Docstoc

SpliceGrapher detecting patterns of alternative splicing from RNA

Document Sample
SpliceGrapher detecting patterns of alternative splicing from RNA Powered By Docstoc
					   Supplementary information for:



SpliceGrapher: detecting patterns of alternative splicing
from RNA-seq data in the context of gene models and EST
data
Mark F. Rogers1 , Julie Thomas3 , Anireddy S.N. Reddy3 , Asa Ben-Hur∗1,2

1 Department of Computer Science, Colorado State University, Fort Collins, Colorado, USA
2 Department of Statistics, Colorado State University, Fort Collins, Colorado, USA
3 Department of Biology and Program in Cell and Molecular Plant Biology, Colorado State University, Fort Collins, CO, USA



Email: Mark F. Rogers - rogersma@cs.colostate.edu; Julie Thomas - thomasju@lamar.colostate.edu; Anireddy S.N. Reddy -
reddy@colostate.edu; Asa Ben-Hur∗ - asa@cs.colostate.edu;

∗ Corresponding   author




   • SR gene case-study
   • Supplementary Tables 1-5

   • Supplementary Figures 1-17




                                                              1
1   Validation with SR Genes
To further validate our SpliceGrapher method, we compared its predictions on 15 A. thaliana serine-arginine
(SR) genes with results from a separate study [32]. Briefly, the authors obtained full-length genomic se-
quences for these genes from TAIR, and used MEGABLAST to find matching ESTs in NCBI’s dbEST
database and full-length cDNAs (fl-cDNA) in the nr nucleotide database. These sequences were fed along
with genomic sequences into an in-house alignment pipeline to assess the extent of AS among the SR genes.
We then used SpliceGrapher to produce predictions for these genes, using the two Arabidopsis short-read
datasets described in the Methods section of the main text. We compared SpliceGrapher’s predictions with
the graphs produced from the EST and fl-cDNA alignments. Supplementary Figures 14-17 show the predic-
tions made by SpliceGrapher and Cufflinks for SR33 (AT1G55310) using the 32nt and 76nt reads. The full
set of output graphs are available at http://combi.cs.colostate.edu/SpliceGrapher/results/at sr genes.php.
Despite having less reads, SpliceGrapher is able to make more predictions using the 76nt reads, and those
predictions are in agreement with the predictions made from ESTs and fl-cDNAs. Although it is clear from
the read coverage of the 76nt reads that there is an IR event, SpliceGrapher was not able to unambiguously
determine which intron was retained, so no prediction was made (Figure 15).


   We compare SpliceGrapher’s predictions with those of Cufflinks on these same genes using the same data
and supplying Cufflinks with the gene models in each case. Using the 32nt data with strong coverage across
the gene (Supplementary Figure 16), Cufflinks predicts some of the same novel AS events as in the ESTs,
but it also predicts an exon-skipping event that is not evident in the gene models, the RNA-Seq data or in
the ESTs. With the 76nt data (Supplementary Figure 17), Cufflinks predicts only the gene models despite
having the same evidence for a novel exon and IR events as SpliceGrapher.


    We note that there is a splice junction in the EST/fl-cDNA graph that is marked as a false-positive site.
The acceptor dimer for this exon is a non-canonical AT site and thus could not be predicted with our AG
splice site classifier. Neither SpliceGrapher nor Cufflinks found evidence for this splice site. Cufflinks uses
a heuristic that looks for GT-AG and GC-AG sites, while SpliceGrapher uses dimer-specific classifiers to
predict splice sites. At this time the gene models and EST data we have for A. thaliana provide only 17
examples of AT sites, too few to train an accurate AT classifier.


