Integration of transcriptomic analysis and multiple machine learning approaches identifies NAFLD progression-specific hub genes to reveal distinct genomic patterns and actionable targets

Using R package “limma”, significantly upregulated genes were identified in NASH compared to NAFL, and then submitted to GO analysis for pathway exploration. R package “GOSemSim” was used to measure the similarity among all the biological processes (BP), and R package “ggtree” was used to visualize the clusters, and similar BPs were clustered into a mutual branch (Fig. 1a). According to the significance, we extracted the top five altered GOBPs and displayed them in a Circos diagram (Fig. 1b). We observed that alterations in two biological features are mainly responsible for the NAFLD progression: extracellular matrix (ECM) organization and cell cycle process.

Two filtering strategies were combined to identify NAFLD progression-specific genes. Firstly, WGCNA was performed with the transcriptome profiling data of 98 NAFLD samples and their sample category (NAFL or NASH) to construct a scale-free co-expression network. A total of 36 gene modules were generated with a power of 6 as the optimal soft threshold (Supplementary Fig. 1). Among these modules, the brown module exhibited a highest correlation with sample category (|r| = 0.61, p = 6e-11) and was considered as “gene module of NAFLD progression” (Fig. 1c). Secondly, the “limma” algorithm was used to identify a total of 378 DEGs between NAFL and NASH samples with a filtering threshold of q value less than 0.01 and fold change (FC) > 2 or < 0.5 (Fig. 1d). Finally, 182 overlapping genes in the intersection of “gene module of NAFLD progression” and “DEGs between NAFL and NASH” were considered as “NAFLD progression-specific genes” (Fig. 1e) and the detailed information is shown in supplementary Table 2. A PPI network was generated to depict the functional and physical linkages among these DEGs, and we observed that the hub (the red circle) in the network is mainly composed of collagen family members (Fig. 1f).

To explore potential targets and applicable drugs for high-risk NAFLD patients, the above-mentioned 182 NAFLD progression-specific genes were submitted to CMap for further investigation. Top 10 compounds with the highest predictive scores and corresponding 7 mode-of-actions (MoAs) were displayed in a dot diagram (Fig. 2g). The 7 MoAs were annotated with “Angiotensin receptor antagonist”, “Aromatic hydrocarbon derivative”, “Aurora kinase inhibitor”, “EGFR inhibitor”, “HDAC inhibitor”, “NFκB pathway inhibitor”, and “Sphingosine kinase inhibitor”. In particular, four compounds named HDAC3-selective, entinostat, mocetinostat, and Merck60 share a common MoA of HDAC inhibitor (HDACi), which indicates HDACi might be a potentially applicable drug for advanced NAFLD patients.

The identified 182 genes concerning NAFLD progression were further analyzed using GO method, and a Circos illustrated that they were mainly enriched in five GOBPs annotated with “ECM organization”, “Blood vessel development”, “Wnt signaling pathway”, “Cell adhesion”, and “Cell morphogenesis” (Fig. 2a). We observed that the five GOBPs include the main biological alterations occurred in NASH, and this result further indicated that the 182 genes could commendably represent the pathological development of NAFLD. Based on the expression profile of 182 NAFLD progression genes and using the NMF algorithm, the 51 NAFLs in the training cohort were divided into two subclusters (C1 = 21, C2 = 30; Fig. 2b). Using ssGSEA algorithm, we found that the quantified scores of “Inflammatory response”, “ECM organization”, and “Cell-cell adhesion” were significantly and progressively elevated from C2 to C1 to NASH (Fig. 2c - e; *** p < 0.001). In particular, positive correlations between “Inflammatory response” and “ECM organization” or “Cell-cell adhesion” were observed in NAFL-C1, C2 and NASH samples (Fig. 2f & g). Furthermore, among all correlations, NASH exhibited the weakest correlation between “Inflammatory response” and “ECM organization” (Fig. 2f).

Considering inflammation as a driver in NAFLD progression, we selected a list of widely acknowledged inflammatory factors (LPAR1, PTPRE, CCR2, CCL20, CLEC5A, CXCL6, ITGB8, PDPN, and GPC3) and showed their expression profiles in NAFL-C1, C2 and NASH (Fig. 2h; *** p < 0.001). Using one-way ANOVA test, we found that all these inflammatory factors were significantly decreased in the NAFL-C2 group compared to either NAFL-C1 or NASH. In addition, we evaluated the abundance of fibrosis-related cell populations including fibroblast and Th1/2 cells in the three groups using xCell algorithm. The fibroblasts abundance was significantly downregulated in NAFL-C2. Th1 cell infiltration was significantly upregulated in NAFL-C2, while no significant difference of Th2 cell infiltration was observed among the three groups (Fig. 2i; * p < 0.05, *** p < 0.001, # not significant). Furthermore, the NAFL-C2 group exhibited the lowest stroma score, while no significant difference of stroma score was observed between NAFL-C1 and NASH (Fig. 2j; ** p < 0.01, *** p < 0.001, # not significant).

