An effective multistage mouse model of esophageal carcinogenesis for preclinical and computational pathology applications

Highlights

• •Define the tumor-promoting role of SOR by activating the Raf-MEK-ERK pathway in SSE.
• •Establish a novel and effective mouse CIMCM of ESCC using 4NQO and SOR.
• •The CIMCM of ESCC markedly reduces the TMB to a level similar to that of human ESCC.
• •Microenvironment plays a critical role in ESCC carcinogenesis.
• •WSIs of the CIMCM of ESCC applied in C-Path enables ESCC multistage diagnosis.

Abstract

The use of carcinogen-induced multistage carcinogenesis animal models of esophageal squamous cell carcinoma (CIMCM of ESCC) is limited by prolonged timelines, high toxicity, and excessive mutational burden. In this study, we report the establishment of an effective mouse CIMCM of ESCC by using 4-nitroquinoline-1-oxide (4NQO) as a carcinogen and sorafenib (SOR) as a tumor promoter. We show that SOR specifically activates the Raf-MEK-ERK signaling pathway in normal esophageal stratified squamous epithelium cells, thereby promoting tumor progression. This CIMCM of ESCC accurately recapitulates the multistage process of ESCC carcinogenesis from precancerous lesions to invasive carcinoma, with shortened time and high efficiency. Pathological, molecular, cellular and multiomic analyses show that the CIMCM of ESCC significantly reduces the tumor mutation burden to levels detected in human ESCC samples, while preserving key genetic driver mutations and abnormal transcriptomic/protein expression profiles. Notably, the CIMCM of ESCC demonstrates that the tissue microenvironment plays an important role in ESCC carcinogenesis, as the application of mechanical injury to the esophageal SSE of the CIMCM results in the inflammatory-related response, site-specific tumor formation and high tumor incidence. Since the CIMCM of ESCC provides valuable samples from different stages of tumor initiation and progression, the pathological whole slide images of the CIMCM of ESCC are applied to the computational pathology, which enables the detection, segmentation and annotation of the ESCC initiation and progression with pathologist-level accuracy. Taken together, this mouse CIMCM of ESCC provides a versatile platform for ESCC early diagnosis, basic and preclinical research and therapeutic strategy.

Introduction

Esophageal cancer is one of the most common malignancies of the gastrointestinal tract and is the seventh leading cause of cancer death worldwide [1,2]. Esophageal cancer mainly manifests as esophageal squamous cell carcinoma (ESCC) in Asia, but due to the lack of obvious symptoms in the early stages and effective diagnostic methods, the majority of ESCC patients are diagnosed in the late stages, resulting in poor prognosis and high mortality [2,3]. Therefore, the development of feasible, effective and reliable preclinical and clinical models/methods for early detection and treatment of ESCC is essential for improving survival.

Like many types of squamous cell carcinoma (SCC), ESCC carcinogenesis is an intricate, multifactorial, and multistep process involving genetic susceptibility, lifestyle factors, and environmental exposures that sequentially progresses from normal esophageal epithelium (NOR) to hyperplasia (HYP) to dysplasia (DYS) to invasive carcinoma (CAR) [4,5]. Over the past two decades, advances in multiomic techniques have identified dozens of cancer driver genes and their associated molecular pathways involved in the carcinogenesis of human ESCC [6,7]. However, considering that the mammalian esophageal stratified squamous epithelium (SSE) is one of the most rapidly renewing tissues in the body, cancer driver gene mutations have also been frequently identified in normal esophageal SSE specimens [8]. These findings indicate that the initiation and progression of ESCC are not only dependent on the accumulation of mutations in cancer driver genes and their associated pathways but may also be influenced by other non-mutagenic factors such as tumor-promoting factors, tissue injury/inflammatory effects and/or environmental/microenvironmental exposures [9]. Thus, accumulating evidence demonstrates that genetic together with epigenetic/environmental alterations contribute to ESCC carcinogenesis, although the precise mechanisms underlying ESCC initiation and progression remain largely unknown.

To better understand the process of carcinogenesis, the design and development of feasible, effective and reliable animal models has proven invaluable in providing cancer research with samples from different stages of tumor initiation and progression that are not readily available from human patients. Currently, the most commonly used animal models include genetically engineered mouse models (GEMMs) and chemical carcinogen-induced multistage carcinogenesis models (CIMCMs) [[10], [11], [12]]. Each type of model has advantages and disadvantages, with GEMM primarily used to study the molecular mechanisms by which one or a few defined genes are involved in the carcinogenesis of a specific tissue/organ, whereas CIMCM is used to study the cancer etiology of a specific tissue/organ for the initiation and progression of multistep carcinogenesis processes [5,13,14]. Furthermore, both individual models and their combination, particularly the CIMCM model, can be used for early cancer detection, therapeutic strategy development, and investigations into the competition of normal/cancer cells, as well as cellular plasticity, stemness, polarity, and disruptions in tissue hierarchy and homeostasis [15,16]. The best example of CIMCM is the dimethylbenzanthracene/12-O-tetradecanoylphorbol-13-acetate (DMBA/TPA)-induced multistage mouse skin carcinogenesis model [17,18], in which DMBA serves as a carcinogen and TPA serves as a promoter that can efficiently and reliably induce skin tumors at the different stages to study the etiology and molecular mechanisms of skin SCC carcinogenesis. However, as TPA is not applicable to internal animal tissues/organs, efficient and reliable CIMCM has not been well established for most internal animal tissues/organs.

Recently, our laboratory reported the successful use of the multi-kinase inhibitor (MKI) sorafenib (also known as SOR, SRF or SORA, hereafter referred to as SOR) as a tumor promoter together with N-nitrosomethylbenzylamine (NMBzA) as a carcinogen to establish a feasible, efficient and reliable NMBzA/SOR-induced rat multistage model of ESCC carcinogenesis [4]. Although the NMBzA-induced rat model has been used to study ESCC initiation, progression, and new therapeutic strategies, it is time-consuming and has a low incidence of malignant tumors [19,20]. Our study dramatically increased the efficiency and consistency of esophageal tumor formation. The application of SOR as a tumor promoter is based on several clinical reports that SOR is an MKI used for several human cancer treatments, however its clinical use had been found to lead to SSE lesions and the development of keratoacanthomas and SCCs [21,22]. These findings are also consistent with the previous pharmacological reports of SOR, in which SOR had paradoxical effects on cell proliferation/differentiation/death in different cell types with different genetic backgrounds in different tissues [23,24]. We showed that, in contrast to inhibiting the growth of some tumor cells, SOR could promote cell proliferation in normal rat esophageal SSE and some cell lines, although the potential underlying mechanism remains to be determined [4]. Meanwhile, the NMBzA/SOR-induced rat ESCC model has several technical drawbacks: 1) NMBzA is not water-soluble and administration of NMBzA requires gavage or intraperitoneal injection; 2) in contrast to mouse models, subsequent genetic manipulation and multiomic analyses are more complex and difficult to achieve in rat models; 3) it is unclear whether SOR can be applied as a tumor promoter with other types of carcinogens besides nitrosamines to induce SCC in different experimental animal species and/or different experimental designs.

