2.1 Selection of datasets and removal of batch effects

The study's prognostic model was created using the training cohort, and its effectiveness was assessed by validating it in the validation cohort. The TCGA-LUAD project, which offers comprehensive clinical and high-throughput data, was selected for the training cohort. The project's expression and other associated data were accessed via the Xena Hub online portal (https://xenabrowser.net/). The validation cohort's data were sourced from the Gene Expression Omnibus (GEO) database, which was accessed through its official website https://www.ncbi.nlm.nih.gov/geo/. Our search was tailored to identify a dataset related to ‘lung adenocarcinoma’, where we filtered out any results that did not contain expression and survival data to create our candidate dataset. We opted for GSE29013, GSE30219, GSE31210, GSE37745 and GSE50081 datasets from GEO. It is essential to highlight that these datasets underwent preprocessing before being used. To carry out preprocessing, we utilized the R package ‘inSilicoMerging’³⁵ to merge them, and we eliminated batch effects using the approach established by Johnson et al.³⁶ The preprocessed GEO data were utilized as the validation cohort.

2.2 Consensus clustering for m6A/m5C/m1A-regulated genes (mRG) subgroups

We selected 10 m1A-regulated genes, which are writer: TRMT61A, TRMT10C, TRMT61B and TRMT6; reader: YTHDF2, YTHDF3, YTHDF1 and YTHDC1; eraser: ALKBH1 and ALKBH3. We selected 13 genes regulated by m5C, which are writer: NSUN7, NSUN5, NSUN6, NSUN4, NSUN3, TRDMT1, DNMT1, NOP2, NSUN2, DNMT3A and DNMT3B; reader: ALYREF; eraser: TET2. We selected 21 m6A-regulated genes, which are writer: KIAA1429, RBM15B, METTL3, CBLL1, METTL14, ZC3H13, RBM15, WTAP; reader: IGF2BP1, LRPPRC, ELAVL1, HNRNPA2B1, HNRNPC, FMR1, YTHDC2, YTHDF2, YTHDF3, YTHDC1 and YTHDF1; eraser: ALKBH5 and FTO. We selected a total of 44 genes, and after removing duplicate ones, the remaining 40 genes were waiting to be dispatched. The R language package ‘limma’³⁷ was employed to verify if the 40 distinct genes exhibited varied expression levels in normal tissues and LUAD tumours. The genes exhibiting differential expression were selected by applying a threshold of FDR < 0.05 for the differential expression analysis. Subsequently, the selected genes were fed into the ‘ConsensusClusterPlus’³⁸ algorithm of the R language package for unsupervised clustering of the LUAD training samples. The optimal number of subtypes was decided by evaluating the value of k and examining whether there were any survival differences between the subtypes. To evaluate variations in survival among mRG clusters, we conducted Kaplan–Meier (KM) analysis, and to exhibit differences between clusters, we employed principal component analysis (PCA). The execution of both KM and PCA analyses was dependent on specific R language packages, namely ‘survival³⁹’, ‘survminer⁴⁰’ and ‘scatterplot3d⁴¹’. Additionally, we employed several R packages, including ‘GSEABase⁴²’, ‘reshape2⁴³’, ‘limma³⁷’, ‘ggpubr⁴⁴’ and ‘GSVA⁴⁵’, to execute the single-sample gene set enrichment analysis (ssGSEA) and generate visualizations. KEGG analysis was then performed using ‘GSVA’ R package on the mRG clusters to discover potential pathways.^46-48 We utilized the ‘limma’ R package with an FDR threshold of less than 0.05 to identify differentially expressed genes related to the mRG clusters (mRG-DEGs) between the clusters.

2.3 Development of mRG-DEG cluster and signature of m6A/m5C/m1A-regulated lncRNAs (mRLncSig)

We categorized patients in the training cohort based on mRG-DEG and generated KM curves to evaluate survival disparities across the mRG-DEG clusters. To assess the level of differentiation among the different clusters, we utilized PCA. Then, we conducted the ssGSEA and generate visualizations. Next, we employed the ‘limma’, ‘GSEABase’, ‘GSVA’ and ‘pheatmap’ R packages to perform GSVA to identify the top significant KEGG pathways among the mRG-DEG clusters. We explored the lncRNA transcripts that were differentially expressed between the mRG-DEG clusters with an FDR threshold of less than 0.05. Subsequently, we conducted univariate Cox and KM analyses on these differentially expressed lncRNAs (DELs) to identify the ones that showed potential prognostic significance with a p-value of less than 0.05. To further decrease the dimensionality of the prognostic DELs and avoid overfitting, we utilized the R language package ‘glmnet’⁴⁹ to implement the LASSO algorithm. By subjecting the DELs to a 10-fold cross-validation test, we obtained a set of lncRNAs and their corresponding coefficients through the LASSO analysis. The risk score of each LUAD was calculated as the sum of the product of each lncRNA's expression level and its corresponding coefficient. The formula is as follows: Risk score = (lncRNA₁ expression*lncRNA₁ coefficient) + (lncRNA₂ expression*lncRNA₂ coefficient) + … + (lncRNA_n expression*lncRNA_n coefficient).

