Mechanical stress upregulates the tumor EMT pathway and promotes tumor metastasis.
We added Matrigel to the scratched LUAD cells seeded on 6-well plates to investigate the influence of mechanical stress to the migration and invasive ability of LUAD cells. The A549 and H1975 cells had enhanced migration ability after adding matrix adhesive and were significantly superior in 24-hour migration ability compared to their counterparts maintained without Matrigel (Fig. 1a). When testing lung adenocarcinoma migration ability through the transwell experiments (Fig. 1b), we also found that the tumor cells with added Matrigel above achieved stronger migration ability and had statistical significance. Similarly, when testing the invasive ability of lung adenocarcinoma cells through the Transwell experiment, we also found that the two types of tumor cells with added matrix glue achieved significantly stronger invasive ability than their counterparts maintained without Matrigel. (Fig. 1b).
The Western blot experiment showed that the EMT pathway of tumor cells was significantly upregulated when adding the matrix gel, manifested as the upregulation of Vimentin and Snail and downregulation of beta-Catenin and E-cadherin (Fig. 1c). Additionallythe immunofluorescence experiments showed that the tumor cells with matrix glue added significantly upregulated Vimentin (Fig. 1d). The Phalloidin experiment demonstrated that the tumor cell cultured with matrix gel changed morphology significantly and were no longer round and continuous, with a significant increase in tentacles (Fig. 1e).
SVM_Score generated based on machine learning effectively predicts the prognosis of lung adenocarcinoma patients.
A heat map between normal and lung adenocarcinoma tissues of TCGA-LUAD is shown in Supplementary Fig. 1a. The upste plot then shows 35 BM genes in all seven datasets we used (Supplementary Fig. 1b). The volcanic map displays the differential genes between normal and lung adenocarcinoma tissues of TCGA-LUAD (Supplementary Fig. 1c). At the same time, we conducted unit correlation analysis on these 35 BM genes and selected 14 BM genes related to the overall survival of lung adenocarcinoma for subsequent modeling by intersecting with differential gene results (Supplementary Fig. 1d). The oncoplt showed the mutations status of these 14 genes in lung adenocarcinoma, with the FBN2 gene has the highest mutation rate of 16%. (Supplementary Fig. 1e).
Our machine learning analysis results indicate that SVM_score, constructed using a Support Vector Machine (SVM), has the highest Concordance Index(C-index), as shown in Fig. 2a. C-index is a statistical measure used to assess a model’s discriminatory power or predictive accuracy, particularly in the context of survival analysis or time-to-event data. A higher C-index indicates a model with good discrimination. Therefore, we selected SVM_score for further analysis. Survival analysis revealed that in TCGA-LUAD (Fig. 2b), GSE31210 (Fig. 2c), GSE50081 (Fig. 2d), GSE3141 (Fig. 2e), GSE26939 (Fig. 2f), GSE30219 (Fig. 2g), and GSE72094 datasets (Fig. 2h), lower SVM_score values corresponded with a shorter overall survival. Additionally, in TCGA-LUAD, GSE31210, GSE50081, and GSE30219 datasets, lower SVM_score values were associated with shorter progression-free survival (Fig. 2i).
SVM_score demonstrated robust prognostic prediction capabilities in TCGA-LUAD, GSE31210, GSE50081, GSE3141, GSE26939, GSE30219, and GSE72094, with C-index values exceeding 0.6 across all seven datasets (Fig. 3a). Furthermore, the area under the curve (AUC) values for one-year overall survival, determined by ROC analysis, exceeded 0.7 in TCGA-LUAD (Fig. 3b), GSE31210 (Fig. 3c), GSE50081 (Fig. 3d), GSE3141 (Fig. 3e), GSE26939 (Fig. 3f), and GSE30219 (Fig. 3g), with GSE30219(Fig. 3h) even surpassing 0.8. The one-year overall survival AUC reached 0.7 when combining all GEO datasets, (Fig. 3i).
The c-index of SVM_score, compared to other commonly used clinical factors in the TCGA-LUAD dataset, is slightly lower than the Stage factor but higher than all other factors (Supplementary Fig. 2a). In the meta-GEO dataset, SVM_score outperforms all other factors, including Stage (Supplementary Fig. 2b). Similarly, in the GSE31210 dataset, SVM_score surpasses all other factors except for Stage (Supplementary Fig. 2c). In the GSE50081 dataset, SVM_score outperforms all other factors (Supplementary Fig. 2d). In the GSE26939 dataset, SVM_score exceeds all other indicators except age (Supplementary Fig. 2e). Finally, in the GSE30219 (Supplementary Fig. 2f) and GSE70294 (Supplementary Fig. 2g) datasets, SVM_score outperforms all other factors.
SVM_Score is associated with tumor proliferation and EMT propensity in TMU LUAD patients.
