Malignant cutaneous melanoma is one of the most common cancers in young adults. During the last decade, targeted and immunotherapies have significantly increased the overall survival of patients with malignant cutaneous melanoma. Nevertheless, disease progression is common, and a lack of predictive biomarkers of patient response to therapy hinders individualized treatment strategies. To address this issue, we performed a longitudinal study using an unbiased proteomics approach to identify and quantify proteins in plasma both before and during treatment from 109 patients treated with either targeted or immunotherapy. Linear modeling and machine learning approaches identified 43 potential prognostic and predictive biomarkers. A reverse correlation between apolipoproteins and proteins related to inflammation was observed. In the immunotherapy group, patients with low pretreatment expression of apolipoproteins and high expression of inflammation markers had shorter progression-free survival. Similarly, increased expression of LDHB during treatment elicited a significant impact on response to immunotherapy. Overall, we identified potential common and treatment-specific biomarkers in malignant cutaneous melanoma, paving the way for clinical use of these biomarkers following validation on a larger cohort.

Significance:

This study identifies a potential biomarker panel that could improve the selection of therapy for patients with cutaneous melanoma.

During the last decade, the introduction of targeted therapy such as BRAFV600E and MEK inhibitors (inhibiting the MAPK signaling pathway) has revolutionized the treatment of cutaneous malignant melanoma (CMM) dramatically by increasing the response and overall survival (OS) rates for patients harboring a BRAF mutation, especially when these drugs are administrated in combination (1, 2). Unfortunately, despite initial remarkable response for most of these patients, the majority of them experienced disease progression due to acquired resistance through multiple mechanisms (3). The 5-year OS rates for CMM patients treated with combinatorial targeted therapy is only 34%, showing that only a subset of them benefit from long-term effects (4).

At the same time, immunotherapy with checkpoint inhibitors (ICI), such as antibodies against either programmed cell death protein 1 [PD-1 (PDCD1), nivolumab, and pembrolizumab] or cytotoxic T-lymphocyte associated protein 4 (CTLA4; ipilimumab; ref. 5), were also introduced into the field. The rationale is to boost the T-cell activation of the immune system against CMM. Nivolumab and pembrolizumab interfere with the ligation of PD-1 on tumor-specific T cells with its immunosuppressive ligand PD-L1 (CD274), whereas ipilimumab targets the CTLA4 on T cells and thereby increases the activation and proliferation of tumor infiltrating T cells, boosting an antitumor response. Nivolumab and pembrolizumab demonstrate a better response rate than ipilimumab and less toxicity (6, 7) with a 5-year OS rate of 34% to 44% (8). Even though data from studies with head-to-head comparison between targeted and immunotherapy are still missing, it seems that the 5-year OS rate is similar for PD-1 inhibitors and the combined targeted therapy (4, 8). Studies comparing combination of BRAF and MEK inhibitors against ipilimumab and anti-PD1 and the best sequencing strategy are still ongoing (NCT02224781, NCT02631447); however, the current data indicate that the combination of ipilimumab and nivolumab represents the best treatment option, with a 5-year OS rate of 52%, 44%, and 26% for the combinatorial treatment (ipilimumab and nivolumab), nivolumab and ipilimumab, respectively (9). The toxicity of this combinatorial treatment represents the main limitation as almost 60% of patients presents severe or very severe (grade 3 and 4) side effects (10).

Currently, it is difficult to predict which patient might benefit from which therapy. Hence, there is an urgent need for biomarkers that would enable stratification of patients for the most effective treatment against CMM, minimizing unnecessary side effects and costs. As of today, no molecular biomarkers are yet employed in the clinical practice to predict response to therapy in CMM, except for BRAF mutation, which is used to predict response to MAPK inhibitors (MAPKi) targeted therapies. Biomarkers can be distinguished between two main categories: prognostic and predictive. The first predicts the outcome regardless of the therapy employed, whereas the second foresees the outcome for a specific treatment. The distinction is vague, as biomarkers often have some degree of prognostic and predictive value simultaneously, with the dominance of one of the two (11). The sources of biomarkers can be biological fluids (i.e., blood, urine) or histologic. The accountability of the latter is questionable due to the known intratumor heterogeneity in CMM (12); hence, the sample might not necessarily mirror the whole tumor. On the basis of this consideration, biological fluids may represent a more reliable and accessible source for biomarkers with minimal invasive procedures, that is, a simple blood sample.

In this work we employed an unbiased proteomics platform, linear modelling, and machine learning approach to investigate the presence of biomarkers that would help to predict the clinical outcome [response and progression-free survival (PFS)] to MAPKi and ICI therapies. We performed a longitudinal study on 109 patients, whose plasma was analyzed at different time points: pretreatment and during treatment, and at disease progression. Our results show that a panel of proteins correlates with the response to therapy and PFS. These could potentially be used as biomarkers to anticipate which patients may have long-term benefits from a specific treatment and hence enable their stratification.

Patient cohort

Plasma samples and clinical data were collected from 109 patients with stage IV CMM before and during treatment with MAPKis and ICIs, as well as at disease progression from a small subset, between March 2012 and August 2017 at the Department of Oncology-Pathology, Karolinska University Hospital, Sweden.

The clinical information included age at treatment start, sex, metastatic classification (M-class) according to the 7th American Joint Committee on Cancer (AJCC) staging edition (13), presence of liver metastasis, lactate dehydrogenase (LDH) level, CRP level, type and line of MAPKi and ICI, response to treatment, and PFS (Table 1).

Table 1.

Clinical characteristics.

MAPKi (n = 44a)ICI (n = 65a)
Gender Male 25 42 
 Female 19 23 
Age (years old) Median (range) 60 (32–86) 66 (23–84) 
M1 stage M1a 11 
(AJCC 7th) M1b 
 M1c 41 45 
LDH (μKat/l) Median (range) 4.6 (2.6–58.3) 4 (1.7–26.5) 
CRP (mg/l) Median (range) Not applicableb 6 (1–299)c 
Therapy BRAFi 21 
 BRAFi+MEKi 23 
 anti-CTLA4 10 
 anti-PD-1 52d 
 anti-CTLA4+anti-PD-1 
Line of treatment 1st 42 52 
 2nd 
 3rd 
Response Disease control 33 41 
 Nonresponder 11 22 
 Not evaluable 2e 
PFS Median (range), days 179 (18–1298)f 262 (5–1983)g 
MAPKi (n = 44a)ICI (n = 65a)
Gender Male 25 42 
 Female 19 23 
Age (years old) Median (range) 60 (32–86) 66 (23–84) 
M1 stage M1a 11 
(AJCC 7th) M1b 
 M1c 41 45 
LDH (μKat/l) Median (range) 4.6 (2.6–58.3) 4 (1.7–26.5) 
CRP (mg/l) Median (range) Not applicableb 6 (1–299)c 
Therapy BRAFi 21 
 BRAFi+MEKi 23 
 anti-CTLA4 10 
 anti-PD-1 52d 
 anti-CTLA4+anti-PD-1 
Line of treatment 1st 42 52 
 2nd 
 3rd 
Response Disease control 33 41 
 Nonresponder 11 22 
 Not evaluable 2e 
PFS Median (range), days 179 (18–1298)f 262 (5–1983)g 

aOne of the patients received two lines of treatment, MAPKi and ICI.

bNo CRP measurements available from patients.

cCRP measurements available from 31 patients.

dThree of 52 patients treated with pembrolizumab plus epacadostat within clinical trial.

eNot evaluable due to premature death.

fThree patients were still responding at study cut-off.

gTwenty patients were still responding at the study cut-off.