To circumvent or resolve these problems, in this study we report the administration of a water-soluble quinoline derivative, 4-nitroquinoline 1-oxide (4NQO), as a carcinogen together with SOR as a promoter, to establish an effective and reliable multistage mouse model of ESCC carcinogenesis (the 4NQO/SOR-induced mouse CIMCM of ESCC) for preclinical and computational pathology (C-Path) applications.

Materials and methods

Mouse CIMCM of ESCC

C57BL/6 J mice drank 4NQO (100 μg/ml) for up to 16 weeks. From week 8, SOR (50 mg/kg) was administered intraperitoneally every other day for 8 weeks. Biweekly esophageal brushing was performed under DSA guidance. Mice were sacrificed, and esophagi were processed for tumor analysis, formalin fixation, or −80°C storage. All animal experiments were approved by the Ethical Committee of National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and their care was in accordance with institution guidelines.

Phosphoprotein microarray

D3 cells cultured in conditioned medium were treated with 10 μM SOR for 2 hours. Phosphorylation levels were assessed using a phosphorylation-specific antibody microarray targeting 157 phosphorylation and 144 non-phosphorylation sites. Data were scanned by Agilent SureScan Dx and analyzed with GenePix Pro 6.0. Phosphorylation or protein expression changes were evaluated by calculating the fold change (FC) between the SOR-treated group and the control group.

EMSS-Net architecture

EMSS-Net comprises an encoder and a decoder. The encoder utilizes Mix Transformer for multi-scale feature extraction. The decoder consists of two steps: Step 1 integrates shallow features via the Shallow Fusion Module and optimizes epithelial segmentation using cross-entropy loss, generating an epithelial mask. Step 2 refines multi-stage segmentation within epithelial regions using the Scale Alignment Module, Multi-Scale Fusion Module, and Gate Module, optimized with cross-entropy loss and MSE loss for improved segmentation accuracy.

Evaluation metrics

To evaluate EMSS-Net's segmentation performance, we used common pixel-wise metrics in medical imaging. Dice coefficient measures overlap with the ground truth (GT), while Pixel Accuracy quantifies correctly segmented areas. Metrics are based on TP (correct target pixels), TN (correct non-target pixels), FP (incorrect target pixels), and FN (missed target pixels).

$$Dice=\frac{2*TP}{2*TP+FP+FN}$$

$$Acc=\frac{TP+TN}{TP+TN+FP+FN}$$

Statistical analysis

Statistical analysis was performed using GraphPad Prism 8.0. Differences were analyzed using unpaired t-test. Statistical significance was set at P < 0.05. Data are presented as mean ± SD unless otherwise stated.

Additional Materials and Methods that are routinely used in the laboratory are listed in the Supplementary Information section.

Results

SOR promotes cell proliferation in normal esophageal keratinocytes by specifically activating the Raf-MEK-ERK pathway

To determine the potential underlying mechanism by which SOR acted as a tumor promoter to drive cell proliferation in normal esophageal keratinocytes, we first administered multiple doses of SOR to C57BL/6 J mice for 28 days, followed by an in vivo 5-Ethynyl-2′-deoxyuridine (EdU) labeling experiment on the esophageal SSE. A detailed experimental procedure and results were shown in Fig. 1A and B. Cell proliferation increased in a dose-dependent manner with SOR administration, reaching a plateau at 50–100 mg/kg in vivo. Based on this observation, we selected 50 mg/kg as the standard dose for all subsequent in vivo experiments. This result was consistent with our previously published results in rats and further demonstrated that SOR administration promoted keratinocyte proliferation in normal murine esophageal SSE [4,15].

Fig. 1.

SOR promotes the proliferation of esophageal cells by activating the Raf-MEK-ERK pathway. (A) Study design. 8-week-old C57BL/6 J mice were randomly assigned to the SOR treatment group or the vehicle control group. SOR or vehicle were administered by i.p. injection at multiple doses of SOR every other day for a total duration of 28 days. EDU was injected at day 28, and mice were sacrificed after 2 h. (B) Representative vertical views of esophageal epithelial whole mounts, stained for EDU after treatment with different concentrations of SOR. Scale bar, 200 μm. (C) The fold change between SOR-treated and negative control of phospho-protein and protein expression in the MAPK pathway was analyzed by phosphor-protein microarray profiles in D3 cells. (D) Immunoblot analysis of MAPK pathway related markers in D3 cells treated with SOR or EGF in the conditional medium. (E) Representative views of IHC staining of mouse esophageal tissues treated with SOR and control. Scale bars, 100 μm. (F) Schematic representation of SOR in driving Raf dimerization within esophageal epithelial cells, activating downstream phosphorylation of MEK and ERK. (G) Immunoprecipitation assays were performed on total cell lysates of D3 cells treated with 10 μM SOR, 10 nM EGF, or negative control, using an anti-B-raf antibody. Experiments in D and G were repeated at least three times.

To better understand how SOR promoted cell proliferation in esophageal SSE cells, we conducted a phosphoproteomic analysis using a phospho-specific antibody microarray targeting 157 phosphorylated proteins covering 16 major signaling pathways, incubated with lysates from normal immortalized rat esophageal D3 cells, which were established in our lab [25], treated with or without SOR (Fig. S1A and B). To obtain maximum effects, D3 cells were first cultured overnight in medium containing 0.5 % fetal bovine serum (FBS) (conditioned medium) and then treated with 10 μM of SOR for 2 h. Subsequently, cell lysates from SOR-treated and untreated samples were subjected to phosphoproteomic analysis. The phosphoproteomic analysis revealed that the significant activation (phosphorylation) of the RAF-MEK-ERK pathway was detected in SOR-treated D3 cells compared to control, SOR-untreated D3 cells (Fig. 1C and Table S1). Other potential SOR-targeted receptor tyrosine kinase (RTK) pathways, such as the VEGFR pathway or RTK downstream pathways besides the Raf-MEK-ERK pathway, such as the AKT pathway and others were not markedly affected in SOR-treated D3 cells (Fig. S1C). These results, together with the data showing that SOR promoted the proliferation of esophageal SSE cells in vivo, indicated that SOR enhanced the proliferation of normal esophageal keratinocytes by specifically activating the RAF-MEK-ERK signaling pathway.