2   Tables
     Source       AS Genes     Intron Retention    Exon Skipping         Alt. 5’           Alt. 3’      Total
 TAIR9               4,029     1,987     (33%)     550      (9%)     1,256 (21%)       2,145 (36%)      5,938
 SpliceGrapher       5,173     2,503     (30%)     701      (8%)     1,769 (21%)       3,312 (40%)      8,285
     novel           1,154       560     (23%)     151      (6%)       513 (21%)       1,167 (49%)      2,391

Supplementary Table 1: SpliceGrapher predictions from the 76nt RNA-Seq data generated in house. Despite
generating fewer reads, improvements in NGS sequencing techniques yield longer, high-quality reads that
make it easier to predict novel AS. As a result, SpliceGrapher was able to make confident predictions for
more novel AS events of all types than with the older, 32nt data. The fraction of each type of event is similar
in both cases. (Alt. 3’=alternative 3’ site; Alt. 5’=alternative 5’ site)




                                                      2
               Relationship Between Read Length and False-Positive Rate
  Read        Total                              Total       Novel
 Length    Alignments Ungapped       Spliced   Junctions Junctions         FP (%)
  32 nt    31,474,618 30,718,012     756,606     46,485      1,803        428 (23.7%)
  76 nt    30,828,836 27,271,200 3,557,636       87,891     11,544      4,490 (38.9%)

Supplementary Table 2: The effect of read length on false positive splice junctions. We ran TopHat on
41,428,058 RNA-Seq reads in A. thaliana. Two TopHat runs were performed: on the original 76nt long
reads, and on reads generated by truncating the reads to 32nt long. For the 76nt reads, the number of spliced
alignments increased dramatically. Many of these were associated with novel splice junctions; however the
false-positive rate was substantially higher than that of the 32nt reads. (RNA-Seq=RNA sequences produced
by next-generation sequencing)


                              AS Associated with False-Positive Sites
                                Exon Skipping       Alt. 5’           Alt. 3’              Total
 Cufflinks (no gene models)
   A. thaliana                     267 (71.4%)       233 (62.8%)      215 (46.7%)        715 (59.3%)
   V. vinifera                     105 (20.2%)        66 (23.8%)      146 (21.1%)        317 (21.3%)
 Cufflinks (with gene models)
   A. thaliana                     327 (12.8%)       291 (50.5%)      278 (48.7%)        896 (24.2%)
   V. vinifera                   1,445 (43.3%)     2,036 (52.3%)    2,447 (48.1%)      5,928 (48.1%)
 TAU (no gene models)
   A. thaliana                     486 (77.0%)       978 (52.1%)    1,080 (34.3%)      2,544 (44.9%)
   V. vinifera                     169 (22.4%)       604 (24.7%)    1,400 (27.3%)      2,173 (26.1%)
 TAU (with gene models)
   A. thaliana                     915 (50.0%)     1,551 (26.2%)    1,543 (16.4%)      4,009 (23.4%)
   V. vinifera                   5,709 (67.1%)     7,060 (23.5%)    8,188 (25.9%)     20,828 (29.9%)

Supplementary Table 3: SpliceGrapher’s classifiers identified many Cufflinks and TAU splice sites as false-
positives. Above is shown the number and proportion of each novel AS event type for which predictions
were influenced by false-positive sites. IR events do not contribute to these statistics because the packages
do not depend on splice sites to resolve them.
(Alt. 3’=alternative 3’ site; Alt. 5’=alternative 5’ site)


                            TopHat Alignments for H. sapiens Data
                Total        Aligned    Ungapped       Spliced    Novel             False-Positive
                Reads         Reads    Alignments Alignments Junctions                Junctions
 Caucasian    35,687,594    23,283,623  21,085,632     2,197,991    7,705                   4,549
 Yoruban      38,302,542    25,707,194  23,424,639     2,357,294    8,576                   4,960

Supplementary Table 4: Summary of TopHat alignments for the paired-end H. sapiens data. SpliceGrapher’s
classifiers were used to construct a database of predicted splice sites that were then used to filter out false-
positive junctions from TopHat alignments. Above is shown the total number of reads used in the alignments,
the number of those that aligned, how they broke down into ungapped and spliced alignments, the number
of novel splice junctions inferred by the alignments, and the number of false-positive junctions.




