The MagBeads Total RNA Extraction Kit (MAJORIVD) was used to isolate total RNA. Shanghai Biotechnology Corp. (Shanghai, China) was responsible for the library construction and sequencing. VAHTS Universal V6 RNA-seq Library Prep Kit for Illumina (Vazyme, Nanjing, China) was used for RNA library preparation and NovaSeq 6000 equipment (Illumina, Carlsbad, CA) for sequencing. Sequencing data was further standardized to both fragments per kilobase of transcripts per million mapped reads and read count values.

Detailed primary antibodies for immunohistochemistry (IHC) were described in the previous study [17]. IHC was performed as described before [20]. Slides were scanned with PANNORAMIC 250 Flash III DX (3DHISTECH Ltd., Budapest, Hungary). The densities of targeted cells were calculated as the mean number of cells/mm². Three researchers were asked to execute the quantification on six randomized fields.

For surface staining, after blocking Fc receptors, single-cell suspensions and white blood cells were stained with fluorescently labeled membrane marker antibodies for 30 minutes at 4°C. For intracellular staining, proteins were stained with antibodies in Intracellular Fixation & Permeabilization Buffer (Thermo Fisher Scientific, Waltham, MA). Detailed primary antibodies for flow cytometry were used as described before [17]. Flowjo v10.0 was used to analyze BD LSRFortessaTM X-20 (BD Bio-sciences, San Jose, CA) fluorescence-activated cell sorting data (Tree Star, Ashland, OR).

All analysis approaches were performed by R software (https://www.r-project.org/). Cox and Kaplan-Meier analyses were performed by “survival” and “survminer” packages. Random forest model construction was performed by “randomForestSRC” and “ggRandomForests” packages. For volcano plot, the normalized enrichment score cutoff was 1.5.

The Kruskal-Wallis H test was used to compare continuous variables between groups. The categorical variables were used for the chi-square test. Spearman’s correlation analysis was utilized for quantitative correlation analysis. The cutoff of CDKs family was set at 50%. The cutoff of CDK5 was optimized and set at 33%. All data processing was performed on the R software platform (R Foundation for Statistical Computing, Vienna, Austria).

The prognostic significance of CDKs families in S1 Table. The latest RCC guidelines recommend TKI+IO combination therapy as the conventional first-line therapy for mRCC patients. However, our ZS-MRCC group showed that mRCC patients exhibited a variety of therapeutic effects (Fig. 1D, E and G). Transcriptomic data proved that expression of CDK5 was enhanced in RCC tissues relative to non-tumorous tissues (TCGA cohort) (p < 0.001) (Fig. 1A). However, we found no significantly different progression-free survival (PFS) between CDK5-low and CDK5-high TxNxM1 RCC patients in the TCGA cohort (data not shown). This may be owing to the fact that their cohorts include patients at all stages and the treatment was different from what it is today. Thus, we assessed the expression of CDK5 in various RCC stages and grades. CDK5 was associated with advanced TNM stage (stage IV) and International Society of Urological Pathology grade in RCC (G4) (Fig. 1B and C). Responders to TKI+IO revealed considerably decreased CDK5 expression (p < 0.05) (Fig. 1D). Meanwhile, the number of patients from low CDK5 expression subgroup responses to TKI+IO were considerably elevated (Fig. 1F and G). Patients with a low CKD5 expression in our ZS-MRCC cohort had a longer PFS (p=0.014) (Fig. 2B) than those with a high CDK5 expression, which was verified in JAVELIN 101 cohort (p=0.040) (Fig. 2C). The cutoff was optimized and set at 33%. Further Kaplan-Meier analysis revealed that patients with a higher CDK5 expression had a worse progression-free survival than those with a lower CDK5 expression, only in the TKI+IO group of the JAVELIN Renal 101 cohort (p=0.04) (Fig. 2C) and in our ZS-MRCC cohort (p=0.014) (Fig. 2B), but not in the TKI monotherapy groups (Fig. 2D). Nevertheless, CDK5 was not predictive factor, as TKI+IO demonstrated better outcome to TKI monotherapy in patients with both high and low CDK5 levels (Fig. 2E and F). Subsequently, both univariate and multivariate Cox regression analysis was performed. Clinical and pathological parameters, including age, sex, histology, International Metastatic RCC Database Consortium (IMDC) group, along with CDK5 expression were incorporated into the Cox regression model. Consequently, we found that CDK5 expression indicated poor prognosis independent of the above clinical and pathological parameters based on PFS (univariate: hazard ratio [HR], 3.199; 95% confidence interval [CI], 1.195 to 8.564; p=0.021; multivariate: HR, 2.841; 95% CI, 1.037 to 7.785; p=0.042) (Fig. 2A). Collectively, these results indicated that the expression of CDK5 could serve as an independent adverse prognosticator for patients with TKI+IO treatment.