2.4 Validation of mRLncSig in a large independent cohort

A risk score was assigned to each LUAD of the cohort and used to divide the cohort into high-risk and low-risk groups based on the median risk score. The predictive ability, accuracy and discrimination of mRLncSig were evaluated using various bioinformatics methods, including Cox analysis,⁵⁰ KM analysis,⁵¹ ROC analysis,⁵⁰ tAUC analysis⁵² and survival nomogram.⁵³ The analysis was conducted in R software, utilizing packages such as ‘timeROC⁵⁴’, ‘survival’, ‘survminer’, ‘rms⁵⁵’ and ‘regplot’. A gene set of immunotherapy-predicted pathways was collected from Hu et al.'s study.⁵⁶ We also collected other gene set, oncogenic signature gene set, from Human MSigDB Collections (C6: oncogenic signature gene sets, v2022.1.Hs updated August 2022, https://www.gsea-msigdb.org/gsea/msigdb/human/collections.jsp#C6). The enrichment scores of these signatures were calculated using the GSVA R package.⁵⁶

2.5 Identification of the role of mRLncSig in the immunological status of LUAD

The R package ‘ESTIMATE’ utilizes the gene expression levels of the training cohort to compute stromal, immune and ESTIMATE scores for individual patients.⁵⁷ We evaluated the correlation between mRLncSig and the above category scores using statistical analysis methods like the Pearson coefficient and the Wilcoxon rank-sum test. With R package ‘IOBR’, immuno-oncology exploration can be facilitated, tumour-immune interactions can be explored, and precision immunotherapy can be expedited.⁵⁸ The R package ‘IOBR’ or its algorithms included, namely CIBERSORT,⁵⁹ CIBERSORT-ABS,⁵⁹ quanTIseq,⁶⁰ TIMER,⁶¹ MCPCounter,⁶² xCell⁶³ and EPIC,⁶⁴ were applied to assess immune-infiltrating levels of every LUAD in the TCGA-LUAD. To assess the relationship between mRLncSig and immune-infiltrating levels, we employed the Pearson coefficient and the Wilcoxon rank-sum test, and the outcomes were presented as lollipop plots and heatmaps. We summarized the findings through Venn and cloud diagrams and assessed the immune function of mRLncSig utilizing the ‘ssGSEA’ function available in the ‘gsva’ R package.

2.6 Identification of mRLncSig's role in immunotherapy and its potential checkpoint targets

Initially, we employed the ‘maftools’ R package to visualize the mutation landscape in LUAD. Our primary emphasis was on the top 20 genes with the most mutations, and we aimed to analyse and exhibit them. We utilized the chi-square test to compare the mutation frequencies of these 20 genes between the low-risk and high-risk groups. TMB is a gauge of the incidence of specific mutations in cancer genes and is increasingly being adopted as an indicator of immunotherapy responsiveness.⁶⁵ To assess TMB rank scores for LUAD cases, we followed established protocols. To evaluate the correlation between the risk score and TMB, we utilized a combination of Pearson's coefficient and Wilcoxon rank sum. The ability of Tumour Immune Dysfunction and Exclusion (TIDE) to replicate two possible mechanisms of tumour-immune evasion can be employed to anticipate the effectiveness of immunotherapy.^66-68 Our primary objective was to determine the correlation between our signature and the TIDE. In our study, we chose a set of 60 immune checkpoints that had been previously investigated, which included 24 inhibitory and 36 stimulatory checkpoints⁶⁹ (Table S1). To evaluate the relationships between our mRLncSig and the 60 selected immune checkpoints, we conducted integration analysis including Pearson coefficient and Wilcoxon rank-sum analyses. We sought to determine if our mRLncSig could serve as a guide for immunotherapy. To this end, we utilized the KM and Cox analysis to assess the outcome predictive value of 60 immune checkpoints. Using a Venn diagram, we summarized the results to identify potential checkpoints with targeting ability relate to that of the mRLncSig. Furthermore, we conducted a search of public databases to locate datasets that include information on immunotherapy to evaluate the influence of the checkpoints highlighted earlier on immunotherapy. The ‘Regulatory Prioritization’ function in the TIDE online tool facilitated our visualization of the results of immunotherapy.⁶⁷

