Potential Sciences launched with UK government money to find drugs with AI

A chemist today picks a target protein, runs real lab tests on candidate molecules, and waits weeks for each answer. A new British startup says it can shrink that wait. It wants to generate the chemistry in a simulator. Then it will train AI models on that output.

Potential Sciences launched on 5 November, WIRED reported. The startup has seed funding in the tens of millions of dollars. SovAI is a named backer. SovAI is the UK government's AI venture fund. Twin Path Ventures is also named. The founders would not give WIRED the exact figure or valuation.

What the company says it will do

Potential Sciences plans to build large chemical datasets. It will run physics-based simulations to make them. The model will learn from interactions between atoms, then small molecules, then proteins and amino acids. Once trained, it should pick drug candidates that are likely to work. The real lab test comes after. The company says this cuts the work.

Hatem Helal, founder and CEO, told WIRED: "We want molecules that are never going to work in the clinic to fail fast."

The rest of the AI drug discovery field trains models on the Protein Data Bank, Helal told WIRED. That dataset was built by hand over decades. Helal says it is too small for the task. His bet is that simulated data can be much larger.

Who backs it, with WIRED's numbers

WIRED reports that SovAI was set up earlier this year. It invests in British AI startups. Its remit is to make them parts of the global AI supply chain. The fund has about $800 million. It can give portfolio companies time on the UK's national AI supercomputer. Joséphine Kant, head of ventures at SovAI, has joined the Potential Sciences board.

Potential Sciences joins Basecamp Research and Isomorphic Labs in SovAI's AI drug discovery bets, WIRED says. CuspAI, in materials discovery, is also in the fund.

Where WIRED stops short

Two scientists in the piece sound a note of caution. Peter Wolynes of Rice University told WIRED that protein engineering today is "at the level of the Romans". Alan Fersht of Cambridge said accurate simulation would be a change. He believes it "will eventually happen". Neither claim about Potential Sciences' own results has been published.

Three facts WIRED does not settle. The exact round size. Who else invested. And how Potential Sciences plans to earn money. Helal told WIRED that part has not been decided.

The practice reading, for a buyer

A team buying software that promises simulation output should ask a simple question. How was the training data made? Is it a measured record, or a simulated one? Potential Sciences is clear that its data is synthetic. Any vendor that mixes the two without saying so is asking the buyer to guess.

A team that does its own lab work has a second question to ask. Even a very good screening model does not change the cost of safety and dose testing in animals and people. The design step is cheap next to the test step. Wolynes told WIRED the design cost is small compared to real-world testing cost. That second cost does not move with better AI.

Source

WIRED, "Finding New Drugs Is Laborious. This AI Startup Is Skipping The Lab Work".