forked from djreiss/cMonkey
-
Notifications
You must be signed in to change notification settings - Fork 0
/
cmonkey-funcs.R
1281 lines (1171 loc) · 66.1 KB
/
cmonkey-funcs.R
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
###################################################################################
## cMonkey - version 4, Copyright (C) David J Reiss, Institute for Systems Biology
## This software is provided AS IS with no warranty expressed or implied. Neither
## the authors of this software nor the Institute for Systems Biology shall be held
## liable for anything that happens as a result of using this software
###################################################################################
## This func. is the meat of cmonkey - run one iteration and update the scores/parameters.
## This is called from cmonkey() - see there for its usage.
## Warning - this relies on lots of global variables, that are updated inside the function and
## the updated variables are returned. Again, see cmonkey() for its usage.
## BIG TODO: make it so all big.memory are updated in-place
cmonkey.one.iter <- function( env, dont.update=F, ... ) {
##if ( ! exists( "clust.changed", envir=env ) ) env$clust.changed <- rep( FALSE, k.clust )
if ( ! exists( "row.memb", envir=env ) ) {
env$row.memb <- t( apply( row.membership[], 1, function( i ) 1:k.clust %in% i ) )
env$col.memb <- t( apply( col.membership[], 1, function( i ) 1:k.clust %in% i ) )
env$row.memb <- matrix.reference( env$row.memb, backingfile="row.memb", backingpath=env$cmonkey.filename )
env$col.memb <- matrix.reference( env$col.memb, backingfile="col.memb", backingpath=env$cmonkey.filename )
} else {
env$row.memb[,] <- t( apply( row.membership[], 1, function( i ) 1:k.clust %in% i ) )
env$col.memb[,] <- t( apply( col.membership[], 1, function( i ) 1:k.clust %in% i ) )
}
tmp <- get.all.scores( ... )
env$row.scores <- tmp$r##[,];
env$mot.scores <- tmp$m; env$net.scores <- tmp$n; env$col.scores <- tmp$c
env$meme.scores <- tmp$ms
if ( ! is.null( tmp$cns ) ) env$cluster.net.scores <- tmp$cns
tmp <- get.combined.scores( quant=T )
env$r.scores <- tmp$r; env$c.scores <- tmp$c; env$n.scores <- tmp$n; env$m.scores <- tmp$m
if ( ! is.null( env$row.scores ) ) attr( env$row.scores, "changed" ) <- FALSE
if ( ! is.null( env$col.scores ) ) attr( env$col.scores, "changed" ) <- FALSE
if ( ! is.null( env$mot.scores ) ) attr( env$mot.scores, "changed" ) <- FALSE
if ( ! is.null( env$net.scores ) ) attr( env$net.scores, "changed" ) <- FALSE
if ( length( tmp$scalings ) > 0 ) {
env$row.scaling[ iter ] <- tmp$scalings[ "row" ]
env$mot.scaling[ iter ] <- tmp$scalings[ "mot" ]
env$net.scaling[ iter ] <- tmp$scalings[ "net" ]
}
## Fuzzify scores a bit for stochasticity! (fuzz should be between 0.2 and 0 (decreasing with iter)
if ( fuzzy.index[ iter ] > 1e-5 ) {
env$r.scores[,] <- env$r.scores[,] +
rnorm( length( env$r.scores[,] ), sd=sd( env$r.scores[,][ row.memb[,] == 1 ], na.rm=T ) * fuzzy.index[ iter ] )
if ( ! is.null( env$c.scores ) ) env$c.scores[,] <- env$c.scores[,] +
rnorm( length( env$c.scores[,] ), sd=sd( env$c.scores[,][ col.memb[,] == 1 ], na.rm=T ) * fuzzy.index[ iter ] )
}
tmp <- get.density.scores( ks=1:k.clust ) ## r.scores, col.scores,
env$rr.scores <- tmp$r; env$cc.scores <- tmp$c ##; rm( tmp )
## OUTPUT
if ( iter %in% stats.iters ) {
env$clusterStack <- get.clusterStack( ks=1:k.clust )
env$stats <- rbind( stats, get.stats() )
cat( organism, as.matrix( stats[ nrow( stats ), ] ), '\n' )
} else {
cat( sprintf( "==> %04d %.3f %.3f %.3f\n", iter, ##%5d sum( row.memb != old.row.memb, na.rm=T ),
mean( row.scores[,][ row.memb[,] == 1 ], na.rm=T ), ##mean( col.scores[ col.memb ], na.rm=T, trim=0.05 ),
if ( ! is.null( mot.scores ) ) mean( mot.scores[,][ row.memb[,] == 1 & mot.scores[,] < 0 ], na.rm=T, trim=0.05 )
else NA,
if ( ! is.null( net.scores ) ) mean( net.scores[,][ row.memb[,] == 1 ##& net.scores < 0
], na.rm=T, trim=0.05 ) else NA ) ) ##, "\n" )
}
## NEW - will it work? -- help shrink big clusters, grow small clusters, both in rows and cols
size.compensation.func.rows <- function( n ) exp( -n / ( attr( ratios, "nrow" ) * n.clust.per.row / k.clust ) )
size.compensation.func.cols <- function( n ) exp( -n / ( attr( ratios, "ncol" ) * n.clust.per.col / k.clust ) )
for ( k in 1:k.clust ) {
tmp <- sum( row.memb[ ,k ] )
if ( tmp > 0 ) env$rr.scores[ ,k ] <- env$rr.scores[ ,k ] * size.compensation.func.rows( tmp )
else env$rr.scores[ ,k ] <- env$rr.scores[ ,k ] * size.compensation.func.rows( cluster.rows.allowed[ 1 ] )
if ( ! is.null( env$cc.scores ) ) {
tmp <- sum( col.memb[ ,k ] )
if ( tmp > 0 ) env$cc.scores[ ,k ] <- env$cc.scores[ ,k ] * size.compensation.func.cols( tmp )
else env$cc.scores[ ,k ] <- env$cc.scores[ ,k ] * size.compensation.func.cols( attr( ratios, "ncol" ) / 10 )
}
}
## Fuzzify it along the lines of fuzzy c-means clustering
## -- see http://en.wikipedia.org/wiki/Data_clustering#Fuzzy_c-means_clustering
## No -- it doesnt affect things - same ordering (and updated memberships are based on ordering)
## but should use these scores to weight the centroids that are selected in the next iteration.
## for this, fuzz should vary between e.g. 10 and 2
##rr.scores <- ( rr.scores / sum( rr.scores, na.rm=T ) )^( 2 / ( fuzz - 1 ) )
##cc.scores <- ( cc.scores / sum( cc.scores, na.rm=T ) )^( 2 / ( fuzz - 1 ) )
if ( ! dont.update ) {
## Make a matrix of m[i,k] = whether row/col i is in cluster k
if ( exists( "row.membership" ) ) {
env$old.row.membership <- row.membership
env$old.col.membership <- col.membership
}
tmp <- get.updated.memberships() ## rr.scores, cc.scores )
env$row.membership <- tmp$r; env$col.membership <- tmp$c
if ( ! is.null( tmp ) ) { env$row.membership <- tmp$r; env$col.membership <- tmp$c }
## PLOTTING
if ( ! is.na( plot.iters ) && iter %in% plot.iters ) {
env$clusterStack <- get.clusterStack( ks=1:k.clust )
try( plotStats( iter, plot.clust=env$favorite.cluster(), new.dev=T ), silent=T ) ## Can be set for your given organism
}
if ( exists( "cm.func.each.iter" ) ) try( cm.func.each.iter(), silent=T ) ## User-defined func. to run each iteration
## Allow temp source file to be sourced (e.g. to change a param in the middle of a run, or print or plot or save
## some intermediate results). If first line of file is '## QUIET' then this is done quietly.
if ( any( cm.script.each.iter != "" ) ) {
for ( f in cm.script.each.iter ) {
if ( file.exists( f ) && file.info( f )$size > 1 ) {
tmp <- readLines( f )
if ( all( substr( tmp, 1, 1 ) == "#" ) ) next ## All commented-out code
if ( tmp[ 1 ] != "## QUIET" ) cat( "Sourcing the script '", f, "' ...\n", sep="" )
try( source( f, echo=tmp[ 1 ] != "## QUIET", local=T ), silent=T )
}
}
}
}
## Note: with the above code, when the env is saved via save.image(), all ff obj's are "closed" but their
## filestores still exist, so you can "open.ff(x)" each of them after the env is re-loaded,
## and that will reconnect them with their files.
if ( get.parallel()$mc ) {
if ( getDoParName() == "doMC" ) { ##require( multicore, quietly=T ) ) { ## Clean up any multicore spawned processes (as doc'ed in mclapply help)
chld <- multicore::children()
if ( length( chld ) > 0 ) { try( { multicore::kill( chld ); tmp <- multicore::collect( chld ) }, silent=T ) }
} else if ( getDoParName() == "doSNOW" && "data" %in% ls( pos=foreach:::.foreachGlobals ) ) {
cl <- get( "data", pos=foreach:::.foreachGlobals ) ## Tricky, eh?
