AI-driven multi-omics drug repurposing nominates AZD7762 as a multitarget inhibitor of IL22RA1 and FAM221A in esophageal squamous cell carcinoma

Abstract

Esophageal squamous cell carcinoma (ESCC) is an aggressive malignancy with limited targeted treatment options and poor clinical outcomes. We developed an AI-driven multi-omics pipeline that links prognostic modeling to multitarget drug repurposing for ESCC. Summary-data-based Mendelian randomization was integrated with bulk transcriptomic datasets to identify esophageal cancer-related druggable genes that are differentially expressed. Cox regression and non-negative matrix factorization were then used to define prognostic genes and molecular subgroups, and a Lasso Cox model with SHapley Additive explanation provided an interpretable prognostic signature. Single-cell RNA sequencing analysis mapped the hub genes interleukin 22 receptor subunit alpha 1 (IL22RA1) and family with sequence similarity 221 member A (FAM221A) to epithelial cell populations and associated them with proliferative and DNA repair programs, supporting their role in tumor progression, supporting their role in ESCC progression. To translate these targets into a therapeutic strategy, we applied machine learning-based drug sensitivity prediction, ADMET-AI toxicity, pharmacokinetic profiling, and molecular docking, which converged on the checkpoint kinase inhibitor AZD7762 (3-(carbamoylamino)-5-(3-fluorophenyl)-N-[(3S)-piperidin-3-yl] thiophene-2-carboxamide) as a promising multitarget inhibitor of IL22RA1 and FAM221A. In vitro assays confirmed that IL22RA1 and FAM221A promote ESCC cell proliferation, migration, and invasion. Taken together, this AI-driven multi-omics framework delivers a prognostic model, defines biologically distinct ESCC subgroups, and nominates AZD7762 as a rational multitarget drug repurposing candidate, providing a precision oncology strategy.

Introduction

Esophageal cancer (EC) is a malignant tumor arising from the esophagus, with an increasing global incidence and high disease burden, and remains associated with poor survival outcomes worldwide¹. Esophageal cancer (EC) is a heterogeneous disease that mainly comprises two major histological subtypes: ESCC and esophageal adenocarcinoma (EAC). In high-incidence regions of Asia, particularly China, ESCC is the predominant subtype and accounts for the majority of EC-related mortality; therefore, this study focuses specifically on ESCC. Despite multimodal management with surgery, radiotherapy, and chemotherapy, ESCC still shows high recurrence rates and poor survival, largely due to the lack of effective targeted therapies and limited biomarker-driven stratification for precision treatment of ESCC. Therefore, there is an urgent unmet clinical need to develop robust prognostic models and rational therapeutic strategies that can guide patient risk stratification and enable improved personalized interventions in ESCC^2,3. Epidemiologically, EC presents significant geographical variations, particularly with a notable prevalence in certain regions of Asia, notably China, where the aggressive subgroup is classified as ESCC^4,5. Therefore, the identification of reliable prognostic indicators is critical for improving patient outcomes and tailoring therapeutic approaches to the disease.

In recent years, artificial intelligence (AI) algorithms have improved the clinical precision and personalization of cancer treatment, significantly enhancing prognostic performance^6,7. In ESCC, AI frameworks improve the accuracy and efficacy of aiding treatment decision-making, precision diagnosis, and prognostic assessment⁸. For example, machine learning algorithms have accelerated the development of molecular prognostic biomarkers⁹. Deep learning algorithms have identified novel potential compounds for the treatment of ESCC, which can potentially improve patient clinical outcomes¹⁰.

Although recent AI studies in ESCC have improved prognostic prediction, most approaches remain limited to single-omics feature selection and do not provide a unified path from biologically supported target discovery to actionable therapeutic prioritization of the identified targets. In particular, drug repurposing efforts are often conducted independently of prognostic modeling, and few studies have explicitly prioritized multitarget candidates that can simultaneously address pathway redundancy and tumor heterogeneity. Here, we present an AI-driven multi-omics framework that integrates Mendelian randomization-guided target discovery, bulk and single-cell transcriptomics, molecular subtyping, interpretable prognostic modeling, and multi-target drug repurposing within a single pipeline. This unified design distinguishes our study from prior ESCC AI studies by directly linking prognostic stratification to rational, mechanism-informed therapeutic nomination. In silico and in vitro studies showed that interleukin 22 receptor subunit alpha 1 (IL22RA1) and family with sequence similarity 221 member A (FAM221A) were hub genes associated with ESCC progression. AZD7762 should be considered a multi-target drug repurposing strategy for patients with ESCC. In this study, we established an AI-driven multi-omics framework that integrates Mendelian randomization, bulk transcriptomics, molecular subtyping, interpretable prognostic modeling, and single-cell validation to move from target discovery to multitarget drug repurposing in ESCC. Using this pipeline, we identified IL22RA1 and FAM221A as key prognostic drivers enriched in epithelial tumor cells, defined biologically distinct ESCC molecular subgroups, and constructed an interpretable Lasso-Cox prognostic model supported by SHAP explanations. Importantly, we translated these findings into a therapeutic strategy by nominating AZD7762 as a rational multitarget repurposing candidate and validated its mechanistic relevance through drug sensitivity prediction, ADMET profiling, molecular docking, and in vitro functional assays. Collectively, our study provides a clinically relevant framework for biomarker-driven stratification and precision therapeutic prioritization in ESCC.

