Will AI transform drug development? Researchers believe it hinges on its application.

      The possibility of employing artificial intelligence in drug discovery and development has generated both enthusiasm and doubt among researchers, investors, and the broader public. Some companies and scientists assert that “artificial intelligence is taking over drug development.” In recent years, the increasing interest in utilizing AI for drug design and optimizing clinical trials has led to a rise in research and funding. AI-driven platforms like AlphaFold, which was awarded the 2024 Nobel Prize for its capability to predict protein structures and create new ones, highlight AI’s potential to speed up the drug development process.

      However, some industry experts caution that AI in drug discovery is “nonsense,” emphasizing that “AI’s potential to hasten drug discovery requires a reality check,” as no AI-generated drugs have proven capable of addressing the 90% failure rate experienced by new drugs in clinical trials. In contrast to its success in image analysis, the impact of AI on drug development remains ambiguous. For each medication available in pharmacies, numerous others have failed during development.

      As pharmaceutical scientists with backgrounds in both academia and the pharmaceutical sector, as well as experience as a former program manager at the Defense Advanced Research Projects Agency (DARPA), we contend that AI in drug development is neither a groundbreaking innovation nor entirely without merit. AI is not a magical solution that can transform any idea into a successful drug. Instead, we perceive it as a tool that, when utilized effectively and competently, might help tackle the fundamental reasons behind drug failures and enhance the overall process.

      Most efforts involving AI in drug development aim to decrease the time and financial investment required to bring a drug to the market, which currently takes 10 to 15 years and costs between $1 billion and $2 billion. But can AI actually revolutionize drug development and enhance success rates?

      Researchers have implemented AI and machine learning at every phase of the drug development process. This ranges from identifying targets within the body to screening potential drug candidates, designing drug molecules, predicting toxicity, and selecting patients who may respond most favorably to drugs during clinical trials.

      From 2010 to 2022, 20 AI-focused startups identified 158 drug candidates, with 15 progressing to clinical trials. Some of these candidates managed to complete preclinical tests and enter human trials in just 30 months, in contrast to the usual 3 to 6 years. This achievement demonstrates AI’s ability to expedite drug development.

      Drug development is a lengthy and expensive endeavor.

      Conversely, while AI platforms may swiftly pinpoint compounds that work effectively on cells in Petri dishes or in animal studies, the potential success of these candidates in clinical trials—which is where the majority of drug failures occur—remains highly uncertain.

      Unlike other domains with ample high-quality datasets available for training AI models, such as image recognition and language processing, AI in drug development is limited by smaller, lower-quality datasets. Creating extensive drug-related datasets involving cells, animals, or humans for millions to billions of compounds is challenging. Although AlphaFold represents a significant advancement in predicting protein structures, its accuracy in drug design is still in question. Even slight alterations in a drug’s structure can significantly influence its performance in the body and, consequently, its effectiveness in treating diseases.

      Survivorship bias

      Like AI, historical advances in drug development such as computer-aided design, the Human Genome Project, and high-throughput screening have improved individual stages of the process over the past four decades, yet drug failure rates have not seen an improvement.

      Most AI researchers can successfully address specific tasks within drug development when given high-quality data and definable questions. However, they often lack a comprehensive understanding of the overall drug development landscape, simplifying challenges into pattern recognition issues and refining individual steps. Meanwhile, many scientists experienced in drug development do not have a background in AI and machine learning. This gap in knowledge can inhibit collaboration and prevent scientists from moving past the mechanics of existing processes to identify the fundamental reasons behind drug failures.

      Current drug development strategies, including those incorporating AI, may have fallen victim to a survivorship bias, focusing excessively on less critical elements while neglecting major issues that lead to failure. This scenario is comparable to repairing the wings of returning aircraft from World War II while ignoring the critical weaknesses in the engines or cockpits of planes that did not return. Researchers often direct too much attention on enhancing individual drug properties instead of addressing the root causes of failure.

      While aircraft that experience wing damage may return, those suffering from engine or cockpit damage are far less likely to make it back.

      The current drug development methodology operates like an assembly line, employing a checklist approach with extensive testing at each process stage. While AI may reduce the time and costs associated with the lab-based preclinical phases of this assembly line, it is improbable that it will enhance success rates in the more expensive clinical phases that involve human testing. The ongoing 90% failure rate of drugs in clinical trials, despite four decades of improvements to the process, highlights this limitation.

      Addressing root causes

      Failures of drugs in clinical trials