Gene-expression signature predicts autoimmune toxicity in metastatic melanoma

Abstract

Objectives

To identify predictive gene-expression signatures for immune-related adverse events (irAEs) in patients with melanoma treated with anti-PD-1 inhibitors, in the adjuvant therapy (AT) and first-line therapy (FLT).

Methods

This retrospective study analyzed baseline whole-blood gene expression profile from 161 patients with resected stage III or unresectable stage III-IV melanoma treated with anti-PD-1 inhibitors. RNA was extracted from baseline peripheral blood samples and profiled using the NanoString nCounter PanCancer IO 360 panel. Gene-expression signatures were identified and validated using cross-validated sparse partial least squares modeling and principal component analysis, then correlated with toxicity occurrence.

Results

A total of 223 and 186 irAEs were observed in the AT and FLT groups, respectively, including arthralgia, colitis, and headache. Distinct gene-expression signatures significantly predicted toxicity occurrence, with variation across therapy settings. Arthralgia was predicted by immune-related and apoptotic gene signatures (eg, SMAD5, FASLG in FLT; ICOS, TGFB2 in AT), while colitis was linked to inflammatory and adhesion-related pathways. In the AT group, headache was associated with genes involved in interferon and adhesion signaling. Across both cohorts, specific signatures predicted overall irAE risk and timing. No events were observed in patients with low-risk signatures over the follow-up period. In the FLT cohort, arthralgia and cutaneous toxicities were positively associated with ORR, while arthralgia, asthenia, colitis, fatigue, and skin-related toxicities correlated with improved disease control rate. No significant association between irAEs and relapse risk was observed in the adjuvant cohort.

Conclusions

Whole-blood gene-expression profiling enables early identification of patients at high risk for irAEs during anti-PD-1 therapy. These predictive biomarkers may guide personalized toxicity monitoring in melanoma treatment.

WHAT IS ALREADY KNOWN ON THIS TOPIC.

WHAT THIS STUDY ADDS

• This study identifies gene-expression signatures predictive of the occurrence and earliness of any and specific irAEs, such as arthralgia, colitis, and headache, in patients treated with anti-PD-1 inhibitors, highlighting distinct mechanisms across treatment settings (adjuvant and first-line therapy).

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• Predictive gene-expression signatures could enable personalized management strategies, reducing irAEs and improving patient outcomes, with potential integration into clinical protocols.

Background

Melanoma is an immunological malignancy with a higher prevalence in immunocompromised patients.¹ Immune checkpoint inhibitors (ICIs), designed to block inhibitory signals of T-cell activation and promote antitumor immune responses, represent a transformative therapeutic strategy that has revolutionized the management of many cancers, particularly advanced metastatic melanoma.

ICIs, specifically anti-CTLA4 and anti-PD-1 inhibitors, have significantly improved outcomes for patients with advanced melanoma and are now considered a standard of care. However, a considerable proportion of treated patients do not achieve long-term benefits and eventually lose tumor control.²

The anti-PD-1 inhibitors nivolumab and pembrolizumab have shown significant efficacy in advanced metastatic melanoma, as demonstrated in trials such as CA209-003,^{3 4} KEYNOTE-001,⁵ and KEYNOTE-006,⁶ with favorable long-term safety profiles.⁷ Their use in the adjuvant setting has been supported by trials in resected stage III–IV melanoma (CheckMate-238,⁸ KEYNOTE-054)⁹ and extended to stage IIB–C melanoma (CheckMate-76K,¹⁰ KEYNOTE-716).^{11 12}

Combination regimens involving ICIs have further improved outcomes. The CheckMate 067 trial^{13 14} confirmed long-term survival benefits with nivolumab and ipilimumab. The OpACIN¹⁵ and OpACIN-neo¹⁶ trials evaluated adjuvant and neoadjuvant dosing strategies, identifying regimen B (ipilimumab 1 mg/kg+nivolumab 3 mg/kg) as better tolerated (20% grade 3–4 immune-related adverse events (irAEs)) compared with other arms.¹⁷ The CheckMate 511 trial^{18 19} supported a safer inverted dose combination.

