AI startup skips lab work to speed drug discovery

Potential Sciences, a UK government-backed startup, is taking an unusual approach to drug discovery. Rather than relying heavily on slow and expensive wet-lab experiments, it wants to train AI models on computer-generated biochemical data. The bet is that synthetic datasets can help models explore chemical space much faster than physical testing alone. Drug discovery has long been constrained by the cost and difficulty of laboratory work. Experiments take time, require specialized equipment, and are hard to scale. If models can learn from simulated data, researchers may be able to narrow down promising compounds before committing to real-world tests. That could reduce the number of experiments needed and speed up early-stage research. The approach also reflects a broader interest in using AI to shortcut bottlenecks in science, where data is often scarce or costly to produce. But there is an obvious risk. Simulated data may not reflect real biology. A model trained on computer-generated examples could learn patterns that look convincing but fail when tested in living systems. That concern is long-standing in computational biology, and it is the central question for Potential Sciences. The company will need to show that its synthetic data is useful enough to guide real experiments, not just produce elegant predictions. Support from the UK government suggests that policymakers see potential in the idea, but the science will ultimately be judged by results. If the approach works, it could change how early drug discovery is organized, with AI models doing more of the initial exploration and labs focusing on the most credible candidates. If it does not, it may serve as a reminder that faster computation cannot fully replace biological reality. The startup is therefore testing both a technical method and a strategic question: how much of drug discovery can be moved from the bench to the computer before the real world has its say.