Karlijn Hummelink

Chapter 5 - Supplemental material: methods 275 A that simultaneously refined training class labels for classifier development at the same time as the classifier itself. This approach is explained in detail in [17]. c. Classifier A: Redevelopment of BDX008 to use current mass spectral preprocessing and feature definitions The development of BDX008 has been outlined in detail in Ascierto et al13. This used different spectral processing and feature definitions than we use here. Hence, an alternative version of this test was developed to allow standardization of spectral processing and mass spectral features across all three classifiers. The redevelopment used spectra from 113 of the 119 serum samples obtained from patients with melanoma prior to treatment with nivolumab originally used to develop BDX008. All 274 mass spectral features defined herein were used in the approach. All other parameters were kept the same as in BDX008 development. d. Sample subsets and parameters used to define classifiers B and C Supplementary Table 7: Parameters used in classifier development of classifiers B and C Classifier C Classifier B Samples used for development 96 development set samples that were not both BDX008+ and from a patients with performance status =0 76 development set samples that were not both BDX008- and in the poor prognosis group of Classifier C k (for kNN) 11 11 # training/test split realizations 625 625 # dropout iterations 20,000 200,000 Averaging over dropout iterations Weight-based Weight-based Filtering metric PFS Hazard ratio between classes PFS Hazard ratio between classes Initial assignment of training classes PFS time < median vs PFS time > median PFS time < median vs PFS time > median MS features used 29 features associated with Immune Response Type 2 (See Supplementary Table 8) All 274 features

RkJQdWJsaXNoZXIy MTk4NDMw