Results

Identification of risk factors for ESCC patients

First, EC-associated druggable targets were enriched using two-sample Mendelian randomization analysis, resulting in 83 targets (Fig. 1A). Next, differential expression analysis of the GEO bulk-seq datasets identified 3719 DEGs, which were then intersected with EC-related druggable targets to enrich nine druggable DEGs (Fig. 1A). In addition, the expression patterns of these nine druggable DEGs in the public dataset were analyzed (Fig. 1B). These nine DEGs molecular functions were discovered by KEGG and GO analyses (Fig. 1C, D). Genetic variation analysis was also performed, and then found that GUF1, IL22RA1, SLC35F5, PRKAB1 and IQCB1 shared with high mutation rate (Fig. 1E). Finally, GUF1, TCF19, FAM221A, IL22RA1, and TFAM were identified as prognostic indicators among the nine druggable DEGs (Fig. 1F).

Fig. 1. Druggable prognosis-associated DEGs identification for ESCC patients.

A Druggable targets illustration via forest plot. B Volcano plot illustration of Druggable DEGs. C GO enrichment analysis. D KEGG enrichment analysis. E Genetic variation analysis. F Univariate cox regression analysis.

Novel NMF-driven ESCC subgroups identification

The NMF clustering algorithm was employed to perform a clustering analysis on the TCGA-ESCC dataset based on the five druggable DEGs. To determine the most optimal method for subclassifying TCGA-ESCC samples for subsequent studies, we adopted criteria based on the co-expression profiles. The results of this study demonstrated that classifying TCGA-ESCC samples into three clusters was the most optimal approach (Fig. 2A, B). After grouping the samples into three clusters, patients in Cluster 3(C3) had a poor prognosis compared with those in Clusters 1(C1) and 2(C2) (Fig. 2C). We assessed the distribution of clinical parameters across the identified clusters and found that the number of patients varied significantly between clusters based on the pathological distant metastasis stage (Fig. 2D). In addition, the TIDE score was enriched among C1, C2, and C3, with C3 showing a high score (Fig. 2E). Finally, we analyzed immune checkpoint expression among these three subgroups (Fig. 2F).

Fig. 2. Molecular subgroups identification of ESCC patients via NMF analysis.

A Rank survey of NMF analysis. B Heatmap of NMF clustering. C Survival differences among the three clusters. D Differences in the distribution of different subgroups in different pathological stages of ESCC. E TIDE analysis among these 3 subgroups. F Immune checkpoint expression analysis among these 3 subgroups.

Prognostic model and variables recognition and validation for ESCC patients

We performed LASSO-cox regression analysis to construct a prognostic model based on five druggable prognostic indicators in the TCGA-ESCC cohort, which showed strong predictive performance (Fig. 3A–C). The SHAP explanation showed that FAM221A and IL22RA1 were the major contributors (Fig. 3D–F). In addition, we cross-validated the efficacy of this model in the TCGA-ESCC and GSE53625 cohorts (Fig. 3G). To evaluate model accuracy, we performed Cox regression and nomogram-based examinations, and the results indicated that our model achieved satisfactory performance (Fig. S1A–D).

Fig. 3. Lasso-Cox regression prognostic model construction.

A–C Lasso-cox regression. D–F SHAP explanations for Lasso-Cox regression model. G Prognostic model evaluation in 2 different datasets.

ESCC tumor microenvironment heterogeneity exploration at single-cell level

We first performed QC and SingleR annotation of the single-cell dataset GSE188900 and identified nine cell types (Fig. S2A–D, Fig. 4A, B). Subsequently, CellChat was applied to infer the communication patterns among these nine cell types. (Fig. 4C). The metabolism patterns among these 9 cell types were also quantified in these 9 cell types (Fig. 4D). IL22RA1 and FAM221A were mainly distributed in epithelial cells, indicating that these two targeted genes mainly regulate epithelial cell functions (Fig. 4E).

Fig. 4. The heterogeneity of ESCC tumor microenvironment and targeted gene distributions.

A Single-cell-level cell types of ESCC obtained via UMAP and t-SNE analysis. B Proportion of cell types. C Cell communication among the 9 cell types, D Metabolic pathways in the 9 cell types. E Distribution of IL22RA1 and FAM221A at ESCC single-cell level.

The heterogeneity of IL22RA1 and FAM221A at single-cell level of ESCC patients

First, to observe the molecular functions of IL22RA1 and FAM221A in epithelial cells, we discovered that IL22RA1 mainly negatively regulated the apical junction. Similarly, FAM221A mainly positively regulated DNA repair (Fig. 5A, B). The pseudotime trajectory of epithelial cells and IL22RA1 the FAM221A temporal expression patterns in epithelial cells were evaluated (Fig. 5C, D).

Fig. 5. Pseudo-time trajectory and IL22RA1 with FAM221A temporal expression patterns in epithelial cells.

A GSEA for FAM221A at single-cell resolution. B Single-cell GSEA analysis of IL22RA1. C Pseudotime analysis of epithelial cell dynamics. D Pseudotime trajectory analysis of IL22RA1 and FAM221A in epithelial cells.

