Background: Although immunotherapy and targeted treatments have dramatically improved the survival of melanoma patients, the intra- or intertumoral heterogeneity and drug resistance have hindered the further expansion of clinical benefits.
Methods: The 96 combination frames constructed by ten machine learning algorithms identified a prognostic consensus signature based on 1002 melanoma samples from nine independent cohorts. Clinical features and 26 published signatures were employed to compare the predictive performance of our model.
Results: A machine learning-based prognostic signature (MLPS) with the highest average C-index was developed via 96 algorithm combinations. The MLPS has a stable and excellent predictive performance for overall survival, superior to common clinical traits and 26 collected signatures. The low MLPS group with a better prognosis had significantly enriched immune-related pathways, tending to be an immune-hot phenotype and possessing potential immunotherapeutic responses to anti-PD-1, anti-CTLA-4, and MAGE-A3. On the contrary, the high MLPS group with more complex genomic alterations and poorer prognoses is more sensitive to the BRAF inhibitor dabrafenib, confirmed in patients with BRAF mutations.
Conclusion: MLPS could independently and stably predict the prognosis of melanoma, considered a promising biomarker to identify patients suitable for immunotherapy and those with BRAF mutations who would benefit from dabrafenib.
Keywords: Machine learning; Melanoma; Multi-omics; Prognosis; Treatment.
© 2022. The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.