We further investigate the SVM_Score in the LUAD from our institute. In the 33 cases of TMU LUAD patients, the one-year overall survival AUC value for SVM_score reached 0.87 (Fig. 4a). Although the C-index of SVM_score may not have achieved statistical significance compared to other clinical factors, possibly due to the limited sample size, it demonstrated a stronger trend than all other clinical factors (Fig. 4b). Survival analysis results revealed that a lower SVM_score was associated with a worse prognosis, both in terms of overall survival (Fig. 4c) and progression-free survival (Fig. 4d). We also assessed whether there were differences in SVM_score among various clinical factors, and the results indicated an association between SVM_score and the lymph node metastasis status of lung adenocarcinoma (Fig. 4e).
Immunohistochemistry results further demonstrated that patients with lower SVM_score had more tumor cell expression Ki67, indicating a higher degree of malignancy consistent and statistically significant in the ten patients (Figs. 4f–g). Additionally, tumor tissues from patients with low SVM_score showed increased expression of Vimentin, signifying higher EMT activation pathway and an increased metastasis propensity(Fig. 4h). Again, this trend was consistent and statistically significant in all the patients (Fig. 4i).
Atomic force microscopy was used to detect mechanical stress within the tumor tissues. Tumor tissues from patients in the lower SVM_score group had higher internal mechanical stress (Figs. 4j-k).
The lower SVM_score is associated with lower immunogenicity.
We employed GSVA to calculate scores for 25 immune-related pathways (Supplementary Appendix 1) in TCGA-LUAD and displayed a heatmap illustrating lower SVM_Scores in tumor tissues associated with reduced immunogenicity (Fig. 5a). For patients with low SVM_Scores, the degree of activation of immune-related pathways is inversely correlated, indicating a lower level of immune pathway activation. We conducted separate analyses of the correlation between SVM_Scores and 27 immune cell types in TCGA and meta-GEO datasets. The results revealed statistically significant correlations between SVM scores and Activated B cell, effector memory CD8 + T cell, immature B cell, CD56dim natural killer cell, immature dendritic cell, macrophage, mast cell, and MDSC in both datasets (Fig. 5b).
Using GSVA, we calculated scores for 50 HALLMARK cancer-related pathways (Supplementary Appendix 2). Our findings demonstrated that patients with lower SVM scores exhibited increased activity in cancer pathways, including the EMT pathway (Fig. 5c). Furthermore, patients with lower SVM scores had higher Tumor Mutation Burden (TMB) (Fig. 5d), and SVM scores exhibited a negative correlation with TMB (Fig. 5e).
We employed ESTIMATE to calculate tumor stromal, immune, estimate, and tumor purity scores. We observed that SVM_Scores were independent of stromal scores (Fig. 5f). Patients with higher SVM_Scores exhibited higher immune scores (Fig. 5g) and estimate scores (Fig. 5h), as well as lower tumor purity (Fig. 5i).
The expression of COL5A1 of myofibroblasts influences the SVM_Score, while myofibroblasts are intricately associated with the microenvironment.
To further investigate SVM_Scores among different cell populations in the LUAD patients, we applied the GSE131907 dataset, which contains single-cell sequencing data from 11 LUAD samples, including 25,369 cells. The t-SNE dimensionality reduction plot displays the types of cells in single-cell sequencing data. (Fig. 6a). We computed an SVM_Score for each cell population and visualized it on the t-SNE plot (Fig. 6b). The results showed that the SVM_Score in cancer-associated fibroblasts (CAFs) significantly differed from other cell types (Figs. 6a-b). The genes contributing to the SVM_Score were also predominantly expressed in CAFs (Fig. 6c). We presented the expression distribution of the top four genes of SVM_Score, COL5A1, GPC3, OGN, and SLT3, on the overall single-cell dimensionality reduction plot, and they were primarily localized in CAFs (Fig. 6d). Therefore, we isolated CAFs and performed a separate t-SNE dimensionality reduction (Fig. 6e). When visualizing the distribution of SVM_Scores, the results indicated that all myofibroblasts have lower SVM_Scores (Fig. 6f). Moreover, myofibroblasts exhibited significantly lower SVM_Scores than all other CAFs, except for FB-like cells, which could not be statistically analyzed due to their limited numbers (Fig. 6g). Among the genes contributing to SVM_Scores, COL5A1 showed the highest expression in myofibroblasts (Fig. 6h). The expression distribution of COL5A1 in CAFs closely resembled the SVM_Score pattern (Fig. 6i). In both the TCGA-LUAD cohort (Supplementary Fig. 3a) and the TMU patients (Supplementary Fig. 3b), COL5A1 is highly expressed in tumor tissues. The findings above regarding COL5A1 suggest that COL5A1 may serve as a pivotal biomarker linking SVM_Scores and mechanical stress.
Following that, we analyzed cellular interactions in the tumor microenvironment, which revealed that myofibroblasts were the most interactive cell type with other cells (Supplementary Fig. 4a). Furthermore, myofibroblasts exhibited the highest interaction intensity with other cells (Supplementary Fig. 4b). We also identified that myofibroblasts primarily interact by secreting Macrophage Migration Inhibitory Factor (MIF) and binding to CD74 on the cell surface of other cells, along with CXCR4 or CD44 (Supplementary Fig. 4c).