                                                      3
                Alternative Splicing Statistics for 20,640 H. sapiens   Graphs
 Source         Intron Retention Exon Skipping            Alt. 5’          Alt. 3’     Total
 Gene Models    11,853     (14%) 42,124 (50%) 15,342 (18%)              15,401 (18%)   84,720
 Caucasian      11,944     (14%) 42,699 (50%) 15,515 (18%)              15,624 (18%)   85,782
   Novel           124     (11%)       575 (52%)          174 (16%)        226 (21%)    1,099
 Yoruban        11,960     (14%) 42,653 (50%) 15,561 (18%)              15,657 (18%)   85,831
   Novel           142     (12%)       529 (46%)          221 (19%)        262 (23%)    1,154

Supplementary Table 5: Summary of AS events detected in the two H. sapiens data sets. Novel events were
predicted at nearly the same rate as they appear in the gene models, with the exception of intron retention,
which is difficult to predict from RNA-Seq data.
(Alt. 3’=alternative 3’ site; Alt. 5’=alternative 5’ site)




                                                     4
             3   Figures



                                            Gene Model for AT5G11030



                           Cufflinks AT5G11030 Prediction (no gene models)




                         Cufflinks AT5G11030 Prediction (with gene models)




                                                TopHat Spliced Reads

                    4                                              3       8                  2



             60                                Bowtie Read Coverage
Read Depth




             50
             40
             30
             20
             10
              0
              3495000               3494000                3493000                3492000               3491000

                                                     Alt. 3'       Skipped Exon

             Supplementary Figure 1: Cufflinks predictions for the AT5G11030 gene from A. thaliana are both fragmented
             due to the low read coverage. (Alt. 3’=alternative 3’ site)


                                                               5
                                            Gene Model for AT5G11030



                              TAU AT5G11030 Prediction (no gene models)



                             TAU AT5G11030 Prediction (with gene models)



                                              Supersplat Spliced Reads


             60                             HashMatch Read Coverage
Read Depth




             50
             40
             30
             20
             10
              0
              3495000               3494000                3493000                 3492000                3491000
                                                           Alt. 3'        SE
                                                           IR

             Supplementary Figure 2: TAU predicts several fragments for the AT5G11030 gene from A. thaliana when it
             has no gene models to work with, but predicts a large number of possibly spurious intron retention events
             when gene models are provided. (Alt. 3’=alternative 3’ site; IR=intron retention, SE=skipped exon)




                                                                6
                                              Gene Model for AT5G11030




                                      SpliceGrapher AT5G11030 Prediction




                                                        Spliced Reads

                     7                                                         8        5          5       12



             60                                         Read Coverage
             50
Read Depth




             40
             30
             20
             10
              0
              3495000                3494000                  3493000                 3492000                   3491000

                                                       Alt. 3'         Skipped Exon

             Supplementary Figure 3: SpliceGrapher prediction for the AT5G11030 gene, where low RNA-Seq coverage
             makes it challenging to predict AS. By using the gene model to establish a context for interpreting these
             data, SpliceGrapher is able to report the original gene model while still avoiding spurious predictions. (Alt.
             3’=alternative 3’ site)



                                                                   7
                                                Gene Model for AT5G22640




                                TAU AT5G22640 Prediction (no gene models)




                                                Supersplat Spliced Reads
                                       21                                 63
                                   1        9    15      7     39   31   1 1    10   6               8   1


                                                HashMatch Read Coverage
             102
Read Depth




             101
             100
                          7530000                7531000                 7532000              7533000
                                                      Alt. 5'                            IR
                                                      False-Positive Junction

               Supplementary Figure 4: TAU prediction along with the corresponding alignments for the A. thaliana gene
               AT5G22640. In this gene TAU did not produce any transcript predictions when we provided it with gene
               models. (Alt. 5’=alternative 5’ site, IR=intron retention)




                                                                    8
                                                      Gene Model for AT3G16400



                             Cufflinks AT3G16400 Prediction (no gene models)



                            Cufflinks AT3G16400 Prediction (with gene models)



                                                      TopHat Spliced Alignments                                        7
                                                                                                                      2
                                                                                                                      2
                                              3                                                                       2
                                         20                          66                                               3
                                                  5                                      2
                                        13                2                                                       4
                                        2                 4                              3
                               18       2         6              4                                    2


                                                         Bowtie Alignments
Read Depth




             103
             102
             101
             100
                             5566500                   5567000                5567500        5568000               5568500

               Supplementary Figure 5: Cufflinks prediction for the gene AT3G16400 in A. thaliana. Cufflinks attempts to
               find a set of transcripts that are consistent with the preponderance of splice-junction evidence. The resulting
               graph likely contains spurious splice forms.




