3.1 Patients’ characteristics
The workflow of this study was showed in Figure 1. From January 2019 to December 2021, a total of 102 patients with NSCLC treated with combination of chemotherapy and ICIs in China-Japan Friendship Hospital were enrolled in this study. 3 patients didn't complete the first evaluation for the first administration of ICIs starting recently. Therefore, a total of 99 patients were enrolled for analysis. Patient baseline clinical characteristics are summarized in Table 1. Of these, there were 55 patients (55.6%) with LUAD, 37 patients (37.4%) with LUSC, the others were large cell carcinomas or mixed NSCLC. The median age was 64 years (58 - 69). 15 patients (15.2%) were female. 4 patients (4.04%) suffered from Stage I to IIIa disease and received ICIs as pre-operative adjuvant therapy. The 99 patients were all in good condition (the range of Eastern Cooperative Oncology Group Performance Status [ECOG-PS] was 0–1 score). 18 patients (18.2%) with LUAD had the recorded KRAS mutations. 58 patients (58.6%) were normal weight, 36 patients (36.4%) were overweight, and only 5 patients (5.05%) were obese. No patients were underweight. The overweight and obese groups were thereby combined as the BMI ≥ 25kg/m2 group. No significant difference between the BMI < 25kg/m2 and the BMI ≥ 25kg/m2 group in the baseline circulating lipids, lipoproteins, and apolipoproteins (Table S1). The mean follow-up period for this study was 9.56 months. 75 patients which had been followed up for more than 6 months and had the primary endpoint (DCB or NDB). The best response to ICIs (non-active or active) of 89 patients was available. irAEs of 99 patients were recorded. Patient demographics and disease characteristics including age, gender, smoking history, and diabetes status were generally balanced between the NDB and DCB arms (Table S2), what’s more, between the non-active and active response arms (Table S3).
3.2 BMI ≥ 25 kg/m2 predicts poor ICIs response in NSCLC
Comparations between groups divided according to the duration of responding to ICIs (NDB or DCB) and best response to ICIs (non-active or active) were performed. NDB patients were more likely to have higher BMI (T-test, 25.71 kg/m2 vs 23.29 kg/m2, p.val = 0.034, Figure 2A). The prevalence of BMI was higher in the cases with the non-active response, but the difference was not statistically signiﬁcant (T-test, 25.64 kg/m2 vs 24.40 kg/m2, p.val = 0.065, Figure 2B).
BMI ≥ 25 kg/m2 was a negative predictor for DCB and active response to ICIs. The OR for NDB was 6.06 (95% CI 2.05 - 19.89, p.val = 0.002) for patients with BMI ≥ 25 kg/m2 by the adjusted for the confounders including age, gender, diabetes, and smoking history. The OR for non-active response to ICIs was 4.37 (95% CI 1.64 - 12.43, p.val = 0.004) for patients with BMI ≥ 25 kg/m2 with the confounders adjusted. The area under the ROC curve (AUC) was 0.72 (95% CI 0.60 - 0.85, Figure 2C) for the model used to predict DCB. AUC was 0.70 (95% CI 0.57 - 0.83, Figure 2D) for predicting active response to ICIs. The calibration curves showed that the multivariate models containing the BMI group tended to underestimate the probability of NDB and non-active response (Figure 2E-F).
Patients with normal weight tended to have a longer duration of ICIs response (Log-rank test, median PFS: 17.8 months vs 4.83 months, p.val = 0.001, Figure 2G). The HR for PFS of BMI ≥ 25 kg/m2 without confounding factors adjusted was 2.54 (95% CI 1.41 - 4.60, p.val = 0.002), the HR was 3.08 (95% CI 1.63 - 5.82, p.val < 0.001) by multivariate Cox model, indicating that BMI ≥ 25 kg/m2 was a negative predictor for the duration of ICIs response. The C-index of the multivariate Cox model was 0.68 (95% CI 0.61 - 0.76). The 3-year AUC was 0.76 and the calibration curve was shown in Figure 2H-I.
