A new Perspective article says generative AI could help scientists decipher the hidden complexities of cancer across images, molecules, and clinical data, opening new avenues for smarter diagnosis, discovery, and treatment.
Perspective: Tackle cancer complexity using generative models. Image credit: Antonio Marca / Shutterstock
Recent Perspective articles published in journals cell We argue that generative models can help address the complexity of cancer.
“Cancer hallmarks” provided a framework for organizing our understanding of cancer biology. They proposed a set of principles that govern the transformation of normal cells into malignant cells and the subsequent progression of cancer. Features represent a reductionist framework that integrates diverse observations and yields valuable insights.
However, an intentionally simplistic framework cannot adequately explain the multifaceted mechanisms of cancer. Therefore, complementary tools are needed 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
Over the years, AI has made great strides in its ability to model complex patterns. Advances in learning algorithms, data availability, and processing power have enabled human-level accuracy or better for some tasks. Applications of AI to cancer include understanding, detection, and intervention. Much of the progress in AI against cancer is in detection.
The development of deep convolutional neural networks has significantly improved image classification performance. Examples include detecting breast cancer using mammography data, classifying skin cancer using lesion images, and detecting lung cancer using computed tomography data. Furthermore, many advances in the understanding of cancer biology have resulted from improvements in its molecular characterization.
As the value of epigenomics, proteomics, transcriptomics, and other omics metrics becomes clearer, there is growing interest in using AI to characterize their high-dimensional output. In this context, basic models represent an important area of development. Single-cell RNA-based models use single-cell RNA-seq data to extract biological signals relevant to downstream tasks.
Additionally, AI holds promise for supporting cancer interventions by guiding or optimizing risk stratification, therapeutic decisions, and patient management. For example, biomarker-based treatment selection models incorporate clinical, imaging, and genomic features to identify patients who may benefit from enhanced therapy.
Generative model that goes beyond cancer characteristics
Cancer features constitute a reductionist framework that trades off structural nuance and complexity. This means that a complex system can be approximated by a simpler model, assuming that it sufficiently captures the fluctuations and dynamics of the original system to be predictable and understandable. However, this tension between comprehensibility and complexity remains a fundamental challenge.
In contrast, generative models take the opposite position from reductive models, prioritizing accuracy and complexity over understanding. The authors propose that generative models can learn the complex dynamics and patterns of cancer directly from data, and thus can be an important tool to complement cancer characteristics. They argue that general-purpose generative models can handle multiple tasks simultaneously and may achieve better performance than specialized models.
This argument builds on capabilities already demonstrated by large-scale generative models, such as unstructured input processing and in-context learning, incomprehensibly complex pattern recognition, and multimodal fusion. Although multimodal generative models have the potential for significant long-term impact, they also have the potential for short-term success, particularly in designing screening, diagnostic testing, and biological, therapeutic, and biomarker discovery pipelines.
The authors also note that current cancer AI systems still have limitations, with modalities often not yet well integrated, relying on narrow, task-specific fine-tuning, and still requiring rigorous validation, uncertainty assessment, and human oversight.
Impact of generative AI on cancer treatment
Generative models represent a new paradigm in cancer research by integrating diverse data sources, modalities, and contextual information. These serve as a constructivist system that extends and ultimately exceeds the capabilities of the cancer features framework. Advances in cancer understanding, detection, and intervention highlight the potential for AI to enhance diagnostic, therapeutic, and prognostic decision-making.
Additionally, multimodal generative models can support mechanistic hypothesis generation, computational perturbation, and experiment prioritization. As integration increases, defining success metrics becomes essential. The impact of AI in the clinic could be measured through outcomes such as patient quality of life and survival rates. The efficiency of the experimental pipeline may reflect the success of the generative model at the translation level.
Nevertheless, to realize the utility of generative models in cancer treatment, it is important to address ethical and practical challenges beyond generative model development. By solving challenges and incorporating feedback, generative models can provide new cancer manifestations, experimentally inferred principles, real-world data, and clinical judgment, and highlight where existing technologies fall short.
This paper emphasizes that these systems need to function as tools to support decision-making and discovery, rather than as autonomous replacements for clinicians and researchers, and that their successful implementation also depends on factors such as infrastructure, workflow integration, privacy, bias, and fair access.
