PRISM-GI public research data subsets

Expression files: rows=samples, columns=available genes in the locked 158-gene background.
sample_id is a public repository identifier. Clinical files: OS time in months; event 1=death, 0=censored.
These are processed, gene-mapped OS-eligible study subsets, not the complete original datasets.
Source-specific author normalization is retained; no extra transformations are applied to GEO values here.
The website ranks values within samples, so monotonic log transformations do not change predictions.
Use ONE expression file (one cohort) per prediction run. Do not upload clinical outcomes to the predictor.
Do not concatenate cohorts before scoring: cohort mean adjustment depends on the source cohort.
All 15 model genes must be measured. Missing non-panel background genes are documented in catalog.json.
GSE28735 is excluded from this download because GSE62452 includes overlapping source samples.
The pancreatic bundle has 207 records; historical 249-record pooled estimates were not recomputed here.
Historical GEO cohorts participated in exploratory model selection and are not untouched validation sets.
Refer to GEO/TCGA original studies, attribution requirements and data-use terms when reusing data.
Research use only. These files and the model are not a clinical diagnostic or treatment device.