IL22RA1 and FAM221A-oriented drug reproposing strategies elaboration for ESCC patients

To obtain a systems-level view of AZD7762 activity, we constructed integrative networks linking the prognostic genes, enriched pathways, candidate drugs, and ESCC molecular subgroups. A radial protein–protein interaction (PPI) map showed that IL22RA1 and FAM221A occupy central positions within a cell cycle and DNA repair–focused module together with GUF1, TCF19, SLC35F5, PRKAB1, and IQCB1 (Fig. 6A). A multilayer network placed AZD7762 upstream of these seven genes and their hallmark programs, including DNA repair, G2M checkpoint, E2F targets, apical junction, EMT, inflammatory response, and IL-6/JAK–STAT signaling (Fig. 6B). A comparison with other candidate compounds demonstrated that AZD7762 had the broadest coverage across IL22RA1, FAM221A, and extended checkpoint/cell cycle targets such as CHEK1, CHEK2, CDK1, and CCNB1 (Fig. 6C). Finally, an integrative systems network summarizing AZD7762 connections to gene targets, pathways, ESCC risk/molecular subgroups, and standard chemotherapeutics highlighted its potential as a multitarget precision therapy for ESCC (Fig. 6D).

Fig. 6. Integrative AI-driven network analysis nominates AZD7762 as a multitarget repurposed agent for ESCC.

A Radial protein–protein interaction network of IL22RA1, FAM221A and the other prognostic genes (GUF1, TCF19, SLC35F5, PRKAB1, IQCB1) with cell-cycle, DNA-repair and EMT-related interactors. B Multilayer network connecting AZD7762 to the seven core genes and their enriched hallmark programmes, including DNA repair, G2M checkpoint, E2F targets, apical junction, EMT, inflammatory response and IL-6/JAK–STAT signaling. C Drug–target network comparing AZD7762 with other candidate compounds (Roscovitine, Palbociclib, Olaparib, MK-8776, VX-970, Cisplatin, Gemcitabine, 5-FU, Paclitaxel, Etoposide), showing that AZD7762 has the most extensive coverage across IL22RA1, FAM221A and extended checkpoint/cell-cycle targets (CHEK1, CHEK2, CDK1, CCNB1). D Large integrative systems network summarizing AZD7762-centered connections among gene targets, hallmark pathways, ESCC molecular/risk subgroups, cell states and other drugs, providing an overview of the AI-driven multitarget repurposing landscape in ESCC.

Structural and mechanistic view of AZD7762 as a multitarget inhibitor of IL22RA1 and FAM221A in ESCC

(Figure 7A) Two-dimensional chemical structure of AZD7762, a checkpoint kinase inhibitor prioritized by the AI-driven multi-omics pipeline as a repurposing candidate for ESCC. (Fig. 7B) Predicted binding mode of AZD7762 in the IL22RA1 binding pocket is shown, along with key interacting residues (Asp150, Tyr170, and Arg200) and a favorable docking score. (Fig. 7C) Predicted binding mode of AZD7762 in the FAM221A binding pocket, with residues such as Phe150, Tyr100, and Arg220 forming stabilizing contacts; together with panel B, this illustrates the multitarget binding of AZD7762. (Fig. 7D) Two-dimensional interaction diagrams summarizing hydrogen bonds and hydrophobic contacts between AZD7762 and IL22RA1 (left) or FAM221A (right), highlighting the main amino acid residues that participate in ligand recognition. (Fig. 7E) Schematic model of AZD7762 action in ESCC cells: AZD7762 inhibits IL22RA1-driven pro-proliferative JAK–STAT signaling and FAM221A-associated cell cycle and DNA repair programmes, leading to reduced proliferation, migration, and invasion and increased apoptosis/DNA damage, consistent with the in vitro functional assays.

Fig. 7. Structural and mechanistic view of AZD7762 as a multitarget inhibitor in esophageal squamous cell carcinoma (ESCC).

A Chemical structure of the checkpoint kinase inhibitor AZD7762. B, C Predicted binding modes of AZD7762 in the IL22RA1 and FAM221A binding pockets. D Two-dimensional interaction diagrams showing key contacts between AZD7762 and residues in IL22RA1 and FAM221A. E Schematic model illustrating how AZD7762 inhibits IL22RA1-driven JAK–STAT signaling and FAM221A-related cell-cycle/DNA-repair pathways, reducing proliferation and migration/invasion while increasing apoptosis/DNA damage in ESCC cells.

After drug sensitivity assessment, we found that AZD7762 emerged as a promising drug repurposing candidate for ESCC (Fig. 8A). ADMET-AI analysis suggested that AZD7762 has favorable predicted pharmacokinetic properties and acceptable toxicity risk with high bioavailability and low toxicity (Fig. 8B). Molecular docking analysis showed that AZD7762 has a favorable binding affinity with FAM221A and IL22RA1(−6.9 kcal/mol and −5.3 kcal/mol, respectively) (Fig. 8C).

Fig. 8. Durg repurposing strategy design for ESCC patients.

A Drug sensitivity analysis. B ADMET-AI analysis. C Molecular docking analysis of FAM221A and AZD7762. D Molecular docking analysis of IL22RA1 and AZD7762.