To ascertain how SOR specifically activated the Raf-MEK-ERK pathway to promote normal esophageal keratinocyte proliferation, we performed immunoprecipitation and immunoblotting experiments with cell lysates from D3 cells treated with 0, 1, 5, and 10 μM of SOR for 2 h after culturing in conditioned medium. As a positive control, D3 cells cultured in conditioned medium were also treated with 10 nM epidermal growth factor (EGF) for 15 min. Immunoblotting revealed that D3 cells treated with 1, 5 and 10 μM SOR exhibited a significant increase in MEK/ERK phosphorylation, but not in EGFR (a RTK pathway) or AKT (the PI3K-AKT pathway) when compared to the control, SOR-untreated cells, indicating that SOR specifically activated the Raf-MEK-ERK pathway. In contrast, D3 cells cultured in conditioned medium treated with 10 nM EGF showed not only a significant increase in EGFR phosphorylation but also significant increases in EGFR downstream MEK/ERK phosphorylation and AKT phosphorylation (Fig. 1D). To verify that SOR-driven cell proliferation depended on MEK-ERK signaling, D3 cells were treated with 10 μM of SOR alone or in combination with the MEK inhibitor trametinib or the ERK inhibitor SCH772984 for two hours after culturing in conditioned medium. As expected, co-treatment with trametinib or SCH772984 abolished ERK phosphorylation and reversed the effect of SOR (Fig. S1D). To further determine whether the cell proliferation induced by SOR could be also blocked by MEK and ERK inhibitors, we examined the expression of Myc mRNA, which is a cell cycle-dependent downstream target of the transcription factor AP-1 (Fos/Jun) regulated by the MEK-ERK signaling pathway. Myc mRNA expression increased significantly after SOR treatment in D3 cells but was suppressed markedly when D3 cells were exposed to trametinib or SCH772984, either alone or in combination with SOR (Fig. S1E). In addition, immunoblotting and immunohistochemistry revealed that mouse esophageal SSE treated with SOR exhibited an increase in the phosphorylation of MEK and ERK in epithelial layers in vivo (Fig. S1F and Fig. 1E). Mechanistically, co-immunoprecipitation (co-IP) experiments demonstrated that, when compared to that in untreated cells, the interaction of endogenous H-Ras/B-Raf was significantly enhanced in D3 cells treated with EGF whereas the interaction of endogenous H-Ras/B-Raf showed no change in D3 cells treated with 10 μM SOR (Fig. 1F and G). In contrast, the coimmunoprecipitation of endogenous B-Raf/Raf-1 (c-Raf) was detected in D3 cells after 10 μM of SOR treatment, indicating that SOR treatment in D3 cells resulted in enhanced interactions of hetero-Raf protein species, which had been shown to specifically activate the Raf downstream MEK-ERK signaling pathway [26]. Thus, these results together with previous reports demonstrated that, SOR treatment in normal esophageal SSE cells only enhanced the interactions of B-Raf/Raf-1 proteins, which could form heterodimers to specifically activate the Raf-MEK-ERK signaling pathway, thereby promoting cell proliferation in normal esophageal keratinocytes [23,24,27].

Establishment of an effective and reliable 4NQO/SOR mouse CIMCM of ESCC

To date, the most commonly used mouse CIMCM for ESCC had been the 4NQO-induced model in which 4NQO is a water-soluble quinoline (non-nitrosamine compound) [5,28]. As with the use of NMBzA alone to induce the rat ESCC model, the chemical carcinogen 4NQO alone was shown to induce upper gastrointestinal malignancies, including head and neck squamous cell carcinoma (HNSCC) and ESCC in mice [5,13,29]. However, the carcinogenesis process of using 4NQO alone in mice was also time-consuming, requiring approximately 16 weeks of 4NQO administration, similar to that of NMBzA administration in rats. Unlike NMBzA, 4NQO was water-soluble and could be administered via the drinking water, but it was highly toxic and the prolonged treatment could result in a very high incidence of mortality in mice [30,31]. To overcome these limitations and to establish an effective and reliable mouse CIMCM for ESCC, we tested whether SOR could also serve as a tumor promoter in combination with 4NQO as a carcinogen, as we had done in rat CIMCM of ESCC using NMBzA/SOR [4].

A detailed experimental design was shown in Fig. 2A and Table S2, where 72 C57BL/6 J male mice were randomly divided into eight groups, including C, C-S, N1, N1-S, N5, N5-S, N16, N16-S. Group C served as the experimental control (no 4NQO or SOR treatment), C-S were administered SOR from week 8 to week 16, while others received 4NQO for durations ranging from 1 to 16 weeks with/without SOR for 8 weeks from week 8 to week 16, respectively. Mice in each group were sacrificed and their esophageal tissues were collected and analyzed at weeks 22, 35, or 42, respectively. In summary, short-term 4NQO treatment (1 week) with/without subsequent SOR administration (N1 and N1-S) or SOR alone (C-S) did not increase esophageal tumor incidence in mice compared to the control (C). In contrast, 5 weeks of 4NQO treatment followed by SOR administration (N5-S) significantly induced gross roughness lesions of esophageal SSE and increased esophageal tumor incidence in mice at the weeks 22, 35, or 42 compared to mice treated with 4NQO alone for 5 weeks (N5) (Fig. 2B). These gross roughness lesions of esophageal SSE and esophageal tumor incidence observed in N5-S closely resembled those observed in mice treated with 16 weeks of 4NQO with/without subsequent SOR (N16 and N16-S). However, unhealthy mice and frequent deaths, which were observed in N16 and N16-S, were rarely seen in N5 and N5-S (Fig. 2C-F). Taken together, these results indicated that shortening 4NQO treatment to 5 weeks followed by 8 weeks of SOR administration in mice could markedly reduce the toxic effects of the prolonged and extensive 4NQO treatment in mice, enabling the establishment of a feasible and effective mouse CIMCM for ESCC.

Fig. 2.

Establishment of mouse CIMCM following varying durations of 4NQO exposure and subsequent SOR treatment. (A) Study design. 72 male C57BL/6 J mice were randomly assigned into eight groups (C, C-S, N1, N1-S, N5, N5-S, N16, N16-S) with varying durations of 4NQO treatment (1, 5, and 16 weeks in drinking water) and/or SOR administration (8 weeks, from week 8 to week 16), the mice were sacrificed at weeks 22, 35, or 42 to evaluate the incidence and progression of esophageal tumors. Mice were treated as follows: Control (C): mice were maintained under normal conditions until the end of the experiment, C-S: injected with SOR, N1: administered 4NQO for 1 week, N1-S: administered 4NQO for 1 week, then injected with SOR, N5: administered 4NQO for 5 weeks, N5-S: administered 4NQO for 5 weeks, then injected with SOR, N16: administered 4NQO for 16 weeks, N16-S: administered 4NQO for 16 weeks, and then injected with SOR. (B) Macroscopic view of the whole esophagus of mice from all groups at weeks 22, 35, and 42. Scale bar: 1 cm. (C-E) Body weight of mice at different time points, (C) at week 22 of mice from group C (n = 4), C-S (n = 4), N1 (n = 6), N1-S (n = 10), N5 (n = 14), N5-S (n = 11, 3 mice were died during the injection procedure, and excluded from the estimate of mortality rate), N16 (n = 7), N16-S (n = 6); (D) at week 35 of mice from group C (n = 4), C-S (n = 4), N1 (n = 6), N1-S (n = 10), N5 (n = 10), N5-S (n = 9), N16(n = 3), N16-S (n = 2). All mice in N16 and N16-S groups were sacrificed at week 35 according to the animal welfare; and (E) at week 42 of mice from group C (n = 4), C-S (n = 4), N1 (n = 4), N1-S (n = 8), N5 (n = 6), N5-S (n = 6); unpaired t-test. (F) The survival rate of mice from experiment initiation to week 35, with exclusion of mice sacrificed at week 22.