Pretreatment plasma samples were available from 98 patients (35 MAPKi and 63 ICI patients), samples during treatment were available from 85 patients (27 MAPKi and 58 ICI patients) and samples at disease progression were available from 30 patients (19 MAPKi and 11 ICI). Pretreatment samples were taken on the same day as treatment start from the vast majority of patients, however a variation between 0 and 20 days was observed and three pretreatment samples were collected the day after treatment started. The number of days between the start of therapy and the during-treatment samples collection ranged between 10 and 64 days (median of 15 days).

The majority of patients (n = 94) were treated outside of clinical trials according to the standard local follow-up schedule with medical visits every 4 to 8 weeks and with radiological evaluation performed every 8 to 12 weeks. For patients outside of clinical trials the therapy response was based on joint evaluation of clinical/radiological investigations evaluated by a team including oncologists and radiologists, according to our clinical routine.

Computed tomography (CT), magnetic resonance imaging and/or positron emission CT tomography were used. Fifteen patients received therapy within clinical trials (1 in MO-25515-NCT01307397, 3 in Columbus-NCT01909453, 1 in Combi-D-NCT01584648, 3 in Checkmate 401-NCT02599402, and 7 in ECHO-301/Keynote-252-NCT02752074). These patients were followed according to the multicenter studies protocols and were radiologically evaluated according to RECIST 1.1 (14).

The group of patients with disease control included individuals with complete response, partial response and stable disease whereas those with progressive disease were classified as nonresponders. The PFS time was calculated from the day of treatment start until the date of progress or death, whatever came first (Table 1).

The study was approved by the Stockholm Regional Ethics Committee, Karolinska Institutet, Sweden, and conducted in accordance with Good Clinical Practice/the Declaration of Helsinki. We obtained informed written consent from the subjects.

Plasma sample preparation and LC-MS/MS analyses

Plasma samples were diluted 1:10 and reduced by DTT and alkylated by CAA and digested by Lys-C/Trypsin. Peptides were labeled by TMT11plex according to the manufacture instructions, mixed 1:1 and fractionated by basic reverse phase prior LC-MS/MS analyses using an Ultimate 3000 nano-LC with an EASY‐Spray ion source connected to a Q Exactive HF Hybrid Quadrupole-Orbitrap Mass Spectrometer (Supplementary Materials and Methods).

Raw MS data were searched with MaxQuant 1.6.1.0 using peptide and protein FDR at 1%. TMT11plex was chosen as a quantification platform. Trypsin/P was chosen as cleavage specificity allowing two missed cleavages. Carbamidomethylation (C) was set as a fixed modification, whereas oxidation (M) was used as variable modifications. The database search was performed with a mass deviation of the precursor ion of up to 4.5 ppm (main search). Data filtering was carried out using the following parameters: minimum peptide length was set to 7 and Andromeda minimum score for modified peptides was set to 40, minimum reporter precursor ion fraction = 0.75.

See Supplementary Materials and Methods for further details.

Cox regression

Cox regression univariate analysis was used to evaluate the effect of different clinical and patient characteristics on PFS. Results are presented with hazard ratio (HR) and 95% confidence interval (95% CI). All analyses were carried out using the statistical software STATA.

Data normalization and peptide filtering

Data were normalized using R for possible loading/pipetting errors and batch effects using an empirical Bayes framework (ComBat from the package sva (15), version 3.28.0), and peptides of hemolysis associated proteins were filtered out using the limma (16) Bioconductor package in R (limma version 3.40.6; see Supplementary Materials and Methods).

Modeling of PFS and protein concentration

Relative protein quantities PFS (days) were log10-transformed and a linear model fitted for each protein, before and during treatment separately, as follows:

formula

For explained variance, omega squared (ω2) was used. P values were corrected per term and sample type (PRE and TRM) using the Benjamini–Hochberg (BH) procedure.

Machine learning

Data from samples taken before and during treatment, for all proteins that were significant in the linear models for either the predictive or prognostic factor, as well as data for stage (AJCC 7th), sex, and age was used to build and evaluate predictive machine learning models. Nine separate random forest (RF) models were created using PRE and TRM data, only PRE, and only TRM respectively, each for combined treatments, only ICI, and only MAPKi therapy, respectively. The models were assessed on their area under the curve (AUC) for the resulting receiver operating characteristic (ROC) curve that was generated by cross validation in the model training. Individual predictors were evaluated using caret's varImp function to calculate variable importance. See Supplementary Materials and Methods for details, Supplementary Fig. S1 shows a summary of the samples used in linear models and machine learning, respectively.

Absolute quantification: sample preparation and LC-MS/MS analyses

A subset of 24 samples were randomized and aliquoted together with sextuplicate of plasma pool. A pool of 19 SIS-PrESTs were added to the samples, which were denatured by adding 1% sodium deoxycholate; reduced for 60 minutes at 37°C in 10 mmol/L dithiothreitol and alkylated for 30 minutes at room temperature in the dark in 50 mmol/L chloroacetamide for 30 minutes consecutively, before digestion by Lys-C/trypsin at 37°C O.N. (see Supplementary Materials and Methods).

Data and code availability

Data are available via ProteomeXchange with identifiers PXD019518 and code for data processing and analysis is available on Code Ocean (https://codeocean.com/capsule/2987719).

Plasma samples, collected from 109 patients with stage IV (AJCC 7th) CMM before and during treatment with either MAPKi or ICI, were analyzed to identify potential biomarkers for outcome to MAPKi and ICI. We employed an untargeted MS approach using the TMT11plex platform, which enables the relative quantification of multiple samples simultaneously in the same batch. In each batch a duplicate of a reference sample consisting of a pool of patient samples was included (Fig. 1; Supplementary Fig. S2; ref. 17). We leveraged this property to obtain relative plasma protein levels between samples in different batches/sets. Overall, we identified 572 proteins, out of which, 272 were quantified in at least 50% of the samples. These were in turn aggregated to 194 genes.

Figure 1.

Overview of the study. Plasma samples and clinical data were collected from 109 patients with stage IV CMM before, during treatment, and at disease progression. Plasma samples were analyzed by nLC-MS/MS, generating relative abundances for 194 plasma proteins. These data have then been used to build predictive models to assess relationship between plasma protein levels and patient outcome, progression free survival, and treatment response in patients treated with either MAPKi or ICI.

Figure 1.

Overview of the study. Plasma samples and clinical data were collected from 109 patients with stage IV CMM before, during treatment, and at disease progression. Plasma samples were analyzed by nLC-MS/MS, generating relative abundances for 194 plasma proteins. These data have then been used to build predictive models to assess relationship between plasma protein levels and patient outcome, progression free survival, and treatment response in patients treated with either MAPKi or ICI.

