Update README.md

README.md

@@ -1,3 +1,34 @@
 ---
-license: apache-2.0
+license: cc-by-4.0
+task_categories:
+  - graph-ml
+tags:
+  - biology
+  - genomics
+  - lung-squamous-cancer
+  - graph-neural-networks
+  - pytorch-geometric
+language:
+  - en
+size_categories:
+  - n<1K
 ---
+
+This dataset combines expression data from TCGA with protein interaction network from STRING. An example implementation of a GCN trained on this dataset is available [here](https://github.com/rizanb/lusc_gnn).
+
+## How
+I built node indices from filtered human ppi in `notebooks/process_human_ppi.ipynb` to make a custom `PyG Dataset`.
+`human ppi -> filter for high confidence (score > 0.7) -> filter to keep only genes in expression dataset -> gene ids -> integer ids -> edge indices`
+GCN expects a graph dataset; each graph represents a patient sample. Nodes represent genes, and edges represent protein-protein interactions between genes. The structure of the graph is the same for all samples (given by edge indices), only the node features (expression levels for each gene) vary.
+
+## Data
+- expression: TCGA LUSC HiSeqV2 gene expression data (178 samples, 20530 genes), UCSC Xena
+- subtype labels: TCGA Nature 2012 paper supplementary data
+- STRING PPI network: *Homo sapiens* interaction data; high-confidence interactions (score > 0.7)
+- class distribution: 
+```
+classical: 65 (36.5%)
+basal: 43 (24.2%)
+secretory: 43 (24.2%)
+primitive: 27 (15.2%)
+```