The RNA-seq data consisting of copy number variation (CNV) data, messenger RNAs (mRNAs), long non-coding RNAs (lncRNAs), somatic mutation data (MuTect2), miRNAs, as well as clinicopathological information of HCC patients were obtained from The Cancer Genome Atlas (TCGA) database (LIHC dataset). The data were collected in December 2021 and preprocessed using the “TCGAbiolinks” package [20]. Additionally, DNA methylation data were acquired from UCSC Xena ¹ and underwent preprocessing procedures consistent with previously documented methodologies [21]. The median beta value corresponding to the CDKL3 gene probes was mapped to its promoter, including 1stExon, 5′ untranslated region, transcription start site (TSS)200, and TSS1500 [22]. The mRNA and lncRNA data, originally presented in fragments per kilobase per million reads (FPKM), underwent a conversion to transcripts per kilobase million (TPM) values, followed by a subsequent log₂ conversion. The mature miRNA (mirbase version 21) data, initially in row count values, were transmuted into reads per million mapped reads (RPM) and subjected to log2 conversion. Inclusion criteria for this study were patients diagnosed with primary solid HCC who were 18 years or older. Exclusion criteria included a history of preoperative adjuvant therapy, lack of survival time data, the recorded survival time <30 days, or multiple samples from a single patient [23]. A total of 330 HCC patients and 47 controls were ultimately enrolled. The clinicopathological information of the HCC patients can be found in Supplementary Table S1.

The investigation into pan-cancer expression profiles was facilitated through the TIMER 2.0 tool ² [24]. The expression difference of CDKL3 between HCC patients and controls in the LIHC dataset was analyzed. Kaplan-Meier (KM) overall survival (OS) analysis with the log-rank test was conducted by dividing HCC patients into the high-CDKL3 (hCDKL3) group and the low-CDKL3 (lCDKL3) group. This division was based on the optimal cut-off expression values determined using the maximally selected rank statistics algorithm. The univariate Cox regression model was applied for survival analysis, and the expression difference of CDKL3 between the hCDKL3 and lCDKL3 groups was investigated. Multivariate Cox analysis was implemented to assess the independence of CDKL3 for OS prediction. Furthermore, a nomogram incorporating other independent parameters for OS prediction was established. Internal validation was executed by computing the adjusted Harrell’s concordance index (C-index) utilizing the bootstrapping approach with 1,000 resamples [23]. The performance of the independent parameters of OS and the nomogram were evaluated using the areas under the curve (AUC) values, decision curve analysis (DCA), and calibration curves [25–27].

To investigate the potential difference in biological function between the two groups, a gene set enrichment analysis (GSEA) was undertaken utilizing the Kyoto Encyclopedia of Genes and Genomes (KEGG) and HALLMARK gene sets derived from the Molecular Signatures Database (MSigDB ³ ) [28]. To establish significance, we conducted a screening process that involved identifying p values < 0.05 and a false discovery rate >0.25. Furthermore, the functional annotation of genes in the KEGG pathway was performed using the online tool KOBAS-i ⁴ [29]. Single-sample GSEA (ssGSEA) was performed to explore the tumor microenvironment (TME) and metabolic characteristics using 29 functional gene expression signatures (Fges), which were recently proposed by Bagaev et al. [30]. Moreover, gene set variation analysis was employed to analyze metabolism-associated signatures, as detailed in established methods [31]. Immune checkpoint gene (ICG) expression discrepancies between the two groups were also analyzed. Somatic copy number alternation (SCNA) burden was assessed, representing the gain or depletion of total gene count at the focal and arm levels. This analysis was carried out using the GISTIC 2.0 module ⁵ [32, 33]. The calculation of tumor mutation burden (TMB) for every patient adhered to the methodology outlined in prior descriptions [34]. The mRNA expression-based stemness index (mRNAsi) for HCC patients was obtained from a published reference [35]. Additionally, the percent splice-in (PSI) value of CDKL3 was retrieved from the SpliceSeq database using default parameters ⁶ [36]. For the missing PSI values, the average value of each event was used to fill them up. Only events with a mean PSI value greater than 0.1 were retained for further analysis in this study. The details for analyzing competitive endogenous RNA regulatory networks and estimating the benefits of immunotherapy and chemotherapy can be found in the Supplementary Material.

The human HCC cell lines SMMC-7721 and HepG2 were acquired from GeneChem Corporation (Shanghai, China) and the Cell Bank of Type Culture Collection of the Chinese Academy of Sciences (Shanghai, China), respectively. Subsequently, RPMI-1640 (Gibco, Waltham, MA, United States) or Dulbecco’s modified Eagle’s medium (DMEM) added with 10% heat-inactivated fetal calf serum (FBS, Gibco, Thermo Fisher Scientific, Inc.) was used to culture these cell lines at 37°C. Concentrated and purified lentiviral particles expressing a short hairpin RNA (shRNA) were purchased from GeneChem Corporation (Shanghai, China). To perform transfection, the control lentivirus (shCtrl) or the target lentivirus (shCDKL3) was added to tumor cells (7 × 10⁴ cells/well) in a six-well plate (Corning Incorporated, United States). After 72 h, the expression level of CDKL3 was assessed using qPCR and Western blot assays. More detailed information regarding the qPCR and Western blot assays can be found in the Supplementary Material.