Close modal

Quality control and data cleaning by excluding biomarkers associated with hemolysis and coagulation

During sampling, hemolysis was visually observed (red discoloration of plasma) in 27 of 214 samples. To avoid miss-assignment of certain proteins as potential biomarkers, a differential expression analysis was conducted to find proteins with: (i) significantly higher levels in hemolyzed samples (BH adjusted P value <0.05); and (ii) proteins that were not differentially expressed between the disease control and nonresponder groups. The resulting protein list was disregarded for further analyses. The discarded proteins were: CA1, CA2, CFL1, HBA1, HBA2, HBB, HBD, PF4, PPBP, PRDX2, UBB, out of which, CA1, CA2, HBA1, HBB, HBD, PRDX2, and UBB have previously been reported to have strong association with erythrocyte contamination (18). Furthermore, platelet and coagulation markers suggested by Geyer and colleagues (18) were assessed to avoid biases in our biomarker study due to sample handling. No proteins associated with platelet contamination were found. Nevertheless, three coagulation markers (FGA, FGB, and FGG) were observed to be highly correlated (>0.98 Spearman correlation). In addition, these three proteins were also strongly correlated to ARHGAP17, ECM1, EFEMP1, F13A1, F13B, FN1, FP565260.1, VWF, and ZNF462 (Spearman correlation >0.6). The above-mentioned proteins were excluded from further analysis, as they would be unreliable as biomarkers because of their strong association with sample handling. After these filtering steps, 183 proteins were left for further analysis (Supplementary Table S1).

Linear models identify and distinguish potential predictive and prognostic biomarkers related to PFS

The proteomic profiles generated by the analyses of samples before (PRE) and during treatment (TRM) were further disentangled and related to clinical data, in terms of different covariates such as sex, age, stage (AJCC 7th), as well as to PFS with the possibility to be regarded as potential prognostic and/or predictive based on whether the proteins level's relationship with PFS is treatment specific or not. We generated individual linear models for each protein and extracted the amount of explained observed variance related to the above-mentioned covariates for protein levels (Fig. 2A). The percent explained variance refers to the fraction of observed variance in protein expression that is explained by the particular covariate. Furthermore, the linear models retrieve estimates of coefficients of the prognostic (treatment-independent effects) and predictive (treatment-dependent effects) factors, which were used to assess the relationship between protein level and PFS. In total, 43 proteins were found to have a significant relationship between PFS and protein level, either for measurements taken before treatment or during treatment. The results show that 33 proteins including APOA1, APOC1, CRP, SAA2, SAA2-SAA4, and LDHB have significant prognostic impact before treatment regardless of age, sex, and stage (AJCC 7th) except for KDM5B (Fig. 2A, top), which in addition showed significant association to age. Moreover, high levels of CRP, SAA2, SAA2-SAA4, and LDHB were related to unfavorable outcome, whereas high levels of APOA1 and APOC1 were associated with better outcome. LDHB, one isoform of LDH, shows similar correlation to PFS as the total pretreatment LDH levels in both treatment groups (Table 2A). Furthermore, the results showed that 18 of the 43 proteins have significant prognostic impact during treatment; CRP and LDHB still remain significantly prognostic, whereas four proteins (apolipoproteins APOA1, APOA4, APOC1, and LBP) showed a predictive value. Importantly, high levels of APOA1, APOA4, APOC1 are more favorable in ICI in comparison to MAPKi therapy, whereas LBP is more unfavorable in ICI than MAPKi (Fig. 2A; Supplementary Tables S2 and S3). In addition, APOA1 and APOC1 were the only proteins that were found to have both prognostic and predictive effects. They were both prognostic favorable for pretreatment and predictively comparably favorable for ICI during treatment. Of note, these results could also be tested in a case–control study to assess their predictive power regarding the disease stages. Furthermore, since no control group was used, the predictive factor signifies the differential effect between the two treatment types.

Figure 2.

Summary of linear models and ANOVA of singular proteins' relationship to patient outcome for protein levels before treatment (PRE) and during treatment (TRM). A, Top, percentage of observed variance in protein concentration and clinical variables explained by PFS. Treatment-independent effects (prognostic; blue), treatment-dependent effects (predictive; red), age (gray), sex (dark red), and stage (cyan). Significance level in each bar is marked as follows: *, P < 0.05; **, P < 0.01; ***, P < 0.001. Bottom, linear model coefficient estimates. The point denotes the mean estimation of the effect, whereas the line spanning indicates the 95% confidence interval. For the prognostic factor (blue), a positive estimate signifies that higher levels are related to better outcome (longer PFS), whereas a negative estimate signifies that higher levels are related to worse outcome (shorter PFS). For the predictive factor (red), a positive estimate signifies a relationship of high levels and comparably better outcome for patients treated with ICI, whereas a negative estimate signifies a relationship of higher levels and comparably better outcome for patients treated with MAPKi. APOA1, APOA4, and APOC1 are more favorable in ICI, whereas LBP is more favorable in MAPKi. Only significant effects are shown (BH adjusted P < 0.05). B, Ward-clustered heatmap showing the Spearman correlation of the 43 proteins in A. To the right of the heatmap, three tracks show whether the protein is (i) secreted according to the Human Protein Atlas, (ii) cancer-related according to the Human Protein Atlas, (iii) favorable/unfavorable, defined by the prognostic effect in the linear models displayed in A.

Figure 2.

Summary of linear models and ANOVA of singular proteins' relationship to patient outcome for protein levels before treatment (PRE) and during treatment (TRM). A, Top, percentage of observed variance in protein concentration and clinical variables explained by PFS. Treatment-independent effects (prognostic; blue), treatment-dependent effects (predictive; red), age (gray), sex (dark red), and stage (cyan). Significance level in each bar is marked as follows: *, P < 0.05; **, P < 0.01; ***, P < 0.001. Bottom, linear model coefficient estimates. The point denotes the mean estimation of the effect, whereas the line spanning indicates the 95% confidence interval. For the prognostic factor (blue), a positive estimate signifies that higher levels are related to better outcome (longer PFS), whereas a negative estimate signifies that higher levels are related to worse outcome (shorter PFS). For the predictive factor (red), a positive estimate signifies a relationship of high levels and comparably better outcome for patients treated with ICI, whereas a negative estimate signifies a relationship of higher levels and comparably better outcome for patients treated with MAPKi. APOA1, APOA4, and APOC1 are more favorable in ICI, whereas LBP is more favorable in MAPKi. Only significant effects are shown (BH adjusted P < 0.05). B, Ward-clustered heatmap showing the Spearman correlation of the 43 proteins in A. To the right of the heatmap, three tracks show whether the protein is (i) secreted according to the Human Protein Atlas, (ii) cancer-related according to the Human Protein Atlas, (iii) favorable/unfavorable, defined by the prognostic effect in the linear models displayed in A.

Close modal
Table 2.

Univariate Cox regression.

