205 Annotation-free deep learning identifies cellular neighborhoods associated with responses to immunotherapy in patients with melanoma

Multiplex immunofluorescence tissue imaging allows for the simultaneous analysis of multiple cell types and their spatial interactions is increasingly being used to identify new tissue-based biomarkers immunotherapy. However, identification signatures typically requires a priori selection types, marker expression levels, interactions, which can be challenging as number features quickly increases with markers. Here, we use deep learning framework, named NaroNet, relevant focal, larger cellular neighborhoods in an unsupervised fashion. This strategy uses image-patches size 25x25μm2 from slides stained CD8, FOXP3, PD-L1, PD-1, CD163, Sox10/S100, where every morphology, texture, interaction was extracted by model. Three were identified discovery cohort (n=56) pretreatment melanoma specimens that predicted response anti-PD1-based therapy. These then interpreted AstroPath platform included: 1) An image patch (25x25 μm2) PD-L1(-) tumor CD163+ cells associated non-response 2) A neighborhood (75x75 CD163+PD-L1+, CD8+, FOXP3+, FOXP3+CD8+, PD-L1+ response; 3) composed CD163+PD-L1(-), non-response. The three applied independent, validation (n=44) achieving AUCs 0.76, 0.70, 0.74 (p=0.01, 0.03, 0.01, respectively). When predictive combined, final AUC 0.81 95% CI [0.67,0.92] validated. In conclusion, annotation-free facilitates microenvironment single-cell resolution immuno-oncology biomarker development.