Building AI Models of Human Cell: Tahoe Therapeutics on Databricks
Summary
- Gigascale cell biology data: This video showcases how Tahoe Therapeutics processes massive single-cell datasets (100M cells across 60,000 conditions) to understand drug-patient interactions at the cellular level using Databricks.
- AI drug discovery: Learn how machine learning predicts drug efficacy and resistance, enabling Tahoe to identify better therapeutics faster and validate drug-patient matches before clinical trials.
- Data-driven therapeutics: Discover how the Tahoe-100M atlas and AI models built on Databricks set a new standard for scalable, data-driven drug discovery, accelerating the path to better treatments for cancer patients.
Building AI Models of Human Cell: Tahoe Therapeutics on Databricks

Talk by Nima Alidoust, Co-Founder, CEO, Tahoe Therapeutics
Discover how Tahoe Therapeutics (formerly Vevo) is generating gigascale single-cell data that map how drugs interact with cells from cancer patients. They are using that to find better therapeutics, and to build AI models that can predict drug-patient interactions on Databricks. Their technology enabled the landmark Tahoe-100M atlas, the world’s largest dataset of drug responses-profiling 100 million cells across 60,000 conditions. Learn how we use Databricks to process this massive data, enabling AI models that predict drug efficacy and resistance at the cellular level. Recognized as the Grand Prize Winner of the Databricks Generative AI Startup Challenge, Tahoe sets a new standard for scalable, data-driven drug discovery.
Databricks Named a Leader in the 2025 Gartner® Magic Quadrant™ for Data Science and Machine Learning Platforms: https://www.databricks.com/blog/databricks-named-leader-2025-gartner-magic-quadrant-data-science-and-machine-learning
Build and deploy quality AI agent systems: https://www.databricks.com/product/artificial-intelligence
See all the product announcements from Data + AI Summit: https://www.databricks.com/events/dataaisummit-2025-announcements
FAQs
What is the Tahoe-100M atlas and why is it significant?
This video presents the Tahoe-100M atlas as the world's largest drug-response dataset: 100 million cells profiled across 60,000 different drug-condition combinations. This massive dataset enables AI models to predict how drugs interact with cells from diverse cancer patients, accelerating drug discovery.
How does Databricks enable this scale of data processing?
The video demonstrates that Databricks' lakehouse architecture and distributed computing enable Tahoe to process gigascale single-cell data efficiently. The platform scales from data ingestion through AI model training, making it feasible to analyze datasets far larger than traditional systems could handle.
How does this AI approach improve drug discovery?
Traditional drug discovery is slow and expensive. By predicting drug efficacy and patient-drug interactions using AI models trained on massive cell data, Tahoe can identify promising therapeutics faster, validate them computationally, and accelerate the path to clinical trials and patient benefit.
What does Tahoe's success demonstrate about Databricks?
Tahoe's Grand Prize Win in the Databricks Generative AI Startup Challenge shows that Databricks' platform is the foundation for cutting-edge AI applications beyond traditional data analytics. This video illustrates how startups can build transformative AI solutions using Databricks' infrastructure.
Full transcript
I'm co-founder and CEO of a company called Tahoe Therapeutics. Um, I'm going to give you a huge overview, an small overview of a lot of the work that we are doing at Tahoe, the data we are generating and the AI models we're building on top of the data. Uh, first I'll tell you a little bit about us. We are accompanied with this uh uh with this small vision of mapping how all of chemistry which basically are drug molecules interact with cells from uh from the from very diverse patients from heterogeneous patients.
So basically what what we're doing is that we are mapping how chemistry perturbs biology and the reason we're doing that is to find new cures and better cures for a variety of uh diseases that are afflicting us humans. uh and we are starting with cancer for that. I'm going to take a huge step back and provide a framework for us to be able to comprehend a little bit the kind of work that we're doing. Um around three decades ago uh we started collecting a huge amount of information, data, text, uh images, videos uh on the internet.
So this became a corpus of human knowledge that we have uh we have collated and we are growing it over and over um for 30 for 30 plus years. Um the thing that we got really good at uh towards the end of the '90s was uh searching through this data and we have gotten even better than that over the last 20 25 years where we have been able to index all of this information. Uh so the information is really at the tip of our finger.
Anytime we want to query it, we want to find some nugget of information that's relevant to the task that we want to um to we want to accomplish, we just search it. We go to the Google and we search it and we can find the information. And as you know, search has gotten a lot better over the last three decades. But one inflection point that started to happen around two years ago, three years ago was that not only were we now capable of um searching through the data that we have collected on the internet, now we are capable of uh going beyond it.
