A new 'Perspective' article says generative AI may help scientists read cancer’s hidden complexity across images, molecules, and clinical data, opening a possible new path to smarter diagnosis, discovery, and treatment.
Perspective: Tackling the complexity of cancer with generative models. Image Credit: Antonio Marca / Shutterstock
A recent Perspective article published in the journal Cell argues that generative models could help address the complexity of cancer.
The “Hallmarks of Cancer” provided a framework to systemize the understanding of cancer biology. They proposed a set of principles dictating the transformation of normal cells into malignant cells and subsequent cancer progression. The hallmarks represent a reductionist framework that has unified diverse observations, yielding valuable insights.
However, an intentionally simple framework cannot adequately explain the multifaceted mechanisms of cancer. Thus, complementary tools are required to capture the complex, multiscale, and multimodal nature of cancer. In this paper, the authors proposed that generative models built on advances in artificial intelligence (AI) can address the complexity of cancer.
AI for Cancer Detection and Biological Understanding
AI has achieved significant strides in its ability to model complex patterns over the years. Advances in learning algorithms, data availability, and processing power have led to human-level or even higher accuracy in some tasks. The applications of AI to cancer include understanding, detection, and intervention. Much of the progress in AI for cancer has been in detection.
The development of deep convolutional neural networks has significantly improved image classification performance. Examples include breast cancer detection using mammographic data, skin cancer classification using lesion images, and lung cancer detection using computed tomography data. Further, many advances in understanding cancer biology have resulted from improvements in its molecular characterization.
As the value of epigenomics, proteomics, transcriptomics, and other -omics measures has become clear, there is growing interest in characterizing their high-dimensional outputs using AI. In this context, foundation models represent a key area of development. Single-cell RNA foundation models use single-cell RNA sequencing data to extract relevant biological signals for downstream tasks.
Furthermore, AI can be promising in assisting cancer intervention by guiding or optimizing risk stratification, therapeutic decisions, and patient management. For example, biomarker-guided treatment selection models incorporate clinical, imaging, and genomic features to identify patients who may benefit from intensified treatment.
Generative Models Beyond Cancer Hallmarks
The Hallmarks of Cancer constitute a reductionist framework, trading off nuance and complexity for structure. This means that a complex system can be approximated by simpler models, assuming that the latter capture enough of the original system's variation and dynamics to be both predictive and intelligible. However, this tension between comprehensibility and complexity remains a fundamental challenge.
In contrast, generative models take an opposite stance to reductive models, prioritizing accuracy and complexity over understanding. The authors propose that generative models could be vital complementary tools to the Hallmarks of Cancer, as they can learn the complex dynamics and patterns of cancer directly from data. They argue that general-purpose generative models can address multiple tasks concurrently, potentially achieving better performance than specialized models.
The argument is based on capabilities already shown by large generative models: unstructured input processing and in-context learning, incomprehensibly complex pattern recognition, and multimodal fusion. While multimodal generative models could have a significant impact in the long term, they could also achieve near-term successes, especially in screening, diagnostic testing, and the design of biological, therapeutic, and biomarker discovery pipelines.
The authors also note that current cancer AI systems remain limited, often because they do not yet integrate modalities well, rely on narrow task-specific fine-tuning, and still require rigorous validation, uncertainty assessment, and human oversight.
Generative AI Implications for Cancer Care
Together, generative models represent an emerging paradigm for cancer research by integrating diverse data sources, modalities, and contextual information. They operate as a constructionist system that extends, and ultimately exceeds, the capacity of the Hallmarks of Cancer framework. Progress in understanding, detecting, and intervening in cancer highlights the potential for AI to augment diagnostic, therapeutic, and prognostic decision-making.
Further, multimodal generative models could support mechanistic hypothesis generation, in silico perturbations, and experimental prioritization. With increased integration, defining metrics for success will be essential. The impact of AI in the clinic could be evaluated through outcomes like patient quality of life and survival rates. The efficiency of experimental pipelines could reflect the success of generative models at the translational level.
Nevertheless, addressing ethical and practical challenges beyond the development of generative models will be crucial to realizing their utility in cancer care. By navigating challenges and incorporating feedback, generative models could provide new signatures of cancer, principles inferred from experiments, real-world data, and clinical decisions, and expose where existing technologies are insufficient.
The paper emphasizes that these systems should function as decision- and discovery-support tools, not as autonomous replacements for clinicians or researchers, and that their successful adoption will also depend on factors such as infrastructure, workflow integration, privacy, bias, and equitable access.
Scientists discover BRCA links to head and neck cancer risks