Machine learning‐derived serum protein signatures as biomarkers for checkpoint inhibition in melanoma

Despite the success of immune checkpoint inhibitors (ICI) in melanoma, a substantial proportion of patients do not respond or develop resistance.1 In order to stratify patients appropriately, the identification of biomarkers that could predict outcome or adverse events remains an unresolved clinical need. In this prospectively designed study, we included 122 melanoma patients (Table 1) treated with ICI at Hannover Medical School, Germany. Serum samples were collected at baseline and 3–4 weeks after treatment initiation. Using electrochemiluminescence, we quantified 54 cytokines and chemokines and analysed both, baseline levels and early dynamic changes and correlated the expression pattern with patient outcome (Figure 1a). Because the therapeutic setting, determined by disease stage, has a significant impact on patient outcomes, we analysed the adjuvant (n = 67) and palliative/metastatic (n = 55) patients separately. Results revealed that 28 of 41 quantifiable serum parameters increased significantly after ICI, including pro-inflammatory cytokines such as IFN-γ, CXCL-10, IL-17, as well as pleiotropic markers IL-10 and sICAM. Dual ICI with Nivolumab plus Ipilimumab had a greater impact on dynamic cytokine changes than PD-1 monotherapy (Figure 1b). Regarding patient outcomes, in the adjuvant cohort, univariate time-dependent Cox regression analyses identified six proteins (FLT, IFN-γ, IL-6, IL-10, SAA and sICAM), whose upregulation was significantly associated with shorter recurrence-free survival (RFS) (Figure 1c). Baseline serum parameters did not show significant correlation with RFS. In the palliative cohort, dynamic changes revealed that a greater post-treatment increase in Tie-2 and IL-1RA correlated with shorter progression-free survival (PFS) (Figure 1c). Here, also elevated values of six pre-treatment parameters (bFGF, PIGF, Tie-2, CCL-2, IL-8, and CXCL-10) were significantly associated with shorter PFS. To improve predictive performance, we developed composite risk scores in combining significant cytokines with a Spearman correlation coefficient <0.4 and applied machine learning. In the adjuvant cohort, the random forest model achieved an AUC of 0.82. Subsequently, in the Kaplan–Meier analysis, we stratified the patients into groups based on the risk score. We showed that melanoma patients with a high-risk score had a significantly impaired RFS compared to patients with a low-risk score (Figure 1e). When analysing the overall survival (OS), we identified six baseline serum proteins (Eotaxin, Eotaxin-3, CXCL-10, MIP-1α, IL-6 and LDH) associated with reduced OS in the adjuvant cohort, whereas only one cytokine (PIGF) was linked to shortened OS in the palliative cohort. By incorporating the significant parameters from the adjuvant cohort into the machine learning models, the resulting risk score demonstrated strong predictive power for OS shown in the Kaplan–Meier curve (Figure 1f). Notably, IL-6 in the adjuvant cohort was significantly correlated with both shorter OS and RFS. This cytokine has well-established immunological roles. IL-6 promotes immunosuppressive mechanisms and myeloid-derived suppressor cell expansion,2 and also, previous studies have demonstrated that elevated IL-6 levels are associated with worse treatment outcomes.3 Our study demonstrates that early dynamic changes in circulating cytokines provide insights into treatment response and early progression. Notably, cytokine predictive power was higher in patients with lower tumour burden (adjuvant cohort), whereas systemic inflammation in the palliative cohort may reduce biomarker stability. Nevertheless, machine learning allowed us to account for inter-patient variability and to improve predictive robustness. This work establishes a clinical rationale for the cytokines mentioned above as candidate biomarkers. Prospective, multicentre studies are needed to confirm these findings and prospectively combine cytokines with other liquid biopsy parameters such as circulating tumour DNA and immune cell phenotyping.4, 5 Mechanistic studies may further elucidate how these cytokine dynamics influence immune resistance. In conclusion, our results suggest that machine learning-based risk scores integrating serum cytokines can predict early recurrence and OS in melanoma patients receiving ICI. Bristol Meyer Squibb-Stiftung Immunonkologie Grant Number: FA 10-010. Katrin Schaper-Gerhardt received research grants from Almirall, Kyowa Kirin and Recordati and received honoraria for advice from Almirall and Sun Pharma outside the submitted work. Nina Flatt received travel support from Therakos ECP Immunomodulation™ and Sun Pharma, and honoraria from Pierre Fabre all outside the submitted work. Viola K. DeTemple received research grants from SUN Pharma and Almirall, as well as honoraria and travel support from SUN Pharma, all outside the submitted work. Imke von Wasielewski received honoraria from Almirall, blueprint medicines, Delcath, IMMUNOCORE; Merck, Takeda, Novartis, BMS, MSD, Sanofi, Stemline, Sun, Merck, Kyowa Kirin and travel support from BMS, MSD, Sanofi, Pierre Fabre, Stemline, Kyowa Kirin. Ralf Gutzmer received honoraria for advice and lectures from BristolMyers Squibb, MerckSharpDohme, Novartis, Merck-Serono, Almirall Hermal, Pierre-Fabre, Sun Pharma, Immunocore, Delcath, Sanofi/Regeneron, Incyte, Janssen, received travel support from SUN Pharma and PierreFabre and received research grants from Novartis, Sanofi/Regeneron, Merck Serono, Amgen, SUN Pharma, KyowaKirin, Almirall Hermal and Recordati. The other authors declare no conflict of interest. Written informed consent of the patients was obtained. The study was approved by the ethics committee in Hannover in accordance with the Declaration of Helsinki (vote Nr 8685_BO_K_2019). The data that support the findings of this study are available from the corresponding author upon reasonable request.