LUAD and lung squamous cell carcinoma (LUSC) gene expression patterns and clinical data were collected from The Cancer Genome Atlas (TCGA), wherein the LUAD dataset included 516 tumor samples and 59 paracancerous tissues, while the LUSC dataset included 501 tumor samples and 108 paracancerous tissues. The Gene Expression Omnibus (GEO) was searched for microdata on the mRNA expression. GSE72094 is a LUAD dataset composed of 442 samples, based on the GPL15048 platform, for external validation of the risk score. GSE74777 is a LUSC dataset based on the GPL17586P platform, containing 107 samples, which were used as an external validation for LUSC. Based on the method described by of Deng et al, we downloaded the lipid metabolism-related genes from the Molecular Signature Database (version 7.0) (24).

R “Limma” (R.4.1.0) was applied to identify differentially expressed genes (DEGs) between tumor tissue and normal tissue with a false discovery rate FDR <0.05 and |log2FC| ≥2 for the assessment of the involvement of significantly different genes in lipid metabolism. Functional enrichment analysis of GO and KEGG was performed using by DAVID (V.6.8 https://david.ncifcrf.gov/tools.jsp).

The PPI of DEGs was predicted using the STRING online platform (http://string-db.org/). Hub genes were calculated with reference to the “Degree” algorithm in the cytohub plugin in Cytoscape (v.3.8.2) and the visualization was calculated.

Based on the lasso algorithm, the prognostic risk scoring formula was constructed with the TCGA dataset as the training set, and the GSE72094 and GSE74777 datasets were used as the external validation datasets for LUAD and LUSC, respectively. Independent prognostic factors were identified by multivariate Cox regression analysis using the survival R package. The patients were assigned to the high- and low-risk groups in accordance the median risk score, and Kaplan-Meier analysis was performed to plot overall survival (OS) and ROC curves in order to assess the prognostic power of the risk score.

To explore the underlying molecular mechanisms of prognostic risk score and HPGDS, we applied the Gene Set Enrichment Analysis(GSEA)of the underlying molecular mechanisms (25, 26). C2, C5, and C6 were searched to identify the oncogenic signatures of KEGG pathways, biological processes, cellular components, molecular functions and dysregulation(P < 0.01,FDR < 0.05).

The 20-nt target DNA sequences preceding a 5,-NGG PAM sequence in the genomic HPGDS and ACSL1 were selected for generating a single guide RNA (sgRNA) for the SpCas9 targets by using the CRISPR design website (https://design.synthego.com/#/). Homo sapiens HPGDS and ACSL1 sgRNA were cloned into the lentiCRISPR v2 vector at the site of BsmBI with sgRNA forward CACCGGCCTCATCTTATGCAAGACT and reverse AAACAGTCTTGCATAAGATGAGGCC, forward CACCGGAAGAGTACGCACGTACTGT and reverse AAACACAGTACGTGCGTACTCTTCC. Definitively, to confirm the DNA sequence, the DNA was successfully cloned, after which the plasmid was sequenced and aligned (BGI, Chongqing, China).

A549 (human non-small cell lung cancer line) was sourced from the American Type Culture Collection (ATCC) and cultured in Dulbecco’s Modified Eagle Medium (DMEM) (Gibco, Carlsbad, CA, USA) supplemented with 10% fetal bovine serum (FBS, vol/vol, Biological Industries) and 1% penicillin/streptomycin (Invitrogen, Grand Island, NY, USA). The cell line was cultured at 37°C under a 5% CO₂ atmosphere. sg-HPGDS, sg-negative control (sg-Con), and sg-ACSL1 plasmids were transfected into A549 cells with the TurboFect Transfection Reagent (Thermo Scientific, Waltham, MA, USA) in accordance with the manufacturer’s instructions. Briefly, the cells with 80% confluence were transfected using the Turbofect reagent in the DMEM medium and selected by puromycin (Invitrogen, San Diego, CA, USA) treatment (1μg/mL) for 4 days after transfection for 48 h. The cells were cultured supplementary with a complete medium until all cells survive.

The sg-HPGDS transfection efficiency was determined from the genomic sequence after cell collection. Briefly, the genomic DNA was extracted with the TIANamp Genomic DNA Kit (TIANGEN BIOTECH, Beijing, China). Genomic DNA was amplified with polymerase chain reaction (PCR), using the following primer (sense): ATACACAAAGAAACTAAGAACTGG, and antisense: ATTCTGTGTGTTCTCTATGCACC. The PCR product was sequenced and aligned (BGI, Chongqing, China). The sg-ACSL1 transfection efficiency was determined by Western blotting. To obtain HPGDS and ACSL1 double knockout cell lines, based on the sg-HPGDS cell lines, the sg-ACSL1 plasmids were transfected into the cells with the Transfection Reagent according to the manufacturer’s instructions.