                                                                          9
                                                     Gene Model for AT3G16400




                                               SpliceGrapher AT3G16400 Prediction




                                                              Spliced Reads



                               12                                  4



                                                              Read Coverage
             102
Read Depth




             101
             100
                              5566500               5567000                 5567500            5568000               5568500



              Supplementary Figure 6: SpliceGrapher predictions for the same gene where Cufflinks predicted many false
              positive splice-junctions. By first filtering out false-positive splice junction alignments, SpliceGrapher is able
              to avoid making spurious predictions.




                                                                       10
                                                    Gene Model for AT3G23450




                                              SpliceGrapher AT3G23450 Prediction




                                                           Spliced Reads




                                                           Read Coverage
             103
Read Depth




             102
             101
             100
                   8410500                   8411000                      8411500                      8412000



              Supplementary Figure 7: SpliceGrapher prediction for a single-exon gene in A. thaliana where the data
              suggest novel alternative splicing but lack the splice junction evidence necessary for making confident pre-
              dictions.




                                                                  11
                                                                Gene Model for AT3G23450




                                                                TAU AT3G23450 Prediction




                                                                    Supersplat Spliced Reads
                                                                                        13         61                            2
                                                                                                 14                      5
                                                                                                                   10
                                                                                                   772            3
                                                                                                    8                            2
                                                                                                                             3
                                                                                92              33 8
                                                                                                  31                                   3 39
                                                                                                                   2                    4 4
                                                                                   4 33                                2
                                                                               3        210                                             23
                                                                                 43 2                 3                2         8
                                                                2 4 6                                                  27
                                                                         14                      22                                      2
                                                                    2 12                        432
                                                                                                                             10
                                                                                                                            262 9
                                                                5 5    3                  36                              2                    3
                                                                                                                                                   2
                                                                                           2
                                                                                        49                                                 2
                                                                                318                                      2 54
                                                                                                                                  3      36
                                                                         14 4 3                   3                       42              11
                                                                         44                 2             6              34             4               2
                                                23          9                                                          6
                                                                     2                212                                      10
                                                                                                                                   2               2   3
                              4                 3                               2                             2                                    3     9
                                                                                2

                                                                HashMatch Read Coverage
Read Depth




             103
             102
             101
             100
                   8410500                       8411000                                                               8411500                                8412000
                                    Alt. 3'                                Alt. 3'/IR                     Alt. 5'/IR/SE                 False-Positive Junction
                                    Alt. 3'/Alt. 5'/IR/SE                  Alt. 3'/SE                     Alt. 5'/SE                    SE
                                    Alt. 3'/Alt. 5'/SE                     Alt. 5'

              Supplementary Figure 8: TAU prediction for the AT3G23450 gene in A. thaliana. TAU attempts to find all
              transcripts that are consistent with the preponderance of splice-junction evidence. The resulting graph is
              likely to contain a substantial number of spurious splice forms. (Alt. 3’=alternative 3’ site; Alt. 5’=alter-
              native 5’ site; IR=intron retention, SE=skipped exon)



                                                                                                12
                     1.0 Arabidopsis GT Site Classifier ROC (AUC=0.97, ACC=0.91)                                                                 1.0            Grape GT Site Classifier ROC (AUC=0.98, ACC=0.93)


                     0.8                                                                                                                         0.8
True positive rate




                                                                                                                            True positive rate
                     0.6                                                                                                                         0.6
                                               ~zero: df=-0.01




                                                                                                                                                                  ~zero: df=-0.06
                     0.4                                                                                                                         0.4


                     0.2                                                                                                                         0.2