We repeated the key experiments of the aforementioned 4NQO/SOR mouse CIMCM for ESCC to ensure that the model was reproducible and reliable. 22 mice were randomly divided into four groups. Mice were treated as follows: 4NQO for 5 weeks (N5); 4NQO for 5 weeks followed by SOR for 8 weeks (N5-S); 4NQO for 16 weeks (N16); or 16 weeks of 4NQO combined with the last 8 weeks of SOR (N16-S) (Fig. 3A). We combined the data from both experiments and found that the body weights of mice in N5 and N5-S maintained normal at week 35 (Fig. 3B and Fig. S2A). In contrast, mice in N16 and N16-S exhibited significant weight loss at this time point compared to control mice. In addition, when compared to controls, the extensive and prolonged 16-week 4NQO exposure resulted in low survival rates. Mice in N5-S survived to the end of the experiment and the survival rate in N5 exceeded 80 %, while it was below 70 % in N16 and N16-S up to week 35 (Fig. 3C and Fig. S2B). Furthermore, mice in N5-S developed gross esophageal SSE roughness lesions and tumors, which were nearly identical in severity to those observed in N16 (Fig. 3D). Nevertheless, as in mice treated with the extended 16 weeks of 4NQO, histopathologic analysis of hematoxylin and eosin (H&E) staining showed that the esophageal epithelia of mice treated with 5 weeks of 4NQO followed by 8 weeks of SOR treatment exhibited the full spectrum of progression from NOR to HYP to DYS and then to CAR (Fig. 3E). Notably, gross anatomical observations and H&E analysis also revealed that when comparing N16/N16-S or N5/N5-S, the combined treatment of 4NQO with SOR significantly increased the tumor burden in each mouse and slightly accelerated the onset of more advanced pathological stages than mice treated with 4NQO alone (Fig. 3F and G, Table S3). Taken together, these results demonstrated that SOR could serve as a tumor promoter, in combination with a non-nitrosamine compound, 4NQO, as a carcinogen to establish an effective and reliable mouse CIMCM for ESCC.

Fig. 3.

5-week 4NQO and 8-week SOR induced mouse CIMCM of ESCC. (A) Study design. 70 male mice were utilized, including both the first and second experiment. Mice were treated with 5 or 16 weeks of 4NQO and with or without subsequent SOR injected for 8 weeks. (B) Body weight of mice in two experiments at week 35 from Group N5 (n = 16), N5-S (n = 15), N16 (n = 7), N16-S (n = 6), unpaired t-test. (E) The survival rate of each group with exclusion of mice sacrificed. (F) Macroscopic view of the whole esophagus of each group. Scale bar,1 cm. (G) Pathological progression from normal esophageal epithelium to ESCC by H&E staining. Scale bar,100 μm (upper); 50 μm (lower). (H) Numbers of tumors per entire esophagus at weeks 22, 35 and 42 of Group N5 and N5-S. N5 (week 22) 0.0 ± 0.0, n = 2; N5-S (week 22) 1.5 ± 0.7, n = 2; N5 (week 35) 0.5 ± 0.5, n = 8; N5-S (week 35) 2.1 ± 1.1, n = 9; N5 (week 42) 1.7 ± 1.2, n = 6; N5-S (week 42) 2.2 ± 1.2, n = 6. unpaired t-test. (I) Percentage of the indicated lesion types present in the esophagus of Group N5 and N5-S at weeks 22, 35 and 42. mean ± SD, **P < 0.01, ****P < 0.001.

Previously, several studies reported multiomic analyses of mouse HNSCC/ESCC models induced by 4NQO alone [5,13]. Whole-exome sequencing (WES) analysis revealed significant overlap in the mutation signatures and driver genes between 4NQO-induced mouse HNSCC samples and human HNSCC samples [13,32,33]. The average mutation burden in 4NQO-treated mouse samples was correlated with the duration of 4NQO treatment. Thus, the mouse samples treated extensively with 4NQO for 10-16 weeks exhibited approximately 1,000–5,000 SNVs, showing a 50-100 times higher mutation burden than that observed in human SCC samples (Fig. S3A), which could raise concerns about the applicability of this model for potential molecular mechanistic studies of SCC carcinogenesis [13]. Given the 4NQO/SOR mouse CIMCM of ESCC described here, which was markedly shortened to just 5 weeks of 4NQO treatment, we assumed that the model would not only be effective and reliable, but also strikingly reduce the mutation burden in the model samples. Therefore, WES was performed using DNA extracted from four randomly selected representative SSE samples from N5-S at the 35-week time point, with the tail DNA serving as an internal control. As shown in Fig. 4A, the mutation burden in samples #1 and #2, which were derived from esophageal SSE with gross roughness lesions (HYP/DYS), was <1 per megabase pair (<1/Mbp), while samples #3 and #4, obtained from esophageal SSE with ESCC (CAR), showed a mutation burden of 1-2 per megabase pair (1-2/Mbp). These results indicated that the mutation burden in samples from the 4NQO/SOR mouse CIMCM of ESCC was markedly reduced compared to that of mice extensively treated with 4NQO alone (50-100/Mbp), and closely matched to the mutation burden observed in human SCC samples [34]. Furthermore, the mutational signatures, driver genes and single base substitution (SBS) signatures detected in these 4 samples showed similarities to those of previously reported 4NQO-induced mouse HNSCC samples and human ESCC samples (Fig. 4A and B, Fig. S3B and C).

Fig. 4.

The WES, bulk-RNA and pathological analysis of samples from Group N5-S at 35 weeks. (A) Representative mutational landscapes of ESCC-Associated driver genes in 4NQO and SOR-induced samples. Samples #1 and #2 were esophageal SSE with gross roughness lesions, and samples #3 and #4 were esophageal SSE with ESCC. (B) Frequency of base substitutions within specific trinucleotide mutational contexts in samples. (C-E) Transcriptomic profiling comparing Group N5-S to Group C samples. (C) Volcano plot illustrating upregulation of inflammatory related genes, (D) Representative KEGG pathway enrichment analyses related to proliferation-differentiation, inflammation, cell death, cell cycle, adhesion, and other processes. and (E) GSEA of inflammatory and proliferation-related signatures. (F) KEGG pathway enrichment analysis of human ESCC samples versus normal tissues. Datasets were obtained from GEO (GSE130078, GSE164158, GSE213565, GSE235537, and GSE32424). (G) IHC staining of mouse esophageal tissues with different lesions. Scale bars, 100 μm. (H) Stacked histograms showing quantification of IHC staining scores (0, negative; 1+, week positive; 2+, median positive; 3+, strong positive). Each column was summarized from at least three visual fields. (I) H&E staining showing inflammatory cell infiltration across NOR, HYP, DYS and CAR. Yellow arrows indicate regions of inflammatory cell infiltration. Scale bar, 200 μm (upper); 75 μm (lower). (J) Quantification of inflammatory cell infiltration (H&E staining) across NOR (n = 3), HYP (n = 6), DYS (n = 6) and CAR (n = 6) from H&E staining. Unpaired t-test, *P < 0.05, **P < 0.01.

