From multi-omics to cancer digital twins: Novel paradigm in cancer research and treatment response

Abstract

Integrating multi-omics in cancer research has expanded our understanding of the intricate molecular mechanisms underlying malignancies. Traditional single-omics approaches, while informative, only capture one molecular layer at a time.¹ However, cancer is a multifaceted disease driven by genetic, epigenetic, metabolic and signalling interactions. In addition, the crosstalk between tumour cells and their environment, whether occurring during malignancies or modulated through therapy, is inherently multiscale. For instance, the crosstalk spans molecular-, cellular- and systems-level processes, operating across temporal and spatial scales. Such complexity renders the tumour ecosystem an archetype of a hierarchical system, requiring integrative approaches to fully comprehend and address.² Multi-omics bridges this gap by integrating diverse data types from the same patients to provide a multidimensional view of the tumour ecosystem. Each of these modalities contributes unique insights into the core tumour; for example, genomics focuses on the mutational landscape, transcriptomics highlights aberrant gene expression, proteomics elucidates protein interactions, epigenomics reveals regulatory mechanisms, metabolomics uncovers metabolic reprogramming and spatial omics map biomolecules directly onto the physical position of tumour niches.³ Together, these multiple modalities resume tumours' molecular and functional dynamics, improving applications such as precision medicine, biomarker discovery, drug target identification and patient stratification (Figure 1).

Integrating multi-omics requires sophisticated computational methods capable of handling high-dimensional datasets.⁴ To tackle these challenges, artificial intelligence (AI), machine learning (ML) and deep learning (DL) have emerged as powerful tools. The AI- and ML/DL-based models facilitate the fusion of multiple modalities into a unified framework, addressing the challenges of disparate data types that vary in size, scales, formats, distributions and noise levels. Its ability to process and harmonize multi-omics data at scale allows this approach to analyse complex datasets, identify hidden patterns and uncover correlations across multi-omics layers often imperceptible to human analysis.⁵ For example, Zhang et al. developed a comprehensive multi-omics platform called COMOS, designed as a non-invasive approach to enhance the diagnosis and prognosis of diffuse large B-cell lymphoma. The authors utilized cell-free DNA (cfDNA) extracted from a single tube of patient blood to analyse several features simultaneously. Using ML algorithms, authors integrated diverse parameters, including nucleosome positioning, CpG island methylation, DNase hypersensitive sites, methylated regions, and copy number alterations, providing a comprehensive view of the cfDNA landscape. COMOS demonstrated a robust accuracy for early diagnosis and predicting responses to chemotherapy compared to standard predicting methods.⁶ The main advantage of non-invasive approaches like COMOS is that they require only one tube of peripheral blood and an AI platform, making them more comfortable for patients compared to biopsy specimens. Also, such approaches hold significant potential for routine clinical implementation, offering a less invasive and less costly alternative for ongoing monitoring and early detection. This paradigm shift has expanded our understanding of cancer and enhanced the ability to diagnose, predict prognosis and assess therapeutic response. However, the dynamic nature of the tumour ecosystem, characterized by clonal evolution, microenvironmental interactions, site-to-site heterogeneity and therapy-induced adaptations, will demand approaches beyond static molecular snapshots.

As patients undergo routine checkups, data from imaging, biopsies, omics profiles and clinical outcomes are continuously collected, highlighting the necessity of a real-time tracking system to assess tumour progression and readjust treatment dosage accordingly. Static models, which rely on fixed assumptions or initial datasets, may fail to capture the dynamic shifts in tumour behaviour or to adapt treatment strategies over time. Instead, based on the accumulated data, adaptive models can integrate new insights to provide more accurate predictions and personalized recommendations.⁷ Therefore, a digital representation of the tumour ecosystem, one that can continuously evolve alongside the patient, has the potential to offer a non-invasive and real-time virtual model of the tumour's biological landscape.

Digital twins (DTs), initially developed in engineering, are virtual replicas of physical systems that simulate real-world processes. Recently, DTs have gained significant attention in healthcare due to their potential to optimize patient care by creating dynamic and personalized models of an individual's health. These models represent a computational model of a patient's disease, developed by integrating multiple modalities such as genomics, imaging, clinical history and real-time monitoring.⁸ Cancer digital twins (CDTs), using the DT concept with a focus on cancer, are initially built as templates derived from historical data and continuously refined through real-time patient-specific information integration. These virtual models can accurately reflect the tumour's current state and predict its future changes, including growth patterns, resistance mechanisms and interactions with the microenvironment.⁹ What began as individual CDTs will eventually grow into substantial cohorts, forming a collective dataset that can be utilized for virtual clinical trials and large-scale population studies. These aggregated cohorts can then be utilized to model diverse treatment pathways, minimizing potential risks and reducing the likelihood of adverse outcomes in real-world clinical settings. Establishing a standardized framework for CDTs will be crucial for integrating multiple CDTs developed across different cancer centres to ensure CDT-based research and applications' consistency, reliability and scalability.

Several CDT templates are currently in development for various cancer types, such as pancreatic, melanoma, lung and breast cancers. These models leverage retrospective multimodal data, such as clinical records, imaging, genomics, histopathology and treatment histories, to develop and refine CDTs. Advanced AI techniques, multiscale and mathematical modelling, and high-performance computing are utilized for data integration into unified models to evaluate tumour trajectories, predict therapeutic resistance and reassign treatment regimens.¹⁰

The importance of multimodal data integration in developing sophisticated models such as CDTs represents a pivotal advancement in cancer research and clinical design. While multi-omics has enhanced our understanding of cancer, the development of AI-powered models and CDTs introduce multidimensional and personalized models that simulate tumour behaviour and treatment responses in real time (Figure 1). However, significant limitations remain, including data acquisition, standardization, quality, bias, privacy, computational methods and the need for robust validation frameworks across diverse patient populations. Addressing these challenges requires interdisciplinary collaboration among experts in oncology, bioinformatics, computational biology and engineering. As these technologies evolve, they have the potential to transform the landscape of precision oncology.

S.S.A. corrected, revised, provided critical feedback, contributed ideas, and finalized the manuscript. R.A. conceived the original idea, wrote, revised, and supplemented the manuscript. All authors have read and agreed to the published version of the manuscript.