The scRNA-seq raw data, comprising 27 PDAC tumor samples (accession ID: GSE205013), were sourced from the publicly available National Center of Biotechnology Information (NCBI) Gene Expression Omnibus (GEO), a public database storing extensive collections of high-throughput functional genomics and expression datasets [7]. The dataset was acquired through 10X Genomics and Illumina NextSeq, HiSeq4000 platform was used to sequence the samples.

To analyze the raw sequencing FASTQ files, the solo submodule of the STAR splice-aware aligner tool was employed. Key parameters included "--soloType CB_UMI_Simple" for defining the structure of cellular barcodes and unique molecular identifiers (UMIs), along with "--soloCBwhitelist", using the 10X Genomics V2 whitelist for barcode validation. Additional settings comprised "--soloBarcodeReadLength 0" to automatically adjust barcode read length, "--soloUMIdedup 1MM_All" to allow deduplication of UMIs within a 1-mismatch threshold, and "--soloCBmatchWLtype" to ensure strict matching of barcodes to the whitelist. UMI filtering was conducted with "--soloUMIfiltering MultiGeneUMI", and cell barcode filtering followed the CellRanger 2.2 approach via "--soloCellFilter CellRanger2.2". The reads were decompressed using "--readFilesCommand zcat", and the maximum output of spliced junctions was set with "--limitOutSJcollapsed 2000000" [8]. The reference genome index for GRCh38 Homo sapiens was generated using the STAR "star-build" command. Post-alignment, UMIs associated with individual cells were quantified, and valid cell barcodes were identified, resulting in a sparse cell-by-gene count matrix. Three output files were generated for each sample in Matrix Market (MTX) format, including the count matrix ("matrix.mtx"), cell barcodes ("barcodes.tsv"), and gene identifiers ("features.tsv").

Additionally, Seurat (version 4.4.0) and scType R packages were used for the quality control and downstream analysis [9]. Low-quality cells were removed using three parameters, including the number of UMIs identified per cell barcode (i.e., library size), the number of genes called per barcode (nFeature_RNA > 200 & nFeature_RNA < 6000) for the filtration of cells expressing too few or too many genes, as fewer than 200 gene expressing cells may represent dead or damaged cells, while cells expressing more than 6,000 genes may represent doublets or multiplets of two or more cells, and the proportion of UMIs mapping to mitochondrial genes (< 10%).

Additionally, the high mitochondrial content was filtered out using the condition "percent.mt < 10", as it can indicate damaged or low-quality cells. Lastly, the low RNA content cells were filtered out using the condition "nCount_RNA > 1000", which can indicate that low RNA content might have undergone apoptosis or may have low-quality data [10]. Additionally, to exclude outlier distributions (no potential doublets/multiples and cells with higher mitochondrial reads > 10%), the visual analysis of the violin plots for each metric was performed.

The raw counts were normalized using the “LogNormalize” method, which involves taking the natural logarithm of the raw numbers and adding a pseudocount of 1. This is done by dividing the result by the total number of counts for that cell. Scaling the expression numbers and adding a pseudo count to account for technical variability improves the comparability of gene expression data across cells.

The highly variable characteristics were subsequently determined using the “findvariablefeatures” function from the Seurat package. Specifically, 2,000 genes expressing substantial very high to very low deviation in expression across different cells were identified. Dimensionality reduction was performed using t-distributed stochastic neighbor embedding (t-SNE) to compute the first 20 principal components (PCs) for clustering and visualization [11]. Additionally, cell clustering was performed using the standard Louvain clustering technique [9].

Cell annotation based on marker genes from the “immune system” was performed using an automated supervised clustering pipeline, utilizing an experimentally validated cell marker database, through the scType R package [12]. The differential expression analysis was subsequently performed to determine significant marker genes (p-value < 0.05 and logFC > 0.25 or logFC < –0.25) through the Wilcox method for each cell type as well as effector CD8+ T-cells. Additionally, module score was calculated based on the exhaustive markers given below by utilizing the AddModuleScore within Seurat package, whereas grouping was done based on each cell type for statistical comparison using the ggpubr package (https://cloud.r-project.org/web/packages/ggpubr/index.html).

The functional enrichment analysis of the effector CD8+ T-cells dysregulated genes was performed using the online tool GeneCodis4. GeneCodis 4 (https://genecodis.genyo.es/) is a web-based tool that integrates information from various biological databases and tools to perform functional enrichment analysis based on different annotation sources, such as Gene Ontology (GO) terms and KEGG pathways. It extracts sets of significant concurrent annotations and assigns a statistical score to evaluate those that are significantly enriched in the input list [13]. Moreover, these pathways were further evaluated through literature review whether they are reported for poor survival by their dysregulation.