Given that Trp53 was the most frequently mutated gene in human ESCC, we assessed Trp53 status in our cohort by Sanger sequencing. Nonsynonymous SNVs in the exon regions of Trp53 were detected in 2 of 9 samples (Fig. S3D), corresponding to a mutation frequency of 22 %, which was similar to the previously reported Trp53 mutation rate of 30 % in 4NQO-induced mouse SSC samples [13]. The lower frequency of Trp53 mutations detected in these 4NQO-induced CIMCM models could be attributed to factors such as sequencing depth or contamination with normal cells, etc. Therefore, we further determined the dysregulation of p53 in 4NQO/SOR-induced ESCC tissues using IHC. When compared to the weak p53 staining observed in normal SSE, ESCC samples displayed significant p53 accumulation in either the cytoplasm or nucleus, indicating that p53 was frequently dysregulated in ESCC cells (Fig. S3E) [35].

The initiation and progression of ESCC/SCC involve multistage carcinogenesis driven by genetic, epigenetic, and environmental alterations, ultimately resulting in transcriptomic and proteomic changes in cancer cells and their surrounding mesenchymal and immune cells. Bulk RNA sequencing (bRNA-seq) and single-cell RNA sequencing (scRNA-seq) had been frequently applied to tissue samples from different stages of preclinical models and human tumors to evaluate, identify and determine plausible key drivers and underlying mechanisms involved in carcinogenesis. Therefore, we performed bRNA-seq from 3 samples of esophageal SSE containing CAR in N5-S at 35 weeks, alongside 3 control samples from Group C (Fig. 4C-E, Fig. S3F-H). Principal component analysis (PCA), volcano plot, Kyoto encyclopedia of genes and genomes (KEGG), gene ontology (GO) and gene set enrichment analysis (GSEA) revealed that the 4NQO/SOR treated esophageal SSE samples containing CAR exhibited differential gene expression primarily related to cell cycle control, proliferation-differentiation homeostasis, adhesion, inflammation, and cell death, compared to controls. These findings aligned with previously published data from 4NQO-induced mouse CIMCM models of HNSCC/ESCC [5,13].

In summary, our study demonstrated that the 4NQO/SOR mouse CIMCM of ESCC was both effective and reliable, strikingly reducing the mutational burden of animal tumors to levels found in human ESCC. Moreover, the model preserved the full spectrum of pathological progression from NOR to CAR, along with genetic driver mutations and abnormal transcriptomic expressions. This model should be well-suited for preclinical applications in prevention, early detection, therapeutic development, and molecular mechanistic studies of ESCC carcinogenesis.

Application of the 4NQO/SOR mouse CIMCM of ESCC for preclinical studies

The successful establishment of an effective and reliable 4NQO/SOR mouse CIMCM of ESCC prompted us to explore its potential preclinical applications. Therefore, we analyzed extensive RNA-seq data of rat and human ESCC samples previously published along with limited data obtained in this study (Fig. 4F and Fig. S4A). Analysis from these studies revealed that genes involved in the biological processes and pathways, such as stem cell maintenance, cell proliferation and differentiation, competition and fitness, polarity, apoptosis, chromatin epigenetic modification, metabolism, inflammation and damage-repair response were commonly deregulated during the ESCC carcinogenesis. To confirm whether these transcriptomic alterations were reflected at the protein level, we performed immunohistochemistry (IHC) on tissue samples from NOR, HYP, DYS, and CAR stages of the 4NQO/SOR mouse CIMCM of ESCC (Fig. 4G and H, Fig. S4B). Representative phenotypic markers and transcription factors were selected to cover the enriched pathways: proliferation–differentiation (Krt14, p63, Sox2, Yap1, Krt13), inflammation (S100a8, S100a9), cell death/oxidative stress (Nrf2), cell-cycle progression (Ki67) and cell adhesion (Col17a1). IHC results corroborated the bRNA-seq data: Krt14, p63, Col17a1, Ki67, Yap1 and Nrf2 were significantly upregulated in basal or suprabasal cells during HYP and DYS stages, peaking in tumor cells at the CAR stage when compared to those in control NOR tissues. Krt13 expression decreased as the disease progressed through the stages. In contrast, high levels of S100a8/S100a9, members of the calcium-binding S100 alarmin family, which typically formed a heterodimer secreted from the SSE into the extracellular space to mediate tissue inflammatory responses, were predominantly expressed in HYP and DYS tissues [5,36]. These findings aligned with previous reports, suggesting that inflammatory responses could facilitate immune cell recruitment and tumor-promoting inflammation during ESCC carcinogenesis [5]. Thus, inflammatory cell infiltrations were observed in H&E-stained sections of HYP, DYS and CAR with a peak in HYP to DYS (Fig. 4I and J). Taken together, these results indicated that inflammatory responses and the dynamics and interactions between the esophageal SSE, the immune system and the tissue microenvironment played important roles in ESCC carcinogenesis.

Based on these findings, we investigated whether we could enhance inflammatory responses at a specific esophageal site by mechanical injury together with the 4NQO/SOR mouse CIMCM of ESCC to achieve a controllable localization for tumor formation as a preclinical application that would be useful for potential early detection, therapeutic strategies and interventional therapies of ESCC. Previous studies showed that inflammation was the initial response during wound healing of skin SSE and then followed by cell proliferation and tissue regeneration in the damaged area of skin SSE [37]. Furthermore, epidemiological studies linked the ingestion of hot or hard foods to esophageal mucosal injury, contributing to esophageal SSE damage and ESCC carcinogenesis [38]. To induce site-specific injury in murine esophageal SSE, we drew inspiration from the subglottic stenosis model [39] and applied mechanical injury to the lower esophagus of mice using a nylon brush, a process termed "brushing". We hypothesized that applying this method to the 4NQO/SOR mouse CIMCM of ESCC would induce a localized and controllable physical injury site in the mouse esophagus, resulting in enhanced tissue inflammation and stimulation of local cell proliferation, thereby controlling the sites for tumor formation.

