A total of 27 patients were considered in the analyses. Most were females, with a mean age at EGC diagnosis of 67 \(\pm\) 11.7 years. Tumors were mainly classified as intestinal type and did not involve lymph nodes. Among patients, 14 were diagnosed with Pen B (52%) and 13 (48%) with Pen A tumors, according to the Kodama’s classification; 9 (33%) and 7 (26%) patients had a recurrence or death as first event, whereas 11 (67%) patients were disease free/alive within 10 years of follow-up. No statistically significant differences were observed between Pen B and Pen A tumors (Table 1). At univariate analyses using Cox PH model, higher age and tumor location in the upper-middle tract were associated with an increased hazard of relapse or death from any cause (p = 0.006 and p = 0.032, respectively), as shown in the forest plot (Fig. 1). Moreover, patients with higher tumor grade seem to have about a double risk of relapsing or dying, even if the small case series does not permit statistical significance.

Forest plot. Univariate Cox regression analysis for disease-free survival. The following reference categories were used: “female” for gender; “Pen B” for Kodama classification; “antrum” for tumor location; “intestinal” for Lauren classification; “G1+G2” for grade; “N0” for lymph node status; “Absent” for lymphovascular invasion. Age, Tumor size, and lymph node ratio are reported as continuous variables. HR hazard ratio, CI confidence interval, p p-value

NGS analysis revealed a total of 904 variants. Among them, 200 variants (22.1%) were predicted as pathogenic/likely pathogenic, 372 (41.2%) as variants of uncertain significance (VUS), and 332 (36.7%) as benign/likely benign. Regarding the effect on the protein, 103 variants (11.4%) were frameshift deletions/insertions, 17 (1.9%) in-frame deletions/insertions, 27 (3.0%) nonsense mutations, 721 (79.8%) missense mutations, and 36 (4%) splice site mutations (Fig. 2a). Regarding mutation type, 73.4% were transitions and 26.6% were transversions (Fig. 2b). No significant differences between Pen A and Pen B tumors as well as relapsed and non-relapsed status were observed.

Mutational characterization of EGCs based on DNA sequencing. a For each patient (x-axis, ordered by TMB) relative frequency of variant types is shown for mutations in exonic regions of genes. b Relative frequency of SNV types. c MSI and TMB status  of EGC patients. Patients are grouped by Kodama classification and relapse status in all subfigures

TMB and MSI were significantly associated with 100% of MSI high (> 20%) patients showing also a high TMB (≥ 10), Fisher’s exact p = 0.002. Although no statistically significant difference was observed, Pen A presented higher median TMB values compared to Pen B (8 [IQ-IIIQ 4.8 – 25.3] vs 5.2 [IQ-IIIQ 2.4 – 11.1] (p = 0.120)), as well as a higher number of MSI subtype tumor, defined according to the selected cut off (p = 0.165) (Fig. 2c and Supplementary Table S2).

Genomic analysis revealed that the most frequently mutated genes were TP53, protein tyrosine phosphatase receptor type T (PTPRT), E3 ubiquitin ligase ring finger protein 43 (RNF43), F-box and WD repeat domain-containing7 (FBXW7), AT-rich interaction domain 1A (ARID1A) and spectrin alpha, erythrocytic 1 (SPTA1), as shown in Fig. 3. Twenty-two percent of patients presented erb-b2 receptor tyrosine kinase 2 (ERBB2) gene alterations, 5/27 (18%) being ERBB2 amplified. Similarly, cyclin E1 (CCN1E) was amplified in 15% of tumors. Of interest, mutations in ARID1A were found only in Pen A tumors (p = 0.006), 4/6 of them being frameshift substitutions classified as “likely pathogenic”, the remaining 2 classified as variants of uncertain significance (Table 2 and Fig. 4). A significant association between ARID1A alterations and high TMB (p = 0.027) and microsatellite instability (p = 0.056) has been observed. No other statistically significant molecular differences between Pen A and Pen B were observed.

The OncoPrint chart.  The heatmap represents genomic alterations including pathogenic variants and VUS, found in at least four patients. Patients are ordered by non-relapsed or relapsed status. The chart was built by ComplexHeatmap R package

ARID1A status in EGC patients. The histogram represents the percentage of ARID1A alterations in the two different subtypes of Pen tumors. wt wild-type, mut mutant

In patients with a high tumor mutational burden, other significantly mutated genes were APC (p = 0.029), and ARID2 (p = 0.029). A significant association was also observed between MSI and ANKRD11 (p = 0.013) and RNF43 mutations (p < 0.0001).

The broad molecular characterization of the EGC series did not reveal any other significant associations with clinico-pathological characteristics, except for TP53 mutations which were more frequently found in intestinal-type EGC and in males (p = 0.041 and p = 0.054) and for the association between the low-density lipoprotein receptor-related protein 1b (LRP1B) mutations and higher age (p = 0.020). Despite not reaching statistical significance, tumors harboring LRP1B alterations seem to have a higher hazard of relapse or death from any cause (HR = 2.73 95% CI: 0.857–8.703 p = 0.089), being mutated mainly in relapsed patients (p = 0.093).