A. Clinical and patient characteristics correlation to response
P value
VariableICIMAPKi
Sex ns 0.001   
Age ns ns   
Stage ns ns   
LDH <0.001 ns   
A. Clinical and patient characteristics correlation to response
P value
VariableICIMAPKi
Sex ns 0.001   
Age ns ns   
Stage ns ns   
LDH <0.001 ns   
B. Clinical and patient characteristics correlation to PFS
ICIMAPKi
VariableHR (95% CI)P valueHR (95% CI)P value
sex 0.95 (0.51–1.76) 0.87 0.74 (0.39–1.39) 0.35 
age 1.02 (0.99–1.04) 0.24 0.98 (0.96–1.01) 0.16 
stage 1.7 (0.86–3.37) 0.13 2.76 (0.83–9.22) 0.1 
LDH 1.27 (1.14–1.42) <0.001 1.04 (1.01–1.07) 0.005 
B. Clinical and patient characteristics correlation to PFS
ICIMAPKi
VariableHR (95% CI)P valueHR (95% CI)P value
sex 0.95 (0.51–1.76) 0.87 0.74 (0.39–1.39) 0.35 
age 1.02 (0.99–1.04) 0.24 0.98 (0.96–1.01) 0.16 
stage 1.7 (0.86–3.37) 0.13 2.76 (0.83–9.22) 0.1 
LDH 1.27 (1.14–1.42) <0.001 1.04 (1.01–1.07) 0.005 

A Spearman correlation analysis retrieves a clear anticorrelated clustering pattern between favorable and unfavorable proteins. The formers include ALB, APOA1, DNAH6, APOC1, APOM, KLKB1, APOA4, SELENOP, GSN, TF, TERT, VPS51, PGLYRP2, FETUB, CLEC3B, and LUM; whereas the unfavorable ones include C9, CRP, LBP, SAA2, SAA2-SAA4, ORM1, SERPINA3, ORM2, S100A8, S100A9, HP, KDM5B, SERPINA1, C3, CFI, C4BPA, C5, ACTB, LDHB, MAN1A1, SERPING1, FCN2, ITIH4, CD44, APOE, and CCDC40 (Fig. 2B, with secretion and cancer gene annotations from the Human Protein Atlas; refs. 19–21). The 43 significant proteins were further investigated to see if a significant difference in protein level could be observed from beginning of treatment to recurrence of disease by performing a paired Wilcoxon test of the protein level between the two time points (Supplementary Fig. S3). This showed that LDHB, FCN2, ORM1, SAA2, and SAA2-SAA4 were significantly increased (BH adjusted P < 0.05) between the beginning of treatment to disease recurrence, whereas FETUB showed a significant decrease. Interestingly, all the proteins that showed an increase also had a significant unfavorable prognostic effect on PFS in the linear models, whereas FETUB showed a significant favorable effect on PFS. Furthermore, FCN2, ORM1, SAA2, and SAA2-SAA4 are all markers of inflammation and/or immune system activation. These two observations combined with an increase of LDHB at the recurrence time point suggests higher tumor load and higher inflammation with disease progression.

In addition, to ascertain to which extent the predictive and prognostic effects we observed were already accounted for by clinical features for disease burden, we classified our patients according to the 8th AJCC (22) staging edition in two groups (M1A/M1B or M1C/M1D), LDH levels (normal or elevated), and the number of metastatic sites (>3 sites or ≤3 sites) and thereafter created two linear models for each of the 43 proteins; one model including both protein expression and clinical data, as well as a reduced model including only clinical features. We then performed an ANOVA F test to assess whether the two models were significantly different, resulting in a total of 15 of 43 unique proteins with a significantly (adjusted P < 0.05) different linear fit when introduced into the model for PRE and/or TRM (Supplementary Table S4). Twelve proteins were significant for PRE (AHSG, C5, CRP, KDM5B, LUM, ORM2, SAA2, SAA2-SAA4, TERT, TF, LDHB, SERPINA3), and five for TRM (CCDC40, FETUB, SERPINA1, LDHB, SERPINA3), suggesting that protein data indeed adds information in addition to clinical features (Supplementary Fig. S4; Supplementary Table S4).

Inflammation and apolipoproteins clusters show inverse correlation and relate to the response and PFS

Because apolipoproteins and inflammation markers were prominent in relation to PFS, we chose to focus on these groups of proteins for further comparison of the three different response groups: (i) complete response (CR); (ii) partial response or stable disease (PR or SD); and (iii) no response (NR). Our data show that in PRE samples the relative expression levels of three apolipoproteins (APOA1, APOC1, APOM1) were higher in patients who had CR, and decreased gradually in patients with PR or SD, or NR, mainly showing significance between CR and NR. Recently a higher body mass index (BMI) has been associated with response to ICI (23), however we found no significant correlation between BMI and apolipoproteins. The three inflammation-associated proteins, CRP, SAA2, and SAA2-SAA4 (24, 25) showed an opposite trend (Fig. 3A). As an internal validation to investigate whether the inflammation markers correspond to clinical measurements, our measured CRP values were correlated with clinically measured CRP, pairing patient samples where clinical CRP had been measured within 14 days upon plasma sampling. The resulting Spearman correlation showed a high correlation (Spearman ρ = 0.82) between the two measurements (Supplementary Figs. S5; Supplementary Table S5).

Figure 3.

Inflammation and apolipoproteins are inversely correlated. A, Levels of apolipoproteins and inflammation markers in subgroups of treatment response. Violin plot of log-scaled pretreatment relative protein level in patients stratified by treatment response (complete response, blue; partial response or stable disease, yellow; nonresponder, red).). Significance level is marked as follows: *, P < 0.05; **, P < 0.01; ***, P < 0.001. B, Spearman correlation between apolipoproteins and inflammation proteins. Ward-clustered (dendrogram not shown) heatmap showing the Spearman correlation between selected proteins. Nonsignificant (P > 0.05) correlations are shown as white. C, Two distinct profiles of apolipoproteins and inflammation. Ward-clustered heatmap of relative pretreatment levels of apolipoproteins and inflammation proteins. Patients cluster in two distinct groups; one of low inflammation and high apolipoproteins (marked dark blue at bottom), and one of high inflammation and low apolipoproteins (marked red at bottom). D, Progression-free survival of the two apolipoproteins/inflammation profiles. Kaplan–Meier plots showing PFS of the two distinct groups for ICI and MAPKi, respectively. Median PFS is marked by a dashed line.

Figure 3.

Inflammation and apolipoproteins are inversely correlated. A, Levels of apolipoproteins and inflammation markers in subgroups of treatment response. Violin plot of log-scaled pretreatment relative protein level in patients stratified by treatment response (complete response, blue; partial response or stable disease, yellow; nonresponder, red).). Significance level is marked as follows: *, P < 0.05; **, P < 0.01; ***, P < 0.001. B, Spearman correlation between apolipoproteins and inflammation proteins. Ward-clustered (dendrogram not shown) heatmap showing the Spearman correlation between selected proteins. Nonsignificant (P > 0.05) correlations are shown as white. C, Two distinct profiles of apolipoproteins and inflammation. Ward-clustered heatmap of relative pretreatment levels of apolipoproteins and inflammation proteins. Patients cluster in two distinct groups; one of low inflammation and high apolipoproteins (marked dark blue at bottom), and one of high inflammation and low apolipoproteins (marked red at bottom). D, Progression-free survival of the two apolipoproteins/inflammation profiles. Kaplan–Meier plots showing PFS of the two distinct groups for ICI and MAPKi, respectively. Median PFS is marked by a dashed line.

Close modal

Driven by this observation, we investigated the relationship between eleven apolipoproteins and four inflammation markers and performed a Spearman correlation analysis of pretreatment protein levels. The results clearly show two distinct anticorrelated clusters: the first includes the four inflammation markers (CRP, SAA1, SAA2, and SAA2-SAA4), whereas the second includes the eleven apolipoproteins (APOA1, APOA4, APOB, APOC1, APOC3, APOC4-APOC2, APOD, APOE, APOF, APOL1, and APOM; Fig. 3B). The two clusters have high intra-cluster correlation but are interestingly inversely correlated to each other.

Furthermore, to investigate if there would be any proteins that could be used as potential biomarker to foresee the response of the patients before any treatment, we performed a clustering analysis of the pretreatment samples (Fig. 3C). The analysis retrieved the very same pattern with two groups of patients: one with high and low relative expression of inflammation proteins and apolipoproteins, respectively, whereas the second presented an inverse trend (Fig. 3C).