In vitro association of IL22RA1 with ESCC neoplastic advancement

To explore the function of IL22RA1 in ESCC, we separately constructed IL22RA1 knockdown or overexpressing stable cell lines derived from KYSE150 cells (Fig. 9A). To evaluate the role of IL22RA1 in cell growth, CCK-8 and clonogenic assays were performed (Fig. 9B–D). IL22RA1 knockdown inhibited the growth of these cells. In contrast, cell growth was significantly enhanced in IL22RA1 overexpressed cells. The wound-healing assay showed that cells with IL22RA1 knockdown exhibited significantly slower wound closure (Fig. 9E). Conversely, IL22RA1 overexpression resulted in a significantly faster wound closure rate (Fig. 9F). Transwell assays revealed that IL22RA1 knockdown markedly decreased the migratory and invasive capabilities of ESCC cells (Fig. 9G). In contrast, IL22RA1-overexpressed ESCC cells showed significantly increased invasion capabilities (Fig. 9H). These results indicated that IL22RA1 promotes the progression of ESCC by enhancing key malignant phenotypes.

Fig. 9. IL22RA1 promotes the malignancy of ESCC.

A The protein levels in KYSE150 cell lines after IL22RA1 knockdown or overexpression. B CCK-8 analysis of KYSE150 cell viability under IL22RA1 silencing or overexpression. C, D Representative images of the clonogenic assay of KYSE150 cell proliferation under IL22RA1 silencing or overexpression. E, F Wound-healing assays evaluating migration of KYSE150 cells under IL22RA1 knockdown or overexpression. G, H Transwell assays evaluating migration and invasion of KYSE150 cells under IL22RA1 knockdown or overexpression. All quantitative data are presented as mean ± SD from at least three independent biological replicates (n = 3) ****P < 0.0001; **P < 0.01; *P < 0.05.

In vitro association of FAM221A with ESCC malignant progression

To further determine the involvement of FAM221A in ESCC, we established stable cell lines with either FAM221A knockdown or overexpression using KYSE150 cells (Fig. 10A). To evaluate the impact of FAM221A on the proliferative capacity of cancer cells, we employed both CCK-8 and clonogenic assays (Fig. 10B–D). The results indicated that FAM221A knockdown led to a marked reduction in cell growth, whereas cells overexpressing IL22RA1 exhibited substantially enhanced growth rates. Furthermore, the wound-healing assay demonstrated that cells with FAM221A knockdown displayed a significantly slower wound closure rate (Fig. 10E). In stark contrast, FAM221A overexpression markedly accelerated wound closure (Fig. 10F). Transwell assays indicated that the knockdown of FAM221A significantly diminished the invasive capabilities of ESCC cells (Fig. 10G). Conversely, ESCC cells overexpressing FAM221A showed a pronounced increase in invasion capabilities (Fig. 10H). Collectively, these data suggest that FAM221A plays a pivotal role in promoting ESCC progression by enhancing cell proliferation, migration, and invasion.

Fig. 10. FAM221A promotes the malignancy of ESCC.

A The protein levels in KYSE150 cell lines after FAM221A knockdown or overexpression. B CCK-8 analysis of KYSE150 cell viability under FAM221A knockdown or overexpression. C, D Representative images of the clonogenic assay of KYSE150 cell proliferation under FAM221A silencing or overexpression. E, F Wound healing assay results of control and sh-FAM221A (or overexpression) KYSE150 cell line. G, H Transwell analysis of KYSE150 cell migration under FAM221A silencing or overexpression. All quantitative data are presented as mean ± SD from at least three independent biological replicates (n = 3) ****P < 0.0001; **P < 0.01; *P < 0.05.

Discussion

ESCC poses a major clinical challenge because of its aggressive nature and poor therapeutic outcomes¹¹. Current therapeutic strategies often fail to account for the heterogeneity of the disease, leading to suboptimal clinical outcomes¹². This underscores the urgent need for innovative approaches to enhance patient management and stratification based on molecular characteristics of the disease. This study provides several advances over previous ESCC prognostic and drug discovery studies. First, rather than relying solely on differential expression or survival association, we used summary-data-based Mendelian randomization (SMR) to prioritize genetically supported druggable targets linked to esophageal cancer risk and integrated these with bulk transcriptomic differential expression to identify prognostically relevant candidate genes. Second, we coupled molecular subtyping, interpretable prognostic modeling (Lasso-Cox with SHAP explanations), and single-cell validation to define biologically grounded ESCC risk groups and pinpoint IL22RA1 and FAM221A as epithelium-associated drivers. Third, we extended the pipeline beyond prognostic prediction by integrating machine learning-based drug sensitivity prediction, ADMET-AI profiling, and molecular docking to nominate AZD7762 as a multitarget repurposing candidate linked to the prognostic drivers. Collectively, this study established a unified AI-driven multi-omics framework that moves from genetically informed target discovery to prognostic stratification and multitarget drug repurposing in ESCC. By employing NMF, we successfully stratified the molecular subgroups. Furthermore, by utilizing interpretable Lasso-Cox regression, we successfully constructed a robust prognostic model and provided convergent evidence that IL22RA1 and FAM221A are associated with ESCC progression and support an oncogenic role in ESCC. Furthermore, the application of machine learning techniques and ADMET-AI facilitated the development of an optimal AZD7762 multi-target therapeutic framework, which was validated through molecular docking studies. Our findings highlight the involvement of FAM221A and IL22RA1 in ESCC progression.