Additionally, novel ICI combinations have emerged. The Relativity-047 trial^{20 21} showed promising results for nivolumab+relatlimab. Triplet therapies are under investigation, including nivolumab+relatlimab+ipilimumab (Relativity-048),^{22 23} spartalizumab+dabrafenib+trametinib (COMBI-I),²⁴ atezolizumab+vemurafenib+cobimetinib (IMspire150),²⁵ and pembrolizumab+trametinib+dabrafenib (KEYNOTE-022).²⁶

While combination therapies have demonstrated enhanced efficacy, they are also associated with an increased incidence of irAEs compared with single-agent treatments.²⁷ IrAEs can affect any organ system, with most being mild and manageable, though some can be life-threatening. Severe irAEs reduce clinical benefits, often necessitate treatment discontinuation—particularly in older patients— and negatively impact overall patient quality of life.²⁸

A comprehensive understanding of the toxicity profiles of ICIs and the frequency of irAE occurrences is essential for early detection and improved management.

Recent studies have reported baseline toxicity markers or transcriptomic, genetic, or cellular toxicity markers. Baseline gene expression differences in key immune pathways were identified in peripheral blood T cells from COMBO-treated patients who developed grade 3–5 irAEs, including a SYK-related gene signature that correctly classified ~60% of such cases, in line with prior findings linking anti-CTLA4 irAEs to a germline variant associated with elevated SYK expression.²⁹

In another study, lichenoid dermatitis (LD) irAE was characterized by activation of the CD14/TLR innate immune pathway, with increased frequencies of CD14⁺ and CD16⁺ monocytes compared with BLK controls, suggesting that CD14/TLR signaling may contribute to LD-irAE pathogenesis.³⁰

However, biomarkers to predict with high accuracy which patients are at risk of developing irAEs are currently unavailable.² This study aims to identify gene-expression signature models as predictive tools for toxicities in patients treated with anti-PD-1 inhibitors as both adjuvant therapy (AT) and first-line therapy (FLT).

Patients and methods

Patients

The analysis included consecutive adult patients with histologically confirmed malignant melanoma. Patients with resected stage III melanoma treated with an anti-PD-1 agent as AT and those with unresectable metastatic melanoma (stage IIIb–IV per the American Joint Committee on Cancer seventh Edition) treated with an anti-PD-1 agent as FLT between April 2016 and July 2020 were eligible. All participants provided written informed consent.

Survival outcomes measures

Tumor response was radiologically evaluated using RECIST V.1.1 criteria, classifying outcomes as complete response (CR), partial response (PR), stable disease (SD) or progressive disease (PD). The disease control rate (DCR) was defined as the proportion of patients achieving CR, PR, or SD lasting longer than 1 year. The objective response rate (ORR) was defined as the proportion of patients achieving CR or PR.

Patients reported AEs, and they were systematically recorded during each follow-up visit.

Bio-humoral analysis

Samples were collected prior to ICI treatment in both groups. Baseline serum lactate dehydrogenase (LDH) levels were assessed. Peripheral blood samples were collected at baseline from patients treated with an anti-PD-1 agent for gene expression profile analysis. RNA was extracted from whole blood using the QIAamp RNA Blood Mini Kit (Qiagen). Quantification and purity were assessed with a NanoDrop spectrophotometer (Thermo Fisher Scientific).

Purified RNA was hybridized and analyzed using the NanoString nCounter system with the PanCancer IO 360 panel, which includes 770 human genes involved in tumor microenvironment-immune response interactions. Gene expression data were normalized with nSolver Version 4.0 software, using External RNA Controls Consortium technical controls and 30 housekeeping genes.

Statistical analysis

Continuous variables were reported as mean±SD or median (IQR), depending on their distribution, assessed using the Shapiro-Wilk normality test. Categorical variables were expressed as percentages. Differences in patient characteristics between those with and without toxicity were analyzed using the t-test or Wilcoxon test for continuous variables and Pearson’s χ² test for categorical variables. Linear associations between continuous variables were evaluated using Pearson’s correlation coefficient for normally distributed data or Spearman’s correlation coefficient for non-normal distributions.

The optimal cut-off for defining the boundary between the absence and presence of toxicity was determined using a cross-validation approach to maximize the area under the curve of the receiver operating characteristic (ROC). To validate the selection of genes most associated with the sparse partial least squares components, principal component analysis was performed. Genes with the lowest explained variance on the first principal component analysis component were excluded.