2.7 Drug selection for patients with high mRLncSig score LUAD

A comprehensive examination of numerous human cancer models was conducted through the initiation of the Cancer Cell Line Encyclopedia (CCLE) project in 2008. The drug sensitivity data utilized in this investigation were sourced from the Cancer Therapeutics Response Portal (CTRP, https://portals.broadinstitute.org/ctrp) and PRISM (https://depmap.org/portal/prism) databases, with the former providing information on 481 compounds from 835 cancer cell lines (CCLs) and the latter assessing 1448 compounds from 482 CCLs. In both datasets, the drug sensitivity was determined by the area under the dose–response curve (AUC), with lower values indicating greater sensitivity. Our study involves the analysis of drug response data from CTRP and PRISM to identify feasible drug candidates from the high-scoring group.⁷⁰ To do this, we compared drug responses between patients with the highest and lowest decile risk scores and used a threshold of log2FC > 0.1 to screen for drugs with lower AUC in high-scoring patients.⁷¹ To choose the target compounds,⁷¹ we performed Spearman correlation analysis with a threshold of r < −0.18 to determine the correlation between drug AUC values and risk scores.

2.8 Connectivity Map (CMAP) to validate drug candidates

Afterward, additional validation analyses were conducted on the results of the drug candidate, which involved reviewing the data of clinical trial and published experimental evidence, and the use of CMap to further confirm its potential in LUAD.⁷¹ The Connectivity Map, or CMap, facilitates drug discovery by creating and analysing massive datasets of altered biological conditions, providing insights into human diseases and accelerating the search for novel therapies.⁷¹ In this study, we employed CMap analysis as a supplementary approach to explore the potential efficacy of the identified drug candidates in LUAD. There were 2429 compounds accessible for analysis on CMap. The top 150 upregulated and top 150 downregulated genes in the differential ranking were chosen after conducting differential analysis on LUAD tumour and normal tissue samples. These selected genes were then taken to the CMap online analysis portal for drug validation. Each compound's CMap result is represented as a value between −100 and 100, with a result closer to −100 indicating a greater potential for therapeutic power.

2.9 Comparing mRLncSig with previous studies

To conclude whether our study is more robust than previous, we searched PubMed using the keywords ‘m1a lncRNA signature lung adenocarcinoma prognosis’, ‘m5c lncRNA signature lung adenocarcinoma prognosis’ and ‘m6a lncRNA signature lung adenocarcinoma prognosis’ to find candidate studies. We included the research that contains a lncRNA signature and the related coefficient. Because most of the candidate studies did not upload raw data or used different or unmentioned data preprocessing methods, therefore, to ensure the standard consistency of the comparison, we use the official TCGA data for analysis here, which are TCGA-LUAD_PanCanAtlas from Genomic Data Commons, Pan-Cancer Atlas (https://gdc.cancer.gov/about-data/publications/pancanatlas), and TCGA-LUAD_Count and TCGA-LUAD_FPKM_UQ from Genomic Data Commons Data Portal (https://portal.gdc.cancer.gov/). For specific comparative analysis, we used Cox regression analysis.

2.10 Using real-time PCR to measure the expression levels of lncRNAs and analyse data from multiple databases to determine if mRLncSig has the potential to impact various cancers

The situation of the target gene in the real world can be described by laboratory data obtained from human samples. The expression level of each mRLncSig lncRNA was investigated by collecting nine pairs of LUAD and adjacent tissues from the clinic.⁷⁰ All patients included in this study did not receive any relevant treatment before collecting samples. The Ethical Review Committee of the First Affiliated Hospital of Zhengzhou University approved our approaches, and informed consent was obtained from all patients before the operation. Tissue samples were immediately frozen and stored in liquid nitrogen after extraction during the surgery.⁷² TRIzol reagent (Invitrogen, Thermo Fisher Scientific Corporation, MA, USA) was utilized to extract total RNA from tissues. Reverse transcription was performed using a PrimeScript™ RT reagent Kit with gDNA Eraser (TAKARA BIO INC., Kusatsu, Shiga, Japan). Real-time PCR was conducted using a SYBR Premix Ex Taq™ II kit (TAKARA BIO INC.) on a CFX Opus 96 Real-Time PCR System (Bio-Rad Laboratories, Hercules, CA, USA). Relative gene expression was calculated automatically using $$$ {2}^{-\Delta \Delta {C}_{\mathrm{t}}} $$$.⁷² Detection of genes between normal and tumour samples was conducted utilizing Student's t-test, with statistical significance established for adjusted p-values below 0.05.⁷²

For our pan-cancer analysis,⁷⁰ we opted for the TCGA pan-cancer data that we obtained from the UCSC database (https://xenabrowser.net/). After downloading the data, we filtered out the haematological tumour data and retained only cancer types that featured both normal and tumour tissues. R packages ‘ggplot2’, ‘clusterProfiler’, ‘ComplexHeatmap’ and ‘limma’ were adopted for the calculation and visualization. Then, we conducted the prognostic ability determination and only cancer types that contain expression and survival data were selected. R packages ‘survival’ and ‘pheatmap’ were used for this approach.

