findMarkersTree error with metaclusters enabled

[hsuknowledge]

I have repeatedly run into this error at threshold 0.3486784401 (0.9^10) with my K-48, L-96 data.

Identifying metaclusters...
Identified 7 metaclusters
Building top-level tree across all metaclusters...
Decreasing classifier threshold to 0.81
(skipped 0.729, 0.6561, and so on)
Building top-level tree across all metaclusters...
Decreasing classifier threshold to 0.3486784401
Building top-level tree across all metaclusters...
Error in DecTree[[AltDec]] : subscript out of bounds

This is where it throws the error:

celda/R/findMarkersTree.R

Line 1804 in bbc7fb1

downSplit <- DecTree[[AltDec]]

With options(error = recover), I inspected the .addAlternativeSplit(tree, features, as.factor(metaclusterLabels)) stack frame.

Enter a frame number, or 0 to exit

1: findMarkersTree(featureMatrix, classes, threshold = 0.43046721)  ## note: jump started from 0.9^8
2: .addAlternativeSplit(tree, features, as.factor(metaclusterLabels))


Selection: 2
Called from: top level
Browse[1]> ls()
 [1] "AltDec"      "AltSplitInd" "class"       "DecTree"     "features"    "group1Vec"   "group2only"  "group2Vec"
 [9] "groupList"   "tree"

Browse[1]> group1Vec
[1] "M5" "M3" "M4" "M7" "M1"

Browse[1]> group2Vec
[1] "M1"     "M2(46)" "M3"     "M4"     "M6"     "M7"     "M5"

Browse[1]> group2only
[1] "M2(46)" "M6"

Browse[1]> AltSplitInd
[1] 1 2

Browse[1]> AltDec
[1] 2

Browse[1]> c(length(tree), length(DecTree), length(DecTree[[1]]))
[1] 1 1 9

Here are the input data. Can you help me diagnose this issue? Thanks.

featureMatrix.csv.gz (4707 KiB)

classes.txt (113 KiB)

[hsuknowledge]

The code path with metaclusters errored out, but if I switch off metaclusters, it can build the tree successfully. Here are two alternative paths that I can take:

• DecTree <- findMarkersTree(featureMatrix, classes, autoMetaclusters = FALSE)
• DecTree <- findMarkersTree(featureMatrix, classes_metaclusters, autoMetaclusters = FALSE)

The first variant takes 10 minutes to run because K=48. In the second variant, I enter the clusters as grouped into metaclusters (which is recovered from the findMarkersTree stack frame), so I can get a basic decision tree around the metaclusters. The results from both cases are accurate.