LASSO logistic regression reveals a mixed MiRNA and serum-marker classifier for prediction of immunotherapy response in liquid biopsies of melanoma patients
Introduction: Cutaneous malignant melanoma suffers from the highest metastasis rate and mortality among different skin cancer entities. However, with emerging immune checkpoint inhibitor (ICI) therapy, prognosis has significantly improved over the last years. To better assess treatment response stab...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2024-12-01
|
| Series: | EJC Skin Cancer |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2772611824002489 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Introduction: Cutaneous malignant melanoma suffers from the highest metastasis rate and mortality among different skin cancer entities. However, with emerging immune checkpoint inhibitor (ICI) therapy, prognosis has significantly improved over the last years. To better assess treatment response stable and reliable biomarkers are needed. Methods: We gathered blood samples of 81 patients with predominantly AJCC Stage III/IV melanoma to evaluate serum markers and plasma-derived miRNAs. A machine learning model was developed to predict immunotherapy response. Serum markers were measured according to standard clinical routines. Expression levels of 61 miRNAs were quantified via flowcytometry. LASSO logistic regression was fit to the data to predict therapy outcome, employing AUROC as the performance metric. Nested cross-validation was used to mitigate overfitting. Results: Plasma-derived miRNA expression exhibited significant association with therapy response for 5 miRNAs: miR-132–3p, miR-137, miR-197, miR-214, miR-514a-3p. Serum markers LDH, CRP, S100 and eosinophile concentration showed significant differences between Responders and Non-Responders. Age and previous anti-BRAF therapy (BRAFi/MEKi) were the only demographic parameters significantly related to therapy outcome. Among six machine learning models tested, a relaxed LASSO approach on the entire dataset performed best (AUC = 0.851). Conclusion: Validation of the relaxed LASSO model in the outer loop of the nested cross validation yielded an AUC of 0.847. This model incorporated expression of a miRNA-quartet, LDH, patient age and prior BRAFi/MEKi. It effectively identifies Responders and Non-Responders with high sensitivity and specificity, presenting promising candidates for the validation of future biomarkers. |
|---|---|
| ISSN: | 2772-6118 |