IL22RA1, the receptor subunit for interleukin-22, is an epithelial-enriched signaling component that activates downstream JAK–STAT pathways and regulates epithelial proliferation, survival, and inflammatory responses. These functions are directly relevant to ESCC biology, where malignant epithelial expansion and inflammation-associated signaling contribute to tumor progression and therapeutic resistance. Consistent with this, our bulk and single-cell analyses mapped IL22RA1 predominantly to epithelial tumor populations and linked its expression to hallmark programs related to proliferation, DNA repair, and epithelial-state regulation, supporting a functional role in ESCC aggressiveness. In contrast, FAM221A (family with sequence similarity 221 member A) is comparatively poorly characterized in cancer biology, and its mechanistic roles have not yet been well established. Therefore, our findings identifying FAM221A as a prognostic driver enriched in epithelial cells and associated with DNA repair and cell cycle programs should be viewed as hypothesis-generating. Further mechanistic studies are required to define its upstream regulators and downstream effectors and its potential suitability as a therapeutic target in ESCC¹³. Studies have shown that IL22RA1 regulates tumor progression via JAK-STAT signaling in pan-cancer analysis^14,15. Its dysregulation in ESCC suggests a dual role in promoting tumor growth and potentially modulating the immune landscape of the tumor microenvironment. FAM221A (Family With Sequence Similarity 221 Member A), a gene implicated in regulating cardiovascular health, its molecular function has not yet been fully elucidated¹⁶. Previous studies and our study highlight the role of the immune microenvironment in prognostic forecasting ESCC patient clinical outcomes in patients with ESCC and its association with tumor progression¹⁷. Therefore, molecular research targeting FAM221A can provide additional insights into ESCC tumorigenesis. In addition, numerous studies have pointed out the AZD7762 therapeutic role in cancer treatment, such as stomach adenocarcinoma, osteosarcoma, and glioma^18–20. Our study suggests that AZD7762 may be a promising therapeutic candidate for ESCC; however, further validation in vivo and in clinical settings is required before translational application.

Overall, this study presents a novel prognostic signature and proposes a promising multitarget drug repurposing strategy, paving the way for improved clinical management of patients with ESCC. However, further validation is required before translational application. The immediate next steps include testing AZD7762 efficacy and mechanism in ESCC patient-derived organoid models and xenograft or genetically engineered mouse models and evaluating its impact on tumor growth, invasion, and survival endpoints in vivo. In parallel, mechanistic studies should be performed to define the downstream signaling consequences of IL22RA1 inhibition and to elucidate the molecular role of FAM221A in ESCC proliferation and DNA repair. Finally, future extensions of this AI-driven pipeline should explicitly incorporate mutational features, longitudinal datasets, and resistance modeling to enable resistance-aware optimization of next-generation multitarget candidates and improve the durability of therapeutic responses. However, our results should be further validated in preclinical and clinical studies to enhance their credibility. From a drug design perspective, simultaneous targeting of IL22RA1 and FAM221A with AZD7762 may help mitigate pathway redundancy and adaptive resistance in ESCC. IL22RA1 signaling converges on JAK–STAT and cell cycle regulation, whereas FAM221A is associated with proliferative and DNA repair programs. A multitarget inhibitor acting on these axes is more likely to retain efficacy in the context of clonal evolution and microenvironmental heterogeneity than a single-target inhibitor. Future extensions of this AI pipeline could explicitly integrate mutational data and resistance trajectories to enable resistance-aware optimization of next-generation candidates.

Methods

Summary-data-based Mendelian Randomization (SMR)

The SMR method integrates genome-wide association studies (GWAS) and expression quantitative trait loci(eQTL) data to identify pleiotropic associations between gene expression and complex traits.²¹. We acquired GWAS summary statistics pertaining to EC available through the IEU Open GWAS database using TwoSampleMR of R software²². A two-sample Mendelian randomization analysis was performed using the TwoSampleMR package of R software, in which the exposure variable was the druggable protein and the outcome of interest was EC²³. Rigorous criteria were employed for the selection of single nucleotide polymorphisms (SNPs) to comply with the independence assumption. Only cis-eQTL instruments were used for SMR (defined as variants located within ±1 Mb of the corresponding gene) to reduce the likelihood of spurious trans-effects in the results. SNPs were required to meet the predefined instrument significance threshold (p < 5 × 10⁻⁸), and linkage disequilibrium (LD) clumping was performed to retain approximately independent variants (r² < 0.001 within a 10,000 kb window). SNPs located in the major histocompatibility complex (MHC) region were excluded because of the complex LD structure. For exposures involving a single SNP, causal effects were estimated using the Wald ratio. For exposures involving multiple SNPs, causal estimates were derived using the inverse-variance weighted (IVW) method as the primary estimator. Heterogeneity was assessed using Cochran’s Q-test. Horizontal pleiotropy was evaluated using the MR-Egger intercept test, and the robustness of the IVW estimate was further assessed using weighted median estimation and leave-one-out analysis. The HEIDI test was applied within the SMR framework to distinguish pleiotropy from linkage, and the Steiger directional test was used to confirm the causal directionality. This involved the exclusion of SNPs linked to the major histocompatibility complex (MHC) region and the management of linkage disequilibrium to mitigate the effects of confounding bias. Furthermore, the Wald ratio method was used to analyze the outcomes of Mendelian randomization, in which the exposure comprised a single SNP²⁴. The MR results for exposures involving multiple SNPs were evaluated using the inverse variance weighting (IVW) method. Finally, the TwoSampleMR R package was used to evaluate heterogeneity and pleiotropic associations. Furthermore, the Steiger directional test was applied to confirm causal directionality²⁴.

