Precision medicine meets machine learning: AI and oncology biomarkers

Artificial intelligence (AI) has become transformational across many industries, and the field of oncology is no exception. According to a 2024 ICON survey, many professionals involved in oncology drug development hold high hopes for AI and are already applying it in a variety of ways, from target identification to clinical study design.

Interestingly, while only 16% of oncology researchers reported using AI in biomarker detection and assessment at the time of the survey, nearly half (49%) identified it as the area where AI and machine learning (ML) could most accelerate oncology drug development.¹

Improving and accelerating the identification of biomarkers that are accurate across target patient populations and useful for precision medicine clinical development and commercialisation is a crucial need in oncology. While AI holds great promise in this field, developers face many challenges in realising its potential. In this article, we examine the current and future role of AI in biomarker discovery and the challenges that developers face in leveraging it.

The need for improved biomarker identification

Biomarker identification has become a cornerstone of oncology therapeutic development, from the initial cancer diagnosis throughout the disease journey, and more recently the prediction of a patient’s likelihood to respond to a given therapy. In many cases, companion diagnostics are a prerequisite for biomarker targeted cancer therapies. In particular, precision oncology development can benefit from biomarkers that guide identification of patients that are most likely to respond to a treatment.

However, cancer presents an extremely difficult challenge, with highly complex microenvironments and a plethora of factors influencing disease progression. As a result, a single biomarker, such as a specific gene mutation, can rarely provide useful, binary information about a patient’s cancer or response to treatment. In addition, the relationships between individual biomarkers can be difficult to establish or understand, with multiple characteristics interacting and impacting one another.

This often makes the discovery of relevant biomarkers an uphill battle for researchers, requiring the assessment of large amounts of data across multiple domains, such as genetics, proteomics, and molecular features. Multiplex or multiomic sequencing technologies like next generation sequencing can provide guidance, but they are insufficient on their own.

The role of AI in biomarker discovery

AI is particularly useful in finding patterns in large datasets. This ability enables the analysis of large quantities of data to detect more complex associations than traditional statistical analyses might identify. As a result, AI has the potential to provide a fuller understanding of tumour biology and to identify viable biomarkers.

There are several subtypes of AI that can be leveraged for biomarker discovery and development. ML uses algorithms and statistical models to analyse and draw inferences from data patterns, enabling learning and adaptation, and thus improving accuracy with more data and experience. Deep learning (DL) is a type of ML that utilises layered neural networks, and can draw characteristics and associations from unstructured data to create accurate outputs.

DL in particular is being explored as a potential method of uncovering oncology biomarkers. For example, by applying DL to RNA-seq, miRNA-seq, and DNA methylation data, researchers found distinct populations with survival differences and subsequently uncovered genes linked to hepatocellular carcinoma survival.² While the clinical applications of this discovery are still being explored, there is a distinct possibility that understanding these signatures could help guide therapeutic decisions.

AI-based biomarkers drawn from analysis of individual tumour and patient data may also help guide treatment. For example, DL is able to assess routine colorectal cancer histology slides for genetic biomarkers such as microsatellite instability, without the need for PCR, sequencing, or immunohistochemical assays.³ This reduces the costs and infrastructure required for biomarker testing, allowing patients to be pre-screened for genetic testing and potentially even bypassing genetic testing entirely.

Future directions in AI oncology biomarkers

There are a number of areas in which the use of AI is being developed and explored, but where it has yet to be fully implemented into clinical practice. Biomarker development that enables patients to be stratified into subgroups for tailored treatment remains a particularly high priority.

Immuno-oncology is an area where AI holds significant potential for guiding tailored treatment. Immunotherapy can be a challenging course of treatment, with clinical benefit observed only in a subset of patients and few clinically validated biomarkers available for patient selection. However, AI is rapidly evolving to help fill this gap. Efforts to apply AI in finding predictive biomarkers for immuno-oncology using various types of datasets, such as genomics, radiomics, or pathomics, are on the rise, with radiomics being the most represented modality.⁴ While much of this research has limited applicability at present, it is an area that is likely to see continued interest and growth.

Another area of growing interest is the use of AI models that integrate data across modalities, including molecular characterisation, imaging, and patient history. This multimodal integration enables AI to assess the broader tumour and disease landscape, identifying relationships and patterns to uncover biomarkers that more accurately predict treatment response. In some cases, this can lead to the creation of meta-biomarkers, which are novel, fused biomarkers that are distinct from individual and unimodal biomarkers.⁴ However, factors such as insufficient performance in clinical studies and limitations in data collection have hindered adoption of multimodal integration, underscoring the need for further development in this area.⁵

Challenges in AI biomarker development

For any type of AI biomarker discovery to move forward, developers must address several challenges. To achieve accurate results, AI must be trained on large, diverse, high-quality datasets. Doing so is the only way to ensure that AI can account for the vast array of factors influencing cancer growth and behaviour, while minimising bias. Depending on the type of data an AI model is intended to process, obtaining appropriate quantities and quality of data can be challenging.

To implement an AI-based biomarker in clinical trials, it must be generalisable not only to the initial test cohort, but also across diverse patient groups. It is also critical to validate an AI tool to ensure it is suitable for clinical use and acceptable to regulators. These processes will require substantial evidence, likely including large-scale prospective trials, but are crucial to demonstrating accuracy and effectiveness.

Another key challenge is the explainability of AI algorithms. In many cases, the way AI — particularly DL — comes to conclusions is a “black box”, offering very little transparency in its reasoning. Nevertheless, understanding the process is key to gaining regulatory acceptance and ensuring the validity and reproducibility of AI findings.

Looking ahead, AI biomarker discovery and development will need to overcome complex obstacles. Even so, AI is uniquely positioned to transform the landscape of oncology therapeutics, offering the potential for improved patient selection and enhancing our ability to develop precision medicines.

About the author

Dr Bea Mann, PhD, is senior director and oncology therapeutic expert at ICON. As the head of the Oncology Drug Development Services group, Dr Mann is responsible for overseeing the scientific and medical oncology clinical development strategies globally. She brings to ICON more than 16 years of pharmaceutical industry experience (Medical Affairs) having worked as a medical & scientific advisor in the field of oncology and haematology at Roche and GSK. She has extensive experience across all phases of development from pre-clinical to commercialisation. Dr Mann joined ICON in 2016 and in her current role has an expert understanding of the oncology research environment and a thorough understanding of the rapidly changing guidelines for the development of anticancer drugs from study design to protocol development. She has a PhD in Medicinal Chemistry & Drug Development from Nottingham University, UK and was a major contributor to the identification, development, and commercialisation of two linker molecules as part of her oncology research.

About ICON

ICON plc is a world-leading clinical research organisation. From molecule to medicine, we advance clinical research providing outsourced services to pharmaceutical, biotechnology, medical device and government and public health organisations. We develop new innovations, drive emerging therapies forward and improve patient lives. With headquarters in Dublin, Ireland, ICON employed approximately 39,900 employees in 95 locations in 55 countries as at June 30, 2025. For further information about ICON, visit: www.iconplc.com.