The subgroup analyses showed that BMI ≥ 25 kg/m2 was an active indicator of DCB and active response to ICIs according to the univariate logistic regression in all subpopulations (Figure S1-2). What’s more, patients with BMI ≥ 25 kg/m2 tended to have a shorter PFS of ICIs in all subgroups (Figure S3).
3.3 High levels of remnant cholesterol predict poor ICIs response in NSCLC
Compared with DCB patients, the prevalence of serum cholesterol, RC, and LDL-C were higher in NDB patients, but the difference was statistically signiﬁcant in RC only (T-test, 0.64 mmol/L vs 0.49 mmol/L, p.val = 0.047, Table S2). Serum cholesterol, triglyceride, phospholipid, and HDL-C had no difference between NDB and DCB patients. Circulating apolipoproteins including apoA1, apoE, apoC2, and apoC3 did not significantly differ between patients with NDB and DCB. Circulating lipids, lipoproteins, and apolipoproteins had no significant difference in the active and non-active groups base on the best response to ICIs in the total NSCLC cohort (Table S3).
Based on the PFS of ICIs, the total group was divided into RC high group and RC low group according to the serum RC levels. Similarly, the total cohort was also grouped by LDL-C level and apoB level. The cut points of RC and LDL-C were selected as 0.82 mmol/L and 2.58 mmol/L respectively. The analogous analyses of the independent and joint effects of apoB containing lipoproteins cholesterol (RC and LDL-C) and apoB levels and BMI were performed. Both continuous and categorical for of each marker were respectively brought into univariate and multivariate logistic regression model with the confounders including age, gender, diabetes, smoking history, and BMI adjusted.
The OR of RC for NDB and the non-active response was shown in Table 2, indicating that high serum RC potentially promoted ICIs resistance which led to NDB and non-active response to ICIs. The AUC of the multivariate model containing RC level (continuous) was 0.78 (95% CI 0.67 - 0.89, Figure 3A) for predicting DCB and 0.74 (95%CI 0.63 - 0.86, Figure 3B) for predicting active response to ICIs. The AUC of the model containing low RC was 0.79 (95%CI 0.68 - 0.91, Figure 3C) and 0.75 (95%CI 0.63 - 0.86, Figure 3D) respectively. The calibration curves showed that the multivariate models RC level and low RC all presented satisfied coherence in predicting NDB probability (Figure 3E-F). When it comes to predicting non-active response, the models tended to overestimate the rates when the incidence of non-active response was low and underestimate the rates when the incidence was high (Figure 3G-H).
The OR of low LDL-C for NDB calculated by univariate logist model was 0.31 (95% CI 0.10 - 0.89, p.val = 0.036), however the OR with the confounders including age, gender, diabetes, smoking history, and BMI adjusted was not statistically signiﬁcant (OR = 0.32, 95%CI 0.09 - 1.03, p.val = 0.065). Low serum LDL-C was a protective indicator for active ICIs response both without and with the confounders adjusted (OR = 0.29, 95%CI 0.09 - 0.81, p.val = 0.026; OR = 0.24, 95%CI 0.06 - 0.77, p.val = 0.023). The AUC was 0.76 (95%CI 0.66 - 0.86), but the calibration ability was unsatisfied (Figure S4A-B).