The enrichment analysis revealed no statistically significant associations after correction for multiple testing between a pathway’s PI score and Kodama classification at the detailed level with 218 GC-related pathways. The results of the statistical comparison of individual pathway PI scores across different subgroups of patients (TMB and MSI class, Kodama classification, relapse status) are presented in Supplementary Table S3 (available in our Zenodo repository). Visual representation of the PI score distribution for all 218 pathways is presented in Fig. 5a. Individual pathways can be grouped according to their corresponding Top Level Pathway in the Reactome database hierarchy, representing overarching biological processes critical for the functioning of organisms. Such grouping allows the identification of general patterns in how increased pathway instability potentially disrupts these biological functions for different subgroups of patients. Upon grouping the pathways according to their Top Level Pathways, it becomes apparent that higher PI score in pathways belonging to “Disease” (p < 0.001), “Immune System” (p < 0.001) and “Signal Transduction” (p < 0.001) are associated with Pen A classification (Fig. 5b). Association between overall increased PI score and Pen A classification remained significant when considering all pathways together.

Pathway instability analysis. a Heatmap representation of the 218 GC pathways identified with the Reactome enrichment analysis. Pathways are sorted in columns and grouped according to their corresponding Top Level pathways, with hierarchical clustering applied within the groupings. TMB and MSI class, as well as relapse status and Kodama classification are annotated on the left side for all patients. b, c PI score distribution across  b Kodama classification subtypes and c relapse status, shown for pathways grouped according to selected top level pathways. Each data point indicates a PI score value for an individual GC-related pathway, for each patient separately. Violin plots with horizontal black lines marking the median values convey the overall distribution of PI scores between different patient groupings. Note that since no information on Kodama classification was available for EGC-TCGA patients, these patients are excluded from comparisons between Pen A and Pen B groups. d, e Two-dimensional t-SNE representation of PI scores obtained from 218 pathways of interest, for the combined case series of EGC and TCGA EGC. Clustering overlaid with patients’ d Kodama classification and e TMB classification

To explore in more depth the molecular landscape of early gastric cancer using a bigger case series, the publicly available dataset of Stomach Adenocarcinomas (STAD) from TCGA––PanCancer Atlas was retrieved using the cBioPortal (https://www.cbioportal.org/) [28] and was added to the analysis. To match EGC, only stage I gastric cancer patients with survival data available were included (n = 54). The clinico-pathological characteristics of the two cohorts were largely comparable. Notably, the EGC cohort of the present study had a higher percentage of female patients, tumors predominantly located in the lower tract and a greater proportion of pN+ tumors, compared to the TCGA cohort (Supplementary Table S4). Moreover, for the patients belonging to the TCGA cohort, the Kodama classification of the tumor was not available.

PI scores were calculated for the GC-related pathways for a combined cohort of EGC and EGC–TCGA patients (n = 81), but no statistically significant association between a pathway’s PI score and the relapse status was observed for any single pathway after correction for multiple testing. However, top level pathways revealed that relapsed patients had significantly higher PI scores in pathways belonging to “DNA Repair” (p < 0.001) and “Cellular responses to stimuli” (p < 0.001), whereas they had significantly lowered PI scores for “Immune System” (p = 0.03) and “Signal Transduction” (p = 0.024) (Fig. 5c). Interestingly, whereas disruptions to the “Immune System” and “Signal Transduction” Top Level pathways are significantly enriched in Pen A patients, the exact opposite is true for relapsed patients. However, it is important to note that relapse status was known for all EGC-TCGA patients, whereas Kodama classification was not available, which is why EGC-TCGA patients are absent from comparisons involving Pen status. Furthermore, as we do not observe the expected association between Kodama classification and relapse status, this should not come off as a surprise.

Moreover, the analysis of the combined cohort permits the identification of four main clusters of patients (Fig. 5d, e). Clusters 1–3 are characterized by a majority of TMB low patients, whereas the smallest Cluster 4 is made up of exclusively TMB low patients (p = 0.16). Clusters 3 and 4 contain only three patients each with available information on Kodama status, and both feature the same split (2 Pen B vs 1 Pen A). Cluster 1 is characterized by a slight majority (6/11) of Pen B patients, whereas the reverse is true for Cluster 2 with 6/10 patients being Pen A (p = 0.44). There were no statistically significant differences among the clusters in the distribution of the main clinico-pathological parameters or disease-free survival (p = 0.607) (Supplementary Figure S1).

A closer look showed that the clusters were characterized by specific signatures of disrupted pathways. For example, Cluster 4 is differentiated by disruptions in the pathways belonging mainly to signal transduction by tyrosine kinases. Clusters 2 and 3 share some significantly disrupted pathways such as those involved in transcription, intracellular signaling, metabolism of proteins, DNA repair and programmed cell death. The remaining Cluster 1 had no specific significantly disrupted pathways (Supplementary Table S5- available at https://doi.org/10.5281/zenodo.10816958).

Looking at which genes were most frequently mutated for each cluster provides further credence that PI score-based clustering stratifies patients into groups with distinct molecular characteristics. All patients belonging to Cluster 1 have mutations in TP53 (100%). After TP53, most often mutated genes are SPTA1 (22.8%), LRP1B (18.2%) and ARID1A (18.2%), roughly following the overall trend of mutated genes for the cohort. The most frequently mutated gene in Cluster 2 is ARID1A (44.44%), which is the 2nd most commonly mutated gene across all clusters with an overall mutation frequency of about 30%. Significant associations for patients with mutated ARID1A and TMB and MSI status were observed in the combined case series, but no significant association has been found with the relapse status (p = 0.7721). Interestingly, no patients from Cluster 2 have mutated TP53. As for Cluster 1, Cluster 3 is TP53 mutated in all 18 patients (100%), followed by LRP1B (33.33%) and EPHA5 (33.33%). ARID1A (22.22%) and SPTA1 (11.11%), otherwise among the most commonly mutated genes, are notably less mutated in Cluster 3. Lastly, Cluster 4 features a unique mutation signature with the most commonly mutated gene being RHOA (100%), followed by TP53 (20%) and RNF43 (20%) (Supplementary Table S6).