Schematic presentation of the workflow of this study. The use radiomics to aid finding novel EV charged miRNAs to allow PDAC diagnostics.

The base line information of clinical parameters of patients enrolled from four centers in this multi-center trial.

Propensity Score Matching (PSM) allows matching data of benign pancreas lesions and PDAC patients from DUH to UMMD & JHC patients according to the age factor, all of DUH patients successful matched similar patients (A). The different radiomic features between the benign lesions and PDAC patients CT images(B). 12 most important radiomic features differentiating between the benign pancreatic lesion and PDAC patientś CT images identified by the Boruta algorithms (C) Four radiomic features were selected by Lasso Regression to build model signature (D&E). Applying the four radiomic features related signature in image analysis show high accuracy in predicting the PDAC manifestation in the WUH test dataset(F).

EV miRNA presenting the risk group stratification based on radiomics signature by WGCNA analysis featuring green mode discovering our key module for further analysis (r=0.21, p=0.047) (A). The number of low abundance miRNA in the entire EVseq dataset cohort is n= 295 (B). Out of those low abundance miRNAs, n=12 present matching candidates differentially expressed in high risk group patients. Alignment to our radiomics feature parameters identified three core miRNAs (hsa-miR-1260b, hsa-miR-151a-3p and hsa-miR-5695) (C). The three key miRNAs show significantly different expression levels in tumor condition, both for serum (D-F) and tissue (G-I).

Ten machine learning algorithms demonstrate that three key miRNAs show a high accuracy to diagnosis PDAC in early stage, no matter in training or test datasets. The best machine learning algorithms is GBM (cutoff:0.75) (A). Three miRNAs prediction ability of training dataset (GSE10106817) in GBM model is 0.978(B). Three miRNAs prediction ability of test dataset (GSE113486 and GSE112264) in GBM model is 0.919, and 0.857, respectively(C&D). Data distribution before removal of batch effect of our center data and GSE109319 dataset (E). Data distribution after remove batch effect of our center and GSE109319 dataset (F). Three EV miRNAs prediction ability to identify cancer of our center data and GSE109319 in GBM model is 0.897(G).

Stratification of abundancy levels of shared mRNAs by Non-negative matrix factorization method allows the clustering of patients into two subtypes (C1 and C2) (A). Patients of the C1 with a poor out outcome in OS and DFS(B&C). C1 subtype patients are characterized with older age and bigger average tumor size, higher number of tumor cell positive lymph nodes as compared to C2 patients (D-F). C1 patients are predominantly female, have tumors with pathological classification marks of perineural invasion and advanced tumor stage (G-I).

C2 subtype is positively associated with elevated levels of transcripts regulating gene pathways encoding for CD8 T cells, cytotoxic lymphocytes, NK cells(A&B). Commonlyknown immune checkpoints also higher expressed in C2 subtypes(D). Aligning in vitro drug sensitvtiy and expression data from the GDSC database to subtype signature, predicts that tumors of C1 patients might be more sensitive to AKT INHIBITOR VIII, Bleomycin, Dasatinib, GNF-2, PF - 562271, Refametinib, BMS-509744 as C2 subtypes.

GO functional enrichment analysis indicated that the C1 subtype is enriched for intermediate filament organization, CC ribosome, as well as MF symporter activity(A-C). Pathway enrichment analysis showed that the C1 subtype was activated with the Reactome fatty acids, Reactome diseases of metabolism, as well as Reactome biological oxidations pathways, and may be inhibited with Reactome apoptosis, Reactome DNA repair, and Reactome signaling by Hippo pathways (D-I).