3.1 Patient characteristics

The critical steps of this study are illustrated in Figure 1. To build our validation cohort, we selected 500 LUADs from TCGA-LUAD. Additionally, we gathered 554 LUAD patients from five datasets in the GEO database (GSE29013, GSE30219, GSE31210, GSE37745 and GSE50081) to augment our validation cohort. The elimination of batch effects associated with the data merging was carried out following the approach described by Johnson et al.,³⁶ The UMAP diagram (Figure 2A) revealed that prior to the elimination of batch effects, the merged data set was segregated, whereas after removing the batch effect, the data sets became intertwined, indicating a successful elimination of the batch effect. The cohorts' status and the clinical baseline information of the patients included in our study are presented in Table 1.

3.2 Constriction of mRG clusters in LUADs using consensus clustering

We selected a total of 44 mRGs as mentioned in the method section, and after removing duplicate ones, there were 40 genes remaining. Table 2 demonstrates that out of the 40 candidate genes, 33 fulfilled our criteria based on the differentially expressed FDR values. These genes were subjected to the consensus clustering algorithm to classify LUAD patients, resulting in two mRG clusters. Figure 2B depicts the KM survival curves of the two clusters, demonstrating significant differences in terms of prognosis, with cluster B having a better outcome than cluster A. Additionally, a distinct separation between clusters A and B is noticeable from the PCA analysis shown in Figure 2C. Based on the ssGSEA analysis depicted in Figure 2D, 16 types of immune cells, activated B cell, activated CD8 T cell, activated dendritic cell, CD56dim natural killer cell, eosinophil, immature B cell, immature dendritic cell, MDSC, macrophage, mast cell, monocyte, natural killer cell, regulatory T cell, T follicular helper cell, Type 1 T helper cell and Type 17 T helper cell, were statistically distributed in two mRG clusters. Moreover, there was a pronounced difference between the two LUAD patient clusters in the aspect of the expression of 33 m6A/m5C/m1A-regulated genes (Figure 2E). To identify the most important KEGG pathways, we compared the two mRG clusters and conducted GSVA analysis (Figure 2F, Table S2). Interestingly, the top 10 ranked pathways were related to spliceosome, base excision repair, RNA degradation, basal transcription factors, lysine degradation, mismatch repair, nucleotide excision repair, aminoacyl-tRNA biosynthesis, homologous recombination and one carbon pool by folate. The dissimilarities observed between two populations can frequently be accounted for by genes that are expressed differently between them. To gain insights into the underlying mechanisms responsible for the divergence between the two mRG clusters, we delved deeper into the genes that were expressed differentially, ultimately identifying 256 mRG cluster-related differentially expressed genes (mRG-DEGs) (Table S3).

3.3 Two mRG-DEG clusters constructed and a mRLncSig generated

After adopting mRG-DEG, a consensus clustering approach was employed to partition the training cohort's LUADs into two distinct mRG-DEG clusters. The prognostic ability of the mRG-DEG clusters was assessed using their KM survival curves, revealing that cluster A had a more favourable prognosis compared to cluster B (Figure 3A). Furthermore, a clear separation between clusters A and B was observed through PCA analysis (Figure 3B). The distribution of 14 immune cells, activated B cell, activated CD4 T cell, CD56bright natural killer cell, eosinophil, gamma delta T cell, immature dendritic cell, mast cell, monocyte, natural killer T cell, neutrophil, plasmacytoid dendritic cell, regulatory T cell, Type 17 T helper cell and Type 2 T helper cell, across different mRG-DEG clusters was differentially visualized through ssGSEA (Figure 3C). Additionally, significant differences in gender and tumour stage distribution were noted among the mRG-DEG clusters, as shown in Figure 3D. We performed GSVA to determine the top significant KEGG pathways of the mRG-DEG clusters (Figure 3E, Table S4) showing that KEGG_CELL_CYCLE, KEGG_DNA_REPLICATION, KEGG_HOMOLOGOUS_RECOMBINATION, KEGG_MISMATCH_REPAIR, KEGG_P53_SIGNALING_PATHWAY, KEGG_OOCYTE_MEIOSIS, KEGG_PROTEASOME, KEGG_ALPHA_LINOLENIC_ACID_METABOLISM, KEGG_NUCLEOTIDE_EXCISION_REPAIR and KEGG_PATHOGENIC_ESCHERICHIA_COLI_INFECTION ranked the most important 10 pathways. We examined the distribution of 33 mRGs, including NSUN2, DNMT3B, NOP2, NSUN5, DNMT3A, HNRNPC, LRPPRC, YTHDF1, ALYREF, HNRNPA2B1, TRMT6, DNMT1, TRMT61B, NSUN6, IGF2BP1, ELAVL1, VIRMA, RBM15, METTL3, ALKBH1, NSUN4, YTHDF2, TRMT10C, TRMT61A, NSUN7, RBM15B, CBLL1, NSUN3, YTHDF3, FMR1, YTHDC1, YTHDC2 and FTO, in mRG-DEG clusters. We found that 24 mRGs, including ALKBH1, ALYREF, ELAVL1, HNRNPA2B1, HNRNPC, IGF2BP1, LRPPRC, NOP2, NSUN2, NSUN3, NSUN5, RBM15, RBM15B, TRMT10C, TRMT6, TRMT61A, CBLL1, DNMT1, DNMT3A, DNMT3B, TRMT61B, VIRMA, YTHDF1 and YTHDF3, were related to the mRG-DEG clusters (Figure 3F). Remarkably, all the 24 mRGs mentioned above exhibited upregulation in cluster B compared to cluster A of mRG-DEGs.

