This episode covers Pisces, a machine-learning framework that augments sparse drug-pair datasets by creating multiple modality-based views per drug to improve prediction of drug synergy, xenograft responses, and drug-drug interactions.
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. Imagine for a 2nd that you're trying to find the perfect pairing of ingredients out of 1000000000s of possibilities.
0:16Right, like an impossibly huge kitchen pantry. Exactly. Now, imagine that in this scenario, the right match actually cures a devastating disease. But, you know, the wrong one does absolutely nothing, or worse, it's highly toxic.
0:30Yeah, which is a terrifying thought. It really is. And finding synergistic drug combinations is exactly this. It's a, well, it's a massive life-saving mathematical headache, but there's a fundamental problem here.
0:42We simply don't have enough data to test them all. No, we really don't. So what really happens when we can teach an AI to view a single drug from 8 completely different perspectives at once. Essentially creating high quality training data out of thin air.
0:55How could this change the way we discover multi-drug treatments for cancer? Well, it forces us to completely rethink how machine learning processes sparse biomedical data. I mean, we are moving from a world where we just accept data shortages to one where we uh, engineer our way right out of them.
1:14Today we celebrate the work of Han Win Shu, Jet Cheng Lin, Luai Wang, Sheng Wang, and their colleagues across the University of Washington, UIUC, single University, Microsoft Research, and the Mayo Clinic, who have advanced our understanding of multimodal data augmentation for drug synergy prediction.
1:31Yeah, their research tackles one of the most stubborn roadblocks in modern pharmacology. Which is what, exactly? Well, drug combination therapy is highly sought after, especially for cancer treatment. When you use synergistic drugs, meaning their combined effect is much greater than just adding their individual effects together, you can dramatically reduce cancer resistance.
1:50Because cancer is incredibly adaptive, right? Exactly. Cancer is so adaptive. If you hit a tumor with just one drug, the surviving cells often mutate and find a workaround. You know, they build an escape route.
2:01But hitting it with a synergistic combination blocks the main pathway and simultaneously cuts off those escape routes. But developing machine learning models to predict these perfect combinations, I mean, that is a massive catch 22.
2:15That is the fundamental hurdle here. To capture the complex nonlinear biological effects of how 2 drugs interact inside a specific human cell. You need an expressive, highly complex AI model. Right, a really smart system.
2:31Yeah, but expressive AI models require massive amounts of training data. Without it, they just memorize the small amount of data they do have. Which is what we call overfitting. Exactly. Overfitting. And then they fail completely when you ask them to make new predictions.
2:45And we simply don't have that massive amount of lab data to train them on. We don't, because the search space grows cubically. Cubically. Wait, explain that for the listener. Sure. So you aren't just testing drug A in a vacuum, you are testing drug A, plus drug B, against a very specific cancer cell line.
3:02Ah I see. So if you have a 1000 drugs, the possible combinations against 100s of different cancer cell lines explode into the billions. Wow. Yeah, it is physically, computationally, and financially impossible to experimentally profile all of these combinations in a wet lab.
3:20So researchers try to feed the AI, different types of data, with a paper calls modalities, to give them more context. Right, exactly. But that introduces another massive hurdle. The missing modality issue.
3:31Yeah. In a perfect world, for every single drug, we would have its complete 3D structure, its full side effect profile, its textual description, its genetic targets, and its chemical formula. But I'm guessing we don't live in a perfect world.
3:45We definitely don't. In reality, we rarely have every piece of data for every drug. If a traditional AI model is designed to analyze 5 different modalities, and a drug is missing just one of them. The model often chokes.
3:58It just can't process it. So what do they do? Researchers usually have to guess and artificially fill in the missing data, which, you know, immediately introduces errors. It's like trying to train a chef to create a complex recipe.
4:11But for some ingredients, you only have a photo. For others, you only have a smell, and for some, you just have a written description. That's a great way to put it If the AI demands to taste every single ingredient before it starts cooking, it's going to starve.
4:24So how do we bypass that? Well, that brings us to the methodology introduced in the sources, a novel machine learning model named Pisces. Pisces, okay. Yeah, Pisces takes a completely different approach by using what they call multimodal data augmentation.
4:39Okay, let's unpack this mechanically. How does prices actually work? It starts by looking at a single drug through 8 distinct modalities. Eight of them. Okay, let me guess a few. Uh, chemical formulas, obviously, and maybe side effects.