Gene-expression signatures were derived through a rigorous cross-validation process with 5000 replicates, specifically designed to minimize classification error and enhance the reproducibility of the results. This cross-validation approach serves as an internal validation strategy, reducing the risk of overfitting and ensuring the robustness of the identified signatures.

Toxicity-free survival was defined as the time from the initiation of systemic therapy to either toxicity occurrence or the last patient contact (censored). For both AT and FLT groups, differences between subgroups with low and high signature values were analyzed over at least 54 months using a log-rank test and visualized with Kaplan-Meier curves. All sampling time points were collected at baseline.

For the classification into “low” and “high” groups, the algorithm used for gene profiling automatically determines an optimal cut-point for each signature score. This cut point is identified by optimizing the association between the signature and the outcome, ensuring that the stratification of patients is statistically significant. The weighted contribution of each gene to the overall signature score was considered to capture both the magnitude and direction of their effect, thereby ensuring that the classification of “high” and “low” groups reflects the combined influence of multiple genes.

Raw data generated in this study are available in Zenodo: https://doi.org/10.5281/zenodo.14102423.

Results

Characteristics at baseline

Among the 161 patients included in the analysis, 75 received anti-PD-1 therapy as AT and 86 as FLT. Demographic and baseline clinical data are summarized in table 1. Overall, 88 patients (55%) were male, and the median age was 60 years.

Table 1. Patient characteristics.

Characteristics	AT (n=75)	FLT (n=86)	Overall (n=161)
Age (years)	55 (19.23)	65 (21.14)	60 (24.08)
Sex (male)	44 (59%)	44 (51%)	88 (55%)
BMI	27 (6.44)	22 (3.86)	24 (6.20)
BMI unavailable	3	1	4
BRAF
Wild-type	38 (60%)	65 (79%)	103 (71%)
Mutation	25 (40%)	17 (21%)	42 (29%)
Unknown	12	4	16
Lactate dehydrogenase
Normal	50 (86%)	34 (56%)	84 (71%)
High	8 (14%)	27 (44%)	35 (29%)
Unknown	17	25	42
M-category
M0	33 (92%)	3 (4%)	36 (30%)
M1a	0 (0%)	12 (14%)	12 (10%)
M1b	1 (3%)	13 (15%)	14 (12%)
M1c	1 (3%)	58 (67%)	59 (48%)
M1d	1 (3%)	0 (0%)	1 (1%)
Unknown	39	0	39
Average glycemia	96 (22)	95.5 (24)	96 (23.5)
Unknown	14	10	24
Central nervous system	0 (0%)	20 (23%)	20 (12%)
Disease control rate	–	45 (52%)	45 (52%)
Objective response rate	–	27 (31%)	27 (31%)
Relapse	29 (39%)	–	29 (39%)

Values are reported as median (IQR) or n (%).

AT, adjuvant therapy; BMI, body mass index; FLT, first-line therapy.

In the AT group, 44 patients (59%) were male, with a median age of 55 years. BRAF wild-type melanoma was observed in 38 patients (60%). In the FLT group, 44 patients (51%) were male, with a median age of 65 years; 65 patients (79%) had BRAF wild-type melanoma, and 20 patients (12%) had brain metastases.

A total of 42 patients (29%) had BRAF mutations, including 25 (40%) in the AT group and 17 (21%) in the FLT group. BRAF status was unknown for 16 patients, distributed as 12 in the AT group and four in the FLT group. The median body mass index (BMI) was 24 in the overall population, 27 in the AT group, and 22 in the FLT group. BMI was unavailable for four patients (three in the AT group and one in the FLT group). Glycemia levels were comparable between the groups, with a mean of 96.00 mg/dL in the overall cohort, 96.00 mg/dL in the AT group, and 95.50 mg/dL in the FLT group.

Baseline serum LDH levels were normal in 84 patients (71%): 50 (86%) in the AT group and 34 (56%) in the FLT group. Elevated LDH levels were observed in 35 patients (29%), including eight (14%) in the AT group and 27 (44%) in the FLT group. LDH levels were unknown for 42 patients (17 in the AT group and 25 in the FLT group).