To construct a prognosis model for LUAD, we initially searched the DEL from two mRG-DEG clusters and identified 4517 DELs. Subsequently, we performed KM and Cox analyses to screen for DELs that met our criteria, leading to the identification of 18 DELs (Table S5). To further refine our findings, we carried out LASSO analysis using these 18 DELs for selection and shrinkage. This analysis enabled the identification of 10 lncRNAs (Figure 4A,B and Figure S1), and we obtained the coefficient for each gene (Table 3). To better comprehend the procedures, we examined and the relationships among them, we depicted them as Sankey diagrams (Figure 4C), with the aid of the mRG clusters, mRG-DEG clusters, risk levels and vital status. These variables may provide greater insight into the analyses we conducted and the associations between them. We also used box plots to show that the risk score distribution within mRG-DEG clusters varied significantly (Figure 4D). After examining the expression pattern of the 33 mRGs in both high- and low-risk groups, we identified 25 genes (ALKBH1, DNMT1, DNMT3A, ELAVL1, FMR1, FTO, HNRNPA2B1, HNRNPC, IGF2BP1, METTL3, NSUN3, NSUN4, NSUN5, NSUN6, NSUN7, RBM15, RBM15B, TRMT61A, TRMT61B, VIRMA, YTHDC1, YTHDC2, YTHDF1, YTHDF2 and YTHDF3) that showed significant differences in expression (Figure 4E). Out of the 25 identified genes, only IGF2BP1 was upregulated, while the remaining genes were downregulated in the high-risk group. Additionally, we have included a display of the correlations between each of the 10 lncRNAs and the 33 mRGs in Figure S2A.

3.4 mRLncSig's stable prognostic power was confirmed by an independent cohort validation

Figure S2B,C display the risk plots we created for the general situation of mRLncSig in the two cohorts. The graphs are partitioned into three sections. The top section presents patients sorted in ascending order of risk score from left to right. The middle scatterplot illustrates the vital status of LUADs using blue for alive and red for dead. Finally, the heatmap at the bottom shows the relative expression levels of the 10 lncRNAs in the mRLncSig signature. The visualization in Figure 5A upper depicts the KM analysis of the training cohort, revealing that LUAD in the high-risk category had poorer survival prospects than low-risk LUAD. These findings align with the results seen in the validation cohort's KM curve illustrated in Figure 5A lower. The KM curve (Figure 5B) depicts the divergence in survival rates of progression-free interval, disease-specific survival and disease-free interval between groups classified as high and low risk. This visualization reveals that the high-risk score group exhibited a lower survival rate. Figure S3A illustrates the prognosis ability of 10 lncRNAs through Kaplan–Meier curves utilizing data from both cohorts. Our findings demonstrate that ITGB1-DT and AC010327.4 consistently have a negative effect on LUAD cases, while LIFR-AS1, AC107464.3, LINC00324, COLCA1, LINC00639, LINC00892, AC093010.2 and AL353622.1 contribute to the prognosis improvement of LUADs.