3.4 Low serum LDL-C predicts long PFS of ICIs in NSCLC
Patients with high RC and LDL-C were more like to respond to ICIs with shorter PFS (Log-rank test, median PFS: 14.53 months vs 3.23 months, p.val = 0.005; median PFS: 16.00 months vs 7.40 months, p.val = 0.010; Figure 4A-B). Considering BMI was a potentially important confounding factor for ICIs response and serum apoB containing lipoproteins level, the multivariate Cox regression of RC and LDL-C was performed (Table S4). Low LDL-C was a statistically significant protective factor for the long PFS with the BMI and other confounders adjusted (HR = 0.43, 95%CI 0.22 - 0.86, p.val = 0.016). The serum LDL-C level in continuous form was a negative indicator for ICIs PFS (HR = 1.50, 95%CI 1.01 - 2.24, p.val = 0.045). AUC of the multivariate model with LDL-C (continuous and category) at 3-year was 0.76 and 0.79 respectively (Figure 4C-D). The calibration curves were pictured in Figure 4E-F. The comparison of Cox models used to predict PFS was performed. Compared with only including the BMI category, the adding of serum LDL-C or LDL-C group was more efficient (Anova-test, p.val = 0.049, p.val < 2e-16). The DCA curves showed that the model with the LDL-C group was superior to the model with serum LDL-C in the continuous form and the model without LDL-C (Figure 4G). Serum RC and RC category did not add the extra value to the multivariate Cox model with the BMI category.
Subgroup analysis was performed according to the BMI group. Patients with low RC and LDL-C showed a trend to respond to ICIs with longer PFS in both groups but were statistically different in the BMI < 25 kg/m2 group (Figure 5A-E). The multivariate Cox models of RC and LDL-C (continuous and category) were performed by adjusting confounders including gender, age, diabetes status, and smoking history. Low levels of LDL-C and RC were protective factors in the BMI < 25 kg/m2 subgroup but not in the BMI ≥ 25 kg/m2 group (Figure 5F).
3.5 ApoB containing lipoproteins predict better response to ICIs in LUAD cohort
In the LUAD cohort, the NDB group seemed to have higher serum cholesterol, LDL-C, and RC without a statistical difference (Figure S5A-C). Patients with non-active response to ICIs had higher serum LDL-C and CHO (T-test, 3.45 mmol/L vs 2.84 mmol/L, p.val = 0.010; 5.38 mmol/L vs 4.46 mmol/L, p.val = 0.017, Figure 6A-B). The prevalence of RC tended to be higher in the non-active group but was not statistically significant (Figure 6C). The univariate and the multivariate models including RC (both continuous and categorical), and LDL-C (both continuous and categorical) in the LUAD subgroup demonstrated that low RC and low LDL-C were protective factors for active ICIs response (Figure 6D). However, similar analyses were performed in the LUSC subgroup, serum apoB containing lipoproteins did not show satisfactory ability in predicting the ICIs response (Table S5).
Patients with low LDL-C and low RC tended to respond to ICIs with longer PFS (Figure 6E-F). The Kaplan-Meier curves of the LDL-C and RC groups were crossed in the LUSC cohort (Figure S5D-E). The univariate Cox analysis diagnosed the low LDL-C and low RC as protective factors for long PFS (Figure 6E-F), but only the HR of low LDL-C was statistically significant with the confounders adjusted (HR = 0.34, 95%CI 0.12 - 0.99, p.val= 0.047).
3.6 Correlation between serum lipids, lipoproteins and apolipoproteins and ICIs related adverse events (irAEs)
Until May 1st, 2022, 38/99 patients (38.4%) occurred irAEs, 9 patients stopped ICIs for irAEs. Patients were grouped according to the occurrence of irAEs. Difference of BMI between the two groups were not significant. The prevalence of circulating lipids, lipoproteins and apolipoproteins was not statistically different between patients with irAEs and those without irAEs. Patients who had suffered from irAEs tended to have higher apoA1/(apoA1+apoB) ratio (0.58 vs 0.54, p.val = 0.040). However, the apoA1 ratio was not an ideal biomarker for predicting irAEs for the diagnosis of the univariate and multivariate models containing apoA1 ratio was unsatisfied.
