A new startup founded by a former Google DeepMind scientist is exiting stealth with $50 million in funding.

Latent Labs is building AI foundation models to “make biology programmable,” and it plans to partner with biotech and pharmaceutical companies to generate and optimize proteins.

It’s impossible to understand what DeepMind and its ilk are doing without first understanding the role that proteins play in human biology. Proteins drive everything in living cells, from enzymes and hormones to antibodies. They are made up of around 20 distinct amino acids, which link together in strings that fold to create a 3D structure, whose shape determines how the protein functions.

But figuring out the shape of each protein was historically a very slow, labor-intensive process. That was the big breakthrough that DeepMind achieved with AlphaFold: it meshed machine learning with real biological data to predict the shape of some 200 million protein structures.

Armed with such data, scientists can better understand diseases, design new drugs, and even create synthetic proteins for entirely new use-cases. That is where Latent Labs enters the fray with its ambition to enable researchers to “computationally create” new therapeutic molecules from scratch.

Simon Kohl (pictured above) started out as a research scientist at DeepMind, working with the core AlphaFold2 team before co-leading the protein design team and setting up DeepMind’s wet lab at London’s Francis Crick Institute. Around this time, DeepMind also spawned a sister company in the form of Isomorphic Labs, which is focused on applying DeepMind’s AI research to transform drug discovery.

It was a combination of these developments that convinced Kohl that the time was right to go it alone with a leaner outfit focused specifically on building frontier (i.e., cutting-edge) models for protein design. So at the tail-end of 2022, Kohl departed DeepMind to lay the foundations for Latent Labs, and incorporated the business in London in mid-2023.

“I had a fantastic and impactful time [at DeepMind], and became convinced of the impact that generative modelling was going to have in biology and protein design in particular,” Kohl told TechCrunch in an interview this week. “At the same time, I saw that with the launch of Isomorphic Labs, and their plans based on AlphaFold2, that they were starting many things at once. I felt like the opportunity was really in going in a laser-focused way about protein design. Protein design, in itself, is such a vast field, and has so much unexplored white space that I thought a really nimble, focused outfit would be able to translate that impact.”

Translating that impact as a venture-backed startup involved hiring some 15 employees, two of whom were from DeepMind, a senior engineer from Microsoft, and PhDs from the University of Cambridge. Today, Latent’s headcount is split across two sites — one in London, where the frontier model magic happens, and another in San Francisco, with its own wet lab and computational protein design team.

“This enables us to test our models in the real world and get the feedback that we need to understand whether our models are progressing the way we want,” Kohl said.

While wet labs are very much on the near-term agenda in terms of validating Latent’s technology’s predictions, the ultimate goal is to negate the need for wet labs.

“Our mission is to make biology programmable, really bringing biology into the computational realm, where the reliance on biological, wet lab experiments will be reduced over time,” Kohl said.

That highlights one of the key benefits to “making biology programmable” — upending a drug-discovery process that currently relies on countless experiments and iteration that can take years.

“It allows us to make really custom molecules without relying on the wet lab — at least, that’s the vision,” Kohl continued. “Imagine a world where someone comes with a hypothesis on what drug target to go after for a particular disease, and our models could, in a ‘push-button’ way, make a protein drug that comes with all of the desired properties baked in.”

In terms of business model, Latent Labs doesn’t see itself as “asset-centric” — meaning it won’t be developing its own therapeutic candidates in-house. Instead, it wants to work with third-party partners to expedite and de-risk the earlier R&D stages.

“We feel the biggest impact that we can have as a company is by enabling other biopharma, biotechs and life science companies — either by giving them direct access to our models, or supporting their discovery programs via project-based partnerships,” Kohl said.

The company’s $50 million cash injection includes a previously unannounced $10 million seed tranche, and a fresh $40 million Series A round co-led by Radical Ventures — specifically, partner Aaron Rosenberg, who was formerly head of strategy and operations at DeepMind.

The other co-lead investor is Sofinnova Partners, a French VC firm with a long track-record in the life sciences space. Other participants in the round include Flying Fish, Isomer, 8VC, Kindred Capital, Pillar VC, and notable angels such as Google’s chief scientist Jeff Dean, Cohere founder Aidan Gomez, and ElevenLabs founder Mati Staniszewski.

While a chunk of the cash will go toward salaries, including those of new machine learning hires, a significant amount of money will be needed to cover infrastructure.

“Compute is a is a big cost for us as well — we’re building fairly large models I think it’s fair to say, and that requires a lot of GPU compute,” Kohl said. “This funding really sets us up to double-down on everything — acquire compute to continue scaling our model, scaling the teams, and also starting to build out the bandwidth and capacity to have these partnerships and the commercial traction that we’re now seeking.”

DeepMind aside, there are several venture-backed startups and scaleups looking to bring the worlds of computation and biology closer together, such as Cradle and Bioptimus. Kohl, for his part, thinks that we’re still at a sufficiently early stage, whereby we still don’t quite know what the best approach will be in terms of decoding and designing biological systems.

“There have been some very interesting seeds planted, [for example] with AlphaFold and some other early generative models from other groups,” Kohl said. “But this field hasn’t converged in terms of what is the best model approach, or in terms of what business model will work here. I think we have the capacity to really innovate.”