To confirm the effect of HPGDS, ACSL1 on A549 cells migration, a total of 1×10⁵ sg-Con, sg-HPGDS, sg-ACSL1, sg-HPGDS and sg-ACSL1 double knockout cell lines diluted in DMEM without FBS were plated into the Transwell chamber (Corning), which was plated in a 24 well plate containing 600 μL of the complete medium. After 24 h, the cells that had migrated to the bottom of the membrane were fixed with 4% paraformaldehyde for 30 min, after which the membrane was stained with 0.1% crystal violet (Sangon Biotech, Shanghai, China). Eight fields were captured randomly under a microscope (Wetzlar, Germany). Crystal violet was eluted with 600 μL of 33% acetic acid, and the OD value was measured at 570 nm.

Western blotting analysis was performed as described previously (27). Briefly, the cells were lysed with 1% SDS lysis buffer containing protease inhibitor cocktail and phosphatase inhibitor cocktail (Apexbio, Houston, USA). The protein concentration was determined by using a BCA protein assay reagent kit (Thermo Scientific). The protein content (18 μg) was separated by SDS-PAGE gels and transferred onto PVDF membranes (Millipore, Billerica, MA, USA). The PVDF membranes were blocked overnight by 5% fat free milk in TBST at 4°C, followed by incubation at 4°C for 12 h with primary antibodies for anti-β-actin (1:5000) (Sigma), anti-β-Tubulin (1:5000) (TransGen Biotech), anti-ACSL1 (1:1000) (Cell Signaling Technology), ACC (1:1000) (Cell Signaling Technology), HK2 (1:1000) (Cell Signaling Technology), ACAA1 (1:1000) (Cell Signaling Technology), E-cadherin (1:1000) (Bioworld), N-cadherin (1:1000) (Bioworld), Twist1 (1:1000) (Cell Signaling Technology), respectively. The membranes were washed with TBST and incubated for 1.5 h with the corresponding HRP-conjugated secondary antibodies. The bands were visualized with ECL reagents (Merck, Billerica, MA, USA). The western blots were analyzed with the Image Lab Software (BIO-RAD, USA), and the program included an application with protein gels; the image exposure time to was set to 16 s.

Based on the TCGA dataset, we analyzed the differential genes of lung adenocarcinoma, among which 1301 genes showed differential expression, including 840 upregulated genes and 461 downregulated genes ( Figure 1A ). Univariate COX prognostic analysis of all genes identified 2416 genes with prognostic significance ( Figure 1C ; Supplementary Table 1 ). Following the method of Zheng et al. (24), we obtained a total of 776 lipid metabolism-related genes from the molecular signature database (version 7.0), KEGG and Reactome databases. After the further intersection with 1301 differential genes and 2416 prognostic genes, we finally found got 12 prognostic differences in lung adenocarcinoma expressed genes, and they were also associated with lipid metabolism ( Figure 1B ).

Further enrichment analysis of 45 genes showed that most of their biological processes were related to redox processes, lipid metabolism processes and lipoxygenase pathways, and were mainly located in the cytosol, endoplasmic reticulum membrane and lipid granules. It is also involved in iron ion binding, heme binding and oxidoreductase activity ( Supplementary Figure 1A ). The most relevant metabolic pathways are arachidonic acid metabolism and peroxisome proliferators-activated receptor (PPAR) signaling pathway ( Supplementary Figure 1B ). more over, 45 lipid metabolism-related differential genes were used to draw the PPI network map ( Supplementary Figure 1C ), and the top ten hub genes, ( Supplementary Figure 1D ) namely PPARG, ALOX15, ALOX5, PTGIS, PTGES, HPGDS, PLA2G1B, ALOX15B, CYP27A1, and CAV1 were visualized and calculated using Cytoscape.

Twelve genes were further analyzed by LASSO-Cox regression analysis. A 9-gene signature was constructed based on the optional λ value. The risk score is defined as Risk score = (−0.0797) *CYP4B1 + (0.1527) * KLF4 + (−0.1243) * DPEP2 + (−0.0165) * PTGDS + (−0.0057) * CYP27A1 + (−0.1551) * ACSS3 + (−0.0444) * HSD17B13 + (−0.0213) * HPGDS + (0.0381) * FA2H ( Figures 2A, B ).The TCGA cohort samples were divided into low- and high-risk groups according to the median cut-off value of the risk score ( Figure 2C ), and the Kaplan-Meier analysis showed that this risk grouping was effective in distinguishing between the good and poor prognosis groups. In other words, the OS of the low-score group in LUAD was statistically better than that of the high-score group (P<0.05) ( Figure 2D ). Time-dependent receiver operating characteristic (ROC) curves constructed to test the accuracy of the prognostic model, and the area under the curve (AUC) was 0.696 for 1 year, 0.675 for 3 years, and 0.646 for 5 years ( Figure 2E ). As shown in Supplementary Figure 2 , these 12 genes have independent survival predictive power.

