Ron Alfa: So we basically opened the lab, we hired a team, we got all the instruments, we started sourcing tumor samples and there was no prior here that any of this would work—like zero. We just started generating data and sourcing human tumors, processing. We built this whole processing pipeline to get the tumors into these arrays and the formats. So you've got like these two-week runs where you're processing two slides and we're just churning data for months and we couldn't even train a model. So we sort of just built all this and then like let's say 18 months later, "Hey, I wonder can we train a model off this?" And then it was not obvious.

Dan Bear: Yeah, there wasn't really like anything major to go off of. I mean there were like transformers developed for single cell data. There just weren't really data sets out there that people had been able to develop on. We do a lot of like custom model building.

R.J. Haniki: Hi there. I'm R.J. Haniki and this is Brandon Anderson. We're the co-hosts of the Late Space Science podcast and today we're really happy to be in the studio with some of the people from Noetic.

Ron Alfa: I'm Ron Alfa, co-founder and CEO of Noetic, physician-scientist by training. My hobbies are making hot takes about AI curing cancer.

Dan Bear: Hi, I'm Dan Bear. I'm VP of AI at Noetic. I'm a biologist by training. I did PhD work in neuroscience and then moved into computational neuroscience, computer vision, self-supervised learning and have been doing AI research at Noetic for the past few years.

Host: Maybe we should start with what is Noetic? Why did you found it? What is the difference between Noetic and the other virtual cell companies?

Ron Alfa: Maybe just start with a little bit of a contrarian thesis, which is really the reason for founding Noetic. We all know the numbers that 90% to 95% of cancer drugs fail in the clinic. Why do they fail? So our thesis is they fail not because we're bad at pharmacology, not because we're bad at target selection. We're actually better at that process than we have ever been in the history of drug development. Most of those drugs fail, we'd argue, is because we're bad at selecting which patients those drugs will work in. And often times you see trials where there is no placebo effect in cancer. Some patients respond to these drugs and if you have a patient that responds, that tells you something that there's some biology that's active there. But you have a problem in patient selection. And so really that's the thesis behind our company: can we build models that can fundamentally understand patient biology from the very beginning and help you position molecules in the right patient population?

Host: So you're actually using the models partly at least to select the patient cohort, not just molecules.

Ron Alfa: And that's sort of the power of the models. Once you've trained these models on patient data, you can use them on both sides of the equation. So you can use them for discovering new targets directly from the patient data which people often refer to as reverse translation. So starting from humans and then trying to understand which targets to go after and then you can use that to develop molecules. But you can also use them directly on patient data. If you have, let's say, a Phase 2 or Phase 3 trial, you can use these models to understand which patients or what underlying biology of the patients in the trial is a predictor of response. And we've been doing a ton of that recently.

Host: Are you doing a lot of like rescuing trials that had a bad effect?

Ron Alfa: We are doing a lot of looking at data from Phase 2, Phase 3 trials and then using the models essentially to run inference on patient biopsies and understand whether there's underlying biology that would help us design the next trial. We haven't shared any of that yet.

Host: So cancer is kind of like infamous in that there are many different types of cancers. Whenever it says like "cure cancer," that is almost a meaningless vacuous statement. So your point is even amongst cancer or a specific type and subtype, there's a bunch of different patient populations that each one of them will respond differently to drugs. And your point is you can figure this out right now—that some sub-population will do well and respond to this drug when you think generally speaking the rest of the population would not.

Dan Bear: Yeah that's exactly right and I would maybe even go further and say like nobody actually knows what the subtypes are. There are cancers that originate in a certain tissue like the lung that have been classified into subtypes based on pathologists looking at them for more than a century. And those subtypes certainly have some connection to the real functional subtypes of disease there. But our thesis is kind of that if you look at the data—a much richer kind of data, the multimodal data that we're generating in our lab—we're going to see that actually what people thought was one subtype of lung cancer is really three distinct subtypes of cancer. And that is going to be critical for figuring out which patient should get which drugs.

Ron Alfa: Whenever you make a new drug, you do a set of experiments in cell culture—cells in a dish. Those cells are often cell lines. These cell lines have existed for 40, 50 years and they're immortalized. So they have genomes that allow them to persist that have abnormal numbers of chromosomes. They have gene expression patterns that don't represent any known cell in like the human body really. These are sort of Frankensteinian cells. So you can do your experiments in these cell lines in a dish or then move these into animal models. But ultimately there's a big gap because they don't translate to patient biology most of the time. So what happens is ultimately no matter what you do pre-clinically, the clinical team says, "Look, we don't really know how to design this trial because none of the data that you've produced gives us any insight on which patients to run." So we're basically going to enroll an open label study—all tumors, all patients—and see where we get signal. Statisticslly, you don't see very many responders and then these molecules get canceled.

Host: So in your Noetic system, you help the pharmaceutical company to characterize patients with a certain genetic or transcriptomic profile, and then you sequence the patient to say "yes, this is a match"?

