Machine learning-based integration develops immunogenic cell death-derived lncRNA signatures to predict lung adenocarcinoma prognosis and immunotherapy response

Genetic characteristics and transcriptional changes of ICD-related genes in LUAD

A large-scale meta-analysis identified 34 ICD-related genes.¹¹First, we analyzed the expression of 34 ICD genes between LUAD samples and normal samples (Fig. S1A), and there were significant differences in the expression of most ICD genes, except for ATG5, IL10, CD8A, and CD8B. Next, we analyzed the location of ICD-related genes in the human genome (Figure S1B). We also evaluated mutations in ICD-related genes in LUAD patients in the TCGA cohort. The results showed that approximately 69.63% (188/270) of LUAD patients had mutations in ICD-related genes, and the study showed the top 20 mutations in ICD-related genes, with the highest frequencies of mutations in TLR4 and NLRP3. (Figure S1C and Figure S1D).

In this study, we also performed GO enrichment analysis of ICD-related genes (Figure S1E). This showed that, from a biological process perspective, the main enrichment is in different receptor activities. Regarding cellular components, cytolytic granules and inflammasome complexes were mainly enriched. From the point of view of molecular function, the main enrichment was in the biological processes of interleukins. Furthermore, KEGG enrichment analysis showed that ICD-related genes were enriched in NOD-like receptor signaling pathway, Toll-like receptor signaling pathway, and necroptosis. (Figure S1F).

Building and validating ICDI signatures

A total of 1,367 distinctive lncRNAs were selected by matching the training and validation datasets for detailed analysis. Consensus cluster analysis was used to divide the TCGA-LUAD dataset into two groups based on high and low expression of ICD-related genes. Subsequently, 473 lncRNAs were identified by performing differential expression analysis (Figure 2A and B). These lncRNAs were then compared with 300 lncRNAs obtained by Pearson correlation analysis (Figure 2C) and 176 ICD-associated lncRNAs were identified (Figure 2D). As a result, 24 ICD-related lncRNAs were finally identified by univariate Cox regression analysis (Supplementary Table 2).

A total of 24 ICD-related lncRNAs were input into a comprehensive machine learning model encompassing the 10 aforementioned methodologies to create a prognostic signature. Figure 3A shows the acquisition of a total of 101 prognostic models. The predictive signature created by the combination of RSF + Ridge had an average C-index of up to 0.674 when analyzing the training and testing cohorts. This signature was identified as an ICDI signature (Figure 3A and B). The resulting equation is as follows (see Supplementary Table 3 for details):

As an elastic net mixing parameter, α was bounded by 0 ≤ α ≤ 1. λ is defined as: \(\uplambda =\frac{1-\alpha }{2}{\Vert \beta \Vert }_{2}^{2}+\alpha {\Vert \beta \Vert }_{1}\).

LUAD patients were classified into two groups based on ICDI score: high score group and low score group. The median was used as the cutoff point. As expected, LUAD patients with low ICDI score showed higher overall survival in TCGA-LUAD, GSE29013, GSE30129, GSE31210, GSE3141, and GSE50081 datasets (Figure 3C).

The 1-, 2-, 3-, 4-, and 5-year AUC values ​​of ICDI signature in the TCGA-LUAD cohort were estimated to be 0.709, 0.678, 0.697, 0.716, and 0.660, respectively (Figure 3D). , indicating that the ICDI signature has promising predictive value for LUAD patients. GSE30219 cohort (0.891, 0.758, 0.744, 0.700, 0.716), GSE31210 cohort (0.750, 0.691, 0.653, 0.677, 0.718), GSE3141 cohort (0.690, 0.716, 0.819, 0.801, 0.729), GSE50081 cohort (0.685, 0.694, 0.712 , 0.638, and 0.639), and the GSE3141 cohort (0.639, 0.697, 0.794, 0.670, and 0.521) (Fig. 3D). Due to insufficient survival data, the GSE29013 cohort only calculates AUC values ​​for 2-, 3-, and 4-year periods. Still, it has strong predictive ability (Figure 3D).