To test our hypothesis, we performed initial experiments, in which C57BL/6 J mice were brushed with a nylon brush on their lower esophagus according to the protocol described in the subglottic stenosis model. Mechanical damage at the brushed sites in the mouse esophageal mucosa was clearly observed after brushing using anatomical microscopy (Fig. 5A-C). We performed RNA-seq using SSE from the esophagus of mice at day 7 after the brushing procedure. Transcriptomic analysis showed a significant upregulation of transcriptions of the acute inflammatory stress response genes, such as S100a8 and S100a9 compared to controls (Fig. 5D). Moreover, KEGG and GSEA analyses revealed significant upregulation of inflammation-related pathways in brush-injured tissues compared to controls, confirming that brushing could cause mechanical damage and effectively induce acute inflammatory responses in the mouse esophagus (Fig. 5E and F). Based on bRNA-seq data obtained before and after brushing, three significantly upregulated inflammation-related genes (S100a8, S100a9, and Mmp13) were selected for IHC validation. All three proteins showed robust positive staining in brushed tissues compared with unbrushed controls (Fig. S5A). Histopathologic analysis of H&E staining showed epithelial hyperplasia at the brushed sites on both day 3 and day 7 compared with unbrushed controls. (Fig. 5G). We then performed immunofluorescence analysis of Ki67 at the unbrushed or the brushed sites of esophageal SSE. As shown in Fig. 5H and I, when compared to the unbrushed sites of esophageal SSE, the brushed sites of esophageal SSE revealed that cell proliferation was markedly increased at the sites during the same time points. Previously, we and others showed that an induction of the acute inflammation in esophageal SSE in vivo or in organoids in vitro could disrupt the maintenance of esophageal SSE tissue homeostasis by significantly dysregulated p63⁺Sox2⁺, p63⁻Sox2⁺, and p63⁻Sox2⁻ cell populations [25,40].Consistently, the co-immunostaining of p63 and Sox2 displayed a significant change in these cell populations (Fig. S5B and C). Taken together, these results demonstrated that mechanical damage to a mouse esophagus by controllable brushing at a specific site could locally induce acute tissue inflammatory responses and stimulate cell proliferation.

Fig. 5.

Molecular and cellular responses to esophageal brushing and its application in the 4NQO/SOR-induced mouse model. (A) Detailed images of the nylon brush used for esophageal brushing. (B) Brushing procedure under the guidance of Digital Subtraction Angiography (DSA). (C) Representative views of esophageal injury following brushing. Scale bars, 500 μm (left); 100 μm (right). (D-F) Transcriptomic profiling comparing SSE with brushing to untreated controls. (D) Volcano plot illustrating upregulation of inflammatory related genes, (E) KEGG pathway enrichment analysis, blue background indicating inflammatory pathways, and (F) GSEA showing upregulation of inflammatory related pathways. (G) H&E staining showing epithelium thickening after brushing. Scale bars, 100 μm (left); 50 μm (right). (H) Representative immunofluorescence images of Ki67 in untreated SSE and brushing SSE collected on day 3 and day 7 post-brushing, Scale bar, 20 μm. (I) Percentage of Ki67⁺ cells in untreated and brushed SSE at days 3 and 7 post-brushing, NC 22.21 ± 0.509, DAY3 22.21 ± 0.509, DAY7 22.21 ± 0.509, n = 3. (J)Study design: Application of the brushing procedure to the 4NQO/SOR-induced CIMCM of ESCC. Mice were treated as follows: C-B: brushing 4 times at weeks 9, 11, 13, and 15. N5-B: administered 4NQO from weeks 1 to 5, and brushing, N5-B-S: administered 4NQO from weeks 1 to 5, injected with SOR from weeks 8 to 16, and brushing. (L) Representative macroscopic view of the whole esophagi from C-B, N5-B, N5-S-B, Scale bar, 1 cm. (M) Tumor incidence and localized tumor formation rates in C-B, N5-B, N5-S-B.

We then applied the brushing procedure to the 4NQO/SOR mouse CIMCM of ESCC. The experimental design was detailed in Fig. 5J, with mice randomly divided into three groups (C-B, N5-B, N5-S-B). Mouse esophageal samples from each group were collected and analyzed at week 35 (Fig. 5K). While brushing alone did not induce gross esophageal lesions in mice (C-B), mice treated with 4NQO plus brushes or 4NQO/SOR plus brushes displayed a higher tumor formation rate when compared to mice treated with 4NQO or 4NQO/SOR alone. More importantly, tumors in mice of N5-B and N5-S-B were formed predominantly in the lower parts of the esophagus where they were brushed (Fig. 5L). These results demonstrated that a localized and controllable physical injury in the mouse esophagus together with the 4NQO/SOR mouse CIMCM of ESCC could enhance tissue inflammation and stimulate cell proliferation, thereby locally promoting ESCC formation at specific sites, ultimately achieving a controllable localization for tumor formation and making the model more convenient for prevention, early detection, potential therapeutic strategies and/or interventional therapies of ESCC.

Computational pathology (C-Path) by CIMCM of ESCC

As discussed in the introduction, the CIMCM provided valuable samples from different stages of tumor initiation and progression that would not be readily available from human patients. Although the early diagnosis and precision medicine of ESCC remained significant challenges, the emergence of C-Path provided new clinical insights and/or methods into these aspects of ESCC [41,42]. Utilizing whole slide images (WSIs) and deep learning strategies, C-Path rapidly, reliably, and accurately established computational models for cancer diagnosis at the pathologist-level [43]. C-Path had excellent performance in distinguishing tissue heterogeneity or clear tissue-level boundary features, which focused on differentiating and annotating distinct and highly heterogeneous pathological images. Moreover, significant progress was made in the segmentation and classification of instances at the cell nucleus level [44,45]. These nuclear segmentation and classification models identified and annotated the nuclei of different cell types, enabling feature extraction for computer-aided diagnosis. However, segmentation and classification of the same cell type, such as epithelial cells in different stages/states remained elusive. As our mouse CIMCM was able to provide multistage images of esophageal carcinogenesis, we decided to develop an epithelium multi-stage segmentation network (EMSS-Net). We trained it with H&E-stained WSIs derived from the mouse CIMCM of ESCC to enable the pathological detection of SSE and provide pixel-level segmentation and annotation of the ESCC initiation and progression stages (NOR, HYP, DYS, and CAR).

To establish the EMSS-Net model for multistage ESCC detection, we constructed a WSI dataset using SSE samples from the present and our previous studies [4]. The dataset included 71 H&E-stained WSIs from mouse esophageal SSE and 26 WSIs from rat esophageal SSE. Combining mouse and rat WSIs increased the sample size and diversity. All WSIs were scanned on a digital pathology microscope at 20× magnification, and board-certified pathologists (Liyan Xue and Huiqin Guo) reviewed these WSIs in collaboration with the first author (Yuxia Fu). Yuxia Fu, Liyan Xue and Huiqin Guo confirmed that the histopathologic morphologies in mouse and rat esophageal SSE on H&E-stained slides were nearly identical. Subsequently, Liyan Xue and Huiqin Guo annotated the NOR, HYP, DYS, and CAR regions within the WSIs at multiple magnifications (4×, 10×, and 20×), and these annotations were used as ground truth (GT) labels for training and testing the EMSS-Net. Notably, in some cases, observations of serial sections stained by p63 immunohistochemistry (IHC) were required. P63 was a critical marker of SSE, widely expressed in basal cells and highly expressed in malignant cells of SCC (Fig. S6A). In addition, we implemented a multi-scale enhancement strategy, in which we retained the original WSIs and simultaneously resized the original WSIs to widths of either 8192 pixels (downscaled) or 25600 pixels (upscaled) but maintained the original aspect ratio by proportionally resizing the height (Fig. S6B). Bilinear interpolation was used to fill in additional pixels during the upscaling process. In this way, the original WSIs and resized WSIs were cropped into patches to provide tissue and cell-level features. Taken together, these WSI preprocessing steps resulted in a total of 3377 patches in the training set for EMSS-Net (Fig. S6C and D).