An extensive literature review used Google Scholar and PubMed databases to shortlist the exhaustive immune marker genes in different T-cell types involved in various pro-cancer pathways leading toward an exhaustive immune system. Among all the genes, only genes reported in the literature to cause an exhaustive immune system in effector CD8+ T-cells were shortlisted for further analysis. To extract information about these genes, several search terms were used, such as “T-cells”, “immune cells”, “T-cell exhaustion Genes”, and “CD8+ T-cell exhaustion”.

GEPIA2 web server (http://gepia2.cancer-pku.cn/), a highly cited and valuable resource for gene expression analysis from TCGA and GTEx databases, was used to validate the gene expression levels of the shortlisted genes identified to be dysregulated in effector CD8+ T-cells PDAC tumor samples [14]. The expression levels of the shortlisted exhaustive genes were validated through GEPIA2 based on their expression levels in normal and tumor samples in pancreatic adenocarcinoma (PAAD).

The Genomic Data Commons (GDC) Data Portal (https://portal.gdc.cancer.gov/) comprising data from TCGA (https://www.cancer.gov/ccg/research/genome-sequencing/tcga) and Therapeutically Applicable Research to Generate Effective Treatments (TARGET) program (https://www.cancer.gov/ccg/research/genome-sequencing/target) was used to identify the clinically reported deleterious and possibly damaging mutations in the shortlisted exhaustive genes [15]; on the other hand, the AlphaMissense database (https://alphamissense.hegelab.org/) which predicts the pathogenicity of the missense variants [16] was utilized to evaluate the pathogenicity of clinically identified mutations from the GDC. The genes that were likely pathogenic among both databases were selected for further analysis.

The 3D structure of the selected protein was retrieved using the AlphaFold Protein Structure Database (https://alphafold.ebi.ac.uk/), which predicts the 3D structure by using the amino acid sequence of the protein [17]. Moreover, its protein name and sequence length were also observed from the UniProt database (https://www.uniprot.org/). Additionally, Maestro 12.0 (version 12.0.012; Schrödinger, LLC, New York, NY) was utilized to generate the mutant model of the selected protein from the wild-type protein structure.

Furthermore, the InterPro database (https://www.ebi.ac.uk/interpro/), which predicts domains and provides functional analysis of the proteins by classifying them into families [18], was utilized to retrieve the domain information of the selected protein.

Lastly, the visualization tool PyMOL (The PyMOL Molecular Graphics System, Version 3.0 Schrödinger, LLC) was utilized to superimpose the mutant model of the selected protein on the wild-type model, resulting in root mean square deviation (RMSD) value and structural differences between both models; additionally, domain regions were also colored.

Anti-inflammatory drugs that have the potential to inhibit the inflammatory reaction proteins were identified through a literature search using search terms such as “anti-inflammatory drugs” and “drugs against inflammatory proteins”. Furthermore, the PubChem database (https://pubchem.ncbi.nlm.nih.gov/), comprising chemical molecules and their activities [19], was utilized to retrieve the Simplified Molecular Input Line Entry System (SMILES) of the anti-inflammatory drugs that were converted to PDB-format through a custom Python script using RDKit (RDKit: Open-Source Cheminformatics Software, https://www.rdkit.org).

The virtual screening of the retrieved anti-inflammatory drugs with the wild-type and the mutant model of the IL7R was performed using the molecular docking tool GNINA utilizing a group of conventional neural networks (CNNs) as a scoring function [20]. The top three best binding affinity-scored complexes were visualized on Maestro 12.0 (version 12.0.012; Schrödinger, LLC, New York, NY) to identify the protein-ligand interactions.

The 2D interaction analysis of the top three highest-affinity complexes was conducted using Discovery Studio Visualizer (BIOVIA, San Diego, CA, USA). PDB-format complexes were imported, and 2D illustrations of interactions between protein residues and the ligand were generated.

Sequencing data processing included mapping and quantifying reads with UMIs, yielding a dataset of 156,559 cells and 36,601 RNA features. After normalization, 2,000 highly variable genes were identified to detect elevated expression within tumor cells, including JCHAIN, COL3A1, COL1A2, COL1A1, CCL18, IGKC, IGHG1, APOC1, SPP1, and OLFM4, shown in Figure 1A.

