One of the holy grails of biology is digitally simulating a living cell. If researchers can use computers to more accurately understand how new medicines would react in the body, that could give them greater confidence when they’re tested on animals and humans.

But while large language models have led to breakthroughs in modeling how proteins act, applying the same technology to simulating all the complexities of an entire cell hasn’t been as fruitful. There’s simply not enough data.

But in February of this year, a startup named Tahoe Therapeutics got one step closer to that goal with the release of Tahoe-100M, a collection of 100 million different datapoints showing how different kinds of cancer cells responded to interactions with over 1,000 different molecules. This type of data–called pertubations–is crucial to training AI models, because information on how cells respond to various molecules improves an algorithm’s ability to predict how they’ll be affected by others.

“We believe that the Tahoe-100M was a Mars landing moment for single-cell datasets,” Tahoe CEO Nima Alidoust, 39, told Forbes.

The company was able to build this dataset less than three years after it was founded thanks to its Mosaic platform, which lets the company take “cells from many different types of patients, from all different organs and then put them together,” rather than the conventional techniques, which test cells from only one individual at a time, explained CSO and cofounder Johnny Yu. “So every time we run an experiment, we're generating massive single cell atlases of which drugs affect which patients.”

“Our core superpower is the ability to generate the massive datasets required for virtual cell models,” Alidoust said. The ability to scale that data production quickly, he added, is Tahoe’s “distinguishing factor” compared to other companies working on AI for drug discovery. It's also foundational to the company’s own goal of building virtual cell models and using them to find new treatments for cancer and other diseases.

Today, Tahoe announced it has raised $30 million in new venture funding, led by Amplify Partners. Other investors include Databricks Ventures, Wing Venture Capital, General Catalyst, AIX Ventures, Mubdala Ventures, Civilization Ventures and Conviction. The investment brings the company’s total funding to $42 million and its valuation to $120 million.

Poor AI predictions have been a source of constant frustration in biotech, said Krish Ramadurai, a partner at AIX Ventures who also sits on Tahoe’s board. “These AI algorithms keep recommending stuff, and then when you go to test it in the wet lab, it all sucks,” he said. The data Tahoe can generate, he said, makes a crucial difference for the accuracy of new models.

Just a few months after Tahoe published its 100 million point dataset, the non-profit research organization Arc Institute released an open-source virtual cell model, State, which used Tahoe-100M as part of its training data. When benchmarked, Arc found that it has twice the accuracy of other AI models–and also beat out the simpler machine learning programs that had previously trounced other foundation models.

That’s a testament to nearly a decade’s worth of work for Tahoe cofounder Yu, 34, who developed the underlying technology for Mosaic, while working in the lab of biochemistry and biophysics professor Hani Goodarzi at the University of California San Francisco.

Alidoust first met Goodarzi, 41, when they were classmates at Princeton. The pair reconnected in 2022 to discuss the idea of founding a company to build virtual cell models. Goodarzi said that an essential piece of such a company would be large-scale data collection, so he brought in Yu.

A month later, the three of them cofounded Tahoe–then called Vevo Therapuetics–along with UCSF researcher Kevin Shokat, 60. The company raised a $12 million seed round in December 2022. The name was changed from Vevo to Tahoe in April of this year after a legal challenge.

With new capital in hand, Tahoe is now focused on building a dataset with over a billion single-cell datapoints to power its own virtual cell models. Armed with its own models and proprietary data, the company is accelerating the development of new medicines to fight cancer. Alidoust said Tahoe currently has a drug candidate against “a major cancer subtype” with which it’s conducting the studies required by the FDA to start testing on humans.

Additionally, Alidoust said, although the company intends to keep its larger datasets proprietary, it does plan to select either a major pharmaceutical company or AI company to share data with. The idea would be to collaborate on either developing new medicines or new drug discovery AI models, giving Tahoe “more shots on goal” for gaining revenue. That partner hasn’t been selected yet, he said, but it is currently working with different companies on smaller projects.

In the meantime, he said, the company will keep working on generating more data for its AI models and proving out its technology. “We say in the company that this is morning in biology,” he said. “We are building. And we hope others are going to build with us as well.”