Merge pull request #19 from jjc2718/revisions

Paper edits for revisions
greenelab · Jun 3, 2024 · 012eed1 · 012eed1
2 parents 9e7f053 + 74a786b
commit 012eed1
Show file tree

Hide file tree

Showing 12 changed files with 12,076 additions and 5,891 deletions.
diff --git a/content/02.main-text.md b/content/02.main-text.md
diff --git a/content/91.supp-info.md b/content/91.supp-info.md
@@ -9,3 +9,9 @@
 ![Top row: Distribution of performance differences when thyroid cancer (THCA) data is held out from training set across seeds/folds, grouped by gene. Bottom row: Distributions of performance differences for genes where THCA is included in training/holdout sets, relative to other cancer types that are included.](images/supp_figure_4.png){#fig:thca_by_gene tag="S4" width="100%" .page_break_before}
 
 ![Performance vs. dropout parameter (first column) and weight decay strength (second column), for EGFR mutation prediction (first row) and KRAS mutation prediction (second row) using a 3-layer fully connected neural network trained on TCGA (blue/orange) and evaluated on CCLE (green).](images/supp_figure_5.png){#fig:nn_dropout_wd tag="S5" width="100%" .page_break_before}
+
+![Performance vs. number of gene expression principal components, used as input to a 3-layer fully connected neural network trained on TCGA (blue/orange) and evaluated on CCLE (green), for EGFR and KRAS mutation status prediction.](images/supp_figure_6.png){#fig:nn_pca tag="S6" width="100%" .page_break_before}
+
+![Performance across regularization parameter values for XGBoost mutation status classification, for generalization from TCGA to CCLE. Top row shows performance for EGFR across varying values of `num_estimators` and `max_depth` (Panel A), and for `max_depth=8` across a range of `num_estimators` (Panel B). Panel C summarizes the distribution of performance comparisons between "best" vs. "smallest good" `num_estimators` (33/71 genes best > smallest good, 17/71 smallest good > best, 20/71 best = smallest good).](images/supp_figure_7.png){#fig:xgboost_perf tag="S7" width="100%" .page_break_before}
+
+![Summary of performance for TCGA to CCLE generalization using 5-layer fully connected neural network, similar to results shown in Figure 5 for 3-layer network. All experiments used expression of top 8000 genes by mean absolute deviation, for computational reasons. In the "best" vs. "smallest good" analysis, 27/71 genes had better performance for the best model, and 17/71 had better performance for the smallest good model, with 26/71 genes where the best and smallest good models were equal.](images/supp_figure_8.png){#fig:deep_nn_perf tag="S8" width="100%" .page_break_before}
diff --git a/content/images/figure_2.png b/content/images/figure_2.png
diff --git a/content/images/figure_2.svg b/content/images/figure_2.svg
diff --git a/content/images/figure_4.png b/content/images/figure_4.png
diff --git a/content/images/figure_4.svg b/content/images/figure_4.svg
diff --git a/content/images/supp_figure_6.png b/content/images/supp_figure_6.png
diff --git a/content/images/supp_figure_6.svg b/content/images/supp_figure_6.svg
diff --git a/content/images/supp_figure_7.png b/content/images/supp_figure_7.png
diff --git a/content/images/supp_figure_7.svg b/content/images/supp_figure_7.svg
diff --git a/content/images/supp_figure_8.png b/content/images/supp_figure_8.png
diff --git a/content/images/supp_figure_8.svg b/content/images/supp_figure_8.svg