Outcomes and toxicity

In the FLT group, the DCR was 52% (45 patients), and the ORR was 31% (27 patients). Relapse occurred in 39% (29 patients) of the AT group. In the FLT cohort, we observed that certain irAEs—specifically arthralgia and cutaneous toxicities—were positively associated with ORR. Furthermore, arthralgia, asthenia, colitis, fatigue, and skin-related toxicities were associated with improved DCR. No significant association between the occurrence of irAEs and relapse risk was observed in the adjuvant cohort. For further details, see online supplemental tables S1–S3.

Toxicities were observed as follows: arthralgia in 27% of patients in both the AT (20 patients) and FLT (23 patients) groups (p=1.00); colitis in 21% of AT patients (16 patients) and 16% of FLT patients (14 patients) (p=0.43); and headache in 16% of AT patients (12 patients) compared with 4% of FLT patients (three patients) (p=0.01). Additional data on all investigated toxicities are available in online supplemental table S4. The majority (approximately 90%) of patients who experienced irAEs developed grade 1 or 2 events, while grade 3 or 4 toxicities were observed in approximately 10% of cases (any toxicity). Detailed ROC curves for determining the optimal cut-off for toxicity detection are presented in online supplemental figures S1–S16.

In both AT and FLT groups, the occurrence of colitis was not significantly correlated with clinical variables, including age, sex, BMI, BRAF mutation, LDH levels, and glycemia. However, a significant positive correlation was observed with a specific gene-expression signature (OR=4.26 (95% CI: 1.68 to 17.53), p=0.01 in the AT group; OR=11.89 (95% CI: 2.66 to 239.24), p=0.02 in the FLT group). For additional details, see online supplemental figures S17 and S18.

In the AT group, headache occurrence was significantly correlated with a specific gene-expression signature (OR=4.26 (95% CI: 1.68 to 17.53), p=0.01), while no significant correlation was found with other clinical variables (online supplemental figure S19).

Arthralgia in the AT group was not significantly correlated with clinical variables, although trends were observed: a negative correlation with sex (OR=0.08 (95% CI: 0.00 to 0.84)) and a positive correlation with BMI (OR=1.38 (95% CI: 1.02 to 2.18), p=0.08).³¹ Similarly, no significant correlation between a specific gene-expression signature and arthralgia was found in the AT group, but a positive trend was noted (OR=3.07 (95% CI: 1.10 to 18.51), p=0.09). In the FLT group, arthralgia was significantly correlated with a specific gene-expression signature (OR=3.98 (95% CI: 1.56 to 16.39), p=0.02). No significant correlations were observed between arthralgia and other clinical variables, except for a negative trend with BRAF mutation (OR=0.05 (95% CI: 0.00 to 0.79), p=0.07). Additional details are available in online supplemental figures S20 and S21.

The transcriptomic analysis of peripheral blood mononuclear cells (PBMCs) obtained at baseline identified distinct sets of optimized gene-expression signatures associated with the occurrence of arthralgia, colitis, and headache in both the AT and FLT groups.

In the AT group, the gene-expression signature predictive of arthralgia included the following representative genes: TGFB2, ICOS, CD40LG, CCR4, and ALDOC. All five genes showed a positive correlation with arthralgia occurrence (figure 1A describes the direction of each gene in the signature). Time-dependent analysis revealed that patients with a low signature value maintained an arthralgia-free survival probability above 75% throughout the study period. In contrast, patients with a high signature value experienced a sharp decline in arthralgia-free survival probability, dropping to 25% within the first 12 months (HR=0.17 (95% CI: 0.07 to 0.42), p<0.0001, figure 1B). Heatmap analysis further demonstrated a correlation within the gene-expression signature, specifically between CD40LG and ICOS (figure 1C).

Figure 1. Investigations of arthralgia occurrence in the adjuvant group. (A) Transcriptomic analysis of peripheral blood mononuclear cells obtained at baseline identified a set of representative genes predicting the occurrence of arthralgia. (B) Arthralgia-free survival for high and low gene signature. (C) Heat map representation. Color gradient toward red indicates stronger gene interaction or coexpression, whereas a gradient toward green denotes weaker interaction.

In the FLT group, the gene-expression signature predictive of arthralgia comprised the genes RB1, SMAD5, FOSL1, FCAR, and FASLG. Among these, increased expression of SMAD5 and FASLG was significantly associated with arthralgia occurrence (figure 2A describes the direction of each gene in the signature). Over time, arthralgia-free survival probability in patients with a low signature value was consistently above 80% despite a slight initial decrease. Conversely, in patients with a high signature value, the probability fell below 50% by 18 months (HR=0.06 (95% CI: 0.01 to 0.48), p=0.00037, figure 2B). Heatmap analysis revealed gene correlations within the signature, including SMAD5 with FASLG and FCAR with RB1 and FOSL1 (figure 2C).