We can actually learn from that data. We can train AI models on that data and use that to go beyond whatever that's already within the data, whatever that's contained within the data that has been accumulated in the internet. And that's the whole revolution associated with language models particularly with the advent of chat GPT and equivalents uh after the uh after the launch in 20 late 2022. Now to take a step back basically what we have done over the last 30 years is what we started with an internet that we built all of us together.
We started you know getting better at searching through the internet and eventually we started to find a way to venture beyond to extrapolate beyond uh what actually is contained in the internet. But again everything that we have accomplished the cornerstone of that has been the internet. The reason we have been able to do that is that we have collected all of this human knowledge in the form of text videos and u and images over the last 30 years.
Now imagine that we have an internet of biology. Um an equivalent kind an anal analogous kind of uh data set that allows us to map um how different genes in different cells from very different humans uh actually function, how they drive disease and how those kinds of biological entities interact with chemical modules which are which are drugs. Basically imagine we have that kind of an internet and basically uh imagine we can do the same kind of things that we accomplished with the text data and the and the image data using this biological data sets that would basically allow us to search within this kind of corpus of biological data to find ideas around how how different genes malfunction and drive disease. how different cells actually you know for example become cancer cells and they they they act in an undesirable way and what are the drug molecules out there that are the chemical entities that we know of that we can use in order to cure them not only can we use that data set for this kind of search I'm outlining outlining again very similar to the kind of AI models that we have been training on language on language over the last few years maybe we can also train AI models on this internet of biology so that we can extrapolate beyond it we can actually find new chemistry, new chemicals that could act as drugs.
Chemicals that we have never synthesized. They have never been in the physical world before. It's a great vision, but sadly the reality is we actually the reality is that we just don't have that interal biology today. The reason for it is that generating that data I was talking about basically how drugs interact with different cells from different patients, it's a very unscalable thing. It's the experiments to run that you have to run in order to generate that data is very tough to do.
And it's they're very tough to scale. Now the to give you a sense of the numbers the information that we have on the internet particularly on the text side the number of tokens that has provided for us in order to train those language models is in the order of tens of trillion tokens. For biological data the drug perturbed you know cells of patients the amount of data that we have is in the order of 10 10 billion. So it's another billions of tokens that we can actually use in order to train these models.
And that basically means that we are in a in a 10,00x gap when it comes to biological data that can help us um you know find new dis find new uh new drugs for new diseases and also to extrapolate beyond them for new chemistry. That's exactly where we come into play. So the core technology that we have invented uh for at Tahoe is the ability to scale those kinds of experiments that map how different uh types of chemistry how different types of um drug molecules interact with cells from very diverse sets of patients and that's really the way that's that's really the technology that has allowed us to scale and actually take a huge step towards building this internet of biology.
In fact, we put it to work towards the end of last year and we generated what's called Taho 100, which is the largest data set, the largest atlas that maps how different drugs interact with cells from different types of patients. It contains 100 million single cell data points. It's equivalent to hundreds of billions of tokens. Um, and these are cells from different patients. It's around 50 different patients. And these cells have been treated by one of 12,000 1,200 uh different treatments.
This corpus of data again hundreds of billions of tokens already. This corpus of data is 50 times larger than all of the publicly available preservative data sets that are out there. Basically data sets that have been measuring how drugs interact with cells uh combined and that those data sets have been generated over the last of over the last decade. We generated this data set in a matter of around two months. So this is a completely new way of doing things.
This is a completely new um kind of paradigm in terms of generating the data that could form the basis of what we call the internet of internet of biology. We are so confident about our ability to generate this data at massive scale that we actually entirely open source open sourced the Tahoe 100 data set and that turned out to be uh a very seminal moment in in biology. This happened around two months three months ago now. uh and since then 45,000 times this data set has been downloaded.
This is a first for the bi the biological data sets of this this scale and on hugging face which is the place that people share data data sets we have been on the top 0. 05 percentile of uh trending data sets out there and this is basically competing with data sets that are being released by anthropic but by open AI by Nvidia in the text uh and generally in the language image and video video domain. So for the first time we have generated biological data sets that are actually providing the same kind of power that other modality of um other modality of data that has enabled that have enabled uh can can actually accomplish.