Our analysis focuses on whether the risk score can independently predict the outcome of LUAD, regardless of other clinical factors. To accomplish this, we conducted univariate and multivariate Cox analyses on all cohorts (Figure 5C). Our Cox model included the risk score and several clinical factors, such as age, gender, race, ethnicity, tumour stage and tumour origin. Notably, our risk score demonstrated strong prognostic capabilities in univariate analyses across all cohorts. In the multivariate Cox analysis, the risk score had a hazard ratio of 3.26, a 95% CI of 2.10–5.08 and a p-value of 1.60e-07 in the training cohort, and a hazard ratio of 2.86, a 95% CI of 1.63–5.02 and a p-value of 2.56e-04 in the validation cohort. These findings indicate that the risk score has significant independent prognostic power. Our multivariate Cox model revealed that the clinical parameter ‘age’ was significantly associated with prognosis in the validation cohort, while no significant association was observed in the training cohort. Furthermore, the prognostic potential of lncRNAs included in mRLncSig was demonstrated in our univariate Cox analysis visualization, as illustrated in Figure S3B. To assess the predictive accuracy of mRLncSig for LUAD outcomes, we generated and analysed ROC curves using data from both the training and validation cohorts, as depicted in Figure 5D. In the training cohort, the AUCs of mRLncSig at 1, 3 and 5 years were 0.722, 0.661 and 0.634, respectively. In the validation cohort, the corresponding AUCs were 0.601, 0.644 and 0.624, respectively. Furthermore, we carried out time-dependent AUC analysis to evaluate the prognostic capacity of our risk score at continuous time intervals (Figure 5E). Our results suggested that our risk score was comparable to the established prognostic standard, ‘tumour stage’. Notably, in our training cohort, the combined AUC of risk score and tumour stage exceeded 0.7 and outperformed tumour staging alone. In the validation cohort, the combined AUC was the most effective predictor at all time points. The time-dependent AUC analysis confirmed that our mRLncSig is a valuable addition to tumour stage. The signature can effectively discriminate high- and low-risk cases in our studied cohorts, as demonstrated in Figure 5F, which presents the PCA results. In addition, a prognostic nomogram that has the potential for clinical use was developed, as illustrated in Figure 5G. The input variables for the nomogram consist of our risk score, as well as clinical parameters such as age, gender and tumour stage. The accuracy of the nomogram's predictions was confirmed by the calibration analysis depicted in Figure 5H.

3.5 Correlations between mRLncSig and the enrichment scores of immunotherapy-predicted pathways and oncogenic signature gene sets

We analysed the correlations between mRLncSig and the immunotherapy-predicted pathways. The top 10 pathways that mRLncSig correlated with were progesterone mediated oocyte maturation, oocyte meiosis, cell cycle, pyrimidine metabolism, p53 signalling pathway, Fanconi anaemia pathway, viral carcinogenesis, homologous recombination, mismatch repair and DNA replication (Figure 5I). As expected, mRLncSig correlated with some of the oncogenic signature gene sets, which the top 10 ranked were MTOR_UP.N4.V1_DN, CSR_EARLY_UP.V1_DN, VEGF_A_UP.V1_DN, CSR_EARLY_UP.V1_UP, CSR_LATE_UP.V1_UP, RPS14_DN.V1_DN, E2F1_UP.V1_UP, TBK1.DN.48HRS_DN, HOXA9_DN.V1_DN and SIRNA_EIF4GI_DN (Figure 5J).

3.6 Identification of mRLncSig's potential in immunological status of LUAD

The progression of cancer is driven by the collaboration between subclonal populations, which comprise cancerous and non-cancerous cells in the tumour microenvironment. This intricate system forms a dynamic ecosystem. Therefore, it is crucial to conduct a comprehensive examination of the tumour microenvironment. In this study, we utilized data from the TCGA cohort and utilized the R package ‘ESTIMATE’ to measure various scores such as immune score, stromal score and ESTIMATE score. Figure 6A–C depicts our visualizations of boxplots and correlation analyses. These visualizations reveal that the ‘ESTIMATE’ algorithm scores were lower in the high-risk population, and the risk score demonstrated a negative correlation with the ‘ESTIMATE’ algorithm scores. Using the seven primary immune algorithms, we assigned immune scores to individuals within the training cohort. Subsequently, we utilized statistical methods such as the Wilcoxon rank-sum test and Pearson correlation coefficient to compare differences and correlations between high and low risks. The results were presented as heatmaps and lollipop plots in Figure 6D,E, respectively. Only significant factors were highlighted in the plots, while detailed information was provided in Table S6. Figure 6F presents a comprehensive analysis that employs a combined Venn diagram and word cloud to visualize the results of the heatmap and lollipop analysis. The analysis identifies cells that are most closely related to our signature, such as CD4 T cells, memory B cells, resting T cells, myeloid dendritic cells and CD8 T cells. Furthermore, Figure 6G illustrates the immune function analysis that reveals the differential distribution of immune function scores between high- and low-risk groups. Notably, chemokine receptors, checkpoint, human leukocyte antigen, T cell co-inhibition, T cell co-stimulation and type 2 interferon response exhibit the most pronounced differences. Taken together, these findings suggest that our signature may be linked to the immune status in LUAD.

