FFPE materials Formalin-fixed paraffin-embedded (FFPE) tumor samples and clinical data were obtained from PCNSL patients enrolled in the HOVON105/ALLG NHL 24 intergroup, multicenter, open-label, randomized phase 3 study (NTR2437 and ACTRN12610000908033) through the HOVON Pathology Facility and Biobank. Muiltiplex imaging Multiplex immunofluorescence was performed on 4-µm-thick formalin-fixed, paraffin-embedded whole tissue sections using the Opal 7-color fluorescence immunohistochemistry (IHC) kit (Akoya biosciences, USA), as previously described. In brief, slides were deparaffinized and rehydrated, followed by a blocking step for endogenous peroxidase using 0.3% H2O2/methanol and fixation with 10% neutral buffered formalin (Leica Biosystems, Germany). Slides were washed in Milli-Q water and 0.05% Tween20 in 1x Tris-Buffered Saline (TBS-T). Antigen retrieval was done by placing the slides in 0.05% ProClin300/Tris–EDTA buffer pH 9.0 in a microwave at 100% power until boiling, followed by 15 min at 30% power. Slides were cooled in Milli-Q water, washed in 1x TBS-T and blocked with Antibody Diluent (Agilent, USA). The slides were then incubated with primary antibody diluted in Normal Antibody Diluent, followed by incubation with the broad spectrum HRP from the SuperPicture Polymer Detection Kit (Life Technologies, USA). Next, the slides were incubated with Opal TSA fluorochromes diluted in an amplification buffer (Akoya biosciences, USA). The primary and secondary antibody complex was stripped by microwave treatment with 0.05% ProClin300/Tris–EDTA buffer at pH 9.0. Finally, DAPI working solution (Akoya biosciences, USA) was applied and the slides were mounted with Prolong Diamond Anti-fade mounting medium (#P36965; Life Technologies). Image processing Stained slides were scanned using the Vectra Polaris Automated Quantitative Pathology Imaging System (Akoya biosciences, USA). From each slide, representative tumor regions and regions at the junction of tumor and surrounding cerebral tissue were selected and multispectral imaging (MSI) images were acquired at 40x resolution. After image capture, the images were spectrally unmixed and analyzed, using supervised machine learning algorithms within Inform 4.2.2. (Akoya biosciences). Cells were assigned into ten different phenotype categories: “PAX5+PD-L1-”, “PAX5+PD-L1+”, “CD163+PD-L1-”, “CD163+PD-L1+”, “CD3+CD8-PD-1-”, “CD3+CD8+PD-1-”, “CD3+CD8-PD-1+”, “CD3+CD8+PD-1+”, “other PD-L1+” or “other”, based on the size of the cells and positivity of markers in the panel.