Growing experience suggests that artificial intelligence could potentially transform how cancer research is designed and conducted
Early-phase clinical trials are really the high-stakes trials in oncology - and in medicine in general - where decisions are made based on a few dozens of patients and are key for drug development.
When artificial intelligence (AI) enters this research area, it is being applied in a domain that is not only high stakes but also very high risk. A key question for researchers now is: how can AI help us?
Despite the topic is currently very hyped and a lot of funding have been invested, we need to remember that AI is not magic - AI is still science. So, some of the solutions it may offer are already grounded in good evidence, while others are still conceptual.
Finding the right patients is the biggest challenge for early phase clinical trials because having no patients – and not having the right one – may ‘kill’ your research. Currently, researchers do manually, trial matching but AI is well suited to this type of task, especially large language models (LLMs). There have already been a number of studies published such as TrialGPT which uses an LLM and AI techniques to automatically match patients to the right trials (Nat Commun. 2024 Nov 18;15(1):9074). Whilst the outcomes are not perfect, performance is quite good when compared to human experts and, importantly, it was shown to reduce waste of time (43% screening time less). Unfortunately, real-world prospective validation has found that the precision and recall were not that good (Nat Commun. 2026 Mar 25;17(1):4472).
Another very hot topic, which is however not yet mature, is the potential to derive some AI endpoints. An example in this area is the AI-Based Histologic Measurement of NASH (AIM-NASH), a tool for non-alcoholic steatohepatitis that is based on deep learning and on computational pathology. It can estimate the degree of liver disease and, very importantly, has been qualified by the U.S. Food and Drug Administration (FDA) in 2025 as an AI-derived endpoint that can be used in clinical trials to demonstrate efficacy of NASH drugs.
This story means that potentially there is a way forward for endpoints that are really derived from AI. That would be, of course, another proof of concept that AI can help us design our clinical trials and move more quicker towards predicting the efficacy of medicines. But still, we have to remind ourselves that biomarkers always need to be validated as highlighted by the ESMO Basic Requirements for AI-based Biomarkers in Oncology (EBAI).
Something that we have to be careful about is the fact that LLMs are very good at analysing text. In early phase clinical trials, they can typically analyse all clinical datasets and flag potential adverse events or signals that need to be detected at very early stages.
Resilience, a spin-off from the Institut Gustave Roussy, Villejuif, France, that has now been deployed in many countries, is an app collecting patient-reported outcomes, with an integrated AI analysis that could potentially detect these early adverse-event signals more quickly and efficiently than researchers, which is particularly important in early-phase trials. While patients already use commercial chatbots, we, as academics, need to develop specialised chatbots that can genuinely improve how patients are monitored in early phase clinical trials.
Another hyped area is now digital twins and synthetic data, which may have a very important role in phase 1 trials. Ideally, when you run a phase 1 trial, you would like to have a control arm. Historical controls have several limitations, but now we have technologies that can help us design digital twins, i.e. virtual models of patients. These in silico representation of patients can be used to simulate counterfactuals, the effects of different drugs, and then assess the outcomes. This may look like magic, but still we have to be sure that there is some science behind it, such as accumulating enough data to train those models so that good and representative synthetic patients can be generated and used as control arms for appropriately matched comparison. A digital twin is really a counterfactual model on which different treatments can be tested and designed, potentially identifying the best treatment for a patient based on this simulation.
Despite the hype, these are proof-of-concept models, which are unlikely to be generalisable when they are not properly designed and trained. Also, the risks of hallucinations – i.e. the model inventing some false data in a very confident way – and bias are always present, especially with LLMs.
Can we, at some point, envision adaptive trials in which each patient is assigned to the right arm in a personalised way with dosing being adapted individually? The idea was conceptualised by some Canadian colleagues, who proposed a ‘prismatic’ approach, a patient-centred trial in which dosing and treatment allocation are adapted based on the patient's multimodal data, measurements, and AI (Cancer Cell. 2025 Apr 14;43(4):597-605). Another example is ADAPT/ARPA-H, from U.S. researchers, which is a blueprint for a learning cancer treatment system where you collect data and adapt the trial according to tumour biology and evolution (Cancer Cell. 2026 Mar 9;44(3):449-454). The model loops, on which decisions are made, could be, at some point, closed by AI itself. An AI agent, in fact, is a LLM that can reason, plan, decide and take action: potentially, it can decide what could be the best treatment based on all those multimodal data. I think that agentic AI maybe will help us take the right decision based on all multimodal data that are analysed by AI.
These are very powerful tools that we should count on. However, AI is not our co-investigator yet and prospective validation remains key to the progress of the field.
[Extract from Dr Julien Vibert’s lecture ‘AI for early-phase clinical trials’ at the Educational session ‘Recent advances in the application of artificial intelligence (AI) in early drug development’, ESMO TAT Asia 2026.]
