AI and molecular studies converge to decode cancer complexity in new research collection

Cancer remains one of the leading causes of death worldwide, and a newly published editorial in Frontiers in Oncology argues that the path forward lies not in applying artificial intelligence to cancer as an afterthought, but in fusing AI directly with cancer biology. The editorial, written by Chandrabose Selvaraj, Krishnendu Bera, Senthilkumar Krishnamoorthi and Sajitha Lulu S, introduces a special issue titled “Decoding Cancer Complexity with Artificial Intelligence and Molecular Level Studies: Informatics-Driven Approaches for Therapeutic Innovation,” which compiles thirteen studies spanning diagnosis, biomarker identification, toxicity assessment, vaccine design, prognosis and the prediction of therapeutic responses.

The editorial frames the collection around a central conceptual shift: away from single-modality analysis and toward multimodal, biologically contextualized intelligence. Cancer, the authors note, is characterized by numerous sequential events, including epigenetic regulation and alterations in genes and proteins that interact extensively with the surrounding tumor environment. The integration of artificial intelligence, machine learning and advanced bioinformatics, they argue, is currently transforming how researchers understand cancer biology, uncover biomarkers, elucidate therapeutic response mechanisms and accelerate the development of personalized therapies.

Among the contributions highlighted is work by Huang and colleagues on AI-driven digital pathology in hepatobiliary and pancreatic tumors. Their discussion covers the significance of whole-slide imaging and the role of deep learning in disease diagnosis, prognosis, molecular prediction and therapeutic response assessment. Crucially, the editorial emphasizes that the authors highlight data heterogeneity, interpretability, prospective validation and the incorporation of clinical workflows as decisive factors, and that superior algorithmic performance alone is insufficient for clinical translation.

That theme is reinforced by Ding and colleagues, whose TriBind-Mamba is described as a tri-model framework incorporating clinical knowledge, whole-slide images and transcriptomic information through selective state-space models. According to the editorial, the incorporation of pathway-level molecular representations and clinical knowledge prompting represents an extensive advancement in computational oncology, pointing toward models that contextualize predictive patterns within clinically significant and biologically interpretable frameworks.

Multimodal fusion features prominently throughout the issue. Zhang and colleagues examine the combination of radiomics, multi-omics, circulating tumor DNA, clinical variables and digital pathology, underscoring the potential of cross-modal associations for diagnosis, treatment-response prediction and prognostic assessment. Peng and colleagues turn to triple-negative breast cancer, a disease marked by pronounced biological heterogeneity among its subtypes, and emphasize transcriptomics, multi-omics, pathology and imaging. Their analysis stresses the importance of maintaining reliability despite missing modalities, institutional variations and shifts in real-world data distribution.

Several studies pair computational prediction with molecular and experimental evidence. Tong and colleagues report a machine-learning-based mitochondrial gene signature for breast cancer, built through extensive computational analysis, immune-filtration analysis, tissue validation, drug prediction, single-cell data analysis, molecular docking and molecular dynamics simulations. The work identifies mitochondrial-linked biomarkers and explores therapeutic molecules, illustrating how advanced computational prediction can be tied back to molecular and experimental validation. Wang and colleagues developed an internally validated machine-learning model to predict early-stage bone metastasis in patients newly diagnosed with prostate cancer, identifying clinical laboratory predictors and applying SHAP-based interpretation to improve model transparency. Alongside an interactive prediction tool, the authors acknowledge that external and prospective validation remains a necessary next step.

Not every new computing paradigm delivered a headline result, and the editorial treats that as a strength. Allena and colleagues performed a comprehensive evaluation of quantum machine-learning architectures for breast cancer classification using clinical and genomic data. Within the tested simulator and dimensionality constraints, classical models using the complete feature space significantly outperformed the quantum approaches evaluated. The editorial describes the contribution as a realistic benchmark rather than an assertion of quantum advantage, underscoring the value of rigorous benchmarking as novel computational approaches enter biomedical research.

Explainability and biological scope also receive attention. Gopalan and Sundararajan built an interpretable graph-attention framework incorporating genomics, proteomics, transcriptomics and chemical descriptors to identify cancer drug responses, correlating model-derived feature importance with pathway enrichment analysis and consistently implicating the PI3K/AKT/mTOR signalling pathway. The editorial notes this demonstrates that explainability can go beyond visualizing model decisions and generate mechanistic hypotheses for experimental validation. Guhanraj’s work on bacteriocins and postbiotic metabolites as potential anticancer agents extends computational oncology beyond conventional small-molecule identification, integrating microbiome biology with AI-based techniques, molecular docking, molecular dynamics, systems biology and nanotechnology.

Toxicology and immunotherapy round out the collection. Sharma and colleagues developed ChemModernBERT, a ModernBERT-based molecular language model for predicting LD50 values, trained on over 1.8 million SMILES strings using standardized datasets, demonstrating the potential of molecular language models in computational toxicity assessment and chemical prioritization while underlining the distinction between acute toxicity prediction and the broader biological processes involved in carcinogenesis. Srivastava summarizes AI-enabled approaches for engineering mRNA cancer vaccines, integrating neoantigen prediction, MHC binding, antigen processing, T-cell recognition, codon optimization, UTR engineering, mRNA structural modeling and lipid nanoparticle design, connecting tumor-specific molecular data with personalized immunotherapy design.

The editorial also points to bibliometric and population-level perspectives. Li and colleagues mapped the global academic landscape of AI applications in basal cell carcinoma, revealing a significant surge in research activity since 2019, with deep learning-based dermoscopic and digital pathology approaches as predominant themes. Separately, Zhang and colleagues explored pathomics-derived tumor-infiltrating lymphocyte scores in colorectal cancer whole-slide images from multiple cohorts, a multicohort design the editorial says is particularly suited to building AI systems capable of capturing biological variation rather than learning patterns restricted to a single institutional population.

Taken together, the editorial concludes that computational oncology does not depend on complex algorithms alone but requires the integration of diverse biological data with clinically relevant samples. Digital pathology reveals spatial features, genomics and transcriptomics explain molecular changes, clinical data provides patient-specific context, and computational chemistry illuminates therapeutic and toxicological properties, with AI able to integrate these dimensions into a unified framework. The authors are candid about the field’s shortcomings: data heterogeneity, limited external validation, interpretability, reproducibility and the lack of prospective clinical evaluation remain notable limitations across the studies. They call for future work focused on multicentre validation, standardized data generation, biologically informed interpretation, privacy-preserving collaboration and clinical evaluation, arguing that the future of oncology lies in building systems that are explainable, clinically relevant, experimentally testable and robust across diverse cancer populations.

Subject of Research: Artificial intelligence and molecular-level studies applied to cancer diagnosis, prognosis and therapeutic innovation

Article Title: Editorial: Decoding cancer complexity with artificial intelligence and molecular level studies: informatics-driven approaches for therapeutic innovation

Article References: Selvaraj, C., Bera, K., Krishnamoorthi, S., & Lulu S, S. (2026). Editorial: Decoding cancer complexity with artificial intelligence and molecular level studies: informatics-driven approaches for therapeutic innovation. Frontiers in Oncology, 16, Article 2008527. https://doi.org/10.3389/fonc.2026.2008527

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.2008527

Keywords: artificial intelligence, cancer, multi-omics, machine learning, digital pathology, biomarkers, mRNA vaccine, drug response, toxicity prediction, computational oncology, prostate cancer, breast cancer