Identification of DEGs and prognostic indicators

The GEOquery R package was used to download ESCC bulk cohorts(GSE161533, GSE17351, GSE77861, and GSE53625) from the Gene Expression Omnibus (GEO) database, along with the corresponding clinical pathological information and whole-genome expression data²⁵. For each dataset, tumor and normal samples were retained only if they had valid expression profiles and unambiguous sample annotations. For GEO datasets, primary ESCC tumor tissues and matched/adjacent normal esophageal tissues were included when available; samples derived from non-ESCC histologies, cell lines, or xenografts were excluded. For GSE53625, only patients with available overall survival information and complete clinical annotation were included for external validation, and cases with missing survival time or survival status were excluded from the analysis. For TCGA-ESCC, samples were included only if they had complete transcriptomic profiles and corresponding overall survival data; cases with missing follow-up information were excluded from prognostic modeling. When integrating multiple GEO cohorts (GSE161533, GSE17351, and GSE77861), samples were excluded if they lacked tumor/normal labels or showed poor quality after normalization and batch correction. The ComBat method, implemented with the R package sva, was used to integrate GSE161533, GSE17531 and GSE77861 datasets²⁶. The integrated datasets and GSE53625 were normalized and standardized using the Limma package of R software²⁷. In addition, probes associated with various molecules were eliminated, and probes linked to identical molecules were preserved only if they exhibited the highest signal intensity. The normalization process for the aggregated dataset and acquisition of differentially expressed genes (DEGs) was conducted using the Limma package within the R environment²⁷. The threshold of significant DEGs in the integrated dataset was | log2FC | > 0.5 and P < 0.05. After combining EC SMR results and DEGs, the expression pattern of these druggable DEGs was visualized in the combined dataset using the ggplot2 package of R software²⁸. These druggable DEGs genetic variation landscape and molecular functions of these druggable DEG were determined by mutation analysis and KEGG and GO enrichment analysis using the clusterProfiler and maftools packages of R software^29,30. We acquired STAR-count data along with relevant clinical details for ESCC from the TCGA database. Subsequently, we downloaded the data in TPM format and applied normalization using the log2(TPM + 1) transformation using the same methodology. To select potential prognostic indicators from druggable DEGs for ESCC, we performed univariate Cox regression using the survival package of R software³¹.

Non-negative definite matrix (NMF)

The methodology of non-negative matrix factorization (NMF) unveils the coefficient module at the transcriptional level, systematically organizing genes and samples to elucidate the underlying architecture of the dataset, thereby aiding in the identification of distinct subgroups³². In this investigation, NMF was employed to delineate subgroups of ESCC within the TCGA-ESCC cohort using the NMF R package. We applied the Brunet algorithm with 100 iterations across each value while exploring clusters (k = 2–4). The selection of the final cluster number was determined using multiple criteria, including the cophenetic correlation coefficient, dispersion, and silhouette width. We performed Gene Set Enrichment Analysis (GSEA) between the molecular subgroups based on the hallmark gene sets provided by MsigDB via the clusterProfiler package of R software. Kaplan-Meier (KM) analysis and clinical difference comparisons illustrated the clinical disparities between the two groups using the survival package of R software. The Tumor Immune Dysfunction and Exclusion (TIDE) framework was applied to predict potential responses to immunotherapy (IT)³³. Furthermore, immune checkpoint expression patterns were also determined among the subgroups.

Lasso regression and SHAP interpretation with prognostic model construction

To construct predictive models for prognosis and identify the most informative variables in ESCC, we applied LASSO-penalized Cox proportional hazards regression within the TCGA-ESCC cohort using the glmnet R package. Gene expression features were standardized prior to modeling. The optimal regularization parameter (λ) was selected by 10-fold cross-validation using the minimum mean cross-validated partial likelihood deviance (lambda.min), and the model stability was assessed across repeated cross-validation runs. The final prognostic risk score was calculated as a linear combination of selected gene coefficients. Model performance was evaluated using Kaplan-Meier survival analysis, time-dependent receiver operating characteristic (ROC) curves, and concordance index (C-index) in the TCGA-ESCC training cohort and independently validated in the GSE53625 cohort. For model interpretability, SHapley Additive Explanations (SHAP) were computed to quantify the marginal contribution of each gene to individual patient risk predictions, and SHAP summary plots were used to rank feature importance and visualize the directionality of effects. All modeling and visualization procedures were performed using standard R implementations, and key parameters (cross-validation folds, feature scaling, and λ selection strategy) are now explicitly reported to improve reproducibility. This approach was subsequently validated using the GSE53625 dataset³⁴. The configuration of the Lasso-Cox algorithm was assessed using the area under the curve (AUC) metric. Following this, for the construction of the prognostic model, we performed Kaplan-Meier (KM) analysis, risk factor assessment, and time-dependent receiver operating characteristic (ROC) analysis in both the TCGA-ESCC and GSE53625 cohorts. Notably, the factors contributing to the LASSO-Cox regression model were scrutinized using the SHAP model. To examine model performance, we performed Cox regression and nomogram-based calculations using the survival and rms packages of R software.