if ( ! is.null( data ) ) stopCluster( cl )
}
}
if ( ! dont.update ) env$iter <- env$iter + 1
invisible( env )
}
## TODO: store/plot indiv. resids and network scores for each separate component
get.stats <- function( mean.func=median ) {
##if ( ! exists( "row.memb" ) ) row.memb <- t( apply( row.membership, 1, function( i ) 1:k.clust %in% i ) )
##if ( ! exists( "col.memb" ) ) col.memb <- t( apply( col.membership, 1, function( i ) 1:k.clust %in% i ) )
changed <- NA
if ( ! is.null( old.row.membership ) )
changed <- sum( row.memb[,] != t( apply( old.row.membership, 1, function( i ) 1:k.clust %in% i ) ), na.rm=T )
cs <- as.list( clusterStack )
resids <- sapply( cs, "[[", "resid" )
if ( is.matrix( resids ) ) resids <- apply( resids, 1, function( r ) mean.func( r[ r != 1.0 ], na.rm=T ) )
else resids <- mean.func( resids[ resids != 1.0 ], na.rm=T )
p.clusts <- sapply( cs, "[[", "p.clust" )
if ( is.matrix( p.clusts ) ) p.clusts <- apply( p.clusts, 1, mean.func, na.rm=T )
else p.clusts <- mean.func( p.clusts, na.rm=T )
out <- data.frame( iter=iter, changed=changed,
row.scores=mean( row.scores[,][ row.memb[,] == 1 ], na.rm=T, trim=0.05 ),
col.scores=mean( col.scores[,][ col.memb[,] == 1 ], na.rm=T, trim=0.05 ),
mot.scores=if ( ! is.null( mot.scores ) ) mean.func( mot.scores[,][ row.memb[,] == 1 ], na.rm=T ) else NA,
net.scores=if ( ! is.null( net.scores ) ) mean( net.scores[,][ row.memb[,] == 1 ], na.rm=T, trim=0.05 ) else NA,
resid=weighted.mean( resids, row.weights, na.rm=T ),
nrow=mean.func( sapply( cs, "[[", "nrows" ), na.rm=T ),
ncol=mean.func( sapply( cs, "[[", "ncols" ), na.rm=T ),
p.clust=if ( ! all( is.na( p.clusts ) ) ) weighted.mean( p.clusts, mot.weights, na.rm=T ) else NA
)
if ( length( resids ) > 1 ) for ( i in names( resids ) ) {
out <- cbind( out, resids[ i ] )
names( out )[ ncol( out ) ] <- paste( "resid", i, sep="." )
}
if ( length( p.clusts ) > 1 ) for ( i in names( p.clusts ) ) {
out <- cbind( out, p.clusts[ i ] )
names( out )[ ncol( out ) ] <- paste( "p.clust", i, sep="." )
}
## net.scores weighted by n.rows of bicluster
if ( length( networks ) > 1 ) {
for ( i in names( net.weights ) ) {
if ( exists( "cluster.net.scores" ) && i %in% colnames( cluster.net.scores ) )
out <- cbind( out, weighted.mean( cluster.net.scores[ ,i ], sapply( cs, "[[", "nrows" ), na.rm=T ) )
else out <- cbind( out, rep( NA, nrow( out ) ) )
names( out )[ ncol( out ) ] <- paste( "net", i, sep="." )
}
if ( exists( "cluster.net.scores" ) && "net.scores" %in% colnames( cluster.net.scores ) )
out[ ,"net.scores" ] <- weighted.mean( cluster.net.scores[ ,"net.scores" ],
sapply( cs, "[[", "nrows" ), na.rm=T )
}
out
}
## Only override global parameters if they were not already set previously ----
## but even if they were, copy them into the 'cmonkey.params' environment
set.param <- function( name, val, env=cmonkey.params, override=F, quiet=F ) {
if ( ! exists( name, envir=env ) || override ) {
if ( ! quiet ) try( { cat( name, "-> " ); str( val, digits.d=9, no.list=T ) } )
assign( name, val, envir=env )
} else {
val <- get( name, envir=env )
if ( ! quiet ) try( { cat( name, "= " ); str( val, digits.d=9, no.list=T ) } )
assign( name, val, envir=env )
}
assign( name, val, envir=parent.frame() )
}
## Make sure all processes are killed via kill(children(),SIGKILL) ??
## Use "parallel.cores" param to determine if parallelization is desired - if it is FALSE or 0 or 1 or NA, then no.
## If it is >1 or TRUE then yes. If TRUE, then determine # of cores via multicore:::detectCores()
## This should allow running on Windows boxes when there is no multicore package.
## Note that with the use of "foreach", this can now run on Windows if we use doSMP instead of doMC !!!
## Set the "foreach.register.backend" function to a different backend if desired.
##is.windows.system <- function() FALSE
foreach.register.backend <- function( par ) { ##, force.snow=is.windows.system() ) {
if ( ! require( foreach ) ) return( NULL )
if ( par > 1 && require( doMC, quietly=T ) ) registerDoMC( cores=par )
##if ( par == 1 ) { registerDoSEQ(); return( NULL ) }
##if ( ! is.windows.system() && ! force.snow ) {
## else if require( doSNOW ) {
## if ( ! "data" %in% ls( pos=foreach:::.foreachGlobals ) ) {
## cl <- makeCluster( rep( 'localhost', par ), "SOCK" )
## registerDoSNOW( cl )
## }
##}
else registerDoSEQ()
}
get.parallel <- function( X=k.clust, verbose=F, para.cores=get( "parallel.cores" ) ) {
if ( is.na( para.cores ) || ( is.logical( para.cores ) && para.cores == FALSE ) ||
( is.numeric( para.cores ) && para.cores <= 1 ) ) {
out <- list( mc=FALSE, par=para.cores, apply=lapply )
if ( verbose ) cat( "NOT PARALLELIZING\n" )
} else {
try( has.multi <- require( multicore, quietly=T ), silent=T )
if ( ! has.multi || ( has.multi && multicore:::isChild() ) ) {
out <- list( mc=FALSE, par=para.cores, apply=lapply )
if ( verbose ) cat( "NOT PARALLELIZING\n" )
} else {
mc <- has.multi && ! multicore:::isChild() && X > 1 && ! is.na( para.cores ) &&
( is.numeric( para.cores ) && para.cores > 1 ) ||
( is.logical( para.cores ) && para.cores == TRUE )
par <- para.cores
out.apply <- lapply
if ( mc ) {
if ( is.logical( par ) && par == TRUE ) par <- multicore:::detectCores() ## all.tests=TRUE )
par <- min( c( X, par, multicore:::detectCores() ) ) ## all.tests=TRUE ) ) )
if ( verbose ) cat( "PARALLELIZING:", par, ": " )
## if ( ! exists( "foreach.register.backend" ) || is.null( foreach.register.backend ) ||
## is.null( foreach.register.backend( par ) ) ) { ##use.foreach ) {
## out.apply <- mclapply
## } else {
foreach.register.backend( par )
if ( verbose ) cat( getDoParName(), getDoParWorkers(), "\n" )
out.apply <- function( list, FUN, ... ) foreach( l=list ) %dopar% { FUN( l, ... ) }
## }
##if ( attr( ratios, "nrow" ) > big.run ) print( gc() ) ## gc() before we spawn new copied processes
} else {
par <- 1
if ( verbose ) cat( "NOT PARALLELIZING:", par, "\n" )
}
out <- list( mc=mc, par=par, apply=out.apply )
}
}
##if ( attr( ratios, "nrow" ) > big.run ) print( gc() ) ## gc() before we spawn new copied processes
if ( is.numeric( out$par ) && ! is.na( out$par ) ) options( cores=out$par ) ## getOption("cores") is the default of how mclapply gets its mc.cores number
else if ( is.na( out$par ) || ( is.logical( out$par ) && out$par == TRUE ) ) options( cores=NULL )
else options( cores=1 )