4:54Those are 2 of them, yep. To be precise, there are 8 specific ways. Pisces represents what a drug is and what it does. Okay, lay them out for us. First, there's smiles, which is a standardized text string representing a molecule structure.
5:07Second, 2D molecular graphs. Third, the actual 3D geometric structures of the molecule. Okay, so a lot of shape data. Right. Fourth, the drug targets, meaning the specific genes or proteins, the drug physically interacts with.
5:23Fifth, textual descriptions, like the summaries you would find in a medical database. Sixth, the known side effects. Ah, there's the side effects. Yep. Seventh, drug sensitivity data, basically how the drug performs independently across different cell lines.
5:39And finally, the drug ontology. Ontology. What does that mean in this context? It's basically how the drug is formally categorized in medical family trees. Okay, that is a wildly diverse set of information.
5:52You have raw chemistry, 3D shapes, human written text, and biological side effects. It's a lot of different languages, so to speak. I have to admit, the math here confuses me a bit. How does an AI even begin to compare a paragraph of text to a 3D geometric shape?
6:07Think of it like mapping a city. You can identify a location using GPF coordinates, or you can use street names or you can use a landmark, the 1st component of Pisces, called the projector, acts like a universal translator.
6:19It takes a 3D shape and a text description, and translates them both into a shared mathematical space. So a set of universal coordinates. So it's translating everything into a master language. It's like taking a witness's written statement and a blurry security photo, translating them both into an abstract sketch, and realizing they're describing the exact same suspect.
6:40is a perfect analogy. How does it actually do that under the hood though? Technically, the projector uses specific neural networks? For the medical text, it uses transformers? Like the tech behind large language models.
6:52Yep, just like that. Much like how an LLM reads sentences to understand context. This transformer reads medical literature to extract the drug's purpose. How fascinating. And for the physical molecules, it uses graph convolutional network.
7:06Brush networks. Yeah, molecules is essentially a graph. The atoms are nodes, and the chemical bonds are the lines connecting them. The network mathematically walks along those bonds to understand the shape.
7:16So everything ends up in that single shared embedding space. Correct everything. Everything gets standardized. Okay, once everything is in that shared mathematical space. Pisces does something incredibly clever to solve our data starvation problem.
7:30The 64 times multiplier. Right. This is the real innovation of what they call the augmenter module. How does it work? Well, let's say you want to predict how drug A and drug B will react together. Pisces doesn't just look at drug A and drug B as single entities.
7:46it has all those different views now. Exactly. Because the projector created up to 8 different views for drug A, and up to 8 different views for drug B, the augmenter pairs them up. It pairs every view of drug A with every view of drug B.
8:00Yes. So 8 times eight. Exactly. It pairs the text description of drug A with the 3D structure of drug B, then the side effects of drug A with the smile string of drug B. Oh, wow. So by doing this pairwise combination, you get 64 distinct augmented views for a single drug pair.
8:17You got it. It effectively expands the original sparse data set 64 times over without needing a single new lab test. Right. It's generating its own training data just by looking at the existing data from different angles.
8:30And this solves the missing data problem perfectly because each of the 64 views is treated as a completely independent instance. Exactly. So if a drug is missing a modality. Let's say drug A doesn't have a known 3D structure.
8:45Right. Then Pisces just skips any pairs that would have used that specific view. It just trains on whatever's available without needing to guess. Yeah, it circumvents the missing modality challenge entirely.
8:56Okay, I have to push back here, though. Sure. If we suddenly multiply our data by 64. Aren't we also multiplying the noise? That is a very valid concern. I mean, what if pairing a drugs side effect profile with another drugs 3D structure yields total garbage.
9:11Not every perspective is gonna be useful for predicting a cancer cure. And that is the most critical vulnerability of data augmentation. If you feed an AI garbage. will learn garbage. Right. So how does Pisces avoid that trap?
9:24That is why Pisces includes a 3rd module called the aggregator. The system does not just average all 64 predictions together plainly. Okay. It uses a concept called noisy label learning. Noisy label learning.
9:37How does that filter out the garbage? Well, it forces the model to be ruthlessly selective. Of the up to 64 predictions generated for a single drug pair, the aggregator only keeps the top high confidence predictions.
9:50How many does it keep? In their experiments, they found that keeping the top 8 works best. So it just tosses out the noisy, low quality combinations. It has to. If you let all 64 voices shout at once, the system gets confused.
10:03That makes total sense. By using a deep learning setup called a resonant architecture. It essentially gives a microphone only to the 8 smartest voices. Ah, a resnet. Yeah, a resonant is particularly good at bypassing clutter and finding the strongest signals in deep neural networks.