Figure 2. Investigations of arthralgia occurrence in the metastatic group. (A) Transcriptomic analysis of peripheral blood mononuclear cells obtained at baseline identified a set of representative genes predicting the occurrence of arthralgia. (B) Arthralgia-free survival for high and low gene signature. (C) Heat map representation. Color gradient toward red indicates stronger gene interaction or coexpression, whereas a gradient toward green denotes weaker interaction.

To predict the occurrence of colitis, a specific gene-expression signature was identified in the AT group, consisting of VEGFB, KRAS, ICAM2, CDKN1A, and CD48 genes. Among these, the expression of KRAS showed a negative correlation with colitis occurrence (figure 3A describes the direction of each gene in the signature). Over time, colitis-free survival in patients with high signature values progressively declined, with the probability falling below 50% by 39 months. In contrast, the probability of colitis-free survival was consistently 100% in the subgroup with low signature values (HR not reached, figure 3B). Heatmap analysis revealed correlations among all genes in the signature, except for KRAS (figure 3C).

Figure 3. Investigations of colitis occurrence in the adjuvant group. (A) Transcriptomic analysis of peripheral blood mononuclear cells obtained at baseline identified a set of representative genes predicting the occurrence of arthralgia. (B) Colitis-free survival for high and low gene signature. (C) Heat map representation. Color gradient toward red indicates stronger gene interaction or coexpression, whereas a gradient toward green denotes weaker interaction.

In the FLT group, the gene-expression signature included TNFSF9, TGFB2, LILRA3, IFIT2, and EIF2AK2. For all of these genes, lower expression levels were correlated with colitis occurrence (figure 4A describes the direction of each gene in the signature). Over time, colitis-free survival in the high signature value subgroup steadily declined, with the probability reaching 50% at 39 months. Conversely, in the low signature value subgroup, colitis-free survival was consistently 100% (HR not reached, figure 4B). Heatmap analysis demonstrated correlations within the signature, particularly between EIF2AK2 and IFIT2, LILRA3 and TGFB2 (figure 4C).

Figure 4. Investigations of colitis occurrence in the metastatic group. (A) Transcriptomic analysis of peripheral blood mononuclear cells obtained at baseline identified a set of representative genes predicting the occurrence of arthralgia. (B) Colitis-free survival for high and low gene signature. (C) Heat map representation. Color gradient toward red indicates stronger gene interaction or coexpression, whereas a gradient toward green denotes weaker interaction.

Moreover, associations of colitis predictive signatures with BRAF and KRAS mutations were analyzed in both settings; however, no associations could be detected concerning BRAF mutations, whereas a negative correlation with KRAS mutation could be observed (for further details, see online supplemental figure S22).

A specific gene-expression signature was identified to predict the occurrence of headache in the AT group, characterized by the representative genes TMEM173, SYK, ICAM2, CD84, and CD244. Among these, lower expression levels of SYK were correlated with headache occurrence (figure 5A describes the direction of each gene in the signature). Time-dependent analysis revealed that in patients with high signature values, the probability of headache-free survival declined below 50% within 42 months. Conversely, in patients with low signature values, the probability of headache-free survival was consistently 100% (HR not reached; figure 5B). Heatmap analysis showed correlations within the signature, specifically between CD244 and CD84, ICAM2, and TMEM173 (figure 5C).

Figure 5. Investigations of headache occurrence in the adjuvant group. (A) Transcriptomic analysis of peripheral blood mononuclear cells obtained at baseline identified a set of representative genes predicting the occurrence of arthralgia. (B) Headache-free survival for high and low gene signature. (C) Heat map representation. Color gradient toward red indicates stronger gene interaction or coexpression, whereas a gradient toward green denotes weaker interaction.

Gene-expression signatures were also correlated with the occurrence of other anti-PD-1 treatment-related toxicities, including asthenia, fatigue, hypothyroidism, pancreatitis, and cutaneous AEs. Detailed results are presented in online supplemental figures S23–S40.

