Journal: Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing

Article Title: Spatial Omics Driven Crossmodal Pretraining Applied to Graph-based Deep Learning for Cancer Pathology Analysis

doi:

Figure Lengend Snippet: A) Data collection protocol for Visium spatial transcriptomics slide. B) Training protocol for spatial omics-driven crossmodal contrastive model; two views are generated per modality, per patch; each view is passed through the corresponding branch of the crossmodal model; embeddings are transformed using a shared projection head; unimodal and crossmodal contrastive losses are applied to output embeddings.

Article Snippet: The collected datasets and the downstream tasks they were used on, are summarized below: Visium spatial transcriptomics slides (n=4; 20,000 spots/patches; Co-Registered Spatial Transcriptomics, H&E WSI): to pretrain contrastive crossmodal model Dartmouth Hitchcock Medical Center (n=708 H&E WSI): used for histological stage prediction and clustering analysis TCGA Cohort (n=350 H&E WSI): used for lymph node metastasis prediction, survival prognostication, and tumor infiltrating lymphocyte (TIL) alignment analysis All analyses were conducted on a machine using a single Nvidia Tesla v100 GPU with 32 gigabytes of VRAM, and 100 gigabytes of RAM.

Techniques: Generated, Transformation Assay