out
}
## Quick and dirty quantile normalization of scores matrices against each other!!!
## Optionally weight the scores matrices!!!
quantile.normalize.scores <- function( scores, weights=NULL, keep.nas=F ) {
## scores is a list of matrices of same dimension
if ( ! is.list( scores ) || sum( ! sapply( scores, is.null ) ) <= 1 ) return( scores )
scores <- scores[ ! sapply( scores, is.null ) ]
##ns <- names( scores )
d <- dim( scores[[ 1 ]] )
dn <- dimnames( scores[[ 1 ]] )
##scores <- sapply( scores, function( s ) as.vector( s[,] ) )
##tmp <- apply( scores, 2, rank, ties="min", na="keep" )
##tmp2 <- apply( scores, 2, sort, na.last=T ) ##sapply( 1:ncol( scores ), function( i ) scores[ tmp[ ,i ], i ] )
tmp2 <- sapply( scores, function( i ) sort( i[,], na.last=T ) ) ##sapply( 1:ncol( scores ), function( i ) scores[ tmp[ ,i ], i ] )
if ( is.null( weights ) ) tmp2.mn <- rowMeans( tmp2, na.rm=T )
else { ##tmp2.mn <- apply( tmp2, 1, weighted.mean, weights, na.rm=T )
for ( i in 1:ncol( tmp2 ) ) tmp2[ ,i ] <- tmp2[ ,i ] * weights[ i ]
tmp2.mn <- rowMeans( tmp2, na.rm=T ) / sum( weights, na.rm=T )
}
rm( tmp2 )
##out <- lapply( 1:ncol( scores ), function( i ) tmp2.mn[ tmp[ ,i ] ] )
##out <- lapply( out, function( i ) { z <- matrix( i, nrow=d[ 1 ], ncol=d[ 2 ] ); dimnames( z ) <- dn; z } )
##names( out ) <- ns
##out
##tmp <- sapply( scores, function( i ) rank( i[,], ties="min", na="keep" ) )
out <- list()
for ( n in names( scores ) ) {
tmp <- rank( scores[[ n ]][,], ties="min", na="keep" )
##scores[[ n ]][,] <-
z <- matrix( tmp2.mn[ tmp ], nrow=d[ 1 ], ncol=d[ 2 ] ) ##tmp[ ,n ] ]
dimnames( z ) <- dn
if ( keep.nas ) z[ is.na( scores[[ n ]][,] ) ] <- NA
out[[ n ]] <- z
}
out ##scores
}
get.all.scores <- function( ks=1:k.clust, force.row=F, force.col=F, force.motif=F, force.net=F,
quantile.normalize=T ) {
mc <- get.parallel( length( ks ) )
## Compute row.scores (microarray data)
if ( force.row || ( row.scaling[ iter ] > 0 && ! is.na( row.iters[ 1 ] ) && iter %in% row.iters ) ) {
##if ( ! exists( "row.scores" ) || is.null( row.scores ) || nrow( row.scores ) != attr( ratios, "nrow" ) ||
## ncol( row.scores ) != max( ks ) ) {
if ( is.null( row.scores ) ) {
row.scores <- matrix( 0, nrow=attr( ratios, "nrow" ), ncol=max( ks ) )
rownames( row.scores ) <- attr( ratios, "rnames" )
row.scores <- matrix.reference( row.scores )
} else row.scores[ ,ks ] <- 0
## TODO: Try to quantile normalize the ratios scores too (as is done w/ mot.scores, net.scores below) but
## this is harder because each different "tmp.row" matrix may have different nrows.
for ( i in names( ratios ) ) {
if ( row.weights[ i ] == 0 || is.na( row.weights[ i ] ) ) next
tmp.row <- do.call( cbind, mc$apply( ks, get.row.scores, ratios=ratios[[ i ]]
) )
tmp <- is.infinite( tmp.row ) | is.na( tmp.row )
if ( any( tmp ) ) tmp.row[ tmp ] <-
quantile( tmp.row[ row.memb[,][ rownames( tmp.row ), ] == 1 & ! tmp ], 0.95 ) ##0 ## No measurement is NOT an NA - it is an observation
tmp <- rownames( row.scores )[ rownames( row.scores ) %in% rownames( tmp.row ) ]
row.scores[ tmp, ks ] <- row.scores[ tmp, ks ] + tmp.row[ tmp, ] * row.weights[ i ]
rm( tmp.row, tmp )
}
attr( row.scores, "changed" ) <- TRUE
}
## Compute col.scores (microarray data)
##if ( ! exists( "col.scores" ) || is.null( col.scores ) || nrow( col.scores ) != attr( ratios, "ncol" ) ||
## ncol( col.scores ) != max( ks ) ) {
if ( n.clust.per.col < k.clust && ## If using all columns, can skip this step...
