Table 1.
Overview of the representation models tested over the different tasks included in our benchmark.
| Tasks | Per-cohort OS | Pan-cancer OS | Gene essentiality | |
|---|---|---|---|---|
| Description | Considering one tumor type at a time, we aim to predict the overall survival of patients based on the RNA-seq learned representations and accurately rank them by their predicted risk42 | Same task as Per-cohort OS but pulling all tumors together and ranking patients’ predicted risk across all indications2 | Given cell line RNA-seq learned representations, predict if a given gene of interest is essential to their survivals3 | |
| Evaluation metric | C-index | C-index | Overall Spearman Correlation: we evaluate if models can rank all of the cell lines/genes combination correctly | Per-DepOI Spearman Correlation: scores are the average of correlations computed per-DepOI |
| Models | ||||
| Identity | x | x | x | x |
| PCA | x | x | x | x |
| AE16 | x | x | x | x |
| scVI38 | x | x | x | x |
| MHAE37 | x | x | x | x |
| MAE39 | x | x | x | x |
| DA-GN41 | x | x | x | x |
| GNN36 | x | x | x | x |
| PreAE3,20 | x | x | x | |
| PreAE finetuned | x | x | x | |
Models tested for a given task are indicated with an “x”. References to existing literature are marked within superscript. Due to the lack of a relevant pretraining dataset for the pancancer task, we did not perform any pretraining experiments on this task.