Importantly, these two groups show a significant difference in PFS for patients treated with ICI, but not for patients treated with MAPKi in a Kaplan–Meier survival analysis (Fig. 3D). The ICI group had a significantly longer PFS (log-rank P value = 0.0013) for the low inflammation group (median = 349 days), compared with the high inflammation group (median = 73 days). A similar difference was also observed for MAPKi (201 days vs. 132.5 days) but the difference was not statistically significant (log-rank P value = 0.11; Fig. 3D).

Machine learning and predictive analysis

To assess the degree of predictive reliability of the proteins identified as potential biomarkers, we employed a machine learning approach. Nine separate random forest models were built using: (i) pretreatment dataset only (PRE); (ii) during treatment dataset (TRM) only; or (iii) both datasets (PRE + TRM), for both MAPKi and ICI and combination thereof, to predict treatment response (disease control vs. nonresponder) in patients (Fig. 4A). The models were created using the 43 proteins that were significant from the above-mentioned linear models (Supplementary Table S3).

Figure 4.

Machine learning analysis. A, Performance of random forest models for response classification. ROC curves showing the performance of the 43 linear model proteins. ROC curve and ROC-AUC with confidence interval are displayed for each model (PRE, including only pretreatment data; TRM, only during-treatment data; PRE + TRM, including both). B, Variable importance for MAPKi and ICI PRE + TRM models compared. Scatter plot showing the relative importance of variables for the two models (PRE, cyan, pretreatment measurement of protein; TRM, red, during-treatment measurement of protein. Age and stage are displayed in black).

Figure 4.

Machine learning analysis. A, Performance of random forest models for response classification. ROC curves showing the performance of the 43 linear model proteins. ROC curve and ROC-AUC with confidence interval are displayed for each model (PRE, including only pretreatment data; TRM, only during-treatment data; PRE + TRM, including both). B, Variable importance for MAPKi and ICI PRE + TRM models compared. Scatter plot showing the relative importance of variables for the two models (PRE, cyan, pretreatment measurement of protein; TRM, red, during-treatment measurement of protein. Age and stage are displayed in black).

Close modal

For models generated using all datasets, the predictive performances were as follows: PRE AUC = 0.71 (CI, 0.70–0.73), TRM AUC = 0.77 (CI, 0.76–0.78), and PRE-TRM AUC = 0.78 (CI, 0.77–0.79). The results show that the created models have a higher predictive performance for the TRM rather than the PRE models. For ICI, the generated models provide better prediction (reliability) in terms of outcome for the TRM: PRE AUC = 0.67 (CI, 0.65–0.68); TRM AUC = 0.76 (CI, 0.74–0.77); and PRE-TRM AUC = 0.74 (CI, 0.73–0.75).

As far as the MAPKi therapy is concerned, the prediction reliability is higher for the PRE rather than TRM: PRE AUC = 0.71 (CI, 0.68–0.74); TRM AUC = 0.60 (CI, 0.56–0.64); and PRE-TRM AUC = 0.73 (CI, 0.70–0.76). It is worth mentioning that the ROC curves for MAPKi therapy show lower performances compared with ICI, probably due to a limited number of patients in our study. In contrast to the proteomic models' performances, an RF model created using only clinical variables [treatment type, sex, age, and stage (AJCC 7th)] had an AUC of 0.63 (CI, 0.62–0.64). Overall, our data show that the TRM-related models were the most predictive, especially in the ICI group (Supplementary Tables S6–S8).

Given the above-mentioned treatment predictions, we retrieved the predictive power rank of each included predictor [protein, stage (AJCC 7th), age, and sex] using the variable importance of each predictor in each of the treatment specific models (Supplementary Table S9). The importance of each variable was plotted for both MAPKi and ICI: TRM ACTB demonstrated to be highly predictive for both therapies, whereas TRM LDHBand TRM SERPING1, PRE SERPINA1, PRE ALB were highly predictive for ICI and MAPKi, respectively (Fig. 4B). Total LDH analysis showed similar results, although it was pretreatment levels (Table 2B). Importantly, the other covariates such as age, sex, M1 stage (AJCC 7th) have less impact on the prediction of clinical outcome for ICI, which enhances the potential of the above-mentioned proteins in monitoring the clinical treatment. However, sex is an important covariate for the MAPKi therapy, which is supported by the univariate data (Table 2B).

Absolute quantification for technical validation

On the basis of the results described in the previous sections, we performed absolute quantification on 14 proteins, including 8 of the 43 proteins found to be related to PFS using the Stable isotope-labeled standard Protein Epitope Signature Tags (SIS PrESTs) technology (26, 27). The purpose was to investigate how the absolute concentration in the plasma samples of the two groups of patients, responders and nonresponders, differed. Moreover, the absolute values might pave the way to define a threshold that could be used for the stratification of patients for treatments.

The SIS PrESTs were spiked directly into the plasma samples, including six replicates of the reference plasma pool at the beginning of the sample processing to eliminate any source of error deriving from downstream sample processing. The results show a good correlation between absolute values with relative levels (mean Spearman correlation = 0.62, CV = 11.6%; Supplementary Figs. S6 and S7). The mean value of these selected proteins for the responders and nonresponders, sampled before undergoing any therapy are presented in Supplementary Table S9 and Supplementary Fig. S8.

Recent studies suggest that high levels of the inflammation marker CRP correlates with worse clinical outcome following treatment with ICI in patients with cancer, including patients with CMM (28). We show that high CRP is associated with adverse clinical outcome both in ICI and MAPKi treated patients with CMM. To our best knowledge, this is the first study showing reverse correlation between inflammation and expression of apolipoproteins in CMM. Importantly, high pretreatment levels of inflammation markers including SAA2 and SAA2-SAA4 and low levels of apolipoproteins including APOM and APOC1 are associated with worse response and shorter PFS in both MAPKi and ICI therapy in patients with CMM.

Our observation of an inverse correlation between proteins involved in inflammation and apolipoproteins suggests that these two classes of biomolecules might be dependent on each other. It has been described that during acute inflammation the expression of APOA1 in liver is reduced/suppressed, whereas the expression of SAA is increased. When SAA is secreted in the blood it binds to high-density lipoproteins (HDL) replacing APOA1 (29). A SAA enriched HDL, with APOA1 replacement, leads to a rapid excretion of APOA1 (30–32). In line with our findings, it has also been reported that APOM and APOC1 are negative inflammatory proteins and downregulated in tissues and in blood during inflammation (33, 34). Notably, it has been reported that an increased secretion of SAA1 in CMM causes an immunosuppressive tumor environment by increasing IL10 secreting neutrophils in the tumor. This would inhibit natural killer (NK) T cells (35), hence it would further support our findings. APOA1 levels in TRM plasma samples were more favorable for ICI than patients with MAPKi-treated CMM in relation to PFS. The complex APOA1/HDL has been shown to increase CD8 positive T cells and antitumor macrophages while decreasing myeloid-derived suppressor cells (MDSC; ref. 36). Decrease of monocytic MDSCs and increase of CD8 effector memory T-cell frequencies have been shown in patients receiving ICI and this positively correlated with survival (37).

