantly borrowed algorithms from natural language processing.
6:34To understand the biology, they basically treated the genome like a language. Wait, like the AI models used for text, predictive speech, and chat bots? Yeah, the underlying math is very similar. They used a specific algorithm called No 2 vec.
6:48No toveca. Imagine dropping a digital census taker into this massive gene map. Nodeveck takes thousands of random walks around the graph. It starts at one gene, rolls the mathematical dice, and jumps to a connected neighbor.
7:02And then it just keeps jumping. Yep, to another and another. It does this 1000000s of times to thoroughly explore the neighborhood and document who is hanging out with whom. Here's where it gets really interesting.
7:13This is exactly like how predictive text or AI language models learn. They know the word bark means something different if it's surrounded by tree and leaf versus dog and bite. Rico is learning the definitions of genes based on their cellular neighbors.
7:27You've nailed it. That's the exact mechanism. Once Rika Dimmer understands the language of the untreated tumor. It then compares it to the language of the tumor after it's been treated and becomes resistant.
7:39Okay. To quantify how much a gene's meaning has changed during that battle, recuiter and calculates a context drift score. It uses 4 highly specific metrics to measure this shift. Let's break those 4 metrics down because I don't want us to just list them like a textbook.
7:56What is recruiter actually measuring here? First, we have the neighbor number score. This is the most straightforward one. It just counts how many completely new genes are now directly connected to our index gene in the resistant tumor.
8:08Oh, like checking a social media profile and realizing a coworker has suddenly dropped their old friends and is only hanging out with the corporate executives. Yeah, it's an immediate red flag. You know something is up.
8:18Second is the graph reach score. This looks beyond the immediate neighbors to the neighbors of those neighbor. Measuring the downstream influence, basically. Exactly. If our gene is a middle manager, Graph Reach asks, how much further can this gene's influence spread through the new network now that the tumor is under stress?
8:37It's like the 6 degrees of Kevin Bacon, but for jeans. Huh, exactly. Third is the neighbor cosine score, which measures the average shift in the overall neighborhood. Explain the cosine part for us, if you don't mind.
8:49How does a geometry term apply to a biological neighborhood? Sure. In this multidimensional map, every gene's position is represented by a mathematical vector, an arrow pointing in a specific direction.
9:00The neighbor cosign score measures if all the errors in a specific neighborhood suddenly turn to point in a new direction. It asks, did the whole block change its functional vibe? Ah, so it's like the gentrification of a cellular neighborhood.
9:14The old structural genes moved out and a bunch of heavy lifting drug pumping genes moved in. The individual gene is still there, but the neighborhood's purpose is transformed. A very apt way to look at it.
9:25And finally, the 4th metric is the index gene cosine score. And that measures the gene itself. Right. It measures how much the vector of the specific gene itself has shifted in this space has its fundamental trajectory change.
9:37So Recutter looks at new direct connections, the spread of broader influence, the gentrification of the neighborhood, and the individual movement of the gene itself. You sum those up, and you get a single context drift score.
9:49A score that tells you. This gene has radically changed its job description to keep the cancer alive. Man, that is elegant. And when the team applied this framework to actual living tumors, the results were incredible.
10:02Highly revealing, yes. They didn't start with simple cell cultures. They went straight to mouse models of a really aggressive childhood brain tumor called choroid plexus carcinoma or CPC. CPC. And these are notoriously difficult tumors to treat, just generally speaking.
10:17Very difficult. The blood brain barrier alone makes getting drugs to them a nightmare. They treated these mice with monotherapies initially. Like just radiation. Either radiation alone or a targeted drug called AZD 1390.
10:31That's an ATM inhibitor designed to make radiation more effective by stopping DNA repair. When the tumors eventually relapse, and became resistant to these single, the team ran the single cell data through recode, and the pipeline revealed a massive shift in the gene networks toward DNA repair capacity.
10:49Which makes perfect biological sense. You damage their DNA with radiation, they build a much better, faster repair crew. Exactly. They adapt to the specific threat. But then the team tested a combination therapy, which they called regimony.
11:03Throwing multiple roadblocks at the tumor. What happened then? This is where Ricudo really proved its worth. When the tumors inevitably failed this aggressive combination therapy, Ricodo detected something entirely unexpected.
11:16Not just another shift in DNA repair. No. The context drift scores pointed to a massive increase in trem 2 positive myloid like tumor cell. Hold on, myelade cells. Those are wet blood cells. That's the immune system, not the tumor itself.
11:29Why would a tumorous gene network drift toward an immune profile? Because the tumor cells were literally rewiring their internal networks to mimic immune cells. Wait, really? Yeah. They aren't just putting on a fake disguise.