                     0.0
                       0.0                                                     0.2   0.4             0.6   0.8   1.0                             0.0
                                                                                                                                                   0.0                                       0.2   0.4             0.6   0.8   1.0
                                                                                     False positive rate                                                                                           False positive rate


                     1.0 Arabidopsis AG Site Classifier ROC (AUC=0.95, ACC=0.87)                                                                 1.0           Grape AG Site Classifier ROC (AUC=0.96, ACC=0.92)


                     0.8                                                                                                                         0.8
True positive rate




                                                                                                                            True positive rate




                     0.6                                                                                                                         0.6
                                                             ~zero: df=-0.01




                                                                                                                                                                           ~zero: df=-0.08




                     0.4                                                                                                                         0.4


                     0.2                                                                                                                         0.2


                     0.0
                       0.0                                                     0.2   0.4             0.6   0.8   1.0                             0.0
                                                                                                                                                   0.0                                       0.2   0.4             0.6   0.8   1.0
                                                                                     False positive rate                                                                                           False positive rate


                     1.0 Arabidopsis GC Site Classifier ROC (AUC=0.97, ACC=0.97)                                                                 1.0           Grape GC Site Classifier ROC (AUC=0.92, ACC=0.92)


                     0.8                                                                                                                         0.8
True positive rate




                                                                                                                            True positive rate




                     0.6                                                                                                                         0.6
                             ~zero: df=-0.02




                                                                                                                                                         ~zero: df=-0.00




                     0.4                                                                                                                         0.4


                     0.2                                                                                                                         0.2


                     0.0
                       0.0                                                     0.2   0.4             0.6   0.8   1.0                             0.0
                                                                                                                                                   0.0                                       0.2   0.4             0.6   0.8   1.0
                                                                                     False positive rate                                                                                           False positive rate


                Supplementary Figure 9: ROC curves that demonstrate the accuracy of the SVM classifiers SpliceGrapher
                generated for GT and GC donor sites and AG acceptor sites in A. thaliana and V. vinifera. (ACC=accuracy;
                AUC=area under the curve; ROC=receiver operating characteristic, SVM=support vector machine)

                                                                                                                       13
                                                             EPSTI1 Gene Model


                                                                       EPSTI1
                                                             EPSTI1 Caucasian

                                                                       EPSTI1
                                            EPSTI1 Caucasian Splice Junctions

                                         38 48 14       15                              12           38      44    19
                                             EPSTI1 Caucasian Read Coverage
    Read Depth




                 80
                 60
                 40
                 20
                  0                                           EPSTI1 Yoruban

                                                                       EPSTI1
                                              EPSTI1 Yoruban Splice Junctions
                                                                                       3                           2
                                         69 89 2        7                              19            73      124   46
                                              EPSTI1 Yoruban Read Coverage
Read Depth




             102
             101
             100
                      43560000         43540000               43520000            43500000       43480000          43460000
                                              Alt. 5'          Intron Retention       Skipped Exon

                  Supplementary Figure 10: Splice graphs for the human gene EPSTI1 using short read data for Caucasian
                  and Yoruban samples. Short-read coverage for exons and splice junctions identified a splice form in the
                  Yoruban data that did not appear in the Caucasian data for this gene. (Alt. 5’=alternative 5’ site)

                                                                      14
                                       Gene Model for AT1G02380




                                                       AT1G02380
                             SpliceGrapher Prediction for AT1G02380




                                                       AT1G02380
                                            Supported Junctions


                                  4                                             5              6


             16                                 Read Coverage
             14
             12
Read Depth




             10
              8
              6
              4
              2
              0
             477000             477500                  478000                  478500
                                                      Intron Retention

       Supplementary Figure 11: A. thaliana gene AT1G02380 provides an example where the known splice form
       is the one with the intron retained (as opposed to the example shown in Figure 1, where the novel splice
       form has the intron as retained). This scenario of intron retention requires different evidence, namely the
       novel splice junction, shown in green in the figure. The boundaries of the exon from the gene model were
       used to infer the boundaries of the exons that flank the new intron.




