Digital Life | AI is revolutionizing pharmaceutical research

(Monrovia, California) The Terray Therapeutics lab is a hive of miniaturized automation. Robots carrying tiny tubes of fluids whir to their stations. Scientists in blue lab coats, sterile gloves and safety glasses monitor the machines.

But the real action happens at the nanoscale: proteins are mixed with chemical molecules on silicon chips, the bottoms of which are shaped like muffin tins. Every interaction—there are millions of them every day—is recorded, generating 50 terabytes of raw data per day, the equivalent of more than 12,000 feature films.

Masses of data

The vast lab produces a wealth of data for AI-powered drug research in Monrovia, California. Terray Therapeutics is part of a wave of startups using artificial intelligence (AI) to develop better drugs faster.

AI learns and improves from massive amounts of data. Its use could move pharmaceutical research from artisanal to automated production.

“Once you get the right data, AI can get really, really good,” says Jacob Berlin, co-founder and CEO of Terray.

Just as consumer AI bots—like ChatGPT and DALL-E—feed on text and images found on the internet, pharmaceutical AI runs on data. This is highly specialized data: molecular information, protein structures, and measurements of biochemical interactions. AI detects patterns in the data to suggest drug candidates: it’s like finding the chemical keys to protein locks.

Because pharmaceutical AI is powered by scientific data, the “hallucinations” common among consumer robots are much less likely. Additionally, any drug candidate undergoes rigorous testing in the lab and in clinical trials before approval.

Terray and other companies are building large, high-tech labs to generate the information to train AI, which enables rapid experimentation, identifying patterns and predicting what might work.

Molecules invented by AI

AI can then digitally design a drug molecule. This design is translated, in a high-speed automated lab, into a real molecule that is tested for interaction with a target protein. The results – positive or negative – are recorded and fed back into the AI ​​software to improve the next design, speeding up the overall process.

Developing a drug takes time and a lot of money, whether it works or not. Studies estimating the total cost of the process—from lab to clinical trials to approval—vary widely. The average is about $1 billion and takes 10 to 15 years. Nearly 90 percent of drug candidates that enter human clinical trials fail, either because they are ineffective or have unexpected side effects.

New pharmaceutical AI companies are banking on their technology to reduce risks, costs and delays.

They are most often funded by the pharmaceutical giants, long-time partners and lenders of small research firms. Typically, these companies take care of the preclinical development of new molecules (which takes between four and seven years without AI). Occasionally, some pharmaceutical AI companies try to bring these molecules to clinical trials themselves (a very expensive process that takes years more), but generally, this is where big pharma takes over.

Smaller AI pharmaceutical companies are paid as subcontractors by big pharma for milestones they reach, which can amount to hundreds of millions over several years. And if a drug is eventually approved and marketed, they are entitled to royalties on sales.

Terray, Recursion Pharmaceuticals, Schrödinger, Isomorphic Labs and others are looking to make breakthroughs. But there are basically two different paths: build big labs or don’t.

Isomorphic, a spin-off of Google DeepMind, Google’s AI subsidiary, believes that the better the AI, the less data is needed. And it’s banking on its software prowess.

In 2021, Google DeepMind released software that accurately predicts the shapes that amino acid chains will take to make proteins. These 3D shapes determine how a protein will function. This has improved understanding of biology and helped with drug discovery, since proteins determine the behavior of all living things.

Predictive models

Last month, Google DeepMind and Isomorphic announced that their latest AI model, AlphaFold 3, can predict how molecules and proteins interact, taking drug design a step further.

“We are focusing on the computational approach,” says Max Jaderberg, head of AI at Isomorphic. “We think there is huge potential to be exploited.”

Terray, like most drug development startups, is the product of years of scientific research combined with recent discoveries in AI.

Terray specializes in small molecule drugs (pills that are convenient to take and inexpensive to produce).

Terray’s impressive lab shows how far we’ve come from the days when data was hand-compiled into Excel spreadsheets and automation didn’t exist: “I was the robot,” says Kathleen Elison, chief scientific officer and co-founder of Terray Therapeutics with the Berlin brothers.

Terray has entered into partnerships with Bristol Myers Squibb and Calico Life Sciences, a subsidiary of Google parent Alphabet that focuses on age-related diseases. Terms of the deals were not disclosed.

To grow, Terray will need more money than the $80 million in venture capital it has already raised, says Eli Berlin, Jacob Berlin’s younger brother. He left a job in venture capital to found Terray, where he serves as CFO and COO. He says he believes AI in pharma can be very profitable.

Terray is developing new drugs for inflammatory diseases, including lupus, psoriasis and rheumatoid arthritis. Berlin says he hopes to have some drugs enter clinical trials in early 2026.

Innovations from Terray and other pharmaceutical AI firms can speed things up, but only to a certain extent.

“The real test for our industry will come in 10 years: will we have a better success rate and will the pharmacopoeia have better drugs?” Berlin said.

This article was published in the New York Times.

Read the original version (in English; subscription required)