Modern oncology is transforming as AI and other technological advances redefine how we understand and treat cancer. The ESMO AI and Digital Oncology congress (held 12–14 November in Berlin, Germany) launched in 2025 to help address this rapid evolution. PSI’s Head of Feasibility and Site ID, Kasia Moscicka, and Associate Director of Feasibility, Juan Ignacio Terraza, provide their thoughts on the highlights of the congress and what AI in oncology means for the future of clinical trials.
Over three days, more than 1,000 participants and 40 speakers gathered in Berlin to discuss cutting-edge research, ethics, and emerging innovations. Understanding the intersection of oncology and technology is crucial to advancing cancer research. Read on to discover our highlights, with a focus on what you should know about AI and digital advances in oncology for your clinical trials.
Clinical trials for AI, AI for clinical trials
Nearly 100 AI trials (primarily in medical imaging and almost half evaluating cancer) were reported between 2018 and 2023 [1]. Beyond research, trial participant perception of AI remains high. For example, an AI-enabled platform delivered significantly higher participant engagement and preventive action rates compared to traditional in-clinic consultation.
AI in oncology trial design
- Tools summarizing literature and extracting efficacy and safety data from abstracts to write background information for trial protocols were demonstrated. Such tools reduce the burden of developing these important sections and improve the precision of the protocol
- Similarly, models and systems generating and evaluating eligibility criteria for oncology trials were also presented. Initial in silico data from one presentation showed that many common exclusion criteria had minimal effects on hazard ratios, meaning sicker patients would be eligible, thereby increasing trial diversity
- AI-powered trial matching platforms were also evaluated at ESMO AI – while not a new development, new modalities were described. These included an open-source AI platform for cancer precision medicine trial matching, summarizing patient EHR, clinical trial eligibility criteria, and ranking for similarity. Another model evaluated when potential patients would be “ready” for a clinical trial (i.e., would fulfil the disease and treatment criteria to enroll)
- Advances in synthetic data development were showcased. Based on real-world data, whole records or specific parameters can now be generated with high fidelity and without the risk of disclosing sensitive data
We are in the ‘Age of Agents.’ Clinical trials can move at the speed of the data and aren’t limited by how fast paperwork can be completed. AI tools amplify human expertise and reduce site burden, allowing research to be faster, safer, and more innovative
AI in oncology drug design
There have been over 67 trials of AI-designed drugs in the past 10 years. AI-designed drugs achieve an astonishing 80–90% success rate in phase 1 trials [2]. Some are already in phase 3 trials, and more are expected in the coming years – from small molecules to antibodies and vaccines.
The cost of running clinical trials for rare diseases or subtypes may steer new discoveries towards cancers with broad biomarker expression. The FDA’s new “plausible mechanism pathway” [3], announced a few days before ESMO AI, can help bypass certain clinical trial data requirements.
AI is reshaping oncology by redesigning clinical trials – helping us predict which therapies are likely to succeed, and refining drug targets with unprecedented precision. The FDA’s new pathway, in particular, will bring life-changing therapies to patients most in need of rapid treatment.
Multimodality and pathology tools in cancer research
Multimodal AI (models capable of integrating different data formats, such as images and text) were center stage at ESMO AI. Developments in multimodal AI will impact screening and tumor boards. For example, traditional trial matching is already being surpassed in research using multimodal AI involving Large Language Models (LLMs), tools evaluating biomarkers, and image analyses. This could reduce some of the burden we see at our sites in matching the right patients to the right trials.
Multimodal AI approaches in oncology are beneficial in trials of tumor-agnostic therapies – generalist models can be trained across multiple cancer types, further driving research toward treating patients based on genetics and biomarkers rather than location or tissue type.
AI models are moving from algorithms to essential tools for research and clinical decision-making, but they must be thoroughly validated to ensure trustworthiness and generalizability. ESMO showcased a range of pathology tools that represent groundbreaking developments in image analysis. With continued research and development, these tools have the potential to make a significant impact. Some pathology tools discussed included:
- PathChat: An FDA breakthrough device-designated visual language assistant using CONtrastive learning from Captions for Histopathology (CONCH) to describe abnormalities in pathology slides
- Transformer-based pathology Image and Text Alignment Network (TITAN): A multimodal model enhancing whole-slide image (WSI) analysis with UNI and CONCH frameworks
- MADELEINE: A machine-learning model using tissue stain relationships for robust WSI representation, integrating pathology imaging and various stain types
- TANGLE: A foundation model employing transcriptomics to enhance H&E-stained WSI visual representation
- Apollo: A comprehensive healthcare system model using multimodal data for clinical and biomedical innovation
Open-source developments were also showcased. These included generative neural network-based histology foundation models and open-source microscope workstations [4-6]. Open-source tools can help reduce trial costs and are more customizable. They often lead to greater innovation and improved data quality, as many researchers contribute improvements from their own perspectives and specialties. These tools can also adapt more quickly to changes in regulatory requirements.
Regulation of AI-powered oncology drug development
Oversight and regulation of AI are increasingly on everyone’s mind.
Use of LLMs in clinical trials
ESMO released its Guidance on the Use of Large Language Models in Clinical Practice (ELCAP) prior to ESMO AI [7]; various discussions and debates over the weekend served to score how important guidelines on transparency, data protection, and clinical responsibility with AI are.
ESMO’s LLM practical guidance concerns:
- Patient-facing applications for education and symptom support, complementing supervised clinical care.
- Healthcare professional-facing tools to aid in decision-making, documentation, and translation.
