The research, led by corresponding authors Zhe Yang and Rudi Mao, with Pengfei Zhang, Ying Zhang and Chen Li as co-first authors, set out to answer a question that has long frustrated oncologists: why do seemingly similar esophageal tumors respond so differently to the same dose of radiation? The answer, the study suggests, lies not in a single mutation or pathway but in the collective behavior of distinct subpopulations of tumor cells, and in the way those cells reshape the immune landscape around them when radiation pressure is applied.

Methodologically, the study is a showcase of how modern computational biology can extract signal from the enormous noise of cancer genomics. The team combined single-cell RNA sequencing, which profiles gene expression in individual cells and thereby reveals the full cellular diversity of a tumor, with bulk RNA sequencing, which measures average gene expression across an entire tissue sample. Each approach has well-known limitations on its own: single-cell data are rich in resolution but expensive and technically demanding, while bulk data are abundant and clinically practical but blur the contributions of individual cell types together. Integrating the two is widely regarded as one of the most powerful strategies in contemporary oncology research, because it allows findings made at single-cell resolution to be tested against large patient cohorts.

To perform that integration, the researchers used an algorithm called scAB, which is designed to fuse transcriptomic datasets of different resolutions and identify genes associated with a specific biological condition, in this case the response of esophageal cancer cells to radiotherapy. The scAB analysis revealed significant cellular and molecular heterogeneity within esophageal squamous cell carcinoma, with distinct subpopulations of tumor cells linked to radiation resistance. In other words, within a single patient’s tumor, there exist cells that are intrinsically better equipped to survive a course of radiation, and these cells carry a recognizable transcriptional fingerprint.

From this pool of candidate genes, the team applied two machine learning methods to narrow the field: random forest and LASSO regression. Both are standard tools for feature selection in high-dimensional biological data, where thousands of genes must be evaluated and only a handful are likely to be genuinely informative. Random forest builds many decision trees on random subsets of the data and identifies genes that consistently predict the outcome of interest, while LASSO applies a penalization technique that drives the coefficients of uninformative genes to zero, leaving only the strongest predictors. The convergence of these two independent algorithms on the same shortlist gave the researchers confidence that the genes they identified were not statistical artifacts.

Two genes emerged from this computational gauntlet: MAP1LC3B and HSD17B10. MAP1LC3B encodes a protein central to autophagy, the cellular recycling process by which cells break down damaged components and, crucially, survive periods of severe stress. Radiation kills cells largely by inducing catastrophic DNA damage and oxidative stress, and autophagy is one of the principal mechanisms by which stressed cells hold themselves together long enough to repair or adapt. HSD17B10, by contrast, encodes an enzyme involved in mitochondrial metabolism, participating in the breakdown of fatty acids and amino acids and influencing the energy economy of the cell. A tumor cell that can rewire its metabolism under radiation stress has a marked survival advantage, which makes HSD17B10 a plausible contributor to treatment resistance.

The functional importance of these two genes was not left at the level of correlation. The researchers carried out experiments both in vitro, using cultured cells, and in vivo, using animal models, to test whether MAP1LC3B and HSD17B10 actually drive the behaviors predicted by the sequencing data. The results supported a dual role: the two genes regulate cell survival under radiation stress and also contribute to tumor metastasis, the process by which cancer cells detach from the primary tumor, travel through the body and establish new colonies. The link between treatment resistance and metastatic potential is one of the most clinically troubling aspects of cancer biology, because it means that surviving radiotherapy may select for cells that are also better at spreading.

Equally significant was what the study found when it examined the tumor microenvironment, the complex ecosystem of immune cells, fibroblasts and blood vessels that surrounds and infiltrates a tumor. Immune infiltration analysis showed altered patterns of immune cell presence in tumors with high expression of the two signature genes, suggesting that the microenvironment itself may contribute to radiation resistance. This observation fits into a broader and rapidly growing body of evidence that the immune context of a tumor is not a passive backdrop but an active participant in treatment response. Radiation does not act on tumor cells in isolation; it also reshapes inflammatory and immune signaling, and a tumor microenvironment that suppresses immune attack can shield residual cancer cells from both radiation and immunotherapy.

The clinical implications of the study rest on a striking expression pattern: MAP1LC3B and HSD17B10 were found to be significantly overexpressed in esophageal squamous cell carcinoma tumor tissues compared with healthy tissue, and high expression of both genes correlated with poor prognosis and radiotherapy resistance. This combination of properties, differential expression in tumors, association with survival outcomes and mechanistic involvement in resistance, is precisely the profile sought in a biomarker. If validated in larger and more diverse patient cohorts, a test based on these two genes could eventually help clinicians identify, before treatment begins, which patients are unlikely to benefit from radiotherapy alone and might instead require intensified multimodal therapy, dose escalation or alternative strategies such as immunotherapy combinations.

The study also points toward therapeutic intervention. Genes that actively promote resistance are, in principle, druggable targets, and inhibiting their function could sensitize tumors to radiation. Autophagy modulation, in particular, has long been explored as a strategy to enhance the efficacy of cancer therapy, and the new findings give that effort a specific, data-driven anchor in esophageal cancer. Metabolic enzymes such as HSD17B10 are similarly attractive targets, although the authors are careful to frame their work as identifying potential biomarkers and informing future strategies rather than delivering a ready-made treatment. The research was supported by funding from the Shandong Medical Association, the Beijing Life Oasis Public Service Center and the Wu Jieping Medical Foundation, and was conducted with institutional ethics approval and informed patient consent.

Esophageal squamous cell carcinoma remains a devastating diagnosis worldwide, with particularly high incidence in parts of East Asia, and resistance to radiotherapy is a major reason why long-term survival rates have been so difficult to improve. What this study demonstrates is that the tools now exist to dissect that resistance at single-cell resolution, to validate the resulting hypotheses with machine learning and laboratory experiments, and to translate the findings into a short list of genes that could guide patient stratification and drug development. The convergence of single-cell genomics, bulk sequencing and machine learning, exemplified by this work, is rapidly becoming the standard playbook for attacking one of oncology’s most stubborn problems: the tumor that refuses to die under the beam. Whether MAP1LC3B and HSD17B10 ultimately become the basis of a clinical test or a therapeutic target will depend on the validation studies that must now follow, but the road map from radiation resistance to its molecular roots has rarely been drawn so clearly.

Subject of Research: Molecular mechanisms of radiotherapy resistance in esophageal squamous cell carcinoma identified through integrated single-cell and bulk RNA sequencing with machine learning

Article Title: Machine learning-driven single-cell and bulk RNA sequencing integration unveils key genes modulating radiosensitivity and tumor progression in esophageal squamous cell carcinoma

Article References: Zhang, P., Zhang, Y., Li, C., Li, M., Liao, J., Liu, T., Zhou, Y., Zhang, Y., Yang, Z., & Mao, R. (2026). Machine learning-driven single-cell and bulk RNA sequencing integration unveils key genes modulating radiosensitivity and tumor progression in esophageal squamous cell carcinoma. Cancer Cell International. https://doi.org/10.1186/s12935-026-04476-z

Image Credits: AI Generated

DOI: 10.1186/s12935-026-04476-z

Keywords: esophageal squamous cell carcinoma, radiotherapy resistance, single-cell RNA sequencing, bulk RNA sequencing, machine learning, MAP1LC3B, HSD17B10, autophagy, tumor microenvironment, immune infiltration, tumor metastasis, biomarkers