11:41They are actively secreting signals and creating a localized immunosuppressive fog. Oh, wow. This fog blinds the body's actual immune system and protects the tumor from the drugs. Conventional analysis missed this entirely.
11:55Because if you just looked at the expression volume of these genes. They weren't throwing up massive red flags. They weren't the loudest voices in the room, but their context griff scores. Their network rewiring were off the charts.
12:07That is deeply sinister. The tumor is weaponizing the body's own immune language to create a safe room for itself. It's incredibly sneaky. So Ricody finds this sneaky rewiring, but, you know, finding it is only after the battle.
12:21If we can see this new network, can we just throw a wrench into that specific new machinery instead of trying to poison the whole cell? That is the ultimate goal, and it's exactly what they did. The team triaged the Jains with the highest context drift scores from this regimen E resistant network.
12:35Looking for a weak spot. Looking for vulnerabilities in this new immune mimicking network that could be targeted by existing drugs, drugs that can cross the blood brain barrier. And RICODAR highlighted a vulnerability, a drug called Dasatnip.
12:50That's a multi-kinace inhibitor, right? I think it's usually used for leukemia, right? Correct. And it happened to perfectly hit the critical nodes of these exact shifting networks in the brain tumor. So they added it in.
13:02By adding the saturnib to the combination, they created a highly effective new treatment regimen for these mice. They basically use the tumor's own escape route as the target for the next weapon. It's like finding out the bank robbers are digging a tunnel, so instead of guarding the vault, You just wait at the end of the tunnel.
13:18That is incredible. It's huge step forward. But of course, as we always have to caveat on this show, mice are not humans, we cure cancer in mice all the time. Did the team validate this pipeline on human data?
13:30They did, and the translation to human clinical data is very compelling. They pivoted to human clinical trial data for triple negative breast cancer or TNBC. TNBC is an incredibly aggressive fume of breast cancer, right?
13:45It lacks the 3 common receptors that most targeted drugs attack. Exactly. They use data from the partner trial. Looking at patients before and immediately after neoadjuvant chemotherapy. Just to clarify for you listening, Neoadjumate means the chemotherapy is given before surgery, with the goal of shrinking the tumor down.
14:05Precisely. They wanted to see if Rico Doir could detect context drift in human patients right after that initial wave of treatment before a full relapse even occurred. Could they see the tumor starting to build its escape tunnel, essentially?
14:18Yes. And the results were striking. Wait, okay, I'm looking at my notes on the paper here, and there was a statistic that absolutely blew my mind. Let me guess the overlap stat. Yes. When the team compared the genes with the highest context drift.
14:30The ones recoding our flag to the genes with the most variable expression levels, the ones we normally look at to classify tumors, there was only a one% overlap. one% It is a staggering finding, truly.
14:42And it highlights the flaw in our traditional approach. We were essentially blind to 99% of the action. The genes doing the heavy lifting to resist the chemotherapy weren't getting louder. They were just quietly changing their networks, and our standard tests completely ignored them.
14:56It proves that relying solely on expression levels gives us a dangerously incomplete picture. By focusing on context drift, Rico DR successfully identified a specific subgroup of these TNBC patients, a cohort they called Group B that had incredibly high context drift scores.
15:14Meaning their tumors were massively rewiring under the stress of the chemo. Exactly. Putting them at a super high risk of treatment failure and relapse. And let me guess. The standard clinical markers we use in breast cancer completely missed them.
15:28entirely. Standard clinical markers, like checking for BRCA mutations, or using the PM 50 classification. PM 50, which looks at the expression of 50 specific genes, right? Right. It showed no significant variants that would have flagged these group B patients as uniquely high risk.
15:43Those standard tests are looking for broken genes or loud genes. But only recoders context drift analysis revealed the silent structural danger. Exactly. So what does this all mean? We are talking about a fundamental paradigm shift in how we approach oncology here.
15:58We are moving from a reactive stance waiting to see what mutates, and trying to block it after the fact, to proactively mapping the social networks of these cells as they adapt in real time. If we connect this to the bigger picture, it explains why Recode is such a massive leap forward compared to older computational algorithm.
16:17Right, like WGCNA. Yeah, take WGCNA, for example. It's a well-known tool for finding clusters of correlated genes. It has been a workhorse in genomics for years, but Recode vastly outperformed it in identifying actionable drug targets in this study.
16:31Why is that, though? What makes recoders so much smarter than the older tools? Because older tools like WGCNA often assume fixed correlations based on broad data sets? Recoda doesn't assume anything. It learns the meaning of the gene context directly from the specific tissue being studied in its current state of stress.
16:49So learns the native language of that specific patient's tumor. Exactly. This opens the door for something unprecedented in oncology. Perspective, real-time clinical trial design. You mean taking a biopsy from a patient, running it through recuters, NLP algorithms, and adjusting the drugs on the fly based on how the tumors rewiring that exact week?
