Protein Diet: Inside The Bezos-Backed Biotech Training AI To Fight Disease

Profluent CEO Ali Madani has raised $106M from Altimeter and Jeff Bezos to generate drug-boosting molecules from new AI models built with the world's largest protein dataset.

Profluent founder Ali Madani sees his startup as his biggest experiment to date. Credit: Profluent

The Upshot

Ali Madani has seen all the chatter about AI being just hype, or failing to deliver real impact inside an organization (hello, infamous MIT report).

In his world, the doubts don’t compute. But Madani operates in an area that’s often siloed from these discussions. It’s wonky enough that we haven’t covered it much in Upstarts, either.

As founder and CEO of Profluent Bio, Madani works every day to use AI to unlock new breakthroughs in that field, biology.

Thousands of commercial and research teams use the AI-based gene editor that Madani and his team shared with the world as an open-source project last year. The startup has recently signed partnerships with companies like Revvity and RSRT to, among other things, find a way to use gene editing to take on Rett syndrome.

---

Help shape Upstarts – tell us about you!

Upstarts has big ambitions for next year. To serve the ecosystem better, we need to know more about you, our readers, and what you value most.

Please fill out this short audience survey. It’s anonymous, takes only a couple of minutes, and will help us do two crucial things: bring you more relevant stories and news; and work better with mission-aligned partners to support our plans. (If you’re interested in sponsorship, you can fill out our interest form.)

---

And AI – not large language models like ChatGPT, but bespoke models that Profluent has built itself using related transformer technology – combined with laboratory testing has proven the big unlock. At Profluent’s Emeryville, CA-based headquarters, engineers sit by a window facing the startup’s wet lab, where colleagues test new proteins hands-on.

“When you generate something from natural language as text, you can just read it, you can get a vibe from it,” Madani says. “Biology isn’t interpretable like that. There’s a real world aspect that’s necessary.”

Like the LLMs, Profluent’s models are working with huge amounts of data. The company claims that its protein catalog, spanning 115 billion unique protein sequences, is the largest of its kind in the world. It equates to about 20 trillion tokens for AI training, a few times more than the amount used to train Google DeepMind’s breakthrough model, AlphaFold 3.