Redesigning oncology clinical trials with agentic AI

Over the last 30 years, critical evolutions in oncology have mirrored those in technology, and so it is with artificial intelligence (AI), which is increasingly transforming the oncology research landscape. Current agentic AI models are already capable of performing multiple different tasks with varying levels of precision and accuracy in a semi-autonomous manner. The technology continues to move at an unprecedented pace: yesterday’s state-of-the art has already been superseded, and the degree of complexity that can be accommodated multiplies month on month.

Last year, the collaboration between a large pharma company and specialised AI companies led to the creation of the largest multimodal foundation model in oncology based on electronic health records, laboratory tests, pathology images, pathology reports, radiology images, radiology reports and unstructured notes from the patient hospital journey. Patient ‘embeddings’ extracted by the self-supervision models help to address a range of important questions, such as predicting who might best respond to a particular treatment and, crucially, how to provide these potential benefits at scale using a system widely accessible to healthcare providers.

These tools can be used to help improve the next generation of treatment options for patients in key tumour areas, leveraging real-life data to inform not just the development of novel therapeutic agents but also ways to combine new and existing therapies. A key benefit of the latest AI systems is that they allow accurate generalisation on relatively small sets of data, thereby facilitating progress at a fraction of the time, workforce and ultimately cost involved with conventional routes of research and development. By streamlining clinical trials and reducing the risk of failure in the transition from phase II to phase III trials, these tools effectively remove the need for large-scale recruitment and can help speed up regulatory processes. Models trained on many millions of patients are already being utilised and indicate potential for early cancer detection and toxicity prediction.

In 2024, the first ever predictive biomarker for an antibody–drug conjugate based on pathology-applied AI, and involving TROP2 normalised membrane ratio, was shown to be effective in directing treatment with datopotamab deruxtecan in patients with non-small cell lung cancer enrolled in the TROPION-Lung01 trial (J Thorac Oncol. 2024;19(Suppl):S2–S3). The model uses the Quantitative Continuous Scoring (QCS) platform, a computational pathology fully supervised AI solution that allows for identification of cancer cells and target protein expression level quantification within the subcellular compartments (i.e., membrane, cytoplasm and nucleus) of cancer cells at single-cell resolution. The development of QCS has been substantially accelerated through the implementation of state-of-the-art frontier foundation models to reduce or eliminate the need for handcrafted annotations during algorithmic development, shortening the development time from an estimated 2.5 years to roughly 5.5 months. These AI-based solutions are highly agile, allowing for the development of multiplex biomarker solutions and the next generation of multimodal biomarkers.

The integration of domain-specific AI models into oncology R&D relies on a complex ecosystem of symbiotic partnerships that recognise the different needs and build on the individual strengths of the various contributors. Pharmaceutical companies and centres without specialist AI knowledge can externally source suitable domain-specific self-supervised foundation models from specialised AI companies and use them to reinforce, fine-tune and augment proprietary data. The expertise of academic institutions, which are at the forefront of research, is integral to the success of networks and the payback is that the relationship will help to accelerate the progress of their own programmes. Finally, consultation with patients and advocacy groups is essential to ensure that the patient’s voice is incorporated at the earliest stages of drug development.

AI is not the future – it is the present. By learning how best to harness its power, we can help increase the speed and efficacy of personalised care delivery.