* Changed default parameters for distinct_colors.
* Fixed bug dealing with subtracting 1 to make the factor indices to be zero based
* Changed EM function to use fastor colSumByGroupChange function to shift counts around after z labels have been altered
* updated documentation
* Added 'src/*.so' statement
* Changed 'split.each.z' to 'cC.splitZ' which utilizes the rowSumByGroupChange function to greatly increase speed in celda_C. However a separate function will be needed for the cells in celda_CG
* changed decomposed matrix to have TS as rows rather than columns (e.g. n.C.by.TS => n.TS.by.C)..
* Adding function for splitting clusters for celda_CG
* Reverted code to making any matrix with genes (G) or transcriptional state (TS) as being the rows. This eliminates the need for the counts.t matrix
* Reverted code to making any matrix with genes (G) or transcriptional state (TS) as being the rows and cells (C) or cell populations (CP) as columns. This eliminates the need for transposing matrices back and forth
* Debugged the new function 'cCG.splitZ' which will perform the cell splits for celda_CG in a much faster and more effecient way
* Refactor rowSumByGroupChange
* Modified the split for celda_G to use the faster rowSumByGroupChange function compared to rowSumByGroup
* Updated celda_CG y splitting function
* Fixed bug in using colSumByGroupChange
* updated NAMESPACE
* Updated testthat to reflect new defaults for distinct_color function
* Needed to added statement to load in previously simulated results
* Added missing lines in testthat for celda_C
* Removed dependency on Rmpfr by using functions from matrixStats package
* Changed default gamma in simulateCells functions to 5. Removed 'precision' argument in perplexity
* added 'seed' parameter to celdaTsne
* Updated NAMESPACE
* Added additional heuristic to cCG.splitY that reduces the size of the original counts matrix in the celda_G cluster splitting with L=2. It does this reduciton by splitting each current cell cluster into 'K.subclusters'. So if K=10 and K.subclusters=10, then the counts matrix will be reduced by clusterig each of the 10-K clusters into 10 additional subclusters (10 x 10 = 100 cells clusters). Therefore instead of performing celda_G on original counts matrix with large number of columns, the matrix will have at most K x K.subcluster columns.
Former-commit-id: b27139bbc2de25edb4c4e5689fa37c8cdb3c3df9
Joshua D. Campbell authored on 22/07/2018 17:49:39
• Sean committed on 22/07/2018 17:49:39
| ... | ... |
@@ -183,8 +183,35 @@ cCG.splitZ = function(counts, m.CP.by.S, n.TS.by.C, n.TS.by.CP, n.by.G, n.by.TS, |
| 183 | 183 |
|
| 184 | 184 |
|
| 185 | 185 |
|
| 186 |
-cCG.splitY = function(counts, y, m.CP.by.S, n.G.by.CP, n.TS.by.C, n.TS.by.CP, n.by.G, n.by.TS, nG.by.TS, n.CP, s, z, K, L, nS, nG, alpha, beta, delta, gamma, y.prob, max.clusters.to.try=10, min.cell=3) {
|
|
| 186 |
+cCG.splitY = function(counts, y, m.CP.by.S, n.G.by.CP, n.TS.by.C, n.TS.by.CP, n.by.G, n.by.TS, nG.by.TS, n.CP, s, z, K, L, nS, nG, alpha, beta, delta, gamma, y.prob, max.clusters.to.try=10, K.subclusters=10, min.cell=3) {
|
|
| 187 | 187 |
|
| 188 |
+ ######################### |
|
| 189 |
+ ## First, the cell dimension of the original matrix will be reduced by splitting each z cluster into 'K.subclusters'. |
|
| 190 |
+ ######################### |
|
| 191 |
+ |
|
| 192 |
+ ## This will not be as big as the original matrix (which can take a lot of time to process with large number of cells), but not as small as the 'n.G.by.CP' with current z assignments |
|
| 193 |
+ z.ta = tabulate(z, K) |
|
| 194 |
+ z.non.empty = which(z.ta > 0) |
|
| 195 |
+ temp.z = rep(0, length(z)) |
|
| 196 |
+ current.top.z = 0 |
|
| 197 |
+ for(i in z.non.empty) {
|
|
| 198 |
+ ix = z == i |
|
| 199 |
+ if(z.ta[i] <= K.subclusters) {
|
|
| 200 |
+ temp.z[ix] = (current.top.z + 1):(current.top.z + z.ta[i]) |
|
| 201 |
+ } else {
|
|
| 202 |
+ clustLabel = suppressMessages(celda_C(counts[,z == i], K=K.subclusters, max.iter=5, split.on.iter=-1, split.on.last=FALSE)) |
|
| 203 |
+ temp.z[ix] = clustLabel$z + current.top.z |
|
| 204 |
+ } |
|
| 205 |
+ current.top.z = max(temp.z, na.rm=TRUE) |
|
| 206 |
+ } |
|
| 207 |
+ |
|
| 208 |
+ ## Decompose counts according to new/temp z labels |
|
| 209 |
+ temp.n.G.by.CP = colSumByGroup(counts, group=temp.z, K=current.top.z) |
|
| 210 |
+ |
|
| 211 |
+ ######################### |
|
| 212 |
+ ## Second, different y splits will be estimated and tested |
|
| 213 |
+ ######################### |
|
| 214 |
+ |
|
| 188 | 215 |
## Identify clusters to split |
| 189 | 216 |
y.ta = tabulate(y, L) |
| 190 | 217 |
y.to.split = which(y.ta >= min.cell) |
| ... | ... |
@@ -194,11 +221,11 @@ cCG.splitY = function(counts, y, m.CP.by.S, n.G.by.CP, n.TS.by.C, n.TS.by.CP, n. |
| 194 | 221 |
m = paste0(date(), " ... Cluster sizes too small. No additional splitting was performed.") |
| 195 | 222 |
return(list(y=y, m.CP.by.S=m.CP.by.S, n.TS.by.CP=n.TS.by.CP, n.CP=n.CP, message=m)) |
| 196 | 223 |
} |
| 197 |
- |
|
| 224 |
+ |
|
| 198 | 225 |
## Loop through each split-able Z and perform split |
| 199 | 226 |
clust.split = vector("list", L)
|
| 200 | 227 |
for(i in y.to.split) {
|
| 201 |
- clustLabel = suppressMessages(celda_G(counts[y == i,], L=2, max.iter=5, split.on.iter=-1, split.on.last=FALSE)) |
|
| 228 |
+ clustLabel = suppressMessages(celda_G(temp.n.G.by.CP[y == i,], L=2, max.iter=5, split.on.iter=-1, split.on.last=FALSE)) |
|
| 202 | 229 |
clust.split[[i]] = clustLabel$y |
| 203 | 230 |
} |
| 204 | 231 |
|
| ... | ... |
@@ -229,7 +256,6 @@ cCG.splitY = function(counts, y, m.CP.by.S, n.G.by.CP, n.TS.by.C, n.TS.by.CP, n. |
| 229 | 256 |
previous.y = new.y |
| 230 | 257 |
} |
| 231 | 258 |
y.to.shuffle = head(order(ll.shuffle, decreasing = TRUE, na.last=NA), n = max.clusters.to.try) |
| 232 |
- |
|
| 233 | 259 |
|
| 234 | 260 |
pairs = c(NA, NA) |
| 235 | 261 |
split.ix = 2 |