The EMSS-Net employed a Mix Transformer as an encoder to extract tissue and cell features of SSE (Fig. S6E). The decoder was designed with Step 1 to distinguish epithelial from non-epithelial regions and Step 2 to identify multi-stage intraepithelial regions in Step1. In Step 1, the extracted feature maps were fed into a shallow fusion module, which integrated epithelial features from shallow layers to preserve detailed spatial characteristics, such as cell shape and tissue texture. In Step 2, multi-scale features, such as cell texture, tissue structure, and histopathological category-specific semantics, were processed by different multi-scale fusion modules. Scale alignment modules aligned features from varying scales to directly fuse these characteristics, ensuring consistent integration across scales. Further details of the EMSS-Net architecture were described in the Materials and Methods section. The output of EMSS-Net consisted of WSIs with predicted annotations of epithelial regions and multi-stage of intraepithelial regions (NOR, HYP, DYS, and CAR).

As shown in Fig. 6A, for the segmentation of epithelial and non-epithelial regions (Fig. S6E; Step 1), the EMSS-Net model could accurately delineate the epithelial regions while excluding the non-epithelial areas. We evaluated the model performance using the class pixel accuracy (Acc) metric, which was widely applied in multi-class segmentation tasks to measure and quantify the proportions of correct segmentation areas with the GT images [46]. Moreover, the Dice coefficient (Dice) was calculated as a supplementary metric to assess the overlap scores between predicted segmentation areas and the GT images [47]. The results showed that the Acc of Step 1 reached 99.69 % (Fig. 6C) and the Dice of Step 1 was 93.04 % (Table S4), indicating that the trained EMSS-Net model could accurately identify esophageal SSE.

Fig. 6.

The outcomes of segmentation and explainability analyses of EMSS-Net. (A) Representative whole-slide segmentation outcome for epithelial and non-epithelial regions, with magnified views of selected regions of interest. (B) Representative whole-slide segmentation results for intraepithelial lesion subtypes, including magnified views of regions of interest. (C) Summary of EMSS-Net Acc for the segmentation of epithelial and non-epithelial regions (Step 1) and multi-stage segmentation of intraepithelial regions (Step 2). (D) Visualization of the explainability algorithm used to evaluate EMSS-Net, which identifies localized attention regions in the multi-segmentation of intraepithelial regions. The attention weights are color-coded, increasing from blue to red, with red indicating a higher contribution.

When applied to multi-stage segmentation of intraepithelial regions (NOR, HYP, DYS, and CAR) in esophageal SSE in Step 2 (Fig. S6E), the trained EMSS-Net model could also effectively predict the overall areas of these regions (Fig. 6B). The Acc for each stage was 76.70 % for NOR, 69.65 % for HYP, 93.19 % for DYS and 96.11 % for CAR, respectively (Fig. 6C). We also assessed the overall performance of the trained EMSS-Net model in Step 2 using the mean pixel Acc (MPA) and the mean Dice (mDice). The MPA was 83.89 % and the mDice was 51.43 % (Table S4). There were several possibilities that could lead to the decreased accuracy of MPA or mDice in the multi-segmentation of intraepithelial regions (NOR, HYP, DYS, and CAR) in esophageal SSE. One would be that in comparison to the binary segmentation tasks, the multi-segmentation tasks were typically more complex, resulting in higher chances of mis-segmentations and lower accuracies. Nevertheless, these results were still well in line with widely accepted accuracy standards of multiple segmentation tasks published recently [48]. Another, more important one, would be that ESCC carcinogenesis was a spatiotemporal process, which showed no clear boundaries between NOR, HYP, DYS, and CAR regions. For instance, although the segmentation performance was lower in the NOR and HYP regions, it was mainly due to the subtle differences in tissue characteristics between these stages, which made boundary identification more difficult.

Hence, we employed explainable artificial intelligence (XAI), a method that could provide rational explanations of a trained model’s decisions [49]. We generated class activation maps (CAMs) to assess how the multiple intra-epithelial regions of NOR, HYP, DYS, and CAR were focused and segmented by the EMSS-Net model. The heatmaps were generated from the four-stage discriminator (Fig. S6E), which highlighted the key features that drove the model decision-making process by the EMSS-Net model in Step 2. The CAMs showed that the EMSS-Net accurately focused on epithelial regions with dense nuclei and epithelial layer structures for the four-stage determination (Fig. 6D). We calculated the overlap ratio between regions with >60 % of the weight in the CAMs and the corresponding GT regions for the representative patches. The overlap ratio also varied widely. However, the majority was >70-90 % (Table S5). Consistent with our speculations, the EMSS-Net model predicted the major regions for each stage but only showed inaccuracies in segmenting the boundaries and transitional zones of NOR, HYP, DYS, and CAR (Fig. 6D). Taken together, these results indicated that by applying the valuable WSIs from our CIMCM samples, the EMSS-Net model could accurately segment the esophageal epithelial and non-epithelial regions and predict four stages of intra-epithelial regions NOR, HYP, DYS, and CAR, thus shedding light on the potential for ESCC early detection by C-Path.

Discussion

In this study, we present the establishment of a novel and effective CIMCM of ESCC, employing 4NQO as the carcinogen and SOR as the tumor promoter for preclinical and C-Path applications. Our findings are consistent with those of previous studies, demonstrating that SOR can function with different types of carcinogens, such as the nitrosamine-type and water-soluble 4NQO carcinogens, to promote ESCC carcinogenesis in rats and mice [4,15]. The cellular and molecular analyses in vivo and in vitro, as well as the phosphoproteomic assay, further elucidate that, although SOR is an MKI with therapeutic efficacy against various cancers via the inhibition of RTKs and/or Raf kinases, it can promote B-Raf/Raf-1 heterodimerization, resulting in specific activation of the Raf-MEK-ERK pathway in normal esophageal SSE cells. These results align with the clinical epidemiology data indicating that cancer patients treated with SOR often paradoxically develop SCCs and with the laboratory experimental data showing that SOR can specifically activate the Raf downstream MEK-ERK signaling pathway in normal cells [21,24]. Taken together, the findings collectively indicate that SOR, functioning as a tumor promoter, activates the Raf-MEK-ERK pathway in SSE, thereby promoting the CIMCM of ESCC carcinogenesis. The potential of SOR to act as a general tumor promoter in the establishment of CIMCM in other stratified epithelial tissues remains to be determined. However, based on the action of SOR in normal cells, we hypothesize that SOR could work in conjunction with other tissue-specific carcinogens, such as N‑butyl‑N(4-hydroxybutyl)-nitrosamine (BBN), to effectively establish rat and/or mouse CIMCM of urothelial carcinoma. We are currently in the process of testing this hypothesis.

