Fig. 5: PreMode and other methods’ performances in genetic mode-of-action prediction tasks.

a Performances of PreMode and baseline methods on gain- and loss-of-function predictions in 9 genes, mean and standard error were calculated from five random training/testing splits. b Performances of PreMode and LoGoFunc in 9 genes, training/testing data was split by presence in LoGoFunc data or not, mean and standard error were calculated from five random splits. c Comparison of PreMode and funNCion in three ion channel genes, training/testing data was split randomly or same as in their paper, mean and standard error were calculated from five random splits. d Ablation analysis of PreMode on gain- and loss-of-function predictions in 9 genes, for each model, mean and standard error were calculated by weighted average of mean and standard error in 9 genes based on dataset sizes. e Few-shot transfer learning of PreMode and random forest method on subsampled training data, mean and standard error were calculated by weighted average of mean and standard error in 9 genes based on dataset sizes. f PreMode performances when trained with G/LoF variants from the same gene (marked as “Gene Only”) or G/LoF data from same domain across itself and other genes (marked as “Protein Family”) or G/LoF data from same domain in other genes (marked as “Protein Family, exclude gene”) and tested on the same testing dataset, mean and standard error were calculated by weighted average of five random splits.