By hematoxylin and eosin and immunohistochemistry, we identified markers of TME components, including effector cells, regulatory cells, regulatory chemicals, angiogenesis, Ki-67, and polybromo 1 (PBRM1) in our ZS-HRRCC cohort (Fig. 3A). Tumor infiltrating lymphocytes (TILs) did not differ significantly between low and high CDK5 expression groups (Fig. 3B), nor were TILs linked with CDK5 as determined by spearman analysis (data not shown). The number of CD8⁺ T cells was remarkably decreased in patients with high CDK5 levels (p < 0.05) (Fig. 3C), and was inversely correlated with the expression of CDK5 (Spearman’s ρ=−0.45, p < 0.001, data not shown). Unexpectedly, CD4⁺ T cells (p=0.720) showed exact opposite trends, compared with CD8⁺ T cells (p < 0.01) (Fig. 3D).

To further investigate the immunologic state of tumor-infiltrating T cells, we utilized flow cytometry on resected nephrectomy samples from our ZS-HRRCC cohort (Fig. 3E). Consistently, CDK5 was negatively associated with CD8⁺ T cells (Spearman’s ρ=−0.49, p < 0.001) (Fig. 3F), and positively linked with CD4⁺ T cells (Spearman’s ρ=0.51, p < 0.001) (Fig. 3G). The CD8⁺ T cells and CD4⁺ T cells percentage among total T cells were enhanced in low CDK5 and high CDK5 groups, respectively (p < 0.05) (Fig. 3F and G).

T cells are crucial anti-tumor mediators that identify and react to tumor-expressed antigens, and they have proved indispensable for cancer immunotherapy. Nonetheless, T cells are not always as anticipated as expected against cancer. T cells attain a dysfunctional or fatigued state defined by sustained expression of inhibitory receptors and a transcriptional state distinct from that of functioning effector or memory T cells. Therefore, flow cytometry was used to evaluate the expression of granzyme B (GZMB) and programmed death-1 (PD-1) on CD8⁺ and CD4⁺ T cells in samples from our ZS-HRRCC cohort. Although the amount of CD8⁺ T cells was reduced in individuals with high CDK5 expression (Fig. 3C and F), a positive connection was identified between GZMB⁺ CD8⁺ T cells/total T cells and CDK5 (Spearman’s=−0.51, p < 0.001) (Fig. 4A). Moreover, PD-1⁺ CD4⁺ T cells/total T cells were strongly linked with CDK5 (Spearman’s=0.44, p < 0.001) (Fig. 4B). A significant positive connection was identified between Eomesodermin (EOMES; Spearman’s=0.52, p < 0.001) (Fig. 5B), lymphocyte activating gene 3 (LAG3; Spearman’s=0.50, p < 0.001) (Fig. 5C), T-cell immunoreceptor with Ig and ITIM domains (TIGIT; Spearman’s=0.53, p < 0.001) (Fig. 5D), T cell factor 1 (TCF1; Spearman’s=0.30, p=0.060) (Fig. 5E), and CDK5. The expression of IHC markers positive cells were normalized to tumor-infiltrating lymphocytes (Fig. 5A). These results indicated a potential correlation between CDK5 and T cell exhaustion mechanisms.

