Require harmony '1.2.0'

- Reenable temporarily broken unit test
- Give helpful error message for outdated harmony version

DESCRIPTION

@@ -34,7 +34,7 @@ Imports:
     MatrixGenerics,
     matrixStats,
     Rcpp,
-    harmony (>= 1.0.3),
+    harmony (>= 1.2.0),
     limma,
     BiocNeighbors
 Suggests: 

NEWS.md

@@ -5,6 +5,7 @@ the `group_by` argument for `find_de_neighborhoods`. (thanks Katha for pushing f
 * The formula parsing automatically detects global variables and adds them to the 
 colData. This avoids problems with the random test / training assignment.
 * Duplicate column names in colData are now longer allowed.
+* Require harmony version >= 1.2.0 (thanks Maija for reporting the problem)
 
 
 # v1.1

R/harmony_wrapper.R

@@ -27,12 +27,7 @@ harmony_init <- function(embedding, design_matrix,
   sigma <- rep_len(sigma, nclust)
   lambda_range = c(0.1, 10)
   if(packageVersion("harmony") < "1.2.0"){
-    harmonyObj <- harmony_new_object()
-    harmonyObj$setup(
-      embedding, phi,
-      sigma, theta, lambda_vec, max.iter.cluster, epsilon.cluster,
-      epsilon.harmony, nclust, block.size, lambda_range, B_vec, verbose
-    )
+    stop("Your 'harmony' version is outdated: ", packageVersion("harmony"), ". Please update to version >= 1.2.0")
   }else{
     alpha <- 0.2
     harmonyObj <- harmony::RunHarmony(embedding, mm_groups, nclust = nclust, max.iter = 0, return_object = TRUE, verbose = FALSE)
@@ -53,23 +48,14 @@ harmony_init <- function(embedding, design_matrix,
 #' @keywords internal
 harmony_new_object <- function(){
   Y <- randn(3, 100)
-  harmonyObj <- if(utils::packageVersion("harmony") >= "1.0.3"){
-    # Harmony ignores 'verbose = FALSE'
-    harmony::RunHarmony(Y, rep(c("a", "b"), length.out = 100), nclust = 2, max.iter = 0, return_object = TRUE, verbose = FALSE)
-  }else{
-    harmony::HarmonyMatrix(Y, rep(c("a", "b"), length.out = 100), do_pca = FALSE, nclust = 2, max.iter.harmony = 0, return_object = TRUE)
-  }
+  harmony::RunHarmony(Y, rep(c("a", "b"), length.out = 100), nclust = 2, max.iter = 0, return_object = TRUE, verbose = FALSE)
   harmonyObj
 }
 
 harmony_init_clustering <- function(harmonyObj, iter.max = 25, nstart = 10){
   stopifnot(is(harmonyObj, "Rcpp_harmony"))
   harmonyObj$Y <- t(stats::kmeans(t(harmonyObj$Z_cos), centers = harmonyObj$K, iter.max = iter.max, nstart = nstart)$centers)
-  if(utils::packageVersion("harmony") <= "1.1.0"){
-    harmonyObj$init_cluster_cpp(0)
-  }else{
-    harmonyObj$init_cluster_cpp()
-  }
+  harmonyObj$init_cluster_cpp()
   harmonyObj
 }
 

tests/testthat/test-align.R

@@ -108,7 +108,6 @@ test_that("check that aligning points works perfectly for low number of points",
 
 
 test_that("check that harmony alignment works as expected", {
-  skip("Harmony v1.1.0 has a regression about small input matrices")
   set.seed(1)
   n_genes <- 10
   n_emb <- 2
@@ -127,16 +126,22 @@ test_that("check that harmony alignment works as expected", {
                    alignment_design = NULL, alignment_design_matrix = design_matrix,
                    use_assay = "foo", is_test_data = rep(FALSE, ncol(mat)))
   gr <- rep(seq_len(n_points), times = 2)
-  fit_al2 <- align_harmony(fit, design = fit$alignment_design_matrix, nclust = n_points, ridge_penalty = 1e-3, verbose = FALSE)
+  suppressWarnings({
+    fit_al2 <- align_harmony(fit, design = fit$alignment_design_matrix, nclust = n_points, ridge_penalty = 1e-3, verbose = FALSE)
+  })
   set.seed(1)
-  harm <- t(harmony::RunHarmony(mat, meta_data = df, vars_use = "tmp", nclust = n_points, lambda = 1e-8, verbose = FALSE))
+  suppressWarnings({
+    harm <- t(harmony::RunHarmony(mat, meta_data = df, vars_use = "tmp", nclust = n_points, lambda = 1e-8, verbose = FALSE))
+  })
 
   expect_equal(fit_al2$embedding[,df$tmp == "a"], fit_al2$embedding[,df$tmp == "b"], tolerance = 1e-2)
   expect_equal(harm[,df$tmp == "a"], harm[,df$tmp == "b"], tolerance = 1e-2)
 
   # Reimplement harmony correction
   set.seed(1)
-  harm_obj <- harmony_init(mat, design_matrix, nclust = n_points, lambda = 1e-8, verbose = FALSE)
+  suppressWarnings({
+    harm_obj <- harmony_init(mat, design_matrix, nclust = n_points, lambda = 1e-8, verbose = FALSE)
+  })
   harm_obj <- harmony_max_div_clustering(harm_obj)
   Z_corr <- harm_obj$Z_orig
   for(k in seq_len(harm_obj$K)){