The HCC cells, following infection, were placed in a 96-well plate (Corning Incorporated, United States) and subsequently cultured for a span of 4 days. To assess cellular proliferative capacity, we performed a daily cell counting kit-8 (CCK-8) assay (Beyotime’s kit from Shanghai, China) as per package guidelines [37]. A colony formation experiment involved plating 2,000 infected cells per well over a 6-well plate. After 2 weeks of incubation, 4% paraformaldehyde solution (Beyotime, Shanghai, China) was employed to fix the tumor cells for 30 min. Subsequently, crystal violet solution (Beyotime, Shanghai, China) at 0.1% was utilized to stain the cells for 20 min. Under a light microscopy, the number of viable colonies (>50 cells/colony) was determined. To perform the apoptosis assay, we plated the infected cells on prepositioned slides in a 12-well plate (Corning Incorporated, United States). After an overnight incubation, the cells were immobilized with a 4% paraformaldehyde solution for 30 min. The nuclei were stained using DAPI (Beyotime, Shanghai, China). The detection of cell apoptotic rate was carried out utilizing a TdT-mediated dUTP nick-end labeling (TUNEL) kit (Beyotime, Shanghai, China), as described previously [38].

For the tumor xenograft assay, six BALB/C nude mice (Henan Scrobes Company, China), aged 5 weeks, were used in each group (n = 6/group). At the Center for Laboratory Animal Research in Anhui Medical University, the mice were housed in a pathogen-free environment. All experiments with animals were done in compliance with institutional guidelines and with appropriate oversight. The Experimental Animal Ethics Committee of Anhui Medical University approved this study (No. 20200695). Subcutaneous transplantation of SMMC-7721 cells stably expressing shControl or shCDKL3 (3×10⁶ cells) was performed on the upper right back of the mice. The size of the tumor (computed as: [(π/6) × (length) × (width) [2]]) and its weight were measured every 4 days over a duration of 4 weeks until the humane euthanasia of mice.

Categorical data were compared utilizing the chi-square or Fisher’s exact test, whereas the Wilcoxon rank-sum test or Kruskal‒Wallis test was employed for the comparison of continuous variables. The correlation was tested using a rank correlation method, either Pearson’s or Spearman’s. R 4.0.1 (The R Foundation for Statistical Computing, Vienna, Austria) and GraphPad Prism 8.0.2 (GraphPad Software, Inc.) were used to perform all statistical tests. Statistical significance was assumed when the p-value was <0.05 (two-tailed) unless otherwise specified.

Analysis of data from TIMER 2.0 revealed that CDKL3 was significantly dysregulated in various solid tumors (Figure 1A). In the TCGA dataset, both unpaired and paired analyses demonstrated markedly upregulated levels of CDKL3 in HCC patients compared to controls (Figures 1B, C). Notably, CDKL3 displayed excellent diagnostic performance for HCC, with an impressive AUC value of 0.924 (Figure 2A). KM analysis revealed that the OS of the hCDKL3 group was lower than the lCDKL3 group (Figures 2B, C). Consistent results were observed for progression-free interval (PFI), disease-free interval (DFI), and disease-specific survival (DSS) (Supplementary Figure S1). Figure 2D illustrates the gene expression differences between the hCDKL3 and lCDKL3 groups. Univariate Cox analysis highlighted the association of CDKL3 level with OS, DSS, DFI, and PFI (Figure 2E). Additionally, we explored how CDKL3 expression related to clinicopathological variables (like tumor grade and family history of cancer), and found a substantial link between the two (Supplementary Figure S2).

CDKL3 was shown to independently function as a predictor of OS in a multivariate Cox regression study (Figure 3A). To aid in clinical decision-making, a nomogram was devised incorporating CDKL3 level and other independent clinicopathological variables (Figure 3B). The predicted and observed survival rates were highly consistent, as shown by calibration curve graphs (Supplementary Figure S3A). Receiver operating characteristic (ROC) curve analysis illustrated that the nomogram outperformed individual factors alone in estimating 1-, 3-, and 5-year OS, as indicated by higher AUC values (0.78, 0.75, and 0.78, respectively) (Supplementary Figure S3B). Furthermore, DCA demonstrated that the nomogram yielded greater net benefits compared to individual factors (Supplementary Figure S3C). Internal validation yielded an adjusted C-index of 0.713, indicating the reliability of the nomogram in prognostic prediction.

