Fig. 1: PACpAInt approach for identification of PDAC molecular tumor subtypes.

a Simplified workflow of the study: a first model is applied to find the tumor area (tumor cells and stroma) (PACpAInt-Neo) followed by a second model predicting either the global tumor cell molecular type (classical vs basal-like) at the slide level (PACpAInt-B/C) or predicting at the tile level (small square 112 µm wide) the nature of the cells (tumor or stroma - PACpAInt-cell type) and their molecular subtype (classical vs basal-like for tumor cells and active vs inactive for stroma) (PACpAInt-Comp), b Description of the cohorts. Discovery cohort (DISC) was composed of 202 patients (surgical specimens) from three centers. A tissue carrot (diameter 600 μm) was taken from a block for RNA profiling. HES slides (at least 2/tumor) were digitized for PACpAInt analysis. In most cases, the tissue carrot and the HES did not come from the same block. The workflow was similar in the first validation cohort BJN_U unmatched (surgical specimens). For the two next validation cohorts (BJN-M matched (surgical specimens) and EUS_Liver (liver metastases, fine-needle biopsies)), the same block was used for RNA extraction after microdissection of the neoplastic area and to generate the HES slide that was digitized and analyzed with PACpAInt. In addition, in the BJN_M matched cohort, all the remaining tumor slides were also digitized for PACpAInt analysis. Finally, in the TCGA_PAAD validation cohort (surgical specimens), in contrast to all the other cohorts, the RNA was extracted from frozen material, not formalin-fixed paraffin-embedded. Similarly to the discovery cohort, the tissue analyzed by RNAseq was not spatially matched with the digitized slides.