Immunosuppressive tumor-infiltrating cells, such as Tregs and macrophages, are putative regulatory mechanisms for T cell exhaustion. The relationship between CDK5 and Treg and macrophage infiltration was examined. There was a positive correlation between Tregs (CD4(+) CD25(+) CD127(−/low) cells) and CDK5 levels (evaluated by flow cytometry, Spearman’s ρ=0.38, p=0.02) (Fig. 4C), which was confirmed by formalin-fixed paraffin-embedded tissues (FoxP3 evaluated by IHC, Spearman’s ρ=0.30, p=0.06) (Fig. 4D). Probably due to the relatively limited cohort size, the p-value of IHC samples was close but not statistically significant. Macrophage infiltration showed no correlation. Macrophages could be polarized into M1- or M2-macrophages under specific TME conditions. Therefore, immunohistochemistry was used to detect M1- or M2-macrophages in the ZS-HRCC cohort. Neither M1 nor M2 macrophages exhibited significant changes in samples with high levels of CDK5. Tregs inhibit the activity of cytotoxic T cells in the TME by the production of transforming growth factor β (TGF-β) [21]. TGF would achieve this by inhibiting the expression of cytolytic gene products, such as granzymes B (GZMB) and K, perforin, FasL, and IFN-γ, from cytotoxic T cells. Spearman analysis showed that the expression of TGF-β was associated with CDK5 expression (Fig. 4E). Besides, through Gene Set Enrichment Analysis, we found that negative regulation of leukocyte proliferation, Tregs differentiation, negative regulation of leukocyte mediated immunity, and negative regulation of T cell–mediated immunity pathways were enriched in high-CDK5 samples (Fig. 4F).

A summary of the chromosomal mutations ranked by CDK5 expression for the JAVELIN-101 cohort was compiled (Fig. 4G). Frequent mutations in localized RCC were observed in advanced RCC, including VHL (von Hippel-Lindau tumor suppressor, 55%), PBRM1 (32%), and SETD2 (SET domain containing 2, histone lysine methyltransferase, 25%). Only mammalian target of rapamycin (mTOR) mutation demonstrated a significant relationship with CDK5 expression (p < 0.05) (Fig. 4G).

The latest EAU Guidelines for RCC recommended TKI+IO combinations as standard first-line therapy and the PFS of the JAVELIN 101 clinical trial suggested that TKI+IO is a preferred option for mRCC patients. However, the therapeutic benefits of TKI+IO varied by individual. Typically, there is always a portion of patients who responded unfavorably to TKI+IO therapy. Thus, it is necessary to build a model that can determine which subgroup responds best or worse to TKI+IO treatment. Random forest, one of the most prominent machine learning algorithms, was implemented. The expression of CDK5, PD-1, PD-L1, CD8A, CD4, GZMB, and GZMK were enrolled as the parameters for model construction. The contribution of each parameter to the final model was further investigated (Fig. 6A). The predictive value of our model was subsequently validated by Kaplan-Meier. The cutoff was set at 75%. The results indicate that patients with a low signature score in our model have a longer PFS when treated with TKI+IO combination therapy (p < 0.001) (Fig. 6C). This model was subsequently verified in our ZS-MRCC cohort (p=0.002) (Fig. 6D).

Furthermore, TKI monotherapy is an alternative first-line treatment for mRCC patients. Application of our model in mRCC to PFS data from JAVELIN Renal 101 showed in the low RFscore arm (≤ 75% signature score) TKI+IO combination therapy was associated with a trend toward longer PFS (HR, 0.55; 95% CI, 0.43 to 0.706; p=0.009) (Fig. 6B), whereas in the high RFscore TKI+IO and TKI alone did not differentiate PFS (HR, 1.077; 95% CI, 0.720 to 1.610; p=0.719). The random forest model showed a better function in predicting the prognosis of TKI+IO against TKI monotherapy for mRCC patients. Ultimately, Kaplan-Meier validated the predictive value of our model (Fig. 6E).