( force.col || ( row.scaling[ iter ] > 0 && ! is.na( col.iters[ 1 ] ) && iter %in% col.iters ) ) ) {
if ( is.null( col.scores ) ) {
col.scores <- matrix( 0, nrow=attr( ratios, "ncol" ), ncol=max( ks ) )
rownames( col.scores ) <- attr( ratios, "cnames" )
col.scores <- matrix.reference( col.scores )
} else col.scores[ ,ks ] <- 0
##col.scores <- matrix( 0, nrow=attr( ratios, "ncol" ), ncol=max( ks ) )
for ( i in names( row.weights ) ) {
if ( row.weights[ i ] == 0 || is.na( row.weights[ i ] ) ) next
tmp.col <- do.call( cbind, mc$apply( ks, get.col.scores, ratios=ratios[[ i ]]
) )
tmp <- is.infinite( tmp.col ) | is.na( tmp.col )
if ( any( tmp ) ) tmp.col[ tmp ] <-
quantile( tmp.col[ col.memb[,][ rownames( tmp.col ), ] == 1 & ! tmp ], 0.95 ) ##0 ## No measurement is NOT an NA - it is an observation
tmp <- rownames( col.scores )[ rownames( col.scores ) %in% rownames( tmp.col ) ]
col.scores[ tmp, ks ] <- col.scores[ tmp, ks ] + tmp.col[ tmp, ] * row.weights[ i ]
rm( tmp.col, tmp )
}
attr( col.scores, "changed" ) <- TRUE
}
## Run meme on each cluster (every meme.iters iterations)
for ( i in names( mot.weights ) ) {
if ( force.motif == "run.meme" || ( mot.scaling[ iter ] > 0 && ! is.na( meme.iters[[ i ]][ 1 ] ) &&
iter %in% meme.iters[[ i ]] && exists( "genome.info" ) && ! no.genome.info ) ) {
if ( mot.weights[ i ] == 0 || is.na( mot.weights[ i ] ) ) next
tmp <- motif.all.clusters( ks, seq.type=i, verbose=T ) ##strsplit( i, " " )[[ 1 ]][ 1 ],
##algo=strsplit( i, " " )[[ 1 ]][ 2 ] )
meme.scores[[ i ]] <- tmp
}
}
## Compute mot.scores from meme output
if ( force.motif == TRUE || force.motif == "run.meme" || ( mot.scaling[ iter ] > 0 && ! is.na( mot.iters[ 1 ] ) &&
##iter %in% c( meme.iters, mot.iters ) && exists( "genome.info" ) && ! no.genome.info ) ) {
iter %in% mot.iters && exists( "genome.info" ) && ! no.genome.info ) ) {
if ( is.null( mot.scores ) ) {
mot.scores <- matrix( 0, nrow=attr( ratios, "nrow" ), ncol=max( ks ) )
rownames( mot.scores ) <- attr( ratios, "rnames" )
mot.scores <- matrix.reference( mot.scores )
} else mot.scores[ ,ks ] <- 0
tmp.mots <- list()
for ( i in names( mot.weights ) ) {
if ( mot.weights[ i ] == 0 || is.na( mot.weights[ i ] ) ) next
tmp.mot <- do.call( cbind, mc$apply( ks, get.motif.scores, seq.type=i ) )
tmp.mot[ is.infinite( tmp.mot ) | is.na( tmp.mot ) ] <- 0 ## No measurement is NOT an NA - it is an observation
if ( quantile.normalize && sum( mot.weights > 0 & ! is.na( mot.weights ) ) > 1 ) tmp.mots[[ i ]] <- tmp.mot
else mot.scores[ ,ks ] <- mot.scores[ ,ks ] + tmp.mot[,] * mot.weights[ i ]
rm( tmp.mot )
}
if ( quantile.normalize && length( tmp.mots ) > 1 ) {
tmp.mots <- quantile.normalize.scores( tmp.mots, weights=mot.weights[ mot.weights != 0 ] )
for ( i in names( tmp.mots ) ) mot.scores[ ,ks ] <- mot.scores[ ,ks ] + tmp.mots[[ i ]][,] * mot.weights[ i ]
rm( tmp.mots )
}
attr( mot.scores, "changed" ) <- TRUE
}
cluster.ns <- NULL
## Compute net.scores from STRING ... add weighted scores for other networks (if exist)
if ( force.net || ( net.scaling[ iter ] > 0 && ! is.na( net.iters[ 1 ] ) && exists( "genome.info" ) &&
iter %in% net.iters ) ) {
##net.scores <- NULL
if ( is.null( net.scores ) ) {
net.scores <- matrix( 0, nrow=attr( ratios, "nrow" ), ncol=max( ks ) )
rownames( net.scores ) <- attr( ratios, "rnames" )
net.scores <- matrix.reference( net.scores )
} else net.scores[ ,ks ] <- 0
tmp.nets <- list()
for ( i in names( networks ) ) {
if ( net.weights[ i ] == 0 || is.na( net.weights[ i ] ) ) next
if ( nrow( subset( networks[[ i ]], protein1 %in% attr( ratios, "rnames" ) & protein2 %in%
attr( ratios, "rnames" ) ) ) <= 0 ) next
tmp.net <- do.call( cbind, mc$apply( ks, get.network.scores, net=networks[[ i ]] ) )
if ( all( is.na( tmp.net ) ) || all( is.character( tmp.net ) ) ) next
##tmp.net[,] <- tmp.net[,] - max( tmp.net[ ! is.infinite( tmp.net ) ], na.rm=T ) -
## abs( diff( range( tmp.net[ ! is.infinite( tmp.net ) ], na.rm=T ) ) ) / 10
tmp.net[ is.infinite( tmp.net ) | is.na( tmp.net ) ] <- 0 ## No edge is NOT an NA - it is an observation
if ( quantile.normalize && sum( net.weights > 0 & ! is.na( net.weights ) ) > 1 ) tmp.nets[[ i ]] <- tmp.net
else net.scores[ ,ks ] <- net.scores[ ,ks ] + tmp.net[,] * net.weights[ i ]
cluster.ns <- cbind( cluster.ns, do.call( c, mc$apply( ks, function( k ) mean( tmp.net[ get.rows( k ), k ],
na.rm=T, trim=0.05 ) ) ) )
colnames( cluster.ns )[ ncol( cluster.ns ) ] <- i
rm( tmp.net )
}
if ( quantile.normalize && length( tmp.nets ) > 1 ) {
tmp.nets <- quantile.normalize.scores( tmp.nets, weights=net.weights[ net.weights != 0 ] )
for ( i in names( tmp.nets ) ) net.scores[ ,ks ] <- net.scores[ ,ks ] + tmp.nets[[ i ]][,] * net.weights[ i ]
rm( tmp.nets )
}
attr( net.scores, "changed" ) <- TRUE
cluster.ns <- cbind( cluster.ns, do.call( c, mc$apply( ks, function( k ) mean( net.scores[ get.rows( k ), k ],
na.rm=T, trim=0.05 ) ) ) )
colnames( cluster.ns )[ ncol( cluster.ns ) ] <- "net.scores"
}
list( r=row.scores, m=mot.scores, ms=meme.scores, n=net.scores, c=col.scores, cns=cluster.ns ) ##r=row.scores,
}
get.rows <- function( k, rm=get("row.membership") ) { out <- unique( rownames( which( rm[] == k, arr=T ) ) );
if ( is.null( out ) ) out <- character(); out }
get.cols <- function( k, cm=get("col.membership") ) { out <- unique( rownames( which( cm[] == k, arr=T ) ) );
if ( is.null( out ) ) out <- character(); out }
#' Create old clusterStack-style object from row.membership and col.membership
#' SD Notes: force is a little mysterious to me.
#'
#' @params ks The clusters to include (DEFAULT: 1:k.clust)
#' @params force If FALSE, will return the existing clusterStack if possible (DEFAULT: F)
#' @params ... Must include atleast "iter"
#' @export
#' @usage clusterStack <- get.clusterStack( ks=1:k.clust, force=F, ... )
get.clusterStack <- function( ks=1:k.clust, force=F, ... ) {
if ( ! force && ! is.null( attr( clusterStack, "iter" ) ) && attr( clusterStack, "iter" ) == iter )
return( clusterStack )
mc <- get.parallel( length( ks ) )
clusterStack <- mc$apply( ks, get.clust, ... ) ###, mc.cores=mc$par )
attr( clusterStack, "iter" ) <- iter
clusterStack
}
#' Get the unpreprocessed ratios matrix (i.e. ratios.raw)
#'
#' @export
#' @usage ratios <- get.unpreprocessed.ratios()
get.unpreprocessed.ratios <- function( ... ) {
return( ratios.raw )
}
cluster.resid <- function( k, rats.inds="COMBINED", varNorm=F, in.cols=T, ... ) {
## FLOC residual number is a good statistic
residual.submatrix <- function( rats, rows, cols, varNorm=F, ... ) {
rows <- rows[ rows %in% rownames( rats ) ]
cols <- cols[ cols %in% colnames( rats ) ]
if ( length( rows ) <= 1 || length( cols ) <= 1 ) return( 1 )
maxRowVar <- attr( rats, "maxRowVar" )
rats <- rats[ rows, cols ]
if ( is.vector( rats ) || any( dim( rats ) <= 1 ) || mean( is.na( rats ) ) > 0.95 ) return( 1 )
d.rows <- rowMeans( rats, na.rm=T )
d.cols <- colMeans( rats, na.rm=T )
d.all <- mean( d.rows, na.rm=T )
rij <- rats + d.all
rij[,] <- rij[,] - matrix( d.cols, nrow=nrow( rij ), ncol=ncol( rij ), byrow=T )
rij[,] <- rij[,] - matrix( d.rows, nrow=nrow( rij ), ncol=ncol( rij ), byrow=F )
##rij[,] <- rij[,] - outer( d.rows, d.cols, '+' ) ## more elegant but slower!