To assess the risk stratification of NAFLD in a quantitative method, we combined different machine learning (ML) approaches to screen for robust biomarkers. The training method of the LOOCV framework was introduced in Fig. 3a, and the iteration number in our study is N = 98 (the sample number in the training cohort). The identified 182 genes regarding NAFLD progression were trained in the RF and SVM algorithms with feature selection of recursive feature elimination (RFE) respectively, and 8 overlapping genes (COL1A1, COL1A2, COL4A1, COL4A2, COL5A1, DTNA, THBS1, and UBD) remained in the outputs of the two ML algorithms. Finally, Lasso logistic regression (LR) analysis was applied on the 8 genes, and only 4 genes (DTNA, COL4A2, UBD, and COL1A2) retained their coefficients. Among them, COL1A2 exhibited the highest coefficient (Fig. 3b). PCA analysis showed a clear separation of NAFL and NASH samples with the expression matrix of the four genes (Fig. 3c). Furthermore, dotplots showed that all four genes were significantly and stepwisely elevated from NAFL-C2 to C1 to NASH samples (Fig. 3d). In the training cohort, the discriminative score was calculated for each sample according to the formula stated in the methods section, and ROC analysis indicated that the score could discriminate NASH from NAFL accurately (Fig. 3e).

A dataset named GSE163211, which contains 88 steatosis samples, 72 NASH with F0, and 82 NASH with F1-4, was used as an external testing cohort. The GSE163211 dataset was produced from the GPL29503 platform of NanoString Technologies and contains RNA-seq data of 800 custom genes. Among the four genes (DTNA, COL4A2, UBD, and COL1A2), only COL1A2 and COL4A2 were detected in the platform. COL1A2 and COL4A2 were significantly upregulated in NASH with fibrosis, and the 2-gene discriminative score was also significantly higher than steatosis and NASH samples without fibrosis (Fig. 3f). In the total of 154 NASH samples, the ROC analysis demonstrated that the 2-gene discriminative score could identify advanced fibrotic samples with a favorable performance (AUC = 0.724, 95% CI = 0.644–0.804; Fig. 3g).

In addition, another dataset named GSE135251 containing 10 normal tissues, 51 NAFLs, and 155 NASHs with F0-4 was used as the second external testing cohort. COL1A2, COL4A2 and DTNA were detected in the platform, and we calculated the “3-gene score” for each sample, and compared the score among normal samples, NAFL, and NASH with different fibrosis levels (F0-4). We observed that the “3-gene score” was significantly and stepwisely elevated from normal to NAFL to NASH with advanced fibrosis levels (p < 0.001; Fig. 3h). Furthermore, we evaluated its discriminative capacity for NAFLD and fibrosis levels using ROC analysis, and the “3-gene score” exhibited favorable performances in discriminating advanced stages (AUC = 0.737, 95% CI = 0.665–0.810; Fig. 3i) and advanced fibrosis levels (AUC = 0.729, 95% CI = 0.650–0.808; Fig. 3j).

To evaluate the biological role of the four discriminative genes in NAFLD and malignant progression, we investigated the expression profile of the four genes in different datasets which consist of normal liver tissues, NASH, and HCC samples. In the GSE164760 microarray dataset, all the four genes were significantly upregulated in HCC samples derived from NASH when compared to healthy liver tissues and NASH samples (Fig. 4a; *** p < 0.001). In the combination of GTEx and TCGA-HCC RNA-seq database, all the four genes were also significantly upregulated in HCC samples compared to donated normal tissues and adjacent normal tissues (Fig. 4b; *** p < 0.001).

In addition, three scRNA-seq datasets including GSE125449, GSE146409 and GSE166635 were used to reveal the components of HCC tumor microenvironment (TME) and the expression characteristics of hub genes. UMAP dimensionality reduction was used to show the distribution and dissimilarity of the cell types involved in HCC TME (Fig. 4c, f, and i). Considering COL1A2 possesses the highest discriminative coefficient among the four genes, we assessed the expression characteristics of COL1A2 (Fig. 4d, g, and j) in different cell types, and all the three violin plots showed that COL1A2 is specifically expressed in fibroblasts, almost not expressed in other cell types (Fig. 4e, h, and k). These findings demonstrated that the risk-stratification gene signature was closely relevant to the progression of NAFLD and HCC, and COL1A2 might play a specific role in fibroblast activation and fibrosis severity.