The present 4NQO/SOR mouse CIMCM of ESCC offers distinct advantages over the rat or mouse CIMCM of ESCC previously reported by us and others [4,50]. Firstly, in contrast to the relatively underdeveloped state of rat genetic/epigenetic analyses and manipulations, mouse genetic/epigenetic analyses and manipulations are well developed and can be widely applied to the multiomic studies in the mouse CIMCM samples. A substantial body of evidence demonstrates that multiomic analyses in normal mice or GEMM treated with a tissue-specific carcinogen(s) can help elucidate the molecular mechanism(s) by which a specific gene/pathway is involved in carcinogenesis in the tissue [51,52]. Thus, the 4NQO/SOR mouse CIMCM of ESCC holds particular promises for elucidating the potential molecular mechanisms and facilitating preclinical research in the field of ESCC carcinogenesis. Secondly, mice treated with 4NQO alone over a 16-week period exhibited signs of poor health and/or mortality, whereas mice in the 4NQO/SOR CIMCM of ESCC demonstrated robust health, comparable to that of control mice. Of particular significance is the reliability, efficiency, and effectiveness of the 4NQO/SOR CIMCM of ESCC. The protocol of the 4NQO/SOR CIMCM of ESCC, involving a 5-week course of 4NQO treatment followed by an 8-week course of SOR treatment, results in a substantial reduction of TMB levels in mice, reaching levels comparable to those observed in human ESCC (20-100 mutations per sample) [53]. In contrast, mice treated with 16 weeks of 4NQO alone exhibit a mutation rate of 3,000-5,000 mutations per sample [13,50]. Consistent with this, a recent study also demonstrated that the number of mutated genes in ESCC cell lines derived from mice exposed to 4NQO for 16 weeks contained TMB levels in excess of 1000 [54]. However, notwithstanding the marked reductions of TMB, multiomic analyses demonstrate that the 4NQO/SOR mouse CIMCM of ESCC preserves key genetic driver mutations and abnormal transcriptomic/protein expression profiles in ESCC samples [55]. The 4NQO/SOR mouse CIMCM of ESCC also highlights that the tissue microenvironment plays an important role in ESCC carcinogenesis, consistent with numerous recent publications [5,56]. Taken together, the 4NQO/SOR mouse CIMCM of ESCC provides a robust platform for the scientific community to investigate the underlying molecular mechanisms of multistep ESCC carcinogenesis.

We employ the 4NQO/SOR mouse CIMCM of ESCC for preclinical investigation with the objective of further disrupting esophageal epithelial homeostasis using a site-specific injury-induced inflammation strategy. The hypothesis is that the site-specific injury-induced inflammation can accelerate early tumorigenesis through an imbalance in proliferation, differentiation, and sustained pro-tumorigenic inflammation. Previously, Yao et al. utilized scRNA-seq to demonstrate that the dynamic remodeling of immune responses, exemplified by S100A8, was important during ESCC progression [5]. Zhou et al. also revealed through multiomic analyses that tumor microenvironment (TME) interactions critically regulated the initiation and progression of the disease [56]. Consistent with these findings and our hypothesis, a mechanical injury to the esophageal SSE of the 4NQO/SOR mouse CIMCM of ESCC results in the inflammatory-related response, site-specific tumor formation and high tumor incidence. Hence, this study underscores the pivotal roles of the TME and the inflammation-related carcinogenesis mechanisms in ESCC. Moreover, it signifies that the 4NQO/SOR mouse CIMCM of ESCC with site-specific injury to induce localized inflammation and tumors can serve as a robust platform for ESCC early detection, carcinogenesis mechanisms, and potential interventional therapies.

We also apply the tissue samples of the 4NQO/SOR mouse CIMCM of ESCC for C-Path, which represents a rapidly evolving field that facilitates the automated analysis of digital histopathology images. This is achieved through the utilization of WSIs and deep learning approaches, which are employed to establish effective models for expeditious, dependable, and precise cancer diagnosis, staging, and grading [57]. The excellent performance of C-Path stems from its ability to recognize heterogeneity among different tissues and discern the distinct edge features present at the tissue level. Significant advancements have been also made in nuclear segmentation and classification, enabling models to identify and annotate the nuclei of various cell types, such as epithelial cells, mesenchymal cells/fibroblasts, and immune cells [45]. These advancements facilitate the extraction of features for computer-aided diagnosis. However, despite the efficacy of nuclear segmentation in distinguishing highly heterogeneous cell types, it remains challenging to segment and classify the same cell type, such as epithelial cells in different stages or states. We have designed an EMSS-Net and trained the EMSS-Net with WSIs of NOR, HYP, DYS and CAR from the 4NQO/SOR mouse CIMCM of ESCC. We demonstrate that the EMSS-Net can achieve promising segmentation performance to detect the initiation and progression of ESCC (NOR → HYP → DYS → CAR). The interpretability analyses reveal that the model predominantly focuses on tissue regions characterized by well-defined cell and tissue structures. However, due to the inherent interpretability limitations of deep learning, it remains to be determined whether the EMSS-Net focuses on and utilizes the nuclear features of WSIs. Recent machine learning–based methods have demonstrated promising results in extracting nuclear features. Kumar et al. successfully distinguished stage II and IV colon tumors by identifying nuclear size (area, perimeter, and major-axis length), nuclear orientation, and local variance in nuclear contrast as the five most discriminative features [57]. Similarly, clinical diagnosis mainly relies on nuclear morphology, highlighting the necessity of incorporating these critical features. While p63 IHC has been employed to guide manual annotation of intraepithelial regions, thereby enhancing model precision in this study, it is not included in the training of the EMSS-Net. Recently, Shamai et al. employed deep learning to predict PD-L1 expression in breast cancer directly from H&E-stained images, training the model with IHC results and H&E images. This approach provides a feasible and efficient decision-support tool in clinical practice [58]. Given that our current model has not incorporated p63 IHC results into its training strategy or algorithm, it is conceivable that the EMSS-Net performance could be significantly enhanced by integrating key nuclear features, IHC markers, and even super-resolution subcellular structures into future training datasets to improve segmentation accuracy [59].

In conclusion, we have developed a reliable, effective and efficient mouse CIMCM of ESCC, combining 4NQO as a carcinogen and SOR as a tumor promoter. This robust platform facilitates investigation of multistep ESCC carcinogenesis. The 4NQO/SOR mouse CIMCM of ESCC offers several advantages, including accelerated tumor initiation and progression, and clinically relevant multiomic profiles. By integrating preclinical manipulations and AI-enhanced histopathological analysis, this model enables systematic dissections of tumor initiation and progression, offering a versatile tool for detecting ESCC in early stages, unraveling tumorigenic mechanisms at defined molecular nodes and accelerating therapeutic innovation in ESCC.

CRediT authorship contribution statement

Yuxia Fu: Writing – original draft, Visualization, Validation, Investigation, Formal analysis, Conceptualization. Guoqing Zhang: Writing – original draft, Investigation. Yue Liu: Investigation. Lei Xu: Writing – original draft. Yuanyuan Hu: Investigation. Liyan Xue: Data curation. Huiqin Guo: Data curation. Yan Fu: Resources. Yigang Cen: Resources. Xiao Li: Resources. Wei Jiang: Writing – review & editing, Writing – original draft, Supervision, Funding acquisition, Conceptualization. Xiying Yu: Writing – review & editing, Writing – original draft, Supervision, Funding acquisition, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.