average.r <- mean( abs( rij ), na.rm = TRUE )
if ( varNorm && ! is.null( maxRowVar ) ) {
##maxRowVar <- attr( rats, "maxRowVar" )
row.var <- mean( apply( rats, 1, var, use = "pairwise.complete.obs" ), na.rm=T )
if ( is.na( row.var ) || row.var > maxRowVar ) row.var <- maxRowVar
average.r <- average.r / row.var
}
average.r
}
inds <- rats.inds
if ( rats.inds[ 1 ] == "COMBINED" ) inds <- names( get( "row.weights" ) )
resids <- sapply( ratios[ inds ], function( rn ) {
if ( in.cols ) residual.submatrix( rn, get.rows( k ), get.cols( k ), varNorm=varNorm )
else residual.submatrix( rn, get.rows( k ), colnames( rn )[ ! colnames( rn ) %in% get.cols( k ) ],
varNorm=varNorm )
} )
if ( rats.inds[ 1 ] == "COMBINED" ) resids <- weighted.mean( resids, row.weights[ inds ], na.rm=T )
if ( rats.inds[ 1 ] != "COMBINED" && length( resids ) < length( inds ) && all( is.na( resids ) ) ) {
resids <- rep( NA, length( inds ) ); names( resids ) <- inds }
resids
}
cluster.pclust <- function( k, mot.inds="COMBINED" ) { ## actually returns a list with e-vals and p-vals
inds <- mot.inds
if ( is.null( inds ) ) return( list( p.clusts=NA, e.vals=NA ) )
if ( mot.inds[ 1 ] == "COMBINED" ) inds <- names( get( "mot.weights" ) )
if ( is.null( inds ) ) return( list( p.clusts=NA, e.vals=NA ) )
rows <- get.rows( k )
p.clusts <- sapply( inds, function( n ) {
##ms <- meme.scores[[ n ]][[ k ]]
##if ( length( rows ) > 0 && ! is.null( ms$pv.ev ) && ! is.null( ms$pv.ev[[ 1 ]] ) )
## mean( log10( ms$pv.ev[[ 1 ]][ rownames( ms$pv.ev[[ 1 ]] ) %in% rows, "p.value" ] ), na.rm=T ) else NA
pvs <- meme.scores[[ n ]]$all.pv
if ( is.null( pvs ) ) return( NA )
rows <- rows[ rows %in% rownames( pvs ) ]
if ( length( rows ) > 0 ) mean( log10( pvs[ rows, k ] ), na.rm=T ) else NA
} )
e.vals <- sapply( inds, function( n ) {
ms <- meme.scores[[ n ]][[ k ]]
sapply( 1:length( ms$meme.out ),
function( i ) if ( length( rows ) > 0 && ! is.null( ms$meme.out ) && ! is.null( ms$meme.out[[ i ]] ) )
ms$meme.out[[ i ]]$e.value else NA )
} )
if ( ! is.matrix( e.vals ) ) e.vals <- t( t( e.vals ) )
if ( mot.inds[ 1 ] == "COMBINED" ) {
p.clusts <- weighted.mean( p.clusts, mot.weights[ inds ], na.rm=T )
e.vals <- apply( e.vals, 2, weighted.mean, mot.weights[ inds ], na.rm=T )
} else if ( mot.inds[ 1 ] != "COMBINED" ) {
if ( length( p.clusts ) < length( inds ) && all( is.na( p.clusts ) ) ) p.clusts <- rep( NA, length( inds ) )
e.vals <- apply( e.vals, 1, function( i ) if ( length( i ) < length( inds ) && all( is.na( i ) ) )
rep( NA, length( inds ) ) else i )
}
if ( ! is.matrix( e.vals ) ) e.vals <- t( t( e.vals ) )
if ( is.matrix( e.vals ) && ncol( e.vals ) != length( inds ) ) e.vals <- t( e.vals )
if ( mot.inds[ 1 ] != "COMBINED" ) names( p.clusts ) <- colnames( e.vals ) <- inds
else e.vals <- as.vector( e.vals )
list( p.clusts=p.clusts, e.vals=e.vals )
}
## Create old clusterStack-style cluster object from row.membership and col.membership
get.clust <- function( k, fill=T, fill.motif=T, seq.type=names( mot.weights ), varNorm=F, ... ) { ##"upstream" ) {
gen.clust <- function( rowNames, colNames=NA, fill=F, motif=F, n.motifs=3, ... ) {
rowNames <- rowNames[ rowNames %in% attr( ratios, "rnames" ) ]
if ( ! is.null( colNames ) && length( colNames ) > 1 && ! is.na( colNames ) )
colNames <- colNames[ colNames %in% attr( ratios, "cnames" ) ]
c.tmp <- list( nrows=length( rowNames ), ncols=length( colNames ), rows=rowNames, cols=colNames,
k=999, p.clust=1.0, e.val=rep( 999, n.motifs ), resid={out=rep(NA,length(row.weights));names(out)<-names(row.weights);resid=out} )
if ( fill && c.tmp$nrows > 0 && c.tmp$ncols > 0 && ! all( is.na( colNames ) ) )
c.tmp$resid <- cluster.resid( k, names( row.weights ), varNorm=varNorm, ... )
return( c.tmp )
}
cols <- get.cols( k ); rows <- get.rows( k )
if ( length( cols ) <= 0 ) cols <- NA
clust <- gen.clust( rows, cols, fill=fill, motif=F, n.motifs=max( unlist( n.motifs ) ) )
clust$k <- k
if ( fill.motif ) {
tmp <- cluster.pclust( k, seq.type )
clust$e.val <- tmp$e.vals
clust$p.clust <- tmp$p.clusts
}
clust
}
## Allow weighting of columns (used in get.row.scores) (this can be "overridden")
get.col.weights <- function( rows, cols, ratios ) {
NA ## Default: unweighted
## if ( length( rows ) == 1 ) return( NA ) ## Below is example code weighting columns by their "index of dispersion" over the rows in the cluster
## rats <- ratios[ rows, cols, drop=F ]
## vars <- apply( rats, 2, var, na.rm=T )
## vars <- vars / ( abs( apply( rats, 2, mean, na.rm=T ) ) + 0.05 )
## vars[ is.na( vars ) | vars == 0 ] <- 1
## weights <- 1 / vars
## weights <- weights / sum( weights, na.rm=T ) * length( weights )
## names( weights ) <- cols
## weights
}
## Allow weighting of rows (used in get.col.scores) (this can be "overridden")
get.row.weights <- function( rows, cols, ratios ) NA
get.row.scores <- function( k, cols=get.cols( k ), for.rows="all", ratios=ratios[[ 1 ]],
... ) {
## Compute scores for ALL rows (over just the cols IN the cluster)
if ( length( k ) <= 0 ) return( NULL )
if ( is.numeric( k[ 1 ] ) ) rows <- get.rows( k )
else rows <- k
if ( for.rows[ 1 ] == "all" ) for.rows <- rownames( ratios ) ##attr( ratios, "rnames" )
rows <- rows[ rows %in% rownames( ratios ) ]
cols <- cols[ cols %in% colnames( ratios ) ]
if ( length( rows ) < 1 || length( cols ) <= 1 ) return( rep( NA, length( for.rows ) ) )
rats <- ratios[ for.rows, cols, drop=F ]
rats.mn <- colMeans( rats[ rows, , drop=F ], na.rm=T )
rats.mn <- matrix( rats.mn, nrow=nrow( rats ), ncol=ncol( rats ), byrow=T )
rats[,] <- ( rats[,] - rats.mn )^2 ## abs(
col.weights <- if ( exists( "get.col.weights" ) ) get.col.weights( rows, cols, ratios ) else NA
if ( is.na( col.weights[ 1 ] ) ) rats <- rowMeans( rats, na.rm=T )
else rats <- apply( abs( rats ), 1, weighted.mean, w=col.weights[ cols ], na.rm=T )
rats <- log( rats + 1e-99 )
return( rats )