Although elevated levels of SAA proteins have been correlated to worse prognosis (38, 39), there is limited evidence for a potential role of SAA as a predictor biomarker for cancer treatment. Increased SAA has been correlated to poor clinical outcome in patients with non–small cell lung cancer receiving EGFR inhibitor (40). Furthermore, it has been demonstrated that high levels of APOA1 may be a potential predictive factor for better outcome to cancer therapy and chemosensitivity in other tumor types, however to our best knowledge no information is available for CMM (30, 32).

The role of the apolipoproteins is controversial in cancer in general as both increased and decreased levels have been associated with an unfavorable clinical outcome (32, 41). However, in different animal models including melanoma, increased levels of APOA1 and APOC1 have been reported to suppress tumor growth, suggesting a protective role (41, 42).

Increased expression of lactate dehydrogenase (LDH) is generally acknowledged as a significant adverse prognostic factor in CMM and included in the American Joint Committee on Cancer staging system (22). In our study, patients with increased expression of the LDH isoform LDHB had shorter PFS to both treatments, whereas patients with higher expression of LDHB during treatment showed worse response to ICI, but not to MAPKi. This finding could be explained by the fact that lactate (LDHB converts pyruvate into lactate) has an immunosuppressive effect on tumor microenvironment by inactivating CD8+ T cells and thereby counteracting the effect of ICI therapy (43, 44). In line with our findings, it has been reported that decreasing circulating levels of CRP and LDH measured at 12 weeks upon treatment with ipilimumab correlated with better response and longer survival (45).

Low levels of serum albumin, a highly abundant and stable protein in the blood, has been found to be an independent adverse prognostic factor in multiple tumor types (46). Furthermore, albumin has also been shown to benefit as a drug carrier (47). Both BRAF inhibitors, vemurafenib and dabrafenib, bind to serum albumin (48, 49) suggesting that high serum albumin levels may increase the transport of BRAF inhibitor to the tumors. These data are in line with our findings where we present pretreatment albumin as a favorable factor and having high impact on response to MAPKi therapy. Furthermore, our machine learning models suggest that both age and disease-stage play a negligible role in classifying patients as responder or nonresponder for both MAPKi and ICI. However, it is worth mentioning that this observation could be contextual for our study, and further validation on larger cohorts is required to properly evaluate their relevance. Although these findings are promising, they are preliminary and require validation in an independent and larger cohort with available RECIST data where these findings can be more accurately investigated in the context of metrics of tumor burden. There are at this point no independent dataset available using a quantitative plasma proteomics approach in cutaneous melanoma in relation to treatment response, indicating a need for further studies in the area.

Weber and colleagues have previously (50) highlighted 29 proteins with significantly different expression between treatment-resistant and treatment-sensitive patients receiving anti–PD-1. Out of these 29 proteins 13 overlapped with our 43 proteins identified to have a significant relationship between PFS and protein abundance. Furthermore, 12 of 13 proteins overlapped completely in terms of which ones showed an unfavorable versus favorable relationship with response and survival. Both studies found that complement factors (C3, C5, C9, CFI) and acute phase reactants (CRP, HP, ITIH4, LBP, SAA, SERPINA1) to be unfavorable, whereas two proteins related to wound healing (GSN, KLKB1) are favorable. However, Weber and colleagues found APOE to be favorable, whereas we observed it to be unfavorable. It is noteworthy that the similarity between our results remains even though a majority of their patients have previously been treated with ipilimumab, whereas the majority of patients in our cohort are in first line treatment. Moreover, our cohort additionally included patients treated with targeted therapy. The fact that 12 proteins are overlapping in our studies despite the differences between our cohorts lends credibility to the 12 overlapping proteins, whereas these differences could also be an explanation for that many proteins do not overlap. A biomarker for treatment response and/or prognosis should optimally be independent of whether the patient has received previous treatment, thus highlighting the 12 overlapping proteins as particularly interesting candidates for further studies.

It is worth mentioning that the majority of our patient cohort (86%) is treatment naïve, which has important implications: (i) the proteomics profiles are not affected by any previous therapy; and (ii) this enables the direct comparison of targeted- towards immunotherapy and the presence of specific biomarkers associated with clinical outcome. This is of great importance since the OS rates are lower for patients who received ICI therapy after induced MAPKi therapy resistance than when given as first line treatment (51–53). One contributing factor could be that MAPKi therapy has been demonstrated to induce alterations of genes including the IPRES (innate anti-PD-1 resistance) signature (54). Although the majority of patients in the ICI therapy group were treated with anti-PD1 alone (75%), some patients were treated otherwise: three received epacadostat in combination with anti-PD1 within a clinical trial, 10 were treated with anti-CTLA4 alone and three received anti-PD1 in combination with anti-CTLA4. This could indeed have altered the protein expression in this cohort of patients, but since a large majority of patients received similar treatment in our cohort, we believe that this effect is limited.

In conclusion, our untargeted proteomics platform has identified a number of prognostic and predictive biomarkers in plasma from patients with CMM receiving MAPKi or ICI therapy. We present a panel of inflammation- and apolipoproteins as potential pretreatment biomarkers for stratification of patients with CMM. This study would also warrant further and more systematic validation on a larger cohort.

S. Egyhazi Brage reports grants from Cancer Research Funds from Radiumhemmet and Knut and Alice Wallenberg Foundation during the conduct of the study. G. Maddalo reports grants from Radiumhemmets, O.E. och Edla Johanssons Foundation, and Lars Hiertas Minne during the conduct of the study. No disclosures were reported by the other authors.

M.J. Karlsson: Conceptualization, data curation, software, formal analysis, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. F. Costa Svedman: Conceptualization, resources, data curation, writing–original draft, writing–review and editing. A. Tebani: Conceptualization, software, formal analysis, supervision, visualization, methodology, writing–original draft, writing–review and editing. D. Kotol: Validation, investigation. V. Höiom: Formal analysis. L. Fagerberg: Supervision. F. Edfors: Supervision. M. Uhlén: Supervision, funding acquisition. S. Egyhazi Brage: Conceptualization, resources, data curation, supervision, writing–original draft, project administration, writing–review and editing. G. Maddalo: Conceptualization, supervision, funding acquisition, methodology, writing–original draft, project administration, writing–review and editing.

The authors thank Karl-Johan Ekdahl for collection of the plasma samples. The authors thank the oncologists Maria Wolodarski, Hildur Helgadóttir, Giuseppe Masucci, Johan Hansson, Hanna Eriksson, and Johan Falkenius and the nurse Lena Westerberg in helping us with the recruitment of patients for this study. G. Maddalo has been awarded grants from O.E. och Edla Johanssons Vetenskapliga Stiftelse (5310‐7132); Swedish Cancer Society (Radiumhemmets; 174212); and Lars Hierta Memorial Foundation. V. Höiom and S. Egyhazi Brage (co-applicant) have been awarded grants from Cancer Research Funds from Radiumhemmets Forskningsfonder (194103 and 174153). S. Egyhazi Brage has additionally been awarded grants from KI funds. M. Uhlén has been awarded grants from Knut and Alice Wallenberg Foundation (2019.0341). Knut and Alice Wallenberg Foundation (2013.0093) has supported the collection of samples.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