Using the WES data of TCGA-HCC and NMF algorithm, we attempted to decipher the mutational patterns of HCC and explore the relationship between COL1A2 and mutation patterns in HCC. The optimal k factorization of 5 was selected (Supplementary Fig. 2), and five mutational signatures were identified and matched in the COSMIC database (Fig. 5a). The three major mutational signatures were annotated with “Defective DNA mismatch repair”, “Aflatoxin exposure”, and “Aristolochic acid exposure”, and the abundance of each mutational signature in TCGA-HCC was shown in a pie chart (Fig. 5b). We reviewed the medical records of the TCGA-HCC cohort and extracted 184 HCC samples with history of alcohol consumption, hepatitis or NAFLD, and the distribution of each mutational signature was depicted in a stacked barplot (Fig. 5c). Using chi-square test, we found that NAFLD-HCC is characterized with high COL1A2 expression (Fig. 5d). With quantiles of COL1A2 mRNA expression, 46 HCC samples were assigned to the COL1A2-lowest and -highest group, respectively. Oncoplots of the two groups demonstrated that CTNNB1 acts as the most frequently mutated gene in the COL1A2-low cohort, with the mutation frequency up to 48% (left panel, Fig. 5e). In contrast, CTNNB1 is rarely mutated in the COL1A2-high cohort (right panel, Fig. 5e). In the integrated analyses of TCGA-HCC, MSK-HCC and INSERM-HCC cohorts, the gene-pair of TP53 and CTNNB1 is shown significantly mutually exclusive (Fig. 5f). Furthermore, COL1A2 mRNA expression is significantly elevated in CTNNB1-wild type HCC samples compared to CTNNB1-mutated ones (Fig. 5g). Among nine representative oncogenic pathways, the WNT signaling pathway is the most frequently affected one in the COL1A2-low cohort (Fig. 5h). GOBP analysis was used to further explore the altered pathways in CTNNB1-WT/COL1A2-high samples. A functional network showed the five most important pathways in CTNNB1-WT/COL1A2-high HCC samples were termed “vasculature development”, “chemotaxis”, “ECM organization”, “positive regulation of locomotion”, and “positive regulation of cell adhesion” (Fig. 5i). These evidences demonstrated that the CTNNB1/COL1A2 axis might play a role in the fibrosis and inflammation severity during NAFLD-HCC progression.

The infiltrating immune and stromal cells were estimated by TIMER, Cibersort, quanTIseq and MCP-counter, and the differences of distribution were also investigated in distinct CTNNB1/COL1A2 groups. As shown in a comprehensive heatmap, most of the immune and stromal cells are enriched in the CTNNB1-WT/COL1A2-high samples, regardless of the TMB level (Fig. 6a). Furthermore, the xCell algorithm inferred the absolute infiltration of a total of 36 cell types involved in the TME, and the chord diagram showed that the infiltrating abundance is significantly higher in the CTNNB1-WT/COL1A2-high and CTNNB1-Mut/COL1A2-high samples than the others (Fig. 2b). Infiltrating scores of CAFs were inferred using EPIC, MCP-counter, and xCell, and we observed that CAFs are significantly enriched in the CTNNB1-WT/COL1A2-high samples, suggesting the distinct fibrosis severity resulted from different classification of CTNNB1-WT and high COL1A2 expression (Fig. 6c - e). On the other hand, we evaluated the levels of representative immune checkpoints including CD274, PDCD1 and TIGIT, and significant differences were observed among different groups, indicating heterogenous tumor immunogenicity and distinct potential response to immunotherapy (Fig. 6f - h). Moreover, the ESTIMATE algorithm was applied to infer the immune infiltration and tumor purity of each sample, and we observed that the CTNNB1-WT/COL1A2-high and CTNNB1-Mut/COL1A2-high groups are labelled with high immune infiltration (Fig. 6i), and a significantly negative correlation (r = -0.916, p < 0.001) between the immune score and tumor purity across the four groups was shown in Fig. 6j. Similarly, the inflammatory response activity exhibits distinct distributions among the four groups (Fig. 6k), and a significantly positive correlation (r = 0.839, p < 0.001; Fig. 6l) with immune score was also observed across the four groups.

To investigate the connection between the hub gene COL1A2 and malignant features in pan-cancer, we quantified the ability of ten cancer hallmarks including EMT, angiogenesis, apoptosis, inflammation, hypoxia, glycolysis, cell cycle progression (CCP), senescence, DNA repair and oxidative phosphorylation using ssGSEA algorithm, and calculated the correlations between COL1A2 expression and the ten hallmarks of cancer (Fig. 7a). Among all the cancer hallmarks, COL1A2 mostly correlated with EMT and angiogenesis in 32 malignant solid cancers (EMT: r = 0.86, p < 2.2e-16; Angiogenesis: r = 0.73, p < 2.2e-16) in the overall TCGA pan-cancer cohort (Fig. 7b & c), or in individual tumor types (Fig. 7d & e). Considering EMT and angiogenesis are two critical biological features which contribute to the initial development of cancer, we reasonably speculated that COL1A2 serves a significant role in malignant progression in pan-cancer.