10:18It integrates those top 8 signals into one final highly accurate synergy score. Okay, so we have this beautifully engineered system. It bypasses the data bottleneck, translates 8 different languages of chemistry and biology into one, multiplies its own training data by 64, and filters out its own noise.
10:37That's the summary, yes. How did this artificially expanded data actually perform when tested against real biological outcomes? Well, we can view their experiments as a series of escalating stress tests.
10:49Okay, where do they start? They 1st tested Pisces on cellline synergy using the GDSC combo data set. And that's a massive database of drug combinations tested on cancer cell lines, right? Yes, exactly.
11:01And Pisces thoroughly outperformed existing models. But let's talk about the toughest scenarios. Predicting synergy for drugs you've already seen is one thing, but biology is unpredictable. What happened when they asked it to predict the unknown?
11:14They evaluated it across different data split settings to test exactly that. Okay. In the setting where they ask the model to predict synergy for completely unseen drug combinations. Meaning the model had never seen these 2 specific drugs paired together during training.
11:28Correct. In that toughest scenario, Pisces improved the F1 score by nearly 24% over the next best approach. Okay, wait, let's clarify the F1 score for the listener. Sure. In machine learning, an F1 score measures accuracy.
11:42Basically, how well the model avoids false alarms while successfully catching the real threats. So Pisces was nearly 24% better at this balancing act than the best AI we had before. Yeah, and that is a massive margin in computational biology.
11:58Unbelievable. And when they asked it to predict outcomes on entirely new cell lines it had never seen, representing new patient profiles, it still improved the F1 score by over 10%. Okay, so the AI works in a virtual Petri dish with 2 drugs.
12:12But biology isn't a Petri dish, and humans often take more than 2 medications. Very true. What happened when the researchers really tried to break Pisces by throwing a 3rd drug into the mix? So the researchers extended Pisces to predict the synergy of 3 drug combinations.
12:26But wait, you said earlier was trained entirely on 2 drug combinations. It was. Even when trained only on pairs and single drugs, Pisces successfully predicted the synergy of triplets. Wow. It achieved an A-U-R-O-C of 0.8525.
12:41Let's translate AUROC really quick. It's essentially a letter grade for predictive power. A .5 is a coin flip and a one.0 is absolute perfection. Right. Sitting over .85 on a task that wasn't even formally trained for, is like a tennis player spontaneously entering a badminton tournament and making the finals.
13:01That's a fun way to think about it. But yeah, it shows the model genuinely understands the underlying biology rather than just memorizing training data. Do we have a specific example of that? We do. For instance, they looked at the combination of a fat nib and tremitaniv, which is known to be effective.
13:16Pisces correctly predicted that adding a 3rd drug, linsitanib, makes it a highly effective 3 drug combination for colon adenocarcinoma. So because the model understood the underlying modality so deeply, it could triangulate that 3rd variable.
13:31That is wild. But of course, a cancer cell line in a plastic dish is very different from a living, breathing creature. Did they escalate the stress test to animal models? They did. They moved from cell lines to patient derived xenographs.
13:45A xenographed, right, where human tumor tissue is implanted into an animal model, like a mouse, to observe how it grows and reox to drugs in a complex living organism. Yes, and Pisces didn't just work in Petri dishes.
13:58It worked in Vivo. It predicted the minimum tumor volume changes over time with a piercing correlation of 0.47. Okay, a piercing correlation measures how closely 2 trends move together. A .47 in complex, messy and view obiology means the model is successfully capturing the true biological trajectory of the tumor.
14:17Basically, the AI's map matches reality. Here's where it gets really interesting for you listening. Pisces didn't just give a single score of, yes, this works or no, it doesn't. No it didn't. It actually visualized the chronological growth of these tumors.
14:30The system mapped out branches in a 2D space showing tumors growing over time. Clinicians can look at this map and physically see drug resistance developing. Yeah, that visual aspect is huge. As the tumor progresses from an early time point to a later one, you can trace its path.
14:46It is not just a static score. It's a dynamic trajectory. And that temporal analysis is something existing approach is simply cannot do. Pisces can take the time point as an additional input and accurately predict tumor volume at unmeasured future time points.
15:02Wow. It gives researchers a window into how the tumor will evolve days or weeks later. It doesn't just predict what cures us, right? It also predicts what hurts us. Yes, they applied Pisces to drug drug interaction or DDI prediction.