Using the scType R package and immune system marker database, cells were annotated and clustered, identifying cell types such as basophils, beta cells, cancer stem cells, effector CD8+ T-cells, endothelial cells, hematopoietic cells, hepatocytes, hematopoietic stem cells/multipotent progenitor cells (HSC/MPP) cells, liver progenitor cells, memory B cells, naive CD4+ T-cells, NK cells, non-classical monocytes, plasmacytoid dendritic cells, and platelets, visualized with t-SNE in Figure 1B. However, effector CD8+ T-cells were chosen for further analysis due to their central role in combating cancer cells. Additionally, these cells play a crucial part in exhaustion mechanisms during cancer progression, contributing to immune system exhaustion. Differential gene expression analysis on effector CD8+ T-cells identified 3,888 overexpressed (logFC > 0.25) and 6,894 underexpressed (logFC < –0.25), a volcano depicting these differentially expressed genes is shown in Figure 1C. The top 25 markers for CD8+ T-cells are given in Table 1.

An extensive literature review identified a total of 17 genes (PRF1, GZMA, IFNG, CD160, TBX21, ZEB2, NR4A2, SLAMF7, CD8A, CD3D, NKG7, CCR7, IL7R, ID2, IL2RB, IL2RG, and TGFB1) primarily involved in the exhaustion of different T-cells, including effector CD8+ T-cells. These identified genes were queried against our data using the ‘grep’ Linux command, resulting in 7 common genes (PRF1, GZMA, CD8A, CD3D, NKG7, IL7R, and IL2RG). Subsequently, these common exhaustive genes were selected for further validation. The dot plot of the shortlisted genes is depicted in Figure 1D which shows overexpression of these markers in effector CD8+ T-cells compared to other cell types. To evaluate the state of CD8+ T cells, we analyzed the expression of known exhaustion markers, including TOX, PDCD1, LAG3, TIGIT, EOMES, and RBPJ. By calculating the average expression of these genes across the entire cell type module, we found that CD8+ T cells exhibited a higher average module score than all other cell types for these markers. This suggests that CD8+ T cells have elevated expression levels of exhaustion-related genes, indicating a more pronounced state of exhaustion compared to other cell types; module score is shown in Figure 1E.

Pathway enrichment analysis was conducted to identify dysregulated pathways associated with the upregulated and downregulated markers of effector CD8+ T-cells. Notably, the upregulated genes were found to be enriched in pathways such as T-cell receptor signaling pathway, Th1 and Th2 cell differentiation, Th17 cell differentiation, and PD-L1 expression and PD-1 checkpoint pathway (Figure 2A). Conversely, pathways such as oxidative phosphorylation and chemical carcinogenesis, reactive oxygen species were found to be enriched for the downregulated genes (Figure 2B). The analysis revealed that the altered expression levels of these marker genes were significantly enriched in pathways associated with poor survival in PDAC patients, as reported in the literature [21–23].

In the analysis of TCGA datasets, these markers demonstrated significantly elevated expression levels in PAAD, with the IL2RG gene exhibiting the highest expression at 61.42 TPM, while the PRF1 gene had the lowest at 2.82 TPM. The CD3D gene showed an expression level of 9.98 TPM, and the GZMA, NKG7, and IL7R genes had expression levels of 7.96, 7.93, and 7.61 TPM, respectively, with the CD8A gene expressing at 4.05 TPM in tumor samples. The exhaustive markers and their respective expression levels in PAAD compared to normal are listed in Table 2.

To integrate our scRNA-seq findings and overexpression of these exhaustive markers with whole genome sequencing data, we analyzed the GDC TCGA PAAD dataset. The analysis concentrated on identifying mutations classified as benign, deleterious, and possibly damaging missense mutations. This investigation revealed that only two genes, PRF1 and IL7R, exhibited such mutations. However, PRF1 showed only one mutation (benign) at the 546 positions where proline was mutated to leucine (P546L), while IL7R showed two mutations (probably damaging and possibly damaging) at 110 and 405 positions where lysine was mutated to asparagine (K110N) and leucine was mutated to methionine (L405M), respectively. Moreover, the percent of cases that are affected by the mutations in IL7R and PRF1 genes are shown in Figure 2C.

Moreover, PRF1 and IL7R genes were further searched on the AlphaMissense database to predict the pathogenicity of PRF1 and IL7R mutations identified in the previous step. The AlphaMissense showed a total of 589 likely pathogenic mutations in the PRF1 gene, but none of those 589 mutations coincided with the mutation (P546L) that we identified in the TCGA dataset. Furthermore, the IL7R gene was observed to have a total of 488 likely pathogenic predicted mutations, whereas only one mutation coincided with our findings, where lysine was mutated to asparagine at position 110. Therefore, only one mutation of the IL7R gene (K110N) was used for further analysis.