- Background institutional systems integrated into data management systems (such as EDCs) for data extraction, summaries, and screening activities, among others.
LLMs are not a substitute for direct patient interactions – they help reduce administrative burden and complement clinical care, including tasks such as data extraction and trial matching. Human oversight of processes and professional clinical decision-making remains paramount. Training LLMs is crucial for producing accurate, appropriate, and human-comparable responses.
Requirements for AI-based biomarkers in oncology
ESMO also published its Basic Requirements for AI-based Biomarkers in Oncology (EBAI) [8], categorizing AI biomarkers by how they were developed and their clinical use:
- Class A: For quantification of a score. The AI-based method is new, and the data is subsequently compared to a previously validated score (e.g., pathology scoring) for confirmation.
- Class B: For indirect measurement of existing biomarkers. The input and method are new – for example, EGFR prediction from H&E staining, aiming to reduce the rate or the need for rapid testing.
- Class C: For outcome prediction. The process is entirely AI-based – for example, using TROP2 NMR-QCS scores to predict responsiveness to a TROP2 antibody-drug conjugate. Class C can be further broken down to C1 (prognostic) and C2 (predictive) biomarkers.
ESMO’s recommendations impact trial design. For example, they recommend holding RCTs for validation of response prediction (C2) biomarkers and performing thorough analytical validation for any AI-based biomarker. This includes training with high-quality data, evaluating generalizability across different populations, and ensuring the comparator is adequate and will remain relevant for the duration of the trial.
The ELCAP and EBAI guidance are greatly appreciated and help remove ambiguities that our sponsors may face in designing their clinical trials. We strongly advise our sponsors to read the guidance. Our teams always work to the most recent guidance to collaborate effectively, ensuring our sponsors’ clinical trials are optimally designed and that data analysis is robustly planned. We also look forward to ESMO guidance on AI agents.
LLMs are not a substitute for direct patient interactions – they help reduce administrative burden and complement clinical care, including tasks such as data extraction and trial matching. Human oversight of processes and professional clinical decision-making remains paramount. Training LLMs is crucial for producing accurate, appropriate, and human-comparable responses.
Ethics of artificial intelligence in cancer research
Transparency in how AI is used in clinical trials is crucial for informed consent. The risks of using AI need careful consideration with human oversight. Obtaining informed consent for data use (for research and development) can be complicated by AI’s complexity, requiring clarity on data usage and compatibility with data storage and privacy laws (such as GDPR in the EU). Once a model has been trained on data, it is almost impossible to remove the data from it.
Protection of patient privacy is paramount – closed systems and synthetic data are best for AI training purposes. A robust privacy policy must be practiced.
In summary
Findings presented at the first ESMO AI congress were not just relevant for those attending – they are relevant for all of us, as they represent the future direction of oncology.
Advances in AI applications for clinical trials (particularly trial design, drug development, and patient matching) improve precision, diversity, and generalizability of trial results. However, regulatory oversight, transparency, and ethical considerations remain critical to ensuring the safe and effective integration of AI into clinical practice, as outlined in recent ESMO publications.
In this new era, trial success hinges on choosing a CRO that truly understands AI and technological advancements. With the support of PSI’s AI-powered tools, your oncology project team feasibility/financial planning platform, tools, your oncology project team can adapt quickly to shifting standards of care, global variability in drug access, and the demands of emerging treatment modalities.
Learn more about PSI's AI experience, and deliver your next oncology trial on time
Abbreviations
EBAI, ESMO Basic Requirements for AI-based Biomarkers in Oncology;
EDC, Electronic Data Capture;
EGFR, Epidermal Growth Factor Receptor;
ELCAP, ESMO Guidance on the Use of Large Language Models in Clinical Practice;
ESMO, European Society for Medical Oncology;
EU, European Union;
FDA, Food and Drug Administration;
GDPR, General Data Protection Regulation;
H&E, hematoxylin and eosin (staining);
HER, Human Epidermal Growth Factor Receptor;
LLM, Large Language Model;
NMR-QCR, Normalized Membrane Ratio by Quantitative Continuous Scoring;
TROP2, Tumor-associated Calcium Signal Transducer 2;
WSI, Whole Side Imaging.
References
1. Han, R., et al., Randomised controlled trials evaluating artificial intelligence in clinical practice: a scoping review. The Lancet Digital Health, 2024. 6(5): p. e367–e373.
2.Kp Jayatunga, M., et al., How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today, 2024. 29(6): p. 104009.
3.Prasad, V. and M.A. Makary, FDA’s New Plausible Mechanism Pathway. New England Journal of Medicine. 0(0).
4. Dolezal, J.M., et al., Slideflow: deep learning for digital histopathology with real-time whole-slide visualization. BMC Bioinformatics, 2024. 25(1): p. 134.
5. Howard, F.M., et al., Generative adversarial networks accurately reconstruct pan-cancer histology from pathologic, genomic, and radiographic latent features. Sci Adv, 2024. 10(46): p. eadq0856.
6. Choudhury, D., et al., Developing a low-cost, open-source, locally manufactured workstation and computational pipeline for automated histopathology evaluation using deep learning. EBioMedicine, 2024. 107: p. 105276.
7. Wong, E.Y.T., et al., ESMO guidance on the use of Large Language Models in Clinical Practice (ELCAP). Annals of Oncology, 2025. 36(12): p. 1447–1457.
8. Aldea, M., et al., ESMO Basic Requirements for AI-based Biomarkers In Oncology (EBAI). Ann Oncol, 2025.