}
## get.row.scores.cpp <- function( rats, rats.mn ) { ##, rank=rank.ties ) {
## if ( TRUE && ! is.loaded( "get_row_scores" ) ) {
## so.file <- "cpp/get_row_scores"
## if ( ! exists( "is.slave" ) || ! is.slave ) {
## require( Rcpp ) ## Dont really need to load the package - but need to make sure it's installed
## code <-
## "RcppExport SEXP get_row_scores( SEXP rats, SEXP rats_mn ) {
## SEXP rl = R_NilValue; // Use this when there is nothing to be returned.
## char* exceptionMesg = NULL;
## try {
## RcppMatrixView<double> r( rats );
## RcppVectorView<double> rm( rats_mn );
## int nr = r.dim1(), nc = r.dim2();
## vector<double> out( nr );
## vector<int> ni( nr ); //, nj( nc );
## for ( int j = 0; j < nc; j ++ ) {
## double rr = rm( j );
## if ( ISNA( rr ) ) continue;
## //nj[ j ] = nj[ j ] + 1;
## for ( int i = 0; i < nr; i ++ ) {
## double rrr = r( i, j );
## if ( ISNA( rrr ) ) continue;
## out[ i ] = out[ i ] + fabs( rrr - rr ); //pow( rrr - rr, 2 );
## ni[ i ] = ni[ i ] + 1;
## }
## }
## for ( int i = 0; i < nr; i ++ ) out[ i ] = out[ i ] / ni[ i ];
## RcppResultSet rs; // Build result set to be returned as a list to R.
## rs.add( \"out\", out );
## rl = rs.getReturnList(); // Get the list to be returned to R.
## } catch( std::exception& ex ) {
## exceptionMesg = copyMessageToR( ex.what() );
## } catch( ... ) {
## exceptionMesg = copyMessageToR( \"unknown reason\" );
## }
## if ( exceptionMesg != NULL ) error( exceptionMesg );
## return rl;
## }"
## source( "Rcpp_compile3.R", local=T ) ## Loads compile.cpp.code() (requires Rcpp package)
## compile.rcpp.code( code, fname=so.file, verbose=T )
## } else {
## dyn.load( paste( so.file, ".so", sep="" ) )
## }
## }
## out <- .Call( "get_row_scores", rats, rats.mn )$out
## out[ out == 0 ] <- NA
## out
## }
## Default is to normalize (diff)^2 by mean expression level, similar to "index of dispersion"
## http://en.wikipedia.org/wiki/Index_of_dispersion
get.col.scores <- function( k, for.cols="all", ratios=ratios[[ 1 ]],
norm.method=c("mean","all.colVars","none")[1], ... ) {
## Compute scores for ALL cols (over just the rows IN the cluster)
if ( length( k ) <= 0 ) return( NULL )
if ( is.numeric( k[ 1 ] ) ) rows <- get.rows( k )
else rows <- k
if ( for.cols[ 1 ] == "all" ) for.cols <- colnames( ratios ) ##attr( ratios, "cnames" )
rows <- rows[ rows %in% rownames( ratios ) ]
if ( length( rows ) <= 1 ) return( rep( NA, length( for.cols ) ) )
rats <- ratios[ rows, for.cols, drop=F ]
row.weights <- if ( exists( "get.row.weights" ) ) get.row.weights( rows, cols, ratios ) else NA
if ( is.na( row.weights[ 1 ] ) ) { ## Default
rats.mn <- matrix( colMeans( rats, na.rm=T ), nrow=nrow( rats ), ncol=ncol( rats ), byrow=T )
} else { ## Custom row weights
rats.mn <- matrix( apply( rats, 2, weighted.mean, w=row.weights[ rows ], na.rm=T ), ncol=ncol( rats ), byrow=T )
}
rats[,] <- ( rats[,] - rats.mn )^2 ## abs( Multiplying is faster than squaring
rats <- colMeans( rats, na.rm=T )
var.norm <- 0.99
if ( norm.method == "all.colVars" ) {
all.colVars <- attr( ratios, "all.colVars" )
if ( ! is.null( all.colVars ) ) var.norm <- all.colVars[ for.cols ]
} else if ( norm.method == "mean" ) {
var.norm <- abs( rats.mn[ 1, ] ) ##0.99 ## Use the mean expr. level (higher expressed expected to have higher noise)
}
##col.weights <- get.col.weights( rows, cols )
##if ( is.na( col.weights ) )
rats <- rats / ( var.norm + 0.01 ) ## default
##!else rats <- colMeans( rats, na.rm=T ) / ( var.norm * col.weights[ cols ] + 0.01 ) ## customized col. weights
##return( log( rats + 1e-99 ) )
rats
}
get.motif.scores <- function( k, seq.type="upstream meme", for.rows="all" ) { ##m=meme.scores$upstream[[ k ]], for.rows="all" ) {
if ( length( k ) <= 0 ) return( NULL )
if ( is.numeric( k[ 1 ] ) ) rows <- get.rows( k )
else rows <- k
if ( for.rows[ 1 ] == "all" ) for.rows <- attr( ratios, "rnames" )
if ( length( rows ) <= 1 || is.null( meme.scores[[ seq.type ]]$all.pv ) ) ##is.null( m ) || is.null( m$pv.ev ) )
return( rep( NA, length( for.rows ) ) )
##m.scores <- log( m$pv.ev[[ 1 ]][ ,"p.value" ] ); names( m.scores ) <- rownames( m$pv.ev[[ 1 ]] )
m.scores <- log( meme.scores[[ seq.type ]]$all.pv[ ,k ] )
m.scores <- m.scores[ for.rows ] ##; names( m.scores ) <- for.rows
return( m.scores )
}
get.network.scores <- function( k, net=networks$string, for.rows="all", p1.col="protein1", p2.col="protein2",
score.col="combined_score" ) {
if ( length( k ) <= 0 ) return( NULL )
if ( is.numeric( k[ 1 ] ) ) rows <- get.rows( k )
else rows <- k
if ( for.rows[ 1 ] == "all" ) for.rows <- attr( ratios, "rnames" )
if ( length( rows ) < 1 ) return( rep( NA, length( for.rows ) ) )
cons <- net[ as.character( net[[ p1.col ]] ) %in% rows, c( p2.col, score.col ), drop=F ] ## Note subset is slower!
if ( is.null( cons ) || nrow( cons ) <= 0 ) return( rep( NA, length( for.rows ) ) )
cons <- cons[ as.character( cons[[ p2.col ]] ) %in% for.rows, , drop=F ]
##cons <- get.cluster.network( rows, net, for.rows )
if ( is.null( cons ) || nrow( cons ) <= 0 ) return( rep( NA, length( for.rows ) ) )
tmp <- tapply( as.numeric( cons[[ score.col ]] ), as.character( cons[[ p2.col ]] ), sum, na.rm=T ) / length( rows )
scores <- rep( NA, length( for.rows ) ); names( scores ) <- for.rows
scores[ names( tmp ) ] <- tmp
return( -log( scores + 1 ) )
}
pareto.adjust.weights <- function( kwin=51, diter=21, max.delta=0.05 ) {
out.scaling <- c( row=row.scaling[ iter - 1 ], mot=mot.scaling[ iter - 1 ], net=net.scaling[ iter - 1 ] )
orig.scaling <- c( row=row.scaling[ iter ], mot=mot.scaling[ iter ], net=net.scaling[ iter ] )
if ( iter < diter/2 ) return( out.scaling )
diter <- min( diter, iter )
kwin <- min( kwin, iter )
avs <- NULL
for ( i in names( out.scaling ) ) {##c( "row", "mot", "net" ) ) {
if ( i == "row" ) col <- "row.scores" ##"resid"
else if ( i == "mot" ) col <- "mot.scores" ##"p.clust"
else col <- paste( i, "scores", sep="." )
tmp <- stats[ ,col ]; tmp <- tmp[ ! is.na( tmp ) ]
if ( length( tmp ) > 0 ) {
av <- runmed( tmp, k=kwin )
dy <- ( av[ length( av ) ] - av[ max( 1, length( av ) - diter ) ] ) / diff( range( av, na.rm=T ) )
if ( ! is.na( dy ) ) out.scaling[ i ] <- out.scaling[ i ] + min( 0.1, dy )
out.scaling[ i ] <- max( out.scaling[ i ], orig.scaling[ i ] )
}
}
out.scaling
}
pareto.adjust.weights.OLD <- function( iter, delta.iter=200, delta.factor=1, n.avg=50, max.delta=0.05 ) { ##0.01
if ( iter == 1 ) return( c( row=row.scaling[ iter ], mot=mot.scaling[ iter ], net=net.scaling[ iter ] ) )
out.scaling <- c( row=row.scaling[ iter - 1 ], mot=mot.scaling[ iter - 1 ], net=net.scaling[ iter - 1 ] )
if ( iter < delta.iter + n.avg + 10 ) return( out.scaling )
all.diffs <- numeric()
for ( i in c( "row", "mot", "net" ) ) {
if ( i == "row" ) col <- "resid"
else if ( i == "mot" ) col <- "p.clust"
else col <- paste( i, "scores", sep="." )
tops <- stats[[ col ]][ (iter-n.avg+1):iter ]
bots <- stats[[ col ]][ (iter-n.avg+1):iter - delta.iter ]
all.diffs[ i ] <- mean( tops - bots, na.rm=T ) / abs( diff( range( stats[[ col ]], na.rm=T ) ) )
}
for ( i in which( all.diffs > 0 & ! is.na( all.diffs ) & ! is.na( out.scaling ) & out.scaling > 0 ) )
out.scaling[ names( all.diffs )[ i ] ] <- out.scaling[ names( all.diffs )[ i ] ] +
min( c( out.scaling[ i ] * max.delta, all.diffs[ i ] * delta.factor ) )