Additionally, we compared the predictive value of the ICDI signature with other clinical variables (Figure 4A). The C-index of ICDI signature was significantly higher than other clinical variables including stage, age, gender, etc.

Comparison of prognostic signatures in LUAD

Machine learning-based gene expression analysis can be leveraged to predict disease outcomes, facilitating early disease screening and research into new treatments. Considerable predictive signs have emerged in recent years. To compare the ICDI signature with published signatures, we searched for papers on LUAD-related disease prediction models. Excluding papers with unclear predictive model formulas and missing corresponding gene expression data in training and validation groups, 102 His LUAD-related predictive signatures were finally registered (Supplementary Table 4 ). These signatures include cuproptosis, ferroptosis, autophagy, epithelial-mesenchymal transition, acetylation, amino acid metabolism, anoikis, DNA repair, fatty acid metabolism, hypoxia, inflammation, N6-methyladenosine, mitochondrial homeostasis, mTOR, etc. included different types of biological processes. Established on TCGA-LUAD, GSE29013, GSE30219, GSE31210, GSE3141, and GSE50081 and compared with the C-index of ICDI, we find that the ICDI signature outperforms the majority signature of each cohort (Figure 4B). .

Potential of ICDI signature as a biomarker for immunotherapy

To investigate the contribution of ICDI features in LUAD TIME, we evaluated the correlation between ICDI features and immune infiltrating cells and immune-related processes. Based on the TIMER algorithm, CIBERSORT algorithm, quantiseq algorithm, MCPcounter algorithm, xCell algorithm, and EPIC algorithm, the ICDI signature correlates with most immune infiltrate cells except for a few (such as activated NK cells and CD8+ naive T cells). (Figure 5A). Based on the ssGSEA algorithm, the ICDI signature was significantly correlated with most immune-related processes (Figure 5B). Based on the ESTIMATE algorithm, the ICDI signature was negatively correlated with StromalScore, ImmuneScore, and ESTIMATEScore and positively correlated with TumorPurity, as expected (Fig. 5C).

Additionally, this study also evaluated the relationship between the ICDI signature and known immunomodulatory factors (CYT, TLS, Davoli_IS, Roh_IS, Ayers_expIS, TIS, RIR, and TIDE) (Figure 5D). The values ​​of most immunomodulatory factors (CYT, TLS, Davoli_IS, Roh_IS, Ayers_expIS, and TIS) were significantly higher in the group with lower ICDI signature score. RIR values ​​and TIDE scores were all significantly higher in the high ICDI signature score group, suggesting a higher possibility of immunological evasion (Fig. 5D). All of these exhibited ICDI signatures that were potential immunotherapy biomarkers.

To further explore the potential of the ICDI signature as an immunotherapy biomarker, this study calculated the ICDI score for each immunotherapy cohort, respectively, and evaluated its predictive value. The results showed that people with lower ICDI scores were more likely to benefit from immunotherapy. (FIG. 6A) Receiver operating characteristic (ROC) analysis performed in this study showed that the ICDI signature exhibited consistent ability to predict the efficacy of immunotherapy-based treatments. This finding was further supported by analysis of immunotherapy datasets including cohort melanoma-GSE78220, STAD-PRJEB25780, and GBM-PRJNA482620, yielding ROC values ​​of 0.771, 0.671, and 0.723, respectively (Figure 6B) .

Impact of ICDI signature on chemotherapy

Chemoresistance is a significant barrier to the efficacy of chemotherapy and targeted therapy in the treatment of advanced lung cancer. We performed an analysis to investigate the drug sensitivity of various chemotherapy agents in vivo. We then compared drug sensitivity using ICDI signatures. Those with low ICDI scores showed a significant increase in sensitivity to erlotinib, gefitinib, docetaxel, and paclitaxel. However, there was no significant change in sensitivity to cisplatin and 5-fluorouracil. (Figure 6C) This study provides guidance for the administration of chemotherapy drugs in LUAD patients.