17:10That is the horizon we are looking at. You map the escape route as it is being built and you block it before the tumor can fully relocate its operations. sounds like science fiction, but it is right here in the data.
17:20However, I have to assume there are bottlenecks. You can't just throw every drug recooter suggests into a human patient because an AI took a random walk through a graph network. No, you absolutely cannot.
17:33And the authors are very clear about the current limitations of the study. The primary bottleneck right now is biological validation. Recoder is incredibly powerful at generating highly probable drug targets, but testing all of these newly predicted targets in living mouse models is incredibly slow, painstaking, and, frankly, expensive work.
17:51So how do we speed that up? Because the computational site is moving at the speed of light, but the biology side is stuck in traffic. The next step requires massive, high throughput CRISPR screens. We need to go into living in Vivo models and use CRISPR to systematically delete hundreds of these hydrift jeans simultaneously.
18:10By knocking them out all at once, we can see which specific deletions cause the resistant tumors to collapse. That will validate recoder's targets much faster than testing one drug at a time. Furthermore, for the human data, like those group B breast cancer patients, we need much longer term follow-up.
18:29We need years of data to definitively prove that the high context drift Rico detected directly correlates with relapse down the line. Okay, let's unpack this. Are you saying that even if a gene doesn't mutate or change its overall volume, its entire function can shift just because it starts interacting with a completely different group of genes.
18:48Oh, you already covered that. But I want to push back a little here on the philosophy of this approach. Here, go ahead. So what does this all mean for the future? If a tumor's gene network is constantly drifting and rewiring under stress, aren't we just playing an infinite game of biological whack-a-mole, or does recorder actually let us corner the tumor?
19:06It is a vital question. And the answer really lies in the limitations of biology itself. A tumor cell is highly adaptable. Sure, but it is not infinitely adaptable. Because it still has limits. Exactly.
19:19It still has to obey the laws of physics and cellular metabolism. It has a finite number of viable networks, it can rely on to survive, without just collapsing under its own mutated weight. By predicting the drift, by seeing exactly which networks the tumor heavily relies on to escape a specific drug, we can strategically design combination therapies that cut off all viable escape routes simultaneously.
19:42So we aren't just playing whack-a-mole. No, we are surveying the room, finding all the exits and boarding up the doors before we even initiate the attack. Boarding up the doors before you enter the room.
19:51That is a powerful image, and it completely changes the landscape of how we view these resilient cancers. They aren't invincible. They just have a playbook we haven't been able to read until now. And now we have the translation matrix to read it.
20:04To summarize the core takeaway for you listening. Cancer resistance isn't just about individual genes mutating or turning on and off like isolated light switches. It's about genes dynamically rewiring their cellular social networks to survive therapy.
20:21By tracking this gene context drift, scientists can uncover invisible resistance mechanisms and preemptively design smarter targeted combination treatments. It really is about reading the room, not just listening to the loudest voice.
20:35And it leaves us with a fascinating, almost philosophical question to consider as we close. What does this mean for our fundamental understanding of cell identity? If a cancer cell can so completely rewire its internal networks and community to survive our best drugs, at what point is it not even the same disease we started treating?
20:54This episode was based on an open access article under the CCBY 4.0 license. You can find a direct link to the paper and a license in our episode description. If you enjoyed this, follow or subscribe in your podcast app and leave a 5 star rating.
21:08If 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.
21:18Thanks for listening, and join us next time as we explore more science base by base. Bright screens late nights, Charles to never sleep in the wired up maze, where the quiet secrets creep. Not louder jeans, different friends they keep Circus rearrange and run away too deep.
21:59Connection, slip the map, we draws his lines, a heater, switching code, expression signs. If we can read the drift between the nose We can cut the road where the relapse goes. Feel the drift, feel the drift, don't let it take control.
22:13We chase the shadow network back into a hole. the rescue padway, break the comeback script. Newcomers, new time, and watch the outcome, flip, flip, flip, flip, flip, flip, flip, flip, flip. Engine DNA repair on fire.
22:56Oh, forget it, the dreams climbing higher, mule like echoes and the dangerous, quiet, so we aim if the hinges knock the whole fire. Lying up the schedules the pressure tight. One drug alone can't fade and vibe, prepare the right locks with the right key tonight, and the system can learn how to stay in the light.
23:16Feel the drip, feel the drip Now we know the sign. Target the context, where the patterns intertwine. Throw the line through the changes make the future stick. New maps, new chances. Watch the resistance.
23:27Quick, quick, quick, quick, quick, quick, quick, quick, quick, quick, quick, quick, quick, quick, quick, quick, quick.
