Background In recent years, studies have indicated that in response to almost any type of anti-cancer therapy, the patient (the host) may generate pro-tumorigenic and pro-metastatic effects. This phenomenon, called host-response, counteracts the anti-tumor activity of the treatment. We have previously shown that machine learning-based plasma proteomic analysis of the host response may serve as a predictive tool for response in non-small cell lung cancer. Here, we present initial results of a host response-based machine learning classifier that predicts clinical outcome in melanoma patients treated with immune checkpoint inhibitors (ICIs).

Methods Plasma samples from melanoma patients (training set; n=32) treated with anti-PD-1 or anti-PD-1 and anti-CTLA-4 combination were obtained at baseline and early on treatment. Response was based on RECIST criteria. Proteomic profiling of the plasma samples was performed using ELISA-based antibody arrays. Machine learning algorithms were used to identify a predictive signature that stratifies between responders and non-responders. The signature was validated on an independent cohort of melanoma patients (validation set; n=14). In addition, advanced bioinformatic analysis was performed in order to identify biological pathways unique to responders and non-responders.

Results A 3-protein signature was identified as a predictor of clinical outcome following immunotherapy with an area under the curve (AUC) of the receiver operating characteristics (ROC) plot of 0.88 (p-value 5.84E-05; confidence interval 0.76 – 1.0), and sensitivity and specificity of 0.65 and 0.95, respectively. This signature was successfully validated with AUC of 0.85 (p-value 0.03; confidence interval 0.63 – 1.0), and sensitivity and specificity of 0.75 and 0.9, respectively. To further explore the biological basis of resistance to immunotherapy, we performed a pathway enrichment analysis. Multiple mechanisms for resistance were identified in the non-responder group, including immunosuppression and inflammation associated pathways. Comparison between the two treatment modalities revealed pathways unique to each treatment that involve extracellular modulation, immunosuppression and processes associated with tumor progression, which may imply important differences between the two regimens.

Conclusions Our results demonstrate that analyzing the host response to ICI therapy using plasma-based proteomic profiling combined with machine learning algorithms serves as a successful approach for predictive biomarker discovery in melanoma. This bioinformatics-based functional analysis provides insights into mechanisms of resistance and may be used to identify potential strategies for improving clinical outcomes.

Ethics Approval The study was approved by the Yale University Institutional Review Ethics Board, approval number 0609001869.