out.scaling
}
## WARNING: This function relies on A LOT of global variables!!!
get.combined.scores <- function( quantile.normalize=F ) {
if ( ! exists( "r.scores" ) || is.null( r.scores ) ) {
r.scores <- row.scores[,]
r.scores <- matrix.reference( r.scores )
}
r.scores[,] <- row.scores[,]
##if ( attr( row.scores, "changed" ) == TRUE ) {
tmp <- r.scores[,] < -20; r.scores[,][ tmp ] <- min( r.scores[,][ ! tmp ], na.rm=T ); rm( tmp ) ## -220
r.scores[,][ is.infinite( r.scores[,] ) ] <- NA
r.scores[,][ is.na( r.scores[,] ) ] <- max( r.scores[,], na.rm=T )
## }
##cat( "HERE: row", r.scores[1,1], "\n" )
if ( ! is.null( mot.scores ) ) {
if ( ! exists( "m.scores" ) || is.null( m.scores ) ) {
m.scores <- mot.scores[,]
m.scores <- matrix.reference( m.scores ) ## not useful as m.scores (and n.scores) are not saved in the env.
} else m.scores[,] <- mot.scores[,]
} else m.scores <- NULL
if ( ! is.null( mot.scores ) && ! is.null( m.scores ) && ( is.null( attr( mot.scores, "changed" ) ) ||
attr( mot.scores, "changed" ) == TRUE ) ) { ## Only need to do this if net scores changed!!
##cat("HERE: mot", mot.scores[1,1], "\n")
tmp <- m.scores[,] < -20; m.scores[,][ tmp ] <- min( m.scores[,][ ! tmp ], na.rm=T ); rm( tmp ) ## effective zero is 10^-20
} ##!else m.scores <- NULL
if ( ! is.null( net.scores ) ) {
if ( ! exists( "n.scores" ) || is.null( n.scores ) ) {
n.scores <- net.scores[,]
n.scores <- matrix.reference( n.scores ) ## not useful as m.scores (and n.scores) are not saved in the env.
} else n.scores[,] <- net.scores[,]
} else n.scores <- NULL
if ( ! is.null( net.scores ) && ! is.null( n.scores ) && ( is.null( attr( net.scores, "changed" ) ) ||
attr( net.scores, "changed" ) == TRUE ) ) { ## Only need to do this if net scores changed!!
##cat("HERE: net", n.scores[1,1],"\n")
n.scores[,] <- n.scores[,] - quantile( n.scores[,], 0.99, na.rm=T ) ## Make it so no edge -> zero score -> no penalization
}
if ( ! is.null( col.scores ) ) {
## c.scores <- NULL
if ( ! exists( "c.scores" ) || is.null( c.scores ) ) {
c.scores <- col.scores[,] * 0
c.scores <- matrix.reference( c.scores )
} else c.scores[,] <- col.scores[,]
if ( is.null( attr( col.scores, "changed" ) ) || attr( col.scores, "changed" ) == TRUE ) {
##cat("HERE: col", col.scores[1,1], "\n")
tmp <- c.scores[,] < -20; c.scores[,][ tmp ] <- min( c.scores[,][ ! tmp ], na.rm=T ); rm( tmp ) ## effective zero ##-220
}
} else c.scores <- NULL
new.weights <- c( row=row.scaling[ iter ], mot=mot.scaling[ iter ], net=net.scaling[ iter ] ) ##numeric()
## Hey, instead of just making each distrib. have the same mean and variance, let's make them have the same
## exact distribution!
if ( quantile.normalize ) {
tmp <- list( row=r.scores, mot=m.scores, net=n.scores )
if ( sum( sapply( tmp, function( i ) ! is.null( i ) ) ) > 1 ) {
wts <- new.weights[ ! sapply( tmp, is.null ) ]
tmp <- quantile.normalize.scores( tmp, weights=wts )
if ( ! is.null( r.scores ) ) r.scores[,] <- tmp$row[,]
if ( ! is.null( m.scores ) ) m.scores[,] <- tmp$mot[,]
if ( ! is.null( n.scores ) ) n.scores[,] <- tmp$net[,]
}
rm( tmp )
}
if ( new.weights[ "row" ] != 1 ) r.scores[,] <- r.scores[,] * new.weights[ "row" ]
if ( ! is.null( m.scores ) ) {
tmp <- ! is.na( m.scores[,] )
r.scores[,][ tmp ] <- r.scores[,][ tmp ] + m.scores[,][ tmp ] * new.weights[ "mot" ]
}
if ( ! is.null( n.scores ) ) {
tmp <- ! is.na( n.scores[,] )
r.scores[,][ tmp ] <- r.scores[,][ tmp ] + n.scores[,][ tmp ] * new.weights[ "net" ]
}
r.scores <- matrix.reference( r.scores )
c.scores <- matrix.reference( c.scores )
if ( ! is.null( n.scores ) ) n.scores <- matrix.reference( n.scores )
if ( ! is.null( m.scores ) ) m.scores <- matrix.reference( m.scores )
invisible( list( r=r.scores, c=c.scores, n=net.scores, m=mot.scores, scalings=new.weights ) )
}
## Hack to get more genes added to really small (<10 rows) clusters
get.density.scores <- function( ks=1:k.clust, plot="none", bw.scale=function( nr ) exp( -nr / 10 ) * 10 ) { ##r.scores, col.scores,
rr <- attr( ratios, "rnames" )
rs <- r.scores ##[,]