Now we talked about the internet and how it led us to the revolutions and search and then later for AI models that can extrapolate beyond the information that's contained in the in the data. So let's actually replicate that for biology. So now that we are getting closer to having that internal biology, can we step towards the same kind of um same kind of ambition and that's what that's what gets me to the exciting announcement that we have to make today.
We are releasing what's called Taho Dive. We partnered with a fantastic team uh from a company called Kepler AI. It's a new startup with a very very superstar team X data bricks folks that are building this this this amazing solution. and we we teamed together to build this natural language interface an AI agent that allows you to interact with this massive data set. I told you about 45,000 times that people have been downloading this data set over the last uh last couple of months.
Most of the people who have been downloading it are ML scientists and bionformmaticians. They are the people who are hardcore programmers. They know actually what they're doing. They can actually download I don't know a data set of um one terabyte data set and manipulate it and query it. But what we wanted to do was to get get this to the hands of cancer biologists and other types of biologists that can actually use this data to drive to hunt for drugs and hunt for targets for drugs.
And that's why we built that's why we built Taho Dive. With Tahoive what you can do is you can um basically use the lang natural language prompt and then you get agents to go and analyze the data, query the data, slice and dice it and also visualize the results. But not only that they they add the context of the entirety of the scientific literature together with the kinds of discoveries that you are making by searching through the data set that we have generated.
What I'm going to do is to give you a demo of how this actually works. This is a recorded demo because we want to make sure this actually goes smoothly. Um first I'm going to give you a little bit of a science talk about you know some small corner of cancer that we have been dealing with. So how do we empower cancer biologists to use uh Tahoe 100 to speak to Tahoe 100 and come up with new cures for disease? Let's talk about uh cancer cells a little bit.
Cancer cells are very snitchy. Uh the one of the ways that they avoid getting killed by our immune system is by reducing the levels of a protein that's called M MHC1 at the surface of the cells. That protein is very crucial because that's basically a tag that appears on the surface of the cell and it tells the immune cells come and kill this cancer cell. Okay. So cancer cells reduce that protein levels. What if we actually find drug molecules that can increase the levels of that protein so that it makes it visible to the immune cell to the immune system and immune system can come and kill those cancer cells.
Now this is exactly what we did like that's literally like we we went to the agents and we said what are the drugs that can increase the level levels of MHC1 but also in addition um that there is another pathway they can actually uh activate in order to increase the efficacy of that stimulation. So we asked the agent to do to find that for us and what I'm going to do right now is to actually go to the agent and I just started from the beginning.
We asked that agent to that question I was just talking about. This is the playback of the analysis that it did that we ran five days ago. What it is doing uh actually let me let me let it finish and then it's thinking right now. It's analyzing. It's doing a lot of code generation. It's reasoning. It's doing a lot of very fancy stuff visualization in the back end. And I'm going to go through after it's done. I'm going to go through exactly what it did in a second.
Now summarizing the findings. So what what the agent is doing is basically first it starts by doing like what any software engineer or any bioformmatician would do. It basically loads the data to a notebook. It creates a notebook. It loads the data to that and then what what it starts doing is to synthesize the question that we have given it. It basically says okay these are the gene sets I have been I have to look at uh for ifn alpha the other kind of pathway I was talking about and then MC MHC1 and after it does that it knows where to go and slice and dice the data and then it goes and evaluates the different drugs and the impact that they have on this um on this particular gene sets and what it does is that eventually it gathers all of the information to give you a sense of what are the drugs that activate both of these different pathways basically what are those drugs that can increase the MHC1 um MHC1 expression and also they can activate interferon interferon alpha which is the other pathway that we were just talking about now what's extraordinary is that the drugs that it came up with um this is this is really incredible there were some drugs that were already known to be doing that they were in the data set and it actually came up and said okay these drugs they're called mech inhibitors they they they can increase MHC1 expression.
So this is a very powerful positive validation. So the agent is actually working. But another thing that it did, it actually came up with new drug compounds that were previously not known to have this kind of effect on the cells from the patient. So basically through this agent, we are discovering new drug compounds that will have the desired impact we have on cancer cells so that they can become visible to immune cells and immune cells can come and kill them.
I'm not going to talk about that part of this this analysis. But what's really cool here is that um it discovered things by digging into a data set by generating code and analyzing it and visualizing it. But it also put it in the context of the broader literature. it actually came up with some ideas of what is the more like level two uh kinds of um um level two kinds of uh mechanisms by which the drug these drugs are are going to be working in order to accomplish the task that we wanted to accomplish.