3.7 mRLncSig participates in immunotherapy and targets immune checkpoints

According to our extensive analysis of mutational characteristics (Figure 7A), TP53 emerged as the most mutated gene, with a frequency of approximately 53.8% within the cohort. Following closely were TTN and MUC16, accounting for 51.0% and 44.2% of the mutations, respectively. Missense mutation was the most frequently observed type of mutation. The Wilcoxon test verified that the LUAD with a higher risk score demonstrated an elevated level of TMB, while Pearson's analysis revealed a positive correlation between TMB and risk score (Figure 7B).

Clinical studies have shown that patients with higher TMB tend to respond better to immune checkpoint blockade therapy, resulting in more long-lasting clinical benefits, including treatment responses and improved survival.^{73, 74} Our findings indicated that patients with high-risk LUADs might respond better to immunotherapy to some extent. The TIDE score is a surrogate biomarker that can be utilized to forecast the likelihood of NSCLC patients responding to immune checkpoint blockade therapies, such as anti-PD1 and anti-CTLA4. Higher TIDE prediction scores are generally associated with increased potential for immune evasion.^66-68 Our study evaluated the potential clinical effectiveness of immunotherapy for LUAD by analysing the TIDE score in conjunction with the high and low values of our mRLncSig score. According to Figure 7C, our model indicates that patients in the high-risk group have lower TIDE scores, suggesting that they are more likely to respond positively to immunotherapy. Our TMB and TIDE analysis results are consistent, indicating that patients in the high-risk group are more likely to benefit from immunotherapy.

We selected 60 immune checkpoint genes for analysis based on previous research. Our analysis, using Pearson's correlation (Figure 7D), revealed that 51 of these genes were significantly associated with our risk score. The top five were CD40LG (coefficient = −0.615267127, p = 2.09E-53), BTLA (coefficient = −0.515468939, p = 2.75E-35), SELP (coefficient = −0.503124012, p = 1.93E-33), IL2 (coefficient = −0.468507858, p = 1.2E-28) and TNFRSF14 (coefficient = −0.46810592, p = 1.36E-28) (Figure 7D). Figure 7E demonstrates the Wilcoxon analysis results indicating the distinctive distribution of 60 checkpoint genes between the high- and low-risk groups. Among them, 47 genes exhibited distribution differences. Additionally, the KM curve shown in Figure 7F was utilized to assess the survival variation of checkpoint genes. The findings revealed that the prognosis of LUAD was affected by eight genes, namely SELP, CD40LG, IL10, IL2, KIR2DL3, CD28, BTLA and TLR4. Furthermore, 15 genes were considered to be related to prognosis, as identified by the univariate Cox regression analysis visualization displayed in Figure 7G. To ensure a well-rounded outcome for each analysis, we utilized a Venn diagram to overlap the results from the plots. Upon examination, we observed that CD40LG, BTLA, SELP, IL2, CD28 and IL10 not only displayed a significant association with our mRLncSig, but also had an impact on LUAD prognosis. Consequently, these genes warrant further investigation (Figure 7H). Figure 7I highlights six checkpoint genes that could have an impact on the immune system and immunotherapy. The immunotherapy cohort represented by the black module ranked these genes based on their abilities, with IL10 having the highest rank, followed by IL2, CD40LG, SELP, BTLA and CD28. These results suggest the possibility of in-depth crosstalk studies beneath our mRLncSig and immunotherapy.

3.8 Discovering potential therapeutic agents for LUAD with high mRLncSig score

The CTRP and PRISM datasets comprise gene expression profiles and drug sensitivity profiles of numerous CCLs, enabling the creation of a drug response prediction model. After eliminating duplicates, these datasets encompass a total of 1770 compounds (Figure 8A and Table S7), with 160 compounds common to both. The roadmap for identifying sensitive drugs for patients with high-risk scores is detailed in Figure 8B. These analyses led to the identification of six CTRP-derived compounds, including paclitaxel, methotrexate, selumetinib, leptomycin B, SB-743921 and PD318088 (Figure 8C), as well as six PRISM-derived compounds, including echinomycin, cabazitaxel, vincristine, gemcitabine, NVP-AUY922 and Ro-4987655 (Figure 8D). Our results demonstrate that the discovered compounds had lower AUC values in the high-risk score group, and there was a negative correlation between their AUCs and the risk score.