Furthermore, a significant positive correlation was observed between specific gene signatures and the occurrence of any toxicity: OR=5.58 (95% CI: 1.43 to 43.12), p=0.04 for the AT group, and OR=1.91 (95% CI: 1.10 to 3.83), p=0.04 for the FLT group (for further details, see online supplemental figures S41 and S42).

From the transcriptomic analysis of PBMCs obtained at baseline, the gene signature for the AT group included PVR, LAMB3, IL2RA, EIF2B4, and C1QB. Lower expression of all genes in this signature was associated with the occurrence of any toxicity (figure 6A describes the direction of each gene in the signature). Time-dependent analysis revealed that in patients with a low signature value, the probability of toxicity-free survival fell below 50% within 6 months and declined further to 10% over 36 months. Conversely, in patients with a high signature value, the probability was consistently high (HR=0.52 (95% CI: 0.31 to 0.88), p<0.001, figure 6B). Heatmap analysis showed correlations among all genes in the signature, except for PVR with IL2RA and LAMB3 with EIF2B4 (figure 6C).

Figure 6. Investigations of any toxicity occurrence in the adjuvant group. (A) Transcriptomic analysis of peripheral blood mononuclear cells obtained at baseline identified a set of representative genes predicting the occurrence of any toxicity. (B) Toxicity-free survival for high and low gene signature. (C) Heat map representation. Color gradient toward red indicates stronger gene interaction or coexpression, whereas a gradient toward green denotes weaker interaction.

In the FLT group, the gene signature consisted of CD3G, GIMAP6, LDHB, MAP3K8, and RPL23. Higher expression of all genes in this signature was associated with the occurrence of any toxicity (figure 7A describes the direction of each gene in the signature). Over time, a significant difference in toxicity-free survival was observed between subgroups with low and high signature values. In the high signature value subgroup, the probability of toxicity-free survival dropped to 10% within 12 months (HR=0.54 (95% CI: 0.31 to 0.94), p<0.003, figure 7B). Heatmap analysis demonstrated correlations within the signature, with the exception of MAP3K8. Specific correlations were observed among RPL23 with CD3G, GIMAP6, and LDHB and among GIMAP6 with CD3G and LDHB (figure 7C).

Figure 7. Investigations of any toxicity occurrence in the metastatic group. (A) Transcriptomic analysis of peripheral blood mononuclear cells obtained at baseline identified a set of representative genes predicting the occurrence of any toxicity. (B) Toxicity-free survival for high and low gene signature. (C) Heat map representation. Color gradient toward red indicates stronger gene interaction or coexpression, whereas a gradient toward green denotes weaker interaction.

Discussion

With extensive experience in gene profile analysis,32,37 this study aimed to investigate the correlation between specific gene-expression signatures and the occurrence of inflammatory toxicities, including arthralgia, colitis, and headache, in two cohorts of patients treated with anti-PD-1 inhibitors.

A higher number of toxicity events was observed in the AT group (223 events, 55%) compared with the FLT group (186 events, 45%). Since both cohorts received the same treatment (anti-PD-1 inhibitors), patient characteristics and tumor etiology likely contributed to the differential incidence of toxicities. In the FLT patient group, a secretion of immunomodulatory factors capable of modulating pathways associated with the occurrence of AEs might occur.

A greater number of genes associated with immune activation pathways were identified in the gene-expression signatures of the AT group (20 genes) compared with the FLT group (12 genes). This difference is presumed to reflect the stronger immune-suppressive mechanisms in the FLT group, which may also promote metastasis progression.³⁸

For the same AEs, distinct gene-expression signatures were identified between the two settings, suggesting different activation pathways leading to toxicity. The identified gene-expression signatures included diverse functional categories: immune activation (20 genes in AT vs 12 genes in FLT), inflammation (12 genes in both AT and FLT), apoptosis (three genes in AT vs one gene in FLT), and tumor-related mechanisms (16 genes in both AT and FLT).

Concerning the observed lack of overlap between AT and FLT for the same toxicities, several biological and methodological factors could contribute to this phenomenon. AT and FLT represent different treatment regimens, each of which may induce toxicity through distinct molecular pathways. Even if the clinical manifestation of toxicity is similar, the underlying biological drivers could differ between the two treatments. Moreover, the response to toxicity may be modulated by treatment-specific factors, such as drug metabolism, immune response, or cellular stress mechanisms, leading to unique gene-expression patterns in each group.