So this is what the agent came up with um and what we did was actually um separately what we had done was we went and tested this in the lab. Sorry, we went and tested this in the lab and so separately we did some manual bionformatics work and we came to the same conclusions from the data. We analyzed the data and we said okay that drug actually works as well through orthogonal analysis and then we went to the lab and in completely independent experiments we validated that.
We have written a blog post that we published last week around on exactly that topic basically using the Taho 100 data set to come up with new imunom modulators which is what the fancy word for what I was just talking about and you can go on on our blog blog. ai to read about that but this is an extraordinary validation of what a scientific agent that has access to this beautiful massive uh very unique data set can actually accomplish.
The agent right now is going to be open uh for uh for beta use. So we are giving access to folks who are more on the scientific side and on on cancer biology and biology side. Go to taho bio. ai/taho dive if you want to get in line for getting access to it. And uh you can I'm happy to actually we're going to stay here me and my colleagues who are sitting in the front row here both from Kepler and from Tahoe. Happy to answer questions do a few more demos for you all to see how the how the agent actually works.
We talked about the internet. So we talked about the internet and we said that first we start searching through it and then we start to learn from it. This is very imperative in the biological world. No matter how many data points we we generate, how many atlases beyond Tahoe 100 we generate, there are still going to be of course much bigger than anything else that's out there, but they're going to be much smaller than the entirety of the universe of the chemical space. there are 10 to the^ of 60 chemicals that you can imagine in um in the universe and we're going to be we are going to fall short of that.
The way to solve that is to extrapolate beyond the data in the same way that language models extrapolated beyond uh beyond the internet data that we are feeding them and that basically means training what's what are called foundation models of cell biology. We are doing that uh in the company what we are doing is that we are getting inspired by the models that have been developed in natural language. the same models that fit child GP etc. in order to train these single cell foundation models that learn how genes function in the presence of other genes in different biological context.
Um in natural language what we were doing is that we were masking some words and we were trying to train a model that infers what the masked words are based on what are the other words that are in the sentence. Here what we are doing is that we are uh basically masking some genes the expression of some genes and we are trying to get the model to infer what the expression of the those genes are based on how the other genes are being expressed and through that we are learning the function of genes.
We are learning how genes function in different types of contexts. Again we didn't have the luxury um in in biology to have access to an internet that provides us the training sets. But now that we are getting closer to having the training data sets, we actually have the luxury of using the models and the architectures and the kind of technology that has been developed to train these kinds of models in order to apply to the to the biological data sets.
The problem here is that these models are very expensive to train and it's not like you train one model and it's going to work. you actually have to train hundreds, thousands of different models to find the architectures that are more suitable for the information content that exists in your data and also hyperparameters that you need to use in order to get uh you know performance out of your model. So initially when we started for a 1 billion parameter model it was costing us around $500,000 to uh to train a to train a model.
That's when we started our collaboration with data bricks. That's where we started talking to the folks from Mosaic ML side of things and we started to really use some of the some of the techniques and some of the algorithms they had developed and the tools that they had developed for natural language and we translated them to the to the models that we are building the in the biological setting and over over I think in the span of around two three months we managed to reduce the cost of training from $500,000 per per model to around 25,000 models.
So it actually gave us the ability as a small startup to train hundreds of different models rather than just you know uh flipping a coin and training one model and hoping it it actually works well. This allowed us to train to train the largest and most performant foundation model of human cell. This is a model that we are going to be releasing um later this month. And this is a model that is again equivalent to GPT models.
But what it's doing is that it's uh it's it's learning the hidden structures within the biological data that we are feeding feeding it. And then we can use it in order to uh deploy towards tasks that are more relevant for what we care about which are drug discovery which is finding new drugs for new targets of disease that were previously unknown. For us this is just the beginning. We have generated hundreds of billions of tokens to train these kinds of models.
And we are searching through this to in order to find therapeutic insights. But over the next few years, we are going to be generating a huge amount more of this data. Much of it is going to be propri prop proprietary to us. And we are using the tools that have already built have already been built for the natural language world. We are adopting them, but we're also building new tools that allow us to digest this information and allow us to train models that can help us extrapolate beyond it.
I'm very uh I'm very thankful for you all to be here and uh and listen to our talk. Again, um you can uh you know go to our website, see more about what we are doing. Go to taho. ai/toodive to uh to get on the weight list to use the model. Uh and also uh please come talk to us uh me and my friends from Keeper and Tahoe to tell you more about what we are building at Tahoe. Thanks so much for your attention.
Learn more about the Databricks Data and AI platform.
The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.