The 3D structure of the wild-type IL7R with a sequence length of 459 amino acids was retrieved from the AlphaFold database by utilizing its accession ID (P16871). Furthermore, the InterPro database predicted two domains, IL-7Ralpha, fibronectin type III (FN3) domain in regions 32–127 and 131–231, respectively. The IL7R protein details from the UniProt and InterPro databases are mentioned in Table S1.

Nonetheless, the wild-type IL7R protein structure was utilized to create its mutant model, K110N. Furthermore, both the wild-type and the mutant (K110N) models were superimposed, revealing an RMSD value of 0.218 Å, indicating a slight structural change in the mutant model (Figure 3A). Additionally, the domains, wild-type residue, and mutated residue were colored to highlight the differences in the wild-type and mutant IL7R protein model, shown in Figure 3B and 3C. Moreover, it was observed that the mutation (K110N) lies in the IL-7Ralpha, FN3 domain, indicating potential role in the pathogenicity of the IL7R protein.

A total of 20 FDA-approved anti-inflammatory drugs were identified through a literature review, including diclofenac, diflunisal, etodolac, fenoprofen, flurbiprofen, ibuprofen, indomethacin, ketoprofen, ketorolac, mefenamic acid, meloxicam, nabumetone, naproxen, oxaprozin, piroxicam, sulindac, tolmetin, celecoxib, rofecoxib, and valdecoxib. The drug names, their PubChem IDs, SMILES, and 2D illustrations of the retrieved drugs are shown in Figure S1 and Table S1.

The virtual screening of the IL7R wild-type and mutant model with the anti-inflammatory drugs predicted various binding affinities ranging from –4.97 kcal/mol to –7.56 kcal/mol. However, the top three best binding affinities of –7.56 kcal/mol, –7.56 kcal/mol, and –7.53 kcal/mol were shown by wild-type IL7R-oxaprozin complex, mutant (K110N) IL7R-oxaprozin complex, and wild-type IL7R-celecoxib complex, respectively. Importantly, it was observed that the anti-inflammatory drug oxaprozin showed the same binding affinities (–7.56 kcal/mol) with both the wild-type and the mutant (K110N) model. Additionally, we have provided all docking results in Table S1.

Moreover, the 2D interaction analysis revealed that the oxaprozin exhibited Pi-Pi T-shaped, Pi-Pi stacked, and conventional hydrogen bond interactions with the wild-type protein residues, where including LYS-161 and VAL-160, were involved in conventional hydrogen bonds, while TYR-159, PHE-213, and TYR-212 were involved in Pi-Pi T-shaped and Pi-Pi stacked interactions. The 2D interactions of the wild-type IL7R-oxaprozin complex are shown in Figure 3C.

Furthermore, oxaprozin also exhibited Pi-Pi stacked, Pi-Pi T-shaped, Pi-sigma, unfavorable donor-donor, and conventional hydrogen bond interactions with the mutant model (K110N) protein residues, where TYR-159 and PHE-213 showed Pi-Pi stacked and Pi-Pi T-shaped interactions, TYR-212 showed Pi-sigma, LYS-161 showed unfavorable donor-donor interactions, and lastly, VAL-162 exhibited conventional hydrogen bond. The 2D interactions of oxaprozin with the mutant model (K110N) are shown in Figure 3D and 3E.

However, the visualization of the top three best-binding affinity complexes showed that oxaprozin and celecoxib exhibited a single hydrogen bond in all three complexes. The oxaprozin exhibited hydrogen bond interaction with protein residues K-161 and V-162 of wild-type and mutant (K110N) models with a bond distance of 1.89 and 2.33, respectively. Finally, celecoxib exhibited a hydrogen bond with the protein residue P-222 of wild-type protein with a bond distance of 1.83. The visualization of the best binding affinity complexes is shown in Figure 4A and 4B.

Nonetheless, it was observed that in all three complexes, the anti-inflammatory drugs (oxaprozin and celecoxib) interacted with the wild-type and the mutant (K110N) model within the FN3 domain region, for both models. Moreover, the FN3 domain is known to be involved in cell adhesion, cell morphology, thrombosis, cell migration, tumor metastasis, inflammation, and embryonic differentiation, indicating that the binding of oxaprozin and celecoxib in this domain can inhibit the crucial function of this domain.