Fix a bug where Z-score-based combination of continuous features was computed incorrectly (but seemed to only have subtle differences from the correct value in most cases)

Author: Nick-Eagles

R/multi_gene_z_score.R

@@ -32,8 +32,8 @@ multi_gene_z_score <- function(cont_mat) {
     cont_mat <- cont_mat[, good_indices, drop = FALSE]
 
     #   For each spot, average Z-scores across all features
-    cont_z <- (cont_mat - colMeans(cont_mat, na.rm = TRUE)) /
-        colSds(cont_mat, na.rm = TRUE)
+    cont_z <- (cont_mat - rep(colMeans(cont_mat, na.rm = TRUE), each = nrow(cont_mat))) /
+        rep(colSds(cont_mat, na.rm = TRUE), each = nrow(cont_mat))
     z_vec <- rowMeans(cont_z, na.rm = TRUE)
 
     return(z_vec)

tests/testthat/test-multi_gene_z_score.R

@@ -40,5 +40,9 @@ test_that(
             multi_gene_z_score(cont_mat),
             "^After dropping features with no expression variation"
         )
+
+        #   Now actually verify the Z-score-based calculation is correct
+        cont_mat = matrix(c(1, 2, 3, -3, 0, 3, 7.5, 9, 10.5), nrow = 3)
+        expect_equal(multi_gene_z_score(cont_mat), c(-1, 0, 1))
     }
 )