The probability of developing arthralgia in patients with resected stage III melanoma was positively correlated with BMI and less likely in male patients. Although the correlation between arthralgia occurrence and the specific gene-expression signature identified was not statistically significant, the marked differences in arthralgia-free survival curves between low-signature and high-signature value subgroups indicated that the gene-expression signature significantly influenced the timing of toxicity occurrence.

The gene-expression signature associated with arthralgia in the AT group consisted of five genes: TGFB2, ICOS, ALDOC, CD40LG, and CCR4, all of which were identified as risk factors for developing arthralgia.

Gene correlation analysis within the signature (online supplemental figure S43) revealed three pathways contributing to arthralgia occurrence in the AT group: the TGFβ pathway, the inducible T-cell costimulator pathway (ICOS), and the G-protein cascade (for further details, see online supplemental figure S44). These findings suggest that distinct molecular mechanisms underlie the development of arthralgia in this patient population.

In patients with unresectable metastatic melanoma (FLT group), a specific gene-expression signature was significantly correlated with the probability of arthralgia occurrence and the timing of its occurrence. This signature included the genes RB1,³⁹ SMAD5,^{40 41} FOSL1,^{42 43} FCAR, and FASLG.⁴⁴ Among these, SMAD5 and FASLG emerged as risk factors for the development of arthralgia, whereas patients with BRAF mutations exhibited a protective trend.

Gene correlation analysis (online supplemental figure S45) revealed three pathways potentially contributing to the occurrence of arthralgia in the FLT group: the TGFβ pathway, the HSP90 pathway, and the extrinsic apoptotic pathway (for further details, see online supplemental figure S46).

In both AT and FLT patient groups, specific gene-expression signatures were significantly correlated with the occurrence of colitis. Throughout the observation period, no colitis events were recorded in patients with low signature values in either group. In the AT group, the gene-expression signature consisted of VEGFB,⁴⁵ KRAS,⁴⁶ ICAM2, CDKN1A,⁴⁷ and CD48. Among these, KRAS was identified as a protective factor against the development of colitis.

Thus, KRAS, nuclear adhesion, and immune system activation pathways were identified to lead to the occurrence of colitis for the AT group (for further details, see online supplemental figure S48).

For additional investigation on the potential influence of specific melanoma mutations in the immune activation background, associations of colitis predictive signatures with BRAF and KRAS mutations were analyzed in both settings; however, no associations could be detected concerning BRAF mutations, whereas a negative correlation with KRAS mutation could be confirmed, which increases the risk of developing colitis (see online supplemental figure S22). The signature is thus predictive only in KRAS wild-type patients. While KRAS mutations significantly affect the cellular biology, this seems not to be the case for BRAF.

All genes within the signature (TNFSF9, TGFB2,⁴⁸ LILRA3, IFIT2,⁴⁹ and EIF2AK2⁵⁰) were identified as protective factors against the development of colitis in the FLT group.

Gene correlation analysis (online supplemental figure S49) suggests that the TGFβ pathway, TNF pathway, and B-cell activation pathway are key mechanisms through which colitis develops in the FLT group (for further details, see online supplemental figure S50). Notably, TNF has been reported to be upregulated in the intestine of patients suffering from colitis after dual ipilimumab and nivolumab treatment.⁵¹

Similar to the results observed for colitis, the occurrence of headache was significantly correlated with a specific gene-expression signature in patients with resected stage III melanoma. No toxicity events were reported during the observation period for patients with low signature values. The signature included risk factor genes, such as TMEM173, ICAM2, CD84, and CD244,⁵² while SYK⁵³ was identified as a protective factor against the occurrence of headache.

Based on gene correlation analysis (online supplemental figure S51), specific activation pathways, including the interferon and CD84 pathways, were identified as contributing to the occurrence of headaches in the AT group. These pathways highlight the molecular mechanisms underlying headache development in this patient population (for further details, see online supplemental figure S52).

In both metastatic and adjuvant settings, specific gene signatures were significantly correlated with the probability of any toxicity occurrence and the timing of toxicity event occurrence. In patients with resected stage III melanoma (AT group), the gene signature included PVR, LAMB3, IL2RA, EIF2B4, and C1QB, all of which were identified as protective factors against toxicity occurrence. Gene correlation analysis (online supplemental figure S53) revealed two distinct activation pathways contributing to the occurrence of any toxicity in the AT group, providing insight into the molecular mechanisms underlying toxicity events in this setting (for further details, see online supplemental figure S54).