Although the 12 candidate compounds demonstrated heightened drug sensitivity in high mRLncSig risk patients, the aforementioned analyses solo are not able to substantiate their efficacy. Therefore, additional multi-dimensional analyses were conducted to evaluate their therapeutic capacity in patients with LUAD. CMap analysis indicated that among these compounds, selumetinib and gemcitabine stood out with CMap scores of <−95, suggesting potential therapeutic benefits for LUADs (Figure 8E and Table S7). Furthermore, we performed an analysis to calculate the fold-change values of drug candidates in tumour and normal tissues. The increased values observed in Figure 8E and Table S8 indicate a higher potential for treating LUAD. In addition to performing a thorough search of the literature, including PubMed (https://pubmed.ncbi.nlm.nih.gov/) and ClinicalTrials.gov (https://clinicaltrials.gov/), we sought experimental and clinical evidence that confirmed or supported the candidate drug. The specific results are presented in Figure 8E and Table S7. Based on the comprehensive analysis and presentation outlined above, as well as its performance in both in silico and in vitro studies, gemcitabine emerged as the most promising drug candidate with the highest potential for treating LUAD.

Based on the screening criteria we set, we found eight studies,^75-82 but one of them was removed because it did not contain coefficient information. Finally, seven studies came into our view^{75, 77-82} (Table 4). To compare previous signatures with ours, we performed Cox regression analysis for OS, DSS and PFS using three formats of official TCGA data (Figure 9), respectively. The analysis confirmed that mRLncSig has solid predictive ability in overall, disease-specific and progression-free survival in three testing cohorts (p < 1.21e-04). In particular, our signature occupies the first place in terms of p-value in the OS and DSS prediction. mRLncSig ranked second in terms of p-value in PFS prediction in TCGA-LUAD_PanCanAtlas and TCGA-LUAD_FPKM_UQ. mRLncSig ranked third in in terms of p-value in PFS prediction in TCGA-LUAD_Count.

3.9 Identification of expression patterns of mRLncSig and its ability in pan-cancer

To assess the real-world effectiveness of 10 signature lncRNAs, we utilized real-time PCR to compare their expression levels in human LUAD tissue (n = 9) and adjacent normal lung tissue (n = 9). The primer sequences for the 10 lncRNAs tested, which include AC010327.4, AC093010.2, AC107464.3, AL353622.1, COLCA1, ITGB1-DT, LIFR-AS1, LINC00324, LINC00639 and LINC00892, are presented in Table 3. In Figure 10A, it is evident that there were distinct expression patterns of all lncRNAs in LUAD tumour samples and normal lung tissues. Specifically, AC010327.4 and ITGB1-DT lncRNAs were upregulated in LUAD tumour tissues, while the other lncRNAs were downregulated. Table 3 contains the primer sequences for the signature lncRNAs, which include AC010327.4, AC093010.2, AC107464.3, AL353622.1, COLCA1, ITGB1-DT, LIFR-AS1, LINC00324, LINC00639 and LINC00892. We conducted real-time PCR on nine pairs of LUAD and adjacent tissues to assess the expression levels of these lncRNAs. The comparison results, as shown in Figure 10A, revealed differential expression of the 10 lncRNAs in tumour and normal tissues. Notably, only AC010327.4 and ITGB1-DT were upregulated, while the remaining eight lncRNAs exhibited decreased expression levels in tumour tissues. It is worth noting that the upregulation of AC010327.4 and ITGB1-DT genes in LUAD tissues is consistent with the findings in Figure S2, which indicated their association with an unfavourable prognosis. On the other hand, the downregulated genes demonstrated protective effects on LUAD prognosis, which further supports the credibility of the gene signatures we discovered and provides guidance for future in-depth investigations.

Starting with pan-cancer expression patterns, we investigated the potential of 10 lncRNAs. To explore the expression variance of the 10 signature lncRNAs, we obtained their expression across 24 cancer types, as depicted in Figure 10B. The plots hinted that the lncRNAs, ITGB1-DT, AC010327.4, COLCA1, LIFR-AS1 and LINCO0892 ranked the different expression ability. The cancer types of KICH, KIPAN, NCSLC and THCA may strongly be impacted by the 10 lncRNAs. To delved deeper into the outcome predictive capabilities of 10 lncRNAs in pan-cancer, we meticulously used data from 33 types of cancers and constructed Cox models. The survival heatmap displayed in Figure 10C showed that the ITGB1-DT and AC010327.4 might have an unfavourable impact on most part of the pan-cancer population. In contrast, the remaining lncRNAs mostly protected the pan-outcomes. Our concise examination of the 10 lncRNAs and their association with pan-cancer reinforces the significance of our mRLncSig. This could potentially guide further investigations in other types of cancers.