15:15Using data sets like drug bank and 2 sides. Pisces successfully predicted adverse side effects and toxic interactions between entirely unseen drugs. Now performing the competition again. Exactly, out performing competing methods.
15:27It was even used to construct a novel network to infer interactions for new drugs based solely on their class in the drug ontology. Okay, so we have this incredible computational success. But how does it translate to a human being sitting in a hospital waiting room? For the listener.
15:44Let's put this in perspective. Historically, finding a new vulnerability in cancer takes decades of lab work and trial and error. Right. And the researchers provided a profound clinical validation of their model to answer exactly that question.
15:57What did they find? They focused on breast cancer, specifically the BRCA cell lines. They took all the synergistic predictions that Pisces made and ran them through a genetic interaction network. Okay. And through this, they identified a brand new breast cancer drug sensitive pathway.
16:12Wait, a completely new vulnerability in the armor of breast cancer, discovered by essentially doing math on existing fragmented data. Yes. This newly discovered pathway involves 19 different genes. 19 genes.
16:25Crucially, it includes genes from the PI3K and EGFRHR families. Oh, wow. In oncology those gene families are notorious. They really are. They are deeply involved in how tumors grow, divide, and resist therapies.
16:38Finding a pathway that targets them simultaneously is a major discovery. But they didn't just find the 19 gene pathway in the computer and call it a day. They cross-referenced it with real human survival data.
16:51They did. They validated this pathway using the TCGA database, which tracks vast amounts of real world patient records and genomic beta. Okay, and what were the results? They looked specifically at patients with ER positive tumors, estrogen receptor positive tumors, which are a major subtype of breast cancer.
17:10They found that patients who naturally had this specific 19 gene pathway activated lived substantially longer. Let me just pause you right there because we really need to appreciate the gravity of that.
17:21They found a statistically significant link to actual human survival. Both overall survival and disease free survival, yes. The math predicted the biology, and the biology predicted patient survival. Exactly.
17:33It strongly suggests this computationally discovered pathway could be used as a clinical biomarker for breast cancer prognosis. That is incredible. If a clinician sees this pathway is activated in a patient, it opens up very specific, highly effective treatment options that target those genes.
17:49It moves us closer to true precision medicine. As incredible as this is, we always have to look at the boundaries of the science. What are the limitations here? Because an AI that multiplies its own data 64 times sounds almost too good to be true.
18:06Well, it raises an important point about the nature of deep learning, because Pisces generates 64 different views and passes them through this complex aggregator to find the best signals, it operates a bit like a black box.
18:18Meaning we know it gets the right answer. But we don't always know how it arrived. Exactly. It is very hard to interpret exactly which of the 8 modalities was the most important for any given prediction.
18:29Oh I see. Did it rely mostly on the 3D structure or the text description or a combination of the two? The model doesn't easily tell us its reasoning. It's like having a brilliant savant in the lab. It gives you the life-saving answer, but when you ask it to show its work and tell you which of the 64 views cracked the case, it stays quiet.
18:47That is the next frontier. Interpreting those results is definitely a major future goal. Any other limitations? Additionally, the current model only optimizes for one task at a time. do you mean by that?
18:59It predicts synergy or predicts side effects, or it predicts xenograph tumor volume. Ah, so it siloed. Right. Future iterations will likely need multitask learning. The goal is an AI that can predict synergies, side effects, and live tumor results simultaneously, giving a much more holistic view of the drug combination in one single pass.
19:18But even with those limitations, what we have here is a massive leap forward. Pisces bypasses the critical bottleneck of sparse biomedical data by virtually augmenting 8 different drug modalities into 64 distinct analytical views.
19:34Yes, it really is a breakthrough. This allows machine learning to accurately predict complex drug synergies, dynamic tumor growth trajectories, and hidden side effects, even for entirely unseen drug combinations.
19:45It proves that we do not necessarily need more lab data to make breakthroughs. Sometimes, we just need to teach our computational models, how to look at the data we already have from multiple perspectives simultaneously.
19:58We started by imagining the search for the perfect pairing of ingredients out of 1000000000s of possibilities, trying to save the chef from starvation. Pisces doesn't just feed the chef. It changes the way the kitchen operates entirely.
20:10It really does. What does this mean for the future of personalized medicine when an AI can simulate 1000s of combination therapies, predicting exactly how a tumor will grow or shrink over time before you even take a single pill?
20:23This 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.
20:37If 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.
20:46Thanks for listening and join us next time as we explore more science, base by base.