Single-cell transcriptomic analysis

We first acquired the ESCC single-cell dataset (GSE188900) from the GEO database. Processing of single-cell RNA sequencing (scRNA-seq) data, including quality control (QC), Reduction, Findmarker, was conducted using the Seurat R package³⁵. Quality control (QC) measures were applied to each cell to remove low-quality droplets and potential doublets. Specifically, cells were retained only if they had 200-6000 detected genes, UMI counts >1000, and mitochondrial transcript proportion <10%, which are commonly used thresholds in tumor scRNA-seq studies to exclude dying cells, empty droplets, and multiplets while preserving biologically meaningful populations. After QC filtering, the retained cells were used for downstream dimensionality reduction, clustering and annotation. To account for potential sample-to-sample technical variation, Seurat-based integration was applied across samples prior to clustering and downstream annotation. All downstream analyses were performed in the integrated expression space. Subsequent to QC, the data were normalized, enabling the identification of 2000 genes exhibiting high variability for further examination. Post-normalization, dimensionality reduction techniques, specifically t-SNE and uniform manifold approximation and projection (UMAP), were employed. Cell-type annotation was achieved using the SingleR algorithm³⁶. Transcript levels of the target genes were assessed in the different annotated cell populations. Intercellular communication networks were deduced using the CellChat package³⁷. Sc metabolism was analyzed using the scMetabolism package of R software³⁸. We conducted single-cell gene set enrichment analysis (ssGSEA) to investigate the functional enrichment of the hub gene at single-cell resolution in accordance with the hallmark gene set obtained from the MSigDB database³⁹. Single-cell monocle2 pseudotime analysis was performed to decipher the role of the targeted gene at the single-cell level⁴⁰.

Drug enrichment and validation

We utilized the pRRophetic package of R software to evaluate the optimal sensitivity drugs and then assessed the therapeutic efficacy using the ADMET-AI database^32,41. To evaluate the binding affinity between the drugs and targets, we performed molecular docking. Subsequently, molecular docking analyses were conducted to assess the binding affinity between the selected drug candidate and central gene protein. In particular, the three-dimensional structures of the key active compounds and target hub proteins were retrieved from the PubChem and Protein Data Bank (PDB) databases. The acquired protein structures were preprocessed using PyMOL software, which involved the incorporation of hydrogen atoms, charge calculations, and elimination of solvent molecules. Following this preparation, molecular docking was performed using AutoDock Vina (v1.2.0) implemented through AutoDockTools (v1.5.7). A semi-flexible docking strategy was adopted, in which the ligand was flexible and the receptor was treated as rigid. The docking search space was defined using a cubic grid box centered on the predicted binding pocket of each target protein, and the grid box dimensions (Å) and center coordinates (x, y, z) were defined to fully cover the predicted binding cavity. Docking was evaluated using the Vina scoring function (binding affinity in kcal/mol), and the top-ranked pose with a biologically plausible binding orientation was retained for the interaction analysis. To improve docking reliability⁴².

Cell lines and culture

The KYSE150(ESCC cell line) was acquired from the ATCC located in Manassas, USA. KYSE150 cells were maintained in RPMI 1640 medium (Hyclone, Cytiva, USA) supplemented with 10% fetal bovine serum (Gibco, Gaithersburg, MD, USA) and 1% penicillin-streptomycin (Gibco, Gaithersburg, MD, USA). Cells were maintained in an environment containing 5% carbon dioxide at 37 °C, with the medium replenished every 2–3 days.

Silencing via shRNA and overexpression

We produced lentiviral particles using 8 μg of the PLKO.1-puro vector in combination with 5 μg of packaging and envelope vectors for co-transfection of HEK293T cells, employing Lipofectamine 2000 (Thermo Fisher Scientific) in accordance with the vendor’s instructions. Following a 48-hour incubation period post-transfection, the lentiviral particles were collected. Subsequently, KYSE150 cells were transduced with lentivirus containing known and control sequences for 24 h. Two days post-infection, virus-transduced cells were selected with 4 μg/ml puromycin (Sigma-Aldrich) for an additional 48 h and then subjected to subsequent assays. The siRNA sequences were synthesized and designed by GenePharma (Shanghai, China), with the specific sequences provided as follows:

shRNA-IL22RA1-1:

5’CCGGGCAGAGAGAATATGAGTTCTTCTCGAGAAGAACTCATATTCTCTCTGCTTTTTG 3’

shRNA-IL22RA1-2:

5’CCGGCCTAAAGGTCAGCTTCAGAAACTCGAGTTTCTGAAGCTGACCTTTAGGTTTTTG 3’

shRNA- FAM221A-1:

5’CCGGGATTCCATAGCTGCTTCACTTCTCGAGAAGTGAAGCAGCTATGGAATCTTTTTG 3’

shRNA- FAM221A-2:

5’CCGGGAGGATGATATGGCTTTCTTTCTCGAGAAAGAAAGCCATATCATCCTCTTTTTG 3’

Colony formation assay

Cells undergoing logarithmic growth were treated with trypsin, resuspended, and seeded in 6-cm dishes at a density of 1000 cells/dish. Following a culture period of 2–3 weeks, the cell colonies were fixed with 4% paraformaldehyde for 20 min, stained with crystal violet staining was performed, and counted. Colony formation efficiency was determined using the following formula: (number of colonies/number of seeded cells) × 100%.

