Histopathological features for GTEx dataset

GTEx histology dataset embeddings (slide-level) We provide embeddings of GTEx slides for 25 models: 18 pathology foundation models, 1 visual-language model text similarity, 5 ImageNet trained models, and 1 fine tuned model. In addition, for some models, embeddings are provided at various tile sizes and resolutions. Each .h5ad file stores one feature vector per slide (unique subject × tissue combination). Tile-level embeddings were mean-pooled per slide, so every has slide-level features only. Naming convention gtex.{tissues}.0.5mpp.{strategy}.{model}.h5ad Field Meaning tissues 4_tissues (6 tissue types) or all_tissues (40 GTEx tissues) 0.5mpp Extraction resolution: 0.5 microns per pixel strategy 224px = single tiling strategy; 000px = three strategies concatenated (see below) model Feature extractor model 224px vs 000px — tiling strategy All models were applied by tiling each slide, extracting features per tile, then mean-pooling to a slide-level vector. The difference is in how many tiling strategies were used. Strategy Meaning Tile sizes Feature layout 224px Features from one tiling strategy 224×224 uni.0000, uni.0001, … 000px Features from three strategies concatenated 224×224 + 448×448 + 894×894 resnet50.224px.000 … resnet50.894px.999 The 000px label indicates variable/mixed pixel sizes — each strategy was mean-pooled independently, then the three vectors were concatenated. Per-strategy feature dimensions Model Per-strategy dim Strategies Total dim alexnet 1000 224, 448, 894 3000 resnet50 1000 224, 448, 894 3000 vgg16 1000 224, 448, 894 3000 convnext_base 1000 224, 448, 894 3000 maxvit_t 1000 224, 448, 894 3000 ccnbg63 1024 224, 448, 894 3072 File groups 4_tissues.224px.* — 19 files Pathology foundation models applied at 224×224 tiles on 4 organs: Brain - Cortex, Colon - Sigmoid, Colon - Transverse, Lung, Skin - Sun Exposed (Lower leg), Skin - Not Sun Exposed (Suprapubic). ~4,800 slides from ~971 donors. File Dim Size Model ccnbg63 1024 0.02 GB Convnext_base fine tuned on GTEx chief 768 0.03 GB CHIEF conch 512 0.01 GB CONCH ctranspath 768 0.03 GB CTransPath gigapath 1536 0.03 GB GigaPath h-optimus-0 1536 0.06 GB H-Optimus-0 h-optimus-1 1536 0.06 GB H-Optimus-1 h0-mini 1536 0.06 GB H0-mini hibou-b 768 0.01 GB Hibou-B hibou-l 1024 0.02 GB Hibou-L midnight 3072 0.06 GB Midnight phikon 768 0.01 GB Phikon phikonv2 1024 0.04 GB Phikon v2 prism 1280 0.02 GB PRISM titan 768 0.01 GB TITAN uni 1024 0.02 GB UNI uni2 1536 0.03 GB UNI v2 virchow 2560 0.05 GB Virchow virchow2 2560 0.05 GB Virchow v2 all_tissues.000px.* — 7 files ImageNet-pretrained CNNs (plus ccnbg63) applied at 3 tiling strategies on all 40 GTEx tissues. ~25,000 slides from ~971 donors. File Dim Size alexnet 3000 0.57 GB ccnbg63 3072 0.57 GB convnext_base 3000 0.57 GB efficientnet_v2_l 2000 0.38 GB maxvit_t 3000 0.57 GB resnet50 3000 0.57 GB vgg16 3000 0.57 GB The ccnbg63 file additionally contains pre-computed dimensionality reductions (PCA, UMAP, etc.) in obsm. all_tissues.224px.uni — 1 file UNI embeddings (1024-d) from 224×224 tiles across all 40 tissues. ~25,000 slides. all_tissues.224px.plip.probs — 1 file PLIP model softmax outputs (150 pathology class probabilities) across all 40 tissues. ~25,000 slides. Common metadata (obs columns) Column Description Age Bracket Donor age bracket (20-29, 30-39, …, 70-79) Hardy Scale Death classification (Fast death — natural/violent, Intermediate, Slow, Ventilator) Pathology Categories Histopathology finding categories Pathology Notes Free-text pathology descriptions Sex Donor sex (female, male) Subject ID GTEx donor identifier Tissue GTEx tissue site name Usage import anndata as adadata = adc.read_h5ad("gtex.all_tissues.0.5mpp.000px.resnet50.h5ad")# adata.X = (25335, 3000) slide-level features# adata.obs = slide metadata