In the metastatic setting, the gene signature included the following risk factor genes: CD3G, GIMAP6, LDHB, MAP3K8, and RPL23. Gene correlation analysis (online supplemental figure S55) identified two activation pathways contributing to the occurrence of any toxicity in the metastatic setting, shedding light on the molecular mechanisms driving these events (for further details, see online supplemental figure S56). Additional information also concerning clinical pathways related to specific gene-expression signatures identified for other toxicities (asthenia, fatigue, hypothyroidism, pancreatitis, cutaneous AE) is reported in online supplemental figures S57–S74.

The same toxicity event seems to be associated with the activation of different signaling pathways in the AT and first-line treatment (FLT) settings. Presumably, this difference arises due to the distinct clinical contexts of the two patient groups: patients in the Adjuvant setting are NED (no evidence of disease), while those in the FLT setting still have residual tumor burden.

The presence of metastatic disease in the FLT group likely influences the tumor microenvironment, which in turn modulates specific signaling pathways. This may explain the emergence of different pathways linked to the same toxicity event. Notably, the occurrence of distinct signaling pathways for identical toxicity events may be indicative of a response to immunotherapy in the FLT group—particularly for arthralgia and cutaneous toxicity, as illustrated in online supplemental tables S1 and S2. This suggests that the tumor microenvironment’s modulation of signaling pathways in FLT patients could serve as a biomarker for therapeutic response.

The mechanisms underlying irAEs in patients receiving ICIs remain incompletely understood but are thought to involve immune dysregulation following checkpoint blockade. PD-1 inhibition disrupts self-tolerance, enhancing immune responses against both tumor and normal tissues. Contributing factors include cross-reactivity of tumor-specific T cells with normal tissue antigens, elevated pro-inflammatory cytokines (eg, IL-17 in colitis), and the presence of autoantibodies, particularly in endocrine toxicities such as thyroiditis.⁴⁹ These diverse mechanisms may underlie the range of irAEs observed in our cohort, including arthralgia, colitis, and headache. The identification and understanding of these processes are crucial for developing predictive biomarkers and therapeutic strategies to mitigate irAEs while preserving the anti-tumor efficacy of ICIs. In line with the aims of this study, we provided specific gene-expression signatures capable of predicting the occurrence of toxicities associated with anti-PD-1 treatment.

Each signature was directly correlated with the corresponding toxicity event, as illustrated in the forest plots. For the first time, we also included the variable of time-to-toxicity in our assessment. Specifically, in the AT arthralgia group, we initially observed a non-significant association between arthralgia and the signature (p=0.09 (online supplemental figure S20). However, when incorporating the time variable, the association became significant, revealing a clear separation between patients with high vs low signature levels. Patients with low signature levels demonstrated superior toxicity-free survival. This finding suggests that the signature is not directly associated with the toxicity event itself but is strongly linked to its earliness. In contrast, in the other groups, the signatures were associated with both the occurrence of toxicity and its earliness, highlighting the predictive relevance of the temporal dimension in our analyses.

While our cross-validation approach as an internal validation strategy may not be ideal, the results are of an exploratory nature and warrant independent clinical validation in the future.

Serious forms of toxicity can lead to therapy suspension, underscoring the importance of early identification. Further investigations are warranted to validate the gene-expression signature model and explore its application in predicting immune-related toxicities in patients treated with anti-PD-1 inhibitors for resected stage III and unresectable or metastatic stage III-IV melanoma.

Conclusions

In this retrospective study, we identified a gene-expression signature model capable of predicting the occurrence of specific toxicities, including arthralgia, colitis, and headache, associated with anti-PD-1 treatment. In the FLT cohort, arthralgia and cutaneous toxicities were positively associated with ORR, and arthralgia, asthenia, colitis, fatigue, and skin toxicities with improved DCR, while no significant link between irAEs and relapse risk was found in the adjuvant cohort. Additionally, we identified a distinct gene-expression signature that can predict the occurrence of any toxicity, irrespective of whether the treatment was administered in the AT or FLT setting.

Data availability statement

Data are available in a public, open access repository.