Blood Protein Panel Flags Early Esophageal Cancer with High Accuracy in Chinese Case-Control Study

Researchers have identified a panel of six serum proteins that, combined in a machine-learning model, distinguished patients with esophageal squamous cell carcinoma (ESCC) from controls with an area under the curve of 0.932 in a derivation dataset and 0.968 in an independent validation dataset. The study, published in BMC Cancer, was a case-control analysis conducted in two independent Chinese populations from Taixing and Jinan, and it focused specifically on biomarkers for early diagnosis of this cancer type.

The team first performed differential analysis of serum proteins across the two populations to find candidates that were consistently altered in ESCC patients. This cross-validation strategy yielded 13 proteins with diagnostic value. To strengthen the biological plausibility of these candidates, the investigators turned to bulk RNA sequencing and single-cell RNA sequencing data, checking whether the proteins’ expression patterns at the tissue and cellular levels matched what they observed in serum. Most of the proteins showed consistent expression patterns at the mRNA and single-cell RNA levels compared with the serum samples.

From the 13 candidate proteins, the researchers applied multiple machine-learning algorithms to select a core set of diagnostic markers and build predictive models. Lasso regression was used to retain the most informative variables, leaving six proteins in the final panel: CDKN1A, ADAMTS15, KLK13, RSPO3, ANXA1, and MetAP2. A gradient-boosted decision tree algorithm built on these six variables showed the best discriminative performance, calibration, and clinical utility among the models tested.

The performance figures came with quantified uncertainty. In the derivation dataset, the model achieved an AUC of 0.932 with a 95 percent confidence interval of 0.894 to 0.970, while in the validation dataset the AUC was 0.968 with a 95 percent confidence interval of 0.944 to 0.992. These intervals indicate that the model’s discriminative ability was estimated with reasonable precision in both datasets, though the confidence intervals themselves reflect the sample sizes of the case-control populations studied.

Beyond diagnosis, the study explored whether the genes encoding the core diagnostic proteins could also inform prognosis. The researchers constructed risk scores from the transcriptome gene set of these proteins and found that the scores demonstrated a certain predictive capability for risk stratification of ESCC patients, using survival data to extend the utility of the biomarker panel beyond detection alone.

ESCC is a disease in which early diagnosis is considered crucial for reducing the burden of illness, and the authors argue that their findings refine and enhance diagnostic capability at the earliest stages of the disease. They suggest the work could open new avenues for timely intervention aimed at improving patient outcomes, although the study itself was designed to evaluate diagnostic and prognostic performance rather than to test whether screening with this panel changes survival.

Several features of the study design are worth noting when interpreting the results. The biomarkers were identified and verified in two independent populations, which reduces the chance that the signal is specific to a single region or cohort. The convergence of serum protein measurements with tissue-level and single-cell transcriptomic evidence provides multiple lines of support for the selected proteins. At the same time, the case-control design means the model was evaluated on groups already classified as cases or controls, and the abstract does not report performance in a prospective screening setting among people without a known diagnosis.

The study, authored by Tongchao Zhang, Ziqian Song, Jinyu Man, and colleagues including senior authors Xiaorong Yang and Ming Lu, was published as open access research on 8 September 2026, with the article shared early as a peer-reviewed, accepted version citable under a permanent DOI and subject to further edits before the final Version of Record.

For a cancer type where detection at an early stage is central to improving outcomes, a blood-based protein panel reaching AUC values above 0.93 in both derivation and validation datasets represents a substantial reported step forward. The next test for such models, as the authors’ framing implies, lies in translating diagnostic and risk-stratification performance into clinical pathways for timely intervention, work that would need to be established in settings beyond the two Chinese populations examined here.

Subject of Research: Serum protein biomarkers and machine-learning diagnostic models for early detection of esophageal squamous cell carcinoma

Article Title: Selection and prediction of targeted serum protein biomarkers for early diagnosis of esophageal squamous cell carcinoma

Article References: Zhang, T., Song, Z., Man, J., Yin, X., Zhang, X., Li, Z., Li, Z., Ye, W., Chen, X., Yang, X., & Lu, M. (2026). Selection and prediction of targeted serum protein biomarkers for early diagnosis of esophageal squamous cell carcinoma. BMC Cancer. https://doi.org/10.1186/s12885-026-16940-0

Image Credits: AI Generated

DOI: 10.1186/s12885-026-16940-0

Keywords: esophageal squamous cell carcinoma, serum protein biomarkers, early diagnosis, machine learning, gradient-boosted decision tree, Lasso regression, case-control study, AUC, single-cell RNA sequencing, risk stratification, BMC Cancer, diagnostic model