Software-inspired medicines are getting closer to prime time
The Economist
Science & technology
Dec 20, 2024 | 758words | ★★★★☆
On average, it now takes ten years and more than $2bn to develop a drug from start to finish. The risks are also steep—less than one-tenth of drug candidates that enter clinical trials are approved by regulators. In the past decade America’s Food and Drug Administration has approved an average of just 53 drugs every year. Proponents of artificial intelligence (AI) say it has the potential to make drug discovery faster and cheaper—something that would be welcomed by pharmaceutical companies and patients alike. In 2025, as a wave of new drugs nears regulatory approval, AI will begin to fulfil that promise.
The process of developing a drug begins with identifying a target, such as a protein or gene, that is associated with a particular disease. Researchers then search for a molecule that can either block or enhance the target’s activity. Once potential molecules are found, they are tested for safety and effectiveness first using computer models, and then on animals. This phase, known as the preclinical stage, can involve screening as many as 1m compounds before selecting just one or two promising candidates. All this can take years, and accounts for nearly a third of the cost of drug development, even before any human testing.
It is in this preclinical phase that drugmakers are using AI to boost their chances. The pharmaceutical industry has used computational models for decades, but AI is changing drug discovery in several ways. First, it improves researchers’ understanding of diseases and their targets by analysing huge amounts of disparate data. Software can also pinpoint promising molecules and fine-tune their structures to boost their success in human trials. Generative AI can go a step further, by dreaming up entirely new molecules to test.
In 2020, AlphaFold 2, a model made by DeepMind, an AI laboratory owned by Google, stunned the scientific world by accurately predicting the structure of nearly every protein in the human body. In May 2024 its successor, AlphaFold 3, expanded its capabilities to other molecules that make up living things: proteins, DNA, RNA and small molecules called ligands. Such models are changing drug development by reducing months of trial-and-error experiments to just hours of computation. Insilico, a startup that uses AI to develop drugs, says its software identified a new drug target and designed a molecule suitable for human trials in only 18 months and at a cost of $2.7m—a fraction of the usual time and expense required.
Breakthroughs in the preclinical phase have now progressed into the business end of drug development. BCG, a consultancy, estimates that about 65 AI-inspired molecules are currently in human trials (see chart). Around a third of these are in the second phase of clinical trials, in which the drug is tested for effectiveness and side-effects. Firms must then decide whether to move forward with more expensive phase-three studies on a larger population. Less than a third of drug candidates make that leap.
During 2025, results from this crucial second phase will be reported for more than half a dozen drugs. Some AI-designed drugs have already stumbled at this stage. Benevolentai and Exscientia, two promising British AI startups, recently reported disappointing results in clinical trials for their drugs targeting eczema and cancer. Despite these setbacks, Christoph Meier of BCG believes that AI-based methods could double the productivity of R&D. And it seems likely that four or five AI-developed treatments, if not more, could go on to phase-three trials in 2025.
This is a small sample, but these medicines point to a profound change in drug development. Although AI has yet to shorten clinical-trial timelines, it is already helping pharmaceutical companies make smarter choices about which molecules to take forward, reducing failure rates and cutting costs. Research by Andreas Bender at Cambridge University, for example, shows that reducing failures in phase-two trials by just 20% could save nearly $450m on a single drug’s development. In computing, Moore’s law has run out of steam. In the pharmaceutical industry, Eroom’s law may soon face a similar fate. ■
新英文外刊
每日精选优质外刊文章
扫码关注我们