The Supplementary Figures S4, S5 presented the mutation-related findings of CDKL3. In terms of the mRNAsi, the hCDKL3 group exhibited a score superior to that of the lCDKL3 cohort (Figure 4A). Furthermore, the CDKL3 expression was positively linked to the mRNAsi score (Figure 4B). Figure 4C demonstrated three alternative splicing events for CDKL3. The hCDKL3 group displayed a higher PSI value for CDKL3−73367−AT and a lower PSI value for CDKL3−73366−AT when compared to the lCDKL3 group. Additionally, the expression of CDKL3 had a positive correlation with the PSI value of CDKL3−73367−AT and a negative correlation with the PSI value of CDKL3−73366−AT (Figures 4D, E). Interestingly, Figure 4F revealed that the hCDKL3 group was significantly enriched in hallmark pathways associated with tumorigenesis, especially the G2/M checkpoint, E2F targets, DNA repair, MYC targets V1 and V2, the P53 pathway, and PI3K−AKT−MTOR signaling. These findings strongly suggest that the hCDKL3 and lCDKL3 groups exhibit distinct genetic features.

Figure 5A demonstrates that the lCDKL3 group is characterized by a significantly decreased enrichment score of the proliferation rate signature and a notably elevated enrichment score of Fges, which is related to the cluster of antitumor immune infiltrates. This enrichment included antitumor cytokines, B cells, NK cells, effector cells, the Th1 signature, and T-cell traffic signatures. Correlation analysis confirmed a positive link between CDKL3 expression and the enrichment score of the proliferation rate signature, and a negative correlation with the enrichment score of antitumor cytokines, B cells, NK cells, effector cells, the Th1 signature, and T-cell traffic signatures (Supplementary Figure S6). Moreover, an investigation into the difference in T-cell exhaustion markers between the lCDKL3 and hCDKL3 groups highlighted that hCDKL3 was associated with increased expression of markers such as BTNL2, CD276, CD40, HAVCR2, LAIR1, LGALS9, NRP1, TNFRSF4, TNFSF9, and VTCN1 (Figure 5B). Additionally, the results of GSEA of KEGG pathways indicated that both the lCDKL3 and hCDKL3 groups were significantly enriched in multiple metabolic pathways (Figure 6). Further exploration of the enrichment scores of metabolism-associated pathways in the four categories (amino acids, carbohydrates, lipids, and others) between the lCDKL3 and hCDKL3 groups revealed significant differences. Notably, the hCDKL3 group exhibited significantly decreased pathway enrichment scores for lipid and other pathways. Similar patterns were observed in most amino acid and carbohydrate pathways, except for the GLUCOSE, PURINE, PYRIMIDINE, and SELENOAMINO ACID metabolism pathways (Supplementary Figures S7A–D). Additionally, the CDKL3 levels were inversely linked to the enrichment score of most differentially enriched pathways (Supplementary Figure S7E). Altogether, these findings suggest that the lCDKL3 and hCDKL3 groups possess distinct TME and metabolism characteristics.

We conducted a thorough screening of 50 protein-coding genes that showed a strong positive correlation with CDKL3, as well as 50 genes that displayed a negative correlation (Supplementary Table S2). Utilizing KOBAS-i, we executed functional annotation of these genes in the KEGG pathway, and observed significant enrichment in various clusters and functions (Supplementary Figure S8). Specifically, these genes were found to be closely associated with metabolism pathways, the TGF-beta signalling pathway, and the PPAR signalling pathway, all of which were considered to impart a substantial role in the growth of HCC (adjusted p-value <0.05). For further analysis, we turned to the TISCH database, where we conducted a single-cell analysis to assess the expression of CDKL3 in different cell types. However, our findings did not reveal any significant differences in CDKL3 expression among these cell types (Supplementary Figure S9). Additional details regarding these results can be found in the Supplementary Material. The results for analyzing competitive endogenous RNA regulatory networks and estimating the benefits of immunotherapy and chemotherapy can be found in the Supplementary Figures S10, S11.

To verify the effectiveness of CDKL3 knockdown, we performed qPCR and western blot assays, which confirmed the stable knockdown of CDKL3 in SMMC-7221 and HepG2 cells (Figures 7A, B). Through the CCK-8 assay, we were able to confirm that the knockdown of CDKL3 significantly reduced the proliferation of SMMC-7221 and HepG2 cells (Figure 7C). Furthermore, colony forming ability was also greatly inhibited in these cells as a result of CDKL3 knockdown (Figure 7D). Notably, the apoptosis rate in SMMC-7221 and HepG2 cells was markedly increased following CDKL3 knockdown, as demonstrated in Figure 8. Additionally, the migration of cells was significantly reduced upon CDKL3 knockdown (Supplementary Figure S12). Moreover, in the shCDKL3 group, the tumors formed by SMMC-7721 cells exhibited smaller volumes compared to those in the shCtrl group (Figure 9). These findings suggest that CDKL3 contributes to the tumorigenesis and advancement of HCC.