Cell proliferation assessment

Cells in the logarithmic growth phase were trypsinized, enumerated, and plated in 96-well plates at a density of 3000 cells/well (n = 6). Following incubation at predetermined intervals at time points up to 120 h (4, 24, 48, 72, 96, and we 120 h), CCK-8 (10 µL of) reagent was added to each well, and the plates were incubated for another 2 h. Absorbance was measured at 450 nm using a microplate reader. The cell proliferation rate was calculated using the formula (A_d − A_blank_d) / (A_4h − A_blank_4h). All data are from triplicate experiments.

Transwell assay

A Transwell chamber (Corning Inc., New York, USA) was used for the experiment. A total of 150 µL of cell suspension comprising 10,000 cells was introduced into each chamber. Next, 600 µL of culture medium containing 10% fetal bovine serum was added to the chamber below. After 24 h of incubation, non-migrated cells were removed, and the cells that had migrated were stained with a 0.25% crystal violet solution for 10 min. The chambers were then rinsed with phosphate-buffered saline (PBS) and imaged. For the in vitro cell invasion assay, Transwell chambers that were pre-coated with a matrix gel at a concentration of 100 micrograms per square centimeter, obtained from Thermo Fisher Scientific, were employed.

Western blotting

After treatment, the cells were harvested by scraping, washed with ice-cold PBS, and lysed with RIPA buffer. Protein lysates were electrophoresed on a 10% Tris-MES gel, transferred to PVDF membranes, and probed sequentially with specific primary antibodies and IRDye-conjugated secondary antibodies (1:1000 dilution). Target proteins were visualized using an Odyssey CLx imaging system (Westburg) with β-actin expression as the loading control. Protein expression was determined by quantifying the target band densities using ImageJ (v1.57). The density values were normalized to those of β-actin. The details of the primary antibodies used in this study are as follows:IL22RA1 (ab5984, ABCAM, USA: 1:2000), FAM221A (RP-92835, THERMO, USA: 1:2500), β-actin (No. 66009-1-Ig, Proteintech Group, Inc, CHINA:1:1000).

Ethics approval and informed consent

This study used only publicly available, de-identified human datasets (GEO, TCGA, and IEU OpenGWAS platform). No participants were recruited, and no identifiable private information was accessed. The experiments were limited to established cell lines (KYSE150 and HEK293T) and did not involve the collection of new human specimens or animal work. Accordingly, ethics approval and informed consent were not required, as determined by the Medical Ethics Committee of Fujian Medical University Union Hospital, Fuzhou, China (waiver/reference number: not applicable).

Statistical analysis

All statistical analyses were performed using R software (R Foundation for Statistical Computing) and GraphPad Prism. For reproducibility, the R version and versions of all key R packages used (including TwoSampleMR, GEOquery, sva, limma, Seurat, NMF, clusterProfiler, maftools, and CellChat) are provided in the final revised manuscript. All public datasets were accessed from GEO, TCGA, and the IEU OpenGWAS platform, and the access dates for each resource are reported. The differences between the two groups were assessed using either the Student’s t-test or the Wilcoxon rank-sum test, contingent upon the distribution of the data. For comparisons involving multiple groups, one-way ANOVA was applied, followed by Tukey’s post-hoc analysis. The relationship between gene expression levels and immune cell infiltration was analyzed using Spearman’s correlation analysis. A two-tailed p-value of less than 0.05 was considered statistically significant. For in vitro experimental validation, all functional assays (CCK-8, colony formation, wound healing, Transwell migration/invasion, and western blotting) were performed using at least three independent biological replicates, and all key experiments were independently repeated to confirm reproducibility. Quantitative results are presented as mean ± standard deviation (SD). Given the nature of cell-based assays, formal randomization and blinding were not applied; however, cells were handled under identical experimental conditions, and outcome quantification was performed using predefined, standardized procedures to minimize subjective bias.

Author contributions

Z.Z. and S.Y. designed the study, performed primary bioinformatics analyses, and drafted the main manuscript text. K.P. and P.Z. conducted the machine learning modeling, NMF clustering, prognostic model construction, and SHAP interpretation. J.Y. processed and analysed the single-cell RNA-seq dataset, including cell-type annotation, CellChat communication mapping, and metabolic profiling. J.L. supervised the molecular docking, ADMET-AI evaluation, and drug repurposing validation pipeline. M.K. designed and supervised all in vitro experiments, including cell proliferation, invasion, wound healing, clonogenic assays, and western blotting. All authors contributed to the data interpretation, critically revised the manuscript, and approved the final version for publication. Z.Z. and S.Y. contributed equally to this study.

Data availability

All datasets used in this study are publicly accessible. GEO bulk RNA-seq datasets (GSE161533, GSE17351, GSE77861, GSE53625) and the single-cell dataset (GSE188900) can be accessed at https://www.ncbi.nlm.nih.gov/geo/. TCGA-ESCC data were obtained from the GDC portal (https://portal.gdc.cancer.gov/), and the GWAS summary statistics were retrieved from the IEU Open GWAS platform (https://gwas.mrcieu.ac.uk/). Supplementary Figure S3, containing raw unprocessed western blot scans corresponding to Figs. 9A and 10A, is provided as source data and is available within the Data Availability section. Code Availability: Code will be available from the corresponding authors upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41698-026-01485-z.