## TODO: adjust bw to be function of iter (is this necessary with fuzzy.index turned on???)
bw.r <- max( diff( range( rs[,], na.rm=T ) ) / 100, 0.001 ) ##"nrd0"
get.rr.scores <- function( k ) {
rows <- get.rows( k )
cols <- get.cols( k )
rsk <- rs[ ,k, drop=T ]
if ( length( rows ) > 0 && length( cols ) > 0 ) {
bw <- bw.r * bw.scale( length( rows ) )
d <- density( rsk[ rows ], na.rm=T, bw=bw, adjust=2,
from=min( rsk, na.rm=T ) - 1, to=max( rsk, na.rm=T ) + 1, n=256 )
p <- approx( d$x, rev( cumsum( rev( d$y ) ) ), rsk )$y
if ( "rows" %in% plot ) { h=hist(rsk,breaks=50,main=NULL,xlab="Combined scores");tmp.scale <- round( attr( ratios, "nrow" ) / length( rows )/4 );hist(rep(rsk[rows],tmp.scale),breaks=h$breaks,col="red",border="red",add=T);hist(rsk,breaks=h$breaks,add=T);lines(d$x,d$y/max(d$y,na.rm=T)*attr(ratios,"nrow")/50,col="blue");lines(sort(rsk),p[order(rsk)]/max(p,na.rm=T)*attr(ratios,"nrow")/50,col="green") }
} else {
p <- rep( 1, attr( ratios, "nrow" ) )
}
return( p / sum( p, na.rm=T ) ) ##* length( rows )
}
if ( ! exists( "rr.scores" ) || is.null( rr.scores ) || any( dim( rr.scores ) != dim( r.scores ) ) ) {
rr.scores <- row.scores[,] * 0
rr.scores <- matrix.reference( rr.scores )
}
mc <- get.parallel( length( ks ) )
rr.scores[,] <- do.call( cbind, mc$apply( ks, get.rr.scores ) )
rr.scores[,][ is.infinite( rr.scores[,] ) ] <- NA
##rownames( rr.scores ) <- attr( ratios, "rnames" )
cc.scores <- NULL ## this is required for n.clust.per.col == k.clust !!! dont delete it!
if ( ! is.null( col.scores ) ) {
cs <- c.scores ##[,]
bw.c <- max( diff( range( cs[,], na.rm=T ) ) / 100, 0.001 )
get.cc.scores <- function( k ) {
cols <- get.cols( k )
rows <- get.rows( k )
csk <- cs[ ,k, drop=T ]
if ( length( cols ) > 0 && length( rows ) > 0 &&
! all( is.na( csk[ cols ] ) ) && ! all( is.infinite( csk[ cols ] ) ) &
! all( csk[ cols ][ ! is.na( csk[ cols ] ) ] == csk[ cols[ ! is.na( csk[ cols ] ) ][ 1 ] ] ) ) {
d <- density( csk[ cols ], na.rm=T, from=min( csk, na.rm=T ) - 1, to=max( csk, na.rm=T ) + 1,
bw=bw.c, adjust=2, n=256 )
p <- approx( d$x, rev( cumsum( rev( d$y ) ) ), csk )$y
if ( "cols" %in% plot ) { h=hist(csk,breaks=50,main=NULL,xlab="Combined scores");tmp.scale <- round( attr( ratios, "ncol" ) / length( cols )/4 ) + 1;hist(rep(csk[cols],tmp.scale),breaks=h$breaks,col="red",border="red",add=T);hist(csk,breaks=h$breaks,add=T);lines(d$x,d$y/max(d$y,na.rm=T)*attr(ratios,"ncol")/50,col="blue");lines(sort(csk),p[order(csk)]/max(p,na.rm=T)*attr(ratios,"ncol")/50,col="green") }
} else {
p <- rep( 1, attr( ratios, "ncol" ) )
}
return( p / sum( p, na.rm=T ) ) ##* length( cols )
}
if ( ! exists( "cc.scores" ) || is.null( cc.scores ) || any( dim( cc.scores ) != dim( c.scores ) ) ) {
cc.scores <- col.scores[,] * 0
cc.scores <- matrix.reference( cc.scores )
}
if ( ! is.null( c.scores ) ) { ##&& ! is.na( c.scores ) ) {
cc.scores[,] <- do.call( cbind, mc$apply( ks, get.cc.scores ) )
cc.scores[,][ is.infinite( cc.scores ) ] <- NA
##rownames( cc.scores ) <- attr( ratios, "cnames" )
}
}
invisible( list( r=rr.scores, c=cc.scores ) )
}
## This reassigning of all rows over all clusters is too much ... problem is crappy clusters end up losing
## ALL rows, and okay clusters get TOO MANY. IDEA: allow only 1 change per gene per iter
## TODO: When updating, don't let number of genes in a cluster get > max.cluster.rows or < min.cluster.rows
get.updated.memberships <- function() { ## rr.scores, cc.scores ) {
##if ( ! is.null( rr.scores ) ) {
##n.rows <- tabulate( row.membership )
rm <- t( apply( rr.scores[,], 1, order, decreasing=T )[ 1:n.clust.per.row, ,drop=F ] ) ##[ iter ]
rm <- t( apply( rm, 1, sort ) ); if ( n.clust.per.row == 1 ) rm <- t( rm ) ##[ iter ]
##cra <- cluster.rows.allowed
for ( i in 1:nrow( rm ) ) {
if ( all( rm[ i, ] %in% row.membership[ i, ] ) ) next
mc <- max.changes[ "rows" ] ## If 0<max.changes<1, it's a "prob of seeing a change" and choose whether to change randomly
if ( mc < 1 && mc > 0 && runif( 1 ) > mc ) next ##!else mc <- 1
for ( ii in 1:mc ) {
if ( sum( ! rm[ i, ] %in% row.membership[ i, ] ) >= mc ) {
if ( any( row.membership[ i, ] == 0 ) ) {
col.change <- which( row.membership[ i, ] == 0 )[ 1 ] ## If any are zero (no cluster assigned) use that index
} else {
ttmp <- tabulate( row.membership[ i, ] ) ## Any clusters have this gene assigned more than once?
if ( any( ttmp > 1 ) ) { ## If so, use that index
col.change <- which( row.membership[ i, ] %in% which( ttmp > 1 ) )[ 1 ]
} else { ## Only make change if it improves the rr.score value
delta <- rr.scores[ i, rm[ i, ] ] - rr.scores[ i, row.membership[ i, ] ]
if ( any( row.membership[ i, ] %in% rm[ i, ] ) ) delta[ row.membership[ i, ] %in% rm[ i, ] ] <- 0
if ( all( is.na( delta ) | delta <= 0 ) ) next
col.change <- which.max( delta )
}
}
if ( exists( "maintain.seed" ) && ! is.null( maintain.seed ) && ! is.null( maintain.seed$rows ) && ## Don't remove row i from cluster
! is.null( maintain.seed$rows[[ as.character( row.membership[ i, col.change ] ) ]] ) &&
rownames( row.membership )[ i ] %in%
maintain.seed$rows[[ as.character( row.membership[ i, col.change ] ) ]] ) next
if ( ! rm[ i, col.change ] %in% row.membership[ i, ] ) row.membership[ i, col.change ] <- rm[ i, col.change ]
}
}
}
##}
if ( ! is.null( cc.scores ) ) {
##n.cols <- tabulate( col.membership )
cm <- t( apply( cc.scores[,], 1, order, decreasing=T )[ 1:n.clust.per.col, ,drop=F ] ) ##[ iter ]
for ( i in 1:nrow( cm ) ) {
mc <- max.changes[ "cols" ] ## If 0<max.changes<1, it's a "prob of seeing a change" and choose whether to change randomly
if ( mc < 1 && mc > 0 && runif( 1 ) > mc ) next ##!else mc <- 1
for ( ii in 1:mc ) {
if ( sum( ! cm[ i, ] %in% col.membership[ i, ] ) >= mc ) {
if ( any( col.membership[ i, ] == 0 ) ) {
col.change <- which( col.membership[ i, ] == 0 )[ 1 ]
} else {
ttmp <- tabulate( col.membership[ i, ] )
if ( any( ttmp > 1 ) ) {
col.change <- which( col.membership[ i, ] %in% which( ttmp > 1 ) )[ 1 ]
} else { ## Only make change if it improves the cc.score value
delta <- cc.scores[ i, cm[ i, ] ] - cc.scores[ i, col.membership[ i, ] ]
if ( all( is.na( delta ) | delta <= 0 ) ) next
col.change <- which.max( delta )
}
}
if ( exists( "maintain.seed" ) && ! is.null( maintain.seed ) && ! is.null( maintain.seed$cols ) && ## Don't remove col i from cluster
! is.null( maintain.seed$cols[[ as.character( col.membership[ i, col.change ] ) ]] ) &&