Notably, myofibroblasts predominantly interacted with other cells through the MIF signaling pathway (Supplementary Fig. 4d) and the MK signaling pathway (Supplementary Fig. 4e). Moving forward, our focus will be directed toward myofibroblasts and COL5A1.
COL5A1 from Myofibroblasts increases tumor invasiveness and upregulates the EMT pathway of tumor cells
Furthermore, immunohistochemistry was applied to the ten FFPE samples of the aforementioned TMU LUAD patients. The results indicated that patients with low SVM_Score exhibit higher expression of alpha-SMA (Fig. 7a), which is used as a marker for activated myofibroblasts, and this trend is consistently observed in all ten patient tissues, with statistical significance (Fig. 7b). Similarly, the low SVM_Score patients demonstrate increased expression of COL5A1 (Fig. 7c), which is consistently significant across all ten patient tissues (Fig. 7d). Our study indicates that increased mechanical stress activates the EMT pathway(Fig. 1c, d), concurrently highlighting the reported association between the secretion of COL5A1 and tissue mechanical stress[33]. Therefore, multicolor immunofluorescence analysis was employed to investigate the relationship between COL5A1 and the tumor cell EMT pathway in the tissues of ten lung adenocarcinoma patients. The results revealed that low SVM_score tissues exhibited higher COL5A1 expression, and near the areas with elevated COL5A1 expression, Vimentin was also highly expressed (Fig. 7e). This trend is consistent across all ten patients and is statistically significant (Fig. 7f, g). Furthermore, in-depth statistical analysis of multicolor immunofluorescence shows that the number of COL5A1-positive cells is directly proportional to the number of vimentin-positive cells with a p-value less than 0.0001 (Fig. 7h).
Simultaneously, two myofibroblast strains, CAF1 and CAF2 (Supplementary Fig. 5a), were extracted from the LUAD tumor tissues. Small RNA interference downregulates COL5A1 in these two myofibroblast strains(Supplementary Fig. 5b,c). When co-cultured with tumor cells, it was observed that tumor cells located near CAFs with reduced COL5A1 expression also exhibited decreased Vimentin expression (Fig. 7i). We employed Transwell assays to investigate alterations in the invasive capacity of lung adenocarcinoma cells co-cultured with cancer-associated fibroblasts (CAFs). The co-cultivation of lung adenocarcinoma cells with CAFs led to an increase in the invasive ability of lung adenocarcinoma cells (Fig. 7j). However, when we knocked down COL5A1 in CAFs or treated them with Sorafenib, the invasive capacity of lung adenocarcinoma cells co-cultured with CAFs decreased compared with scrambled transfected CAFs (Fig. 7k).
Sorafenib attenuates the tumor-promoting effect of COL5A1 from myofibroblast
Previous studies have reported that Sorafenib decreases the expression of collagen and fibronectin genes, ultimately contributing to the reduction of tumor stroma stiffness and concurrently alleviating intertumoral stress[21, 22]. To investigate whether Sorafenib attenuates COL5A1 from myofibroblasts, we first analyzed drug sensitivity in TCGA-LUAD, indicating a direct proportionality between the IC-50 and SVM_Score for Sorafenib, suggesting its potential use in treating patients with a poorer prognosis characterized by low SVM scores (Fig. 8a). Subsequently, we treated two myofibroblast cell lines with Sorafenib, revealing its inhibitory effect on COL5A1 expression (Fig. 8b). The IC-50 values for Sorafenib in these two myofibroblast cell lines were 4.133 µM/L and 3.955 µM/L, respectively (Supplementary Fig. 5d). Therefore, we selected a nonlethal concentration of 200nM/L of Sorafenib for further treatment. CCK-8 experiments demonstrated that A549 and H1975 cells had IC-50 values of 2.76 µM/L and 1.852 µM/L for cisplatin, respectively (Fig. 8c). Posttreatment with 200nM/L Sorafenib, there was minimal change in their IC-50 values for cisplatin (Fig. 8d).
Subsequent co-culturing of A549 and H1975 cells with the two CAF cell lines resulted in a nearly twofold increase in their IC-50 values for cisplatin (Fig. 8e, f). However, after treatment with 200nM/L Sorafenib, the IC-50 values reverted to their original levels (Fig. 8e, f).
Concurrently, we established organoid models using patient-derived tissues from the same patients as the two CAF cell lines. When co-cultured with CAF cells, the organoids resisted cisplatin at 2uM/L (Fig. 8g). Conversely, co-culture with COL5A1-knockdown CAF cells rendered the organoids vulnerable to 2uM/L cisplatin (Fig. 8g). Similarly, treatment with 200nM/L Sorafenib did not restore resistance to cisplatin when co-cultured with CAF cells (Fig. 8g).