3.7 High expression of APOB potentially predicted poor ICIs response in the validation cohorts
Expression of APOB which encodes apoB proteins was tested for validation in GSE126044, and GSE135222 cohorts. 16 advanced NSCLC patients including 7 LUAD and 9 LUSC treated by Nivolumab were included in GSE126044. Expression of APOB tended to be higher in the NDB group in the LUAD group despite without statistical difference (Wilcox test, p.val = 0.191, Figure S6A). The 16 patients were divided into two groups according to the expression of APOB. LUAD patients with high expressed APOB tended to have shorter PFS (Log-rank test, 2.22 months vs 3.5 months, p.val = 0.018, Figure S6B). High APOB in LUAD tended to predict a short PFS of ICIs (HR = 2.86, 95%CI 0.82 - 10.02, p.val = 0.100). The tendency did not exist in the total cohort and the LUSC cohort (The total group: HR = 1.05, 95%CI 0.76 - 1.46, p.val = 0.77). 27 advanced NSCLC patients treated with anti-PD-1/PD-L1 were included in GSE135222. Expression of APOB was more likely higher in the NDB group (Wilcox-test, p.val = 0.356, Figure S6C). The low APOB group tended to have a longer PFS in the NSCLC patients without a statistical difference (Log-rank test, 2.95 months vs 1.47 months, p.val = 0.239, Figure S6D). Patients with higher APOB potentially indicated shorter duration of response to ICIs (HR = 1.08, 95%CI 0.99 - 1.18, p.val = 0.073).
3.8 High expression of APOB corresponded to suppressed TME in LUAD
391 LUAD patients without EGFR sensitive mutations and 494 LUSC patients were respectively classified into three clusters in the TCGA cohorts (Figure S7A-D). CYT represented the immune response of the tumor, which decreased from C2 to C3 (C2>C1>C3, Figure S7 B, D). C3 cluster was more likely to be an immunologically cold or non-inflamed tumor and was excluded since it potentially belongs to ICIs resistant subset. C2 cluster presented as an immunologically hot phenotype and was more likely to benefit from immunotherapies. C1 cluster couldn’t be classified in C2 or C3 cluster for the high immune heterogeneity which tended to be the most frequent immunophenotype in routine clinic conditions.
277 patients with LUAD samples and 479 LUSC patients in C1 and C2 clusters were divided into APOBhigh and APOBlow groups according to the expression of APOB. According to the classification, APOBlow expressed more major histocompatibility complex class I (MHC-I) molecules (Figure 7A). The infiltration of T helper 2 cells (Th2) cells and effect memory CD4+ T cells (emCD4) and NK cells were also higher in the APOBhigh group (Figure 7A). TME estimated by MCPcounter showed that the fibroblasts and monocytic lineage cells were highly infiltrated in APOBhigh group, which is more likely to lead to immunosuppression by forming a barrier around the tumor cells (Figure 7B). Likewise, according to the cell component derived from the methylation data, despite the total immune cells showing no statistical difference, the fibroblasts were higher in the APOBhigh group (Figure 7C).
When it comes to the LUSC cohort, APOBhigh group tended to have higher infiltrated immune cells which potentially explain the reason why the difference of ICIs response in LUSC according to serum apoB containing proteins level was not consistent with LUAD (Figure S8).
Functional enrichment analysis between the APOBhigh and APOBlow groups in the TCGA-LUAD cohort showed that apart from metabolic syndrome-associated pathways, many types of cancer-related pathways were upregulated in APOBhigh group. TGF beta signaling pathway, Wnt signaling pathway, and adipocytokine signaling pathway were upregulated in APOBhigh patients, which indicated the potential crosstalk of metabolism and cancer (Figure 8A). Differential gene analysis demonstrated that the apoB containing lipoproteins receptors including low-density lipoprotein receptor-related protein 5 (LPR5) and LPR6 which involve in the endocytosis of apoB containing lipoproteins and mediate the downstream Wnt signaling were upregulated in the non-inflamed TME and NDB patients. LDLR gene was not highly expressed in the NDB patients and patients with non-inflamed TME (Figure 8B-C).