Potential Sciences wants to speed up drug discovery with AI

British startup Potential Sciences aims to speed up drug discovery with AI models trained on computer simulations of molecules. The company wants to filter out unsuitable candidates before costly lab tests. It launched on November 5 and raised tens of millions of dollars from government fund SovAI, Twin Path Ventures and other investors. The founders have not disclosed the exact amount or the company’s valuation. The approach could help find treatments for rare diseases, but it is not yet clear whether the models can predict drug-protein interactions accurately enough or how much faster new drugs will move through testing.

British startup Potential Sciences wants to discover drugs with AI

Simulations instead of some lab experiments

Potential Sciences plans to build large chemical datasets using physical principles. The company intends to simulate interactions first between atoms, then between small molecules, and later between complex amino acids and proteins. It wants to train AI models and prediction systems on this data to find drugs for specific diseases.

The company’s founder and CEO, Hatem Helal, describes the goal this way: molecules that will not work in clinical practice should be ruled out quickly. That requires more accurate predictions of how they interact with human biology. If the approach works, pharmaceutical companies may find it more worthwhile to develop drugs for rare conditions that have previously gone untreated.

Over time, the company hopes to develop more personalized drugs: combinations of molecules selected to account for the biology of a particular person.

Why the company needs new data

Many labs working on AI-driven drug discovery train models on the Protein Data Bank, a public database of three-dimensional atomic structures. Researchers have been adding experimental results to it for decades. Helal says the data is not yet sufficient to reliably speed up the search for promising drugs.

He describes the scale of the task in terms of the number of possible combinations of atoms: by his estimate, the atoms in the universe could be assembled into around 10⁶⁰ molecules that could potentially be given to a person. Finding a drug in that space comes down to how quickly promising candidates can be identified and the rest ruled out.

That is why Potential Sciences wants to generate synthetic data by simulating interactions between chemical compounds. The larger the dataset, the more complex the models and prediction systems the startup expects to train.

Scientists see promise in this approach, but acknowledge the scale of the accuracy required. Alan Fersht, emeritus professor of chemistry at the University of Cambridge and a specialist in protein engineering, says that accurately calculating interactions between every molecule in a cell and changes in their structure would make it possible to predict where molecules bind and would transform drug development. He hopes this will become achievable over time.

Peter Wolynes, a professor of chemistry at Rice University, expects progress but says protein engineering is still at an early stage compared with what could ultimately be achieved.

SovAI funding and computing resources

The startup raised tens of millions of dollars in seed funding from SovAI, the UK government’s venture fund, Twin Path Ventures and other investors. The founders have not disclosed the exact amount or Potential Sciences’ valuation.

SovAI’s portfolio also includes drug discovery companies Basecamp Research and Isomorphic Labs, as well as materials discovery startup CuspAI.

The British government established SovAI earlier this year. The fund is meant to invest in local startups that could become important links in the global AI supply chain and give the UK geopolitical leverage. It has around $800 million at its disposal; portfolio companies may also get access to the national AI supercomputer.

Josephine Kant, SovAI’s head of venture capital and a member of Potential Sciences’ board, says the startup’s focus on the physics of chemistry sets it apart from other companies. She says access to computing resources will make it possible to run calculations at a scale that would otherwise be out of reach for a small company.

The main constraints

The AI drug discovery market is already crowded, with many entrepreneurs and scientists trying to tackle the same problems from different angles. Even with support from the British government, Potential Sciences will face strong competition.

Rapidly identifying promising candidates does not eliminate the time and expense of testing their safety and effectiveness. Wolynes says the cost of developing a drug is incomparable to the expense of finding out how it affects mice, seriously ill patients and healthy people. So faster design may not be the main factor determining how soon a drug reaches the market.

There is also a risk of misuse. Tools that can identify molecules with low toxicity could also be used to find highly toxic substances, Helal says.

Potential Sciences has not yet decided how it will make money from the technology it is developing or whether it will make its models freely available, given the risk of misuse. Helal considers drug discovery one of the important problems worth trying to solve with AI, although the field has already seen many promising developments that have not produced notable results.

Dataist's take: Simulations could expand the search space and reduce the number of candidates that reach lab testing. But they do not remove drug development’s main constraints: testing for safety and effectiveness, as well as the risk that the same methods could be used to create dangerous substances. Questions also remain about the accuracy of the predictions and Potential Sciences’ business model.