Patients with LUAD from the GEO dataset were involved in the validation of the risk model. Based on prognostic information from 442 patients enrolled in GSE72094, patients in the GEO cohort were divided into low- and high-risk groups based on the median risk score of the TCGA cohort, similar to the TCGA cohort. Kaplan-Meier analysis showed that overall survival was higher in the low-risk group than in the high-risk group (P < 0.01), regardless of whether the optimal cutoff ( Figure 3A ), median ( Figure 3C ), or quartile ( Figure 3D ). Due to incomplete 5-year survival data for patients in GSE72094, only an AUC of 0.0.67 at 1 year and 0.6 at 3 years was validated ( Figure 3B ).

To elucidate the underlying molecular mechanisms of risk scoring, we performed a GSEA comparison between high-risk and low-risk groups in the 442 patients in the data GSE72094. In the low-risk group, four oncological features of base excision repair, cell cycle, mismatch repair, and p53 signaling pathway, and a lipid metabolism feature of glyoxylate and dicarboxylate metabolism were enriched. In the high-risk group, the enriched KEGG pathway mainly focused on various cardiac disease associations (including dilated cardiomyopathy, virtual myocarditis, and cardiac muscle contraction, etc.). However, no significant enrichment was found in the oncological features ( Supplementary Figure 3 ).

As previously described, we performed differential analysis on TCGA data of lung squamous cell carcinoma (TCGA-LUSC) ( Figure 4A ; Supplementary Figure 4A ) (LogFC = 2, FDR <0.05), and extracted genes related to lipid metabolism. A total of 542 upregulated genes and 830 downregulated genes was found in lung squamous cell carcinoma, including 12 and 35 genes related to lipid metabolism, respectively ( Figures 4B, C ). To further select the lipid metabolism-related genes with prognostic significance and to avoid the unreproducibility of a single dataset, we performed a univariate COX regression analysis on the above 47 important genes in the LUSC dataset GSE74777. The results showed that HPGD, B4GALNT1, HPGDS, LPL, SGMS2, SLC44A4, and MFSD2A were prognostically significant lipid metabolism-related genes ( Figure 4D ). Moreover, we also performed routine bioinformatics analysis of GO, KEGG and PPI for these 47 important genes (we consider count<20% as invalid enrichment). In terms of GO, they were enriched in two biological processes, oxidation-reduction process and lipid metabolic process, and were mainly located in the cytosol, extracellular exosome and endoplasmic reticulum membrane. However, related molecular functions were not enriched. Additionally, the results showed that the main pathway enrichments were metabolic pathways and arachidonic acid metabolism ( Supplementary Figure 4B , Supplementary Table 2 ). Subsequently, we drew the PPI map of these 47 genes ( Supplementary Figure 4C ), and further analyzed the Hub genes among them using Cytoscape (v 3.8.0). The results showed that the top ten genes were PPARG, HPGDS, ALOX5, FABP4, CYP27A1, ALOX15B, PLA2G1B, LPL, CYP4F3 and PTGIS ( Figure 4E ). Notably, HPGDS simultaneously appeared in the HUB gene of the previous TCGA-LUAD dataset, the lasso regression prognostic model, the HUB gene of the TCGA-LUSC dataset, and the prognostic gene of LUSC ( Figure 4F ).

To explore the underlying molecular mechanism of HPGDS, we performed a GSEA comparison between groups with different HPGDS expression. In terms of the KEGG pathway, the high expression group was enriched in lysosome, whereas proteasome was enriched in the low expression group ( Figure 5A ). Moreover, 15 oncological signatures including HOXA9, STK33, MTOR, RPS14, PGF, CSR, and YAP1 were enriched in the high expression group; however, no significantly enriched oncological signatures were found in the low expression group ( Figure 5B ). GO terms focused on ribosomal and mitochondrial function correlations. These enriched KEGG pathways and GO terms revealed molecular alterations in the HPGDS high expression group and were closely associated with metabolism. The results of GSEA are shown in Supplementary Table 3 .

Using CRISPR gene editing technology, we obtained the HPGDS knockout A549 cell line (Appendix 1). We examined differential expression of lipid metabolism-related proteins by western blotting. Compared with non-knockout cells, the expression of ACSL1, ACC, ACAA1, and HK2 was significantly increased after HPGDS knockout, which indicated that HPGDS knockout promoted lipid biosynthesis, and HPGDS can be associated with lipid metabolism ( Figure 6A ).

As shown in Figure 6B , the knockdown of HPGDS resulted in a decrease in E-cadherin and an increase in N-cadherin and TWIST1. Moreover, the migratory ability of A549 cells was examined using a transwell assay. A549 cells in the HPGDS knockout group were much higher than those in the control group ( Figure 6C ). These results INDICATED the role of HPGDS in the malignant progression of LUAD.

Rescue assays indicated that ACSL1 knockdown reversed the pro-migration effects of HPGDS knockdown. The knockdown of HPGDS promoted migration response by upregulating the expression of the lipid metabolism key enzymes ACSL1 and ACC ( Figure 7 ).