Drug discovery involves searching an immense range of possible molecules and testing which ones might work against a disease. British startup Potential Sciences wants to make that search faster by using physics simulations to create chemical data for artificial intelligence systems, reducing reliance on lab experiments during early discovery.
Building data through simulation
Many AI drug discovery efforts build on the Protein Data Bank, a public collection of 3D atomic structures that researchers have expanded over decades with experimental results. Potential Sciences founder and CEO Hatem Helal argues that the available dataset is too small to train models that can reliably speed the search for promising drug candidates.
The company plans to generate synthetic data by simulating how chemical compounds interact. It says the work will begin with basic physical principles applied to atoms, then extend to small molecules and more complex amino acids and proteins. Potential Sciences hopes that a larger pool of simulated interactions will help it train systems that predict how candidate molecules behave in human biology.
Helal describes drug discovery as a search through an enormous space of possible molecules. In his account, the challenge is to navigate that space and eliminate options that are unlikely to work. Better predictions, he says, could help unsuitable candidates fail earlier, before researchers spend more time and resources on them.
Potential gains, with a major scientific hurdle
If the simulations and predictions prove useful, the approach could help researchers identify targeted candidates more quickly. Helal also argues that lowering experimentation costs could make it more practical for pharmaceutical companies to pursue treatments for rare conditions that have gone untreated.
Potential Sciences has also described a longer-term possibility: medicines made from molecular combinations tailored to a person's biology. That prospect depends on the technology accurately predicting how molecules bind to disease-related proteins. The article presents that capability as an open question, and achieving it would require a significant improvement.
Researchers see promise in the underlying idea. Alan Fersht, an emeritus professor of chemistry at the University of Cambridge who specializes in protein engineering, says accurate calculations of molecular interactions and structural fluctuations could reveal where molecules bind and transform drug development. His optimism is about what such simulations might eventually enable; it does not establish that the startup's system can already make those predictions.
Peter Wolynes, a professor of chemistry at Rice University, expects improvement but describes current protein engineering as still far from its potential. His comments underline the gap between generating more data and reliably predicting how a candidate will behave.
Discovery is only one stage
Even if AI helps researchers find promising candidates faster, it cannot remove the later work of testing safety and effectiveness. The article notes that testing in animals and people still takes time and money, and Wolynes argues that figuring out what a drug will do in those settings can be a larger bottleneck than designing it.
That distinction matters for expectations. Faster candidate selection could change an early stage of research without ensuring that a treatment reaches patients sooner. Each candidate still needs to be evaluated for its effects, and a prediction about a molecular interaction is not itself proof that a drug is safe or effective.
Funding, competition and responsibility
Potential Sciences launched on November 5 and says it has seed funding in the tens of millions of dollars from SovAI, the UK government's AI venture capital fund, and Twin Path Ventures, among other investors. The founders did not disclose the precise funding figure or the company's valuation. SovAI has roughly $800 million at its disposal and can offer portfolio companies access to the UK's national AI supercomputer.
The startup joins other companies pursuing AI-based drug discovery, including Basecamp Research and Isomorphic Labs, while CuspAI is focused on materials discovery. Potential Sciences will compete with entrepreneurs and academics trying to address similar research bottlenecks through different approaches. Its physics-based focus and access to computing resources may support its work, but they do not settle whether its predictions will outperform alternatives.
There is also a safety concern. Tools that help identify effective treatments could potentially be turned toward designing dangerous chemicals. Helal acknowledges that technology capable of reducing toxicity could also be used to increase it. The company has not settled on a revenue model or decided whether to make its models freely available, in part because of the risk of misuse.
For now, Potential Sciences is pursuing a clear but difficult aim: use simulated chemistry to give AI more data and improve the search for drug candidates. Whether that leads to better treatments depends on the accuracy of its predictions, the results of later testing and how the technology is managed.