Ron Alfa: I would say we are even less biased than that. We want the model to learn, let's say from lung cancers, how many different therapeutically relevant subtypes of lung cancers there are just from self-supervised learning from the data. Those subtypes could be driven by large genetic changes, they could be driven by immune changes, they could be driven by any biology that the model is learning in the process of training.

Dan Bear: The biomarkers that people have been using are biased towards simplicity—does the patient have this particular mutation or stain for this single protein. But there's no reason to think that biology or the biology of cancer is that simple. The hypothesis really is here: again, if you were to carve nature at its joints and figure out what's really going on, is there these five subtypes that the correlation there between which patients you give a particular drug and whether you have success is much stronger.

Host: You mentioned the lab, you do a lot of data generation in the lab. Why do you think that that versus using existing public repositories is appropriate?

Ron Alfa: Another hot take I have in AI and bio is you're sort of not at the order of magnitude of data that you are in other spaces of building training models. And so it becomes really hard to brute force these problems just by collecting data. PDB was designed and has been built over the past 50 years. It's not an accident that that data set exists. In bio, you really need to be intentional about the data that you generate and how you generate it. You have to have some foresight around well what are the models we're going to want to train. That's why we've taken this approach.

Dan Bear: A good comparison is to the ImageNet data set which kicked off the deep learning revolution in computer vision. ImageNet is at least 1.2 million images, very carefully curated. These are high quality images, not like random images from the internet.

Ron Alfa: With the data that we're generating, we're around that scale right now. But people have gone much larger in image and language data sets for LLM. So we think that we need to get the data up to that scale before we can really see the meaningful progress on the algorithm side.

Host: What is the data you're collecting? My understanding is you use some pretty specialized instruments and gathering very specific data sets.

Ron Alfa: From first principles how would we do this? So you probably want tissue level biology. 1. Pathology H&E: That's what every patient gets. It's how pathologists identify different cellular structures. 2. Protein Stains (Multiplex Immunofluorescence): We want to have some layer of cell biology, knowing about immune cells and their subtypes. The immune environment of the tumor is a core constituent of whether a patient responds. 3. Spatial Transcriptomics: Spatially resolvable RNA. We get the RNA in a spatially resolved pattern for the same cells. You have between 1,000 or 19,000 different genes. It's like an image except instead of being an RGB image that has three color channels, now all of a sudden it has like 20,000 colors. It's like a very meaty computer vision problem.

Host: What is a "virtual cell"? How do you turn that big pile of data into useful knowledge?

Ron Alfa: One view is we want to be able to simulate all the biochemical processes in a cell. I think that's an interesting intellectual pursuit, but I don't think we have all the modalities of data you would need. I tend to see the virtual cell problem as something more practical. We're trying to make drugs that work in patients. Really what we want to do is understand cell biology in some heuristic that's useful for making drugs. The simplistic thing that we're doing is really just the model can simulate many cells in different contexts and allow you to run some simulations in that regime.

Dan Bear: I think that modeling data that comes from a patient is much more likely to translate to what happens when you give a patient a drug than something that's happening in cell culture.

Host: You not only have patient models, but also a mouse platform?

Dan Bear: We are running in vivo perturbations using a platform based in mouse called Perturb-map. This is a platform for generating highly multiplexed knockouts of individual genes. You end up with mice that have tumors that are barcoded that have a hundred different genetic perturbations in them. We can actually use that to validate our models and ask if what the models are predicting in humans via simulation is actually borne out in a mouse system.

Ron Alfa: We're in silico humanizing the mouse. All the outputs in terms of the transcript from the mouse are in the form of the human genes. Apparently mouse H&E looks enough like human H&E that models think it's perfectly valid H&E. It makes predictions about whether this is immune hot versus cold and those predictions are accurate. It's amazing to me as a biologist that you show it some mouse histology and it's able to say these five different tumor genotypes all look like they have the same phenotype.

Host: You did a big deal recently. I think you have the distinction of being one of the only AI for bio companies that is making money.

Ron Alfa: We announced a deal with GSK where we licensed them OctoVC, our virtual cell foundation model. It's a $50 million deal includes an upfront payment, milestones, and an annual model licensing fee. GSK can use our models to do simulations and therapeutic discovery, but they can also fine-tune the models on their data. What's unique about this deal is the substrate is not a molecule, it's actually a model.

Host: Why do you think there's appetite for this suddenly? It seems like only a year or two ago that bio was dying.

Ron Alfa: I think pharma increasingly want to be able to access models not just for one collaboration on one program—they want to be able to access the technology across the whole pipeline. Startups like us can make the bet that you benefit from generating all of this data in a uniformized way.

Host: For small biotech startups or listeners, do you have any advice?

Dan Bear: Imagine trying to train a foundation model on not enough data. You can't just take something off the shelf and expect that you're going to hit the threshold of GPT-3 like usefulness. It takes some conviction.

Ron Alfa: Everyone should be excited about biology. We're just in the first inkling of the ChatGPT moment for bio, but it's very much just the very beginning. Would love for people to just be more stoked about learning about applications of machine learning in biological sciences because these are the problems that are going to massively impact humanity in the next 10 years.