1.
Long
GV
,
Stroyakovskiy
D
,
Gogas
H
,
Levchenko
E
,
de Braud
F
,
Larkin
J
, et al
Combined BRAF and MEK inhibition versus BRAF inhibition alone in melanoma
.
N Engl J Med
2014
;
371
:
1877
88
.
2.
Robert
C
,
Karaszewska
B
,
Schachter
J
,
Rutkowski
P
,
Mackiewicz
A
,
Stroiakovski
D
, et al
Improved overall survival in melanoma with combined dabrafenib and trametinib
.
N Engl J Med
2015
;
372
:
30
9
.
3.
Luebker
SA
,
Koepsell
SA
. 
Diverse mechanisms of BRAF inhibitor resistance in melanoma identified in clinical and preclinical studies
.
Front Oncol
2019
;
9
:
268
.
4.
Robert
C
,
Grob
JJ
,
Stroyakovskiy
D
,
Karaszewska
B
,
Hauschild
A
,
Levchenko
E
, et al
Five-year outcomes with dabrafenib plus trametinib in metastatic melanoma
.
N Engl J Med
2019
;
381
:
626
36
.
5.
Ribas
A
. 
Tumor immunotherapy directed at PD-1
.
N Engl J Med
2012
;
366
:
2517
9
.
6.
Robert
C
,
Schachter
J
,
Long
GV
,
Arance
A
,
Grob
JJ
,
Mortier
L
, et al
Pembrolizumab versus ipilimumab in advanced melanoma
.
N Engl J Med
2015
;
372
:
2521
32
.
7.
Larkin
J
,
Chiarion-Sileni
V
,
Gonzalez
R
,
Grob
JJ
,
Cowey
CL
,
Lao
CD
, et al
Combined nivolumab and ipilimumab or monotherapy in untreated melanoma
.
N Engl J Med
2015
;
373
:
23
34
.
8.
Topalian
SL
,
Hodi
FS
,
Brahmer
JR
,
Gettinger
SN
,
Smith
DC
,
McDermott
DF
, et al
Five-year survival and correlates among patients with advanced melanoma, renal cell carcinoma, or non-small cell lung cancer treated with nivolumab
.
JAMA Oncol
2019
;
5
:
1411
20
.
9.
Larkin
J
,
Chiarion-Sileni
V
,
Gonzalez
R
,
Grob
JJ
,
Rutkowski
P
,
Lao
CD
, et al
Five-year survival with combined nivolumab and ipilimumab in advanced melanoma
.
N Engl J Med
2019
;
381
:
1535
46
.
10.
Hodi
FS
,
Chiarion-Sileni
V
,
Gonzalez
R
,
Grob
JJ
,
Rutkowski
P
,
Cowey
CL
, et al
Nivolumab plus ipilimumab or nivolumab alone versus ipilimumab alone in advanced melanoma (CheckMate 067): 4-year outcomes of a multicentre, randomised, phase 3 trial
.
Lancet Oncol
2018
;
19
:
1480
92
.
11.
Sechidis
K
,
Papangelou
K
,
Metcalfe
PD
,
Svensson
D
,
Weatherall
J
,
Brown
G
. 
Distinguishing prognostic and predictive biomarkers: an information theoretic approach
.
Bioinformatics
2018
;
34
:
4139
.
12.
Grzywa
TM
,
Paskal
W
,
Włodarski
PK
. 
Intratumor and intertumor heterogeneity in melanoma
.
Transl Oncol
2017
;
10
:
956
75
.
13.
Balch
CM
,
Gershenwald
JE
,
Soong
SJ
,
Thompson
JF
,
Atkins
MB
,
Byrd
DR
, et al
Final version of 2009 AJCC melanoma staging and classification
.
J Clin Oncol
2009
;
27
:
6199
206
.
14.
Eisenhauer
EA
,
Therasse
P
,
Bogaerts
J
,
Schwartz
LH
,
Sargent
D
,
Ford
R
, et al
New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1)
.
Eur J Cancer
2009
;
45
:
228
47
.
15.
Leek
JT
,
Johnson
WE
,
Parker
HS
,
Fertig
EJ
,
Jaffe
AE
,
Zhang
Y
, et al
SVA: surrogate variable analysis
. 
2020
.
16.
Ritchie
ME
,
Phipson
B
,
Wu
D
,
Hu
Y
,
Law
CW
,
Shi
W
, et al
limma powers differential expression analyses for RNA-sequencing and microarray studies
.
Nucleic Acids Res
2015
;
43
:
e47
.
17.
Brenes
A
,
Hukelmann
J
,
Bensaddek
D
,
Lamond
AI
. 
Multibatch TMT reveals false positives, batch effects and missing values
.
Mol Cell Proteomics
2019
;
18
:
1967
80
.
18.
Geyer
PE
,
Voytik
E
,
Treit
PV
,
Doll
S
,
Kleinhempel
A
,
Niu
L
, et al
Plasma proteome profiling to detect and avoid sample-related biases in biomarker studies
.
EMBO Mol Med
2019
;
11
:
e10427
.
19.
Uhlen
M
,
Karlsson
MJ
,
Hober
A
,
Svensson
AS
,
Scheffel
J
,
Kotol
D
, et al
The human secretome
.
Sci Signal
2019
;
12
:
eaaz0274
.
20.
Uhlen
M
,
Fagerberg
L
,
Hallstrom
BM
,
Lindskog
C
,
Oksvold
P
,
Mardinoglu
A
, et al
Proteomics. Tissue-based map of the human proteome
.
Science
2015
;
347
:
1260419
.
21.
Uhlen
M
,
Karlsson
MJ
,
Zhong
W
,
Tebani
A
,
Pou
C
,
Mikes
J
, et al
A genome-wide transcriptomic analysis of protein-coding genes in human blood cells
.
Science
2019
;
366
:
eaax9198
.
22.
Gershenwald
JE
,
Scolyer
RA
. 
Melanoma staging: American Joint Committee on Cancer (AJCC) 8th Edition and Beyond
.
Ann Surg Oncol
2018
;
25
:
2105
10
.
23.
Cortellini
A
,
Bersanelli
M
,
Buti
S
,
Cannita
K
,
Santini
D
,
Perrone
F
, et al
A multicenter study of body mass index in cancer patients treated with anti-PD-1/PD-L1 immune checkpoint inhibitors: when overweight becomes favorable
.
J Immunother Cancer
2019
;
7
:
57
.
24.
Kushner
I
. 
The acute phase response: an overview
.
Methods Enzymol
1988
;
163
:
373
83
.
25.
Lin
HY
,
Tan
GQ
,
Liu
Y
,
Lin
SQ
. 
The prognostic value of serum amyloid A in solid tumors: a meta-analysis
.
Cancer Cell Int
2019
;
19
:
62
.
26.
Edfors
F
,
Forsstrom
B
,
Vunk
H
,
Kotol
D
,
Fredolini
C
,
Maddalo
G
, et al
Screening a resource of recombinant protein fragments for targeted proteomics
.
J Proteome Res
2019
;
18
:
2706
18
.
27.
Hober
A
,
Edfors
F
,
Ryaboshapkina
M
,
Malmqvist
J
,
Rosengren
L
,
Percy
AJ
, et al
Absolute quantification of apolipoproteins following treatment with omega-3 carboxylic acids and fenofibrate using a high precision stable isotope-labeled recombinant protein fragments based SRM assay
.
Mol Cell Proteomics
2019
;
18
:
2433
46
.
28.
Iivanainen
S
,
Ahvonen
J
,
Knuuttila
A
,
Tiainen
S
,
Koivunen
JP
. 
Elevated CRP levels indicate poor progression-free and overall survival on cancer patients treated with PD-1 inhibitors
