Vertu wants executives to pay $6,880 for an AI agent — here’s how it actually performs

A recent competition evaluating artificial intelligence models designed to predict how the human body processes drugs has revealed that larger, more complex AI systems do not necessarily yield better results. The findings, released after a contest focused on the prediction of drug metabolism through the pregnane X receptor (PXR), suggest that the field of AI-driven drug discovery is entering a new phase, one where efficiency and precision may outweigh sheer computational power. The competition, which tested AI models against their ability to predict whether a drug candidate would interact with the PXR receptor, highlighted a shift away from the previous focus on solving grand challenges like protein folding. In 2020, the Critical Assessment of Structure Prediction (CASP) competition brought AlphaFold into the spotlight, leading to its creators receiving a Nobel Prize. At the time, accurately predicting protein structures was considered nearly impossible, yet AlphaFold achieved remarkable success. However, the current emphasis in AI research appears to be moving toward practical applications in drug development. The PXR receptor plays a crucial role in drug metabolism. When activated, it triggers the production of an enzyme known as CYP3A4, which is responsible for breaking down roughly half of all currently marketed drugs. If a drug candidate activates PXR prematurely, it can lead to rapid elimination from the body or dangerous interactions with other medications. Traditionally, drug developers identify these issues late in the development process, often requiring costly revisions or even abandoning promising compounds altogether. An AI capable of accurately predicting PXR activation early on could significantly streamline this process. The competition attracted entries from multiple institutions and companies working on AI models tailored for drug discovery. While some teams opted for large-scale neural networks with millions of parameters, others developed smaller, more specialized algorithms. Surprisingly, several of the top-performing models were relatively modest in size compared to their counterparts. This outcome suggests that the complexity of an AI system does not automatically translate to superior performance in real-world scenarios involving biological data. Researchers involved in the competition emphasized that the results reflect a growing understanding of the limitations of deep learning approaches in pharmacology. Many of the best-performing models incorporated domain-specific knowledge, such as biochemical pathways and molecular interactions, that traditional AI systems often overlook. By integrating this information directly into their architectures, these models demonstrated greater accuracy in predicting PXR responses than purely data-driven approaches. The competition also underscored the importance of interpretability in AI models used for drug discovery. Larger models, while powerful, tend to function as black boxes, making it difficult for scientists to understand why certain predictions are made. Smaller models, particularly those designed with explainable AI techniques, offered clearer insights into the decision-making process. This transparency is critical in regulatory environments where approval agencies require detailed justifications for clinical decisions based on AI outputs. Industry experts noted that the findings could influence future directions in AI research. Rather than pursuing ever-larger models, researchers may begin focusing on optimizing existing frameworks for specific tasks. This approach aligns with broader trends in AI, where specialization and efficiency are increasingly valued over raw computational capacity. For pharmaceutical companies, the implications are clear: investing in AI tools that provide accurate, interpretable predictions could accelerate drug development and reduce costs associated with late-stage failures. As the field continues to evolve, further competitions and studies are expected to refine the criteria for evaluating AI performance in drug metabolism. Researchers are already exploring hybrid models that combine the strengths of different approaches, aiming to balance predictive accuracy with computational efficiency. These efforts represent a pivotal moment in the integration of AI into healthcare, where practical outcomes may ultimately determine the success of technological advancements.