.
ESMO Open
2019
;
4
:
e000531
.
29.
De Buck
M
,
Gouwy
M
,
Wang
JM
,
Van Snick
J
,
Proost
P
,
Struyf
S
, et al
The cytokine-serum amyloid A-chemokine network
.
Cytokine Growth Factor Rev
2016
;
30
:
55
69
.
30.
Georgila
K
,
Vyrla
D
,
Drakos
E
. 
Apolipoprotein A-I (ApoA-I), immunity, inflammation and cancer
.
Cancers
2019
;
11
:
1097
.
31.
van der Westhuyzen
DR
,
de Beer
FC
,
Webb
NR
. 
HDL cholesterol transport during inflammation
.
Curr Opin Lipidol
2007
;
18
:
147
51
.
32.
Ren
L
,
Yi
J
,
Li
W
,
Zheng
X
,
Liu
J
,
Wang
J
, et al
Apolipoproteins and cancer
.
Cancer Med
2019
;
8
:
7032
43
.
33.
Feingold
KR
,
Shigenaga
JK
,
Chui
LG
,
Moser
A
,
Khovidhunkit
W
,
Grunfeld
C
. 
Infection and inflammation decrease apolipoprotein M expression
.
Atherosclerosis
2008
;
199
:
19
26
.
34.
Fuior
EV
,
Gafencu
AV
. 
Apolipoprotein C1: its pleiotropic effects in lipid metabolism and beyond
.
Int J Mol Sci
2019
;
20
:
5939
.
35.
De Santo
C
,
Arscott
R
,
Booth
S
,
Karydis
I
,
Jones
M
,
Asher
R
, et al
Invariant NKT cells modulate the suppressive activity of IL-10-secreting neutrophils differentiated with serum amyloid A
.
Nat Immunol
2010
;
11
:
1039
46
.
36.
Zamanian-Daryoush
M
,
Lindner
D
,
Tallant
TC
,
Wang
Z
,
Buffa
J
,
Klipfell
E
, et al
The cardioprotective protein apolipoprotein A1 promotes potent anti-tumorigenic effects
.
J Biol Chem
2013
;
288
:
21237
52
.
37.
de Coaña
YP
,
Wolodarski
M
,
Poschke
I
,
Yoshimoto
Y
,
Yang
Y
,
Nyström
M
, et al
Ipilimumab treatment decreases monocytic MDSCs and increases CD8 effector memory T cells in long-term survivors with advanced melanoma
.
Oncotarget
2017
;
8
:
21539
53
.
38.
Findeisen
P
,
Zapatka
M
,
Peccerella
T
,
Matzk
H
,
Neumaier
M
,
Schadendorf
D
, et al
Serum amyloid A as a prognostic marker in melanoma identified by proteomic profiling
.
J Clin Oncol
2009
;
27
:
2199
208
.
39.
Moshkovskii
SA
. 
Why do cancer cells produce serum amyloid A acute-phase protein?
Biochemistry
2012
;
77
:
339
41
.
40.
Milan
E
,
Lazzari
C
,
Anand
S
,
Floriani
I
,
Torri
V
,
Sorlini
C
, et al
SAA1 is over-expressed in plasma of non small cell lung cancer patients with poor outcome after treatment with epidermal growth factor receptor tyrosine-kinase inhibitors
.
J Proteomics
2012
;
76 Spec No.
:
91
101
.
41.
Zamanian-Daryoush
M
,
DiDonato
JA
. 
Apolipoprotein A-I and cancer
.
Front Pharmacol
2015
;
6
:
265
.
42.
Sun
Y
,
Zhang
J
,
Guo
F
,
Zhao
W
,
Zhan
Y
,
Liu
C
, et al
Identification of apolipoprotein C-I peptides as a potential biomarker and its biological roles in breast cancer
.
Med Sci Monit
2016
;
22
:
1152
60
.
43.
Mishra
D
,
Banerjee
D
. 
Lactate dehydrogenases as metabolic links between tumor and stroma in the tumor microenvironment
.
Cancers
2019
;
11
:
750
.
44.
Felix
J
,
Cassinat
B
,
Porcher
R
,
Schlageter
MH
,
Maubec
E
,
Pages
C
, et al
Relevance of serum biomarkers associated with melanoma during follow-up of anti-CTLA-4 immunotherapy
.
Int Immunopharmacol
2016
;
40
:
466
73
.
45.
Simeone
E
,
Gentilcore
G
,
Giannarelli
D
,
Grimaldi
AM
,
Caracò
C
,
Curvietto
M
, et al
Immunological and biological changes during ipilimumab treatment and their potential correlation with clinical response and survival in patients with advanced melanoma
.
Cancer Immunol Immunother
2014
;
63
:
675
83
.
46.
Gupta
D
,
Lis
CG
. 
Pretreatment serum albumin as a predictor of cancer survival: a systematic review of the epidemiological literature
.
Nutr J
2010
;
9
:
69
.
47.
Hoogenboezem
EN
,
Duvall
CL
. 
Harnessing albumin as a carrier for cancer therapies
.
Adv Drug Deliv Rev
2018
;
130
:
73
89
.
48.
Kim
G
,
McKee
AE
,
Ning
YM
,
Hazarika
M
,
Theoret
M
,
Johnson
JR
, et al
FDA approval summary: vemurafenib for treatment of unresectable or metastatic melanoma with the BRAFV600E mutation
.
Clin Cancer Res
2014
;
20
:
4994
5000
.
49.
Suo
Z
,
Xiong
X
,
Sun
Q
,
Zhao
L
,
Tang
P
,
Hou
Q
, et al
Investigation on the interaction of dabrafenib with human serum albumin using combined experiment and molecular dynamics simulation: exploring the binding mechanism, esterase-like activity, and antioxidant activity
.
Mol Pharm
2018
;
15
:
5637
45
.
50.
Weber
JS
,
Sznol
M
,
Sullivan
RJ
,
Blackmon
S
,
Boland
G
,
Kluger
HM
, et al
A serum protein signature associated with outcome after anti-PD-1 therapy in metastatic melanoma
.
Cancer Immunol Res
2018
;
6
:
79
86
.
51.
Kreft
S
,
Gesierich
A
,
Eigentler
T
,
Franklin
C
,
Valpione
S
,
Ugurel
S
, et al
Efficacy of PD-1-based immunotherapy after radiologic progression on targeted therapy in stage IV melanoma
.
Eur J Cancer
2019
;
116
:
207
15
.
52.
Amini-Adle
M
,
Khanafer
N
,
Le-Bouar
M
,
Duru
G
,
Dalle
S
,
Thomas
L
. 
Ineffective anti PD-1 therapy after BRAF inhibitor failure in advanced melanoma
.
BMC Cancer
2018
;
18
:
705
.
53.
Carlino
MS
,
Long
GV
,
Schadendorf
D
,
Robert
C
,
Ribas
A
,
Richtig
E
, et al
Outcomes by line of therapy and programmed death ligand 1 expression in patients with advanced melanoma treated with pembrolizumab or ipilimumab in KEYNOTE-006: a randomised clinical trial
.
Eur J Cancer
2018
;
101
:
236
43
.
54.
Hugo
W
,
Zaretsky
JM
,
Sun
L
,
Song
C
,
Moreno
BH
,
Hu-Lieskovan
S
, et al
Genomic and transcriptomic features of response to anti-PD-1 therapy in metastatic melanoma
.
Cell
2016
;
165
:
35
44
.