                                                          15
               A                                                                       B


                                                                           C
                                                         acceptor                               donor




Supplementary Figure 12: The method SpliceGrapher uses to resolve an alternate acceptor site when a short-
read exon (C) extends beyond another exon (B). If the exon C is not directly flanked by splice junctions
(known or predicted from RNA-Seq data), SpliceGrapher will attempt to extend the exon to the nearest
splice junctions, upstream and downstream. If such an extension is not supported by the data, the exon is
considered unresolved.




                     A                                    B                            C




               D                                                                       E
                                                      F



Supplementary Figure 13: Two scenarios in which SpliceGrapher will predict a skipped exon. In the top
panel, a novel splice junction between exons A and C provides evidence that exon B is skipped in some
transcripts. In the lower panel a novel exon, F, is contained within an intron in the graph. The new exon
is flanked by novel splice junctions within the same intron, so SpliceGrapher uses the acceptor site from the
upstream junction to resolve the exon’s 5’ boundary and the donor site from the downstream junction to
resolve its 3’ boundary. If SpliceGrapher is unable to resolve either boundary, the exon is unresolved.




                                                    16
                                                    Gene Model for AT1G55310




                                                   ESTs/fl-cDNAs for AT1G55310




                                      SpliceGrapher AT1G55310 Prediction (32nt reads)




                                                             Spliced Reads


                                                                                    4           13
                            31                 2                      5                         47   135



                                                            Read Coverage
             500
             400
Read Depth




             300
             200
             100
               0           20630500          20631000              20631500             20632000           20632500

                                                      Alt. 3'         False-Positive Junction
                                                      Alt. 3'/SE      IR

              Supplementary Figure 14: Comparison between the gene models, EST/fl-cDNA alignments and SpliceGra-
              pher’s prediction the SR gene AT1G55310 in A. thaliana. From the RNA-Seq data, SpliceGrapher is able
              to predict one of the novel AS events in the ESTs. Note that SpliceGrapher identified a false positive splice
              junction in one of the EST alignments. It is a non-canonical AT site that is 2nt downstream of an AG dimer.



                                                                     17
                                                      Gene Model for AT1G55310




                                                     ESTs/fl-cDNAs for AT1G55310




                                       SpliceGrapher AT1G55310 Prediction (76nt reads)




                                                                 Spliced Reads


                                                                            15                         34
                            48                 64                  8                      16           28   65




             250                                                 Read Coverage
             200
Read Depth




             150
             100
              50
               0           20630500           20631000                 20631500                  20632000        20632500

                                                    Alt. 3'            False-Positive Junction       SE
                                                    Alt. 3'/SE         IR

              Supplementary Figure 15: Comparison between the gene models, EST/fl-cDNA alignments and SpliceGra-
              pher’s prediction for the SR gene AT1G55310 in A. thaliana. Here SpliceGrapher is able to predict more of
              the novel alternative splicing events found in the ESTs despite weaker read coverage than in the previous
              example. The longer reads (76nt) make it easier to identify unique locations for spliced reads. However, it is
              still not possible to identify which of three possible intron retention events are represented in the RNA-Seq
              data. (EST=expressed sequence tag; fl-cDNA=full-length complementary DNA, RNA-Seq=RNA sequences
              produced by next-generation sequencing)
                                                                      18
                                                     Gene Model for AT1G55310




                                                    ESTs/fl-cDNAs for AT1G55310




                                         Cufflinks AT1G55310 Prediction (32nt reads)




                                                          TopHat Spliced Reads

                                                                                                   3
                                                                                                15     133
                           84                  2                                                                        4



                                                         Bowtie Read Coverage
             500
             400
Read Depth




             300
             200
             100
               0           20630500          20631000           20631500                  20632000           20632500

                                                   Alt. 3'      False-Positive Junction       SE
                                                   Alt. 3'/SE   IR

              Supplementary Figure 16: Comparison between the gene models, EST/fl-cDNA alignments and the Cufflinks
              prediction for the SR gene AT1G55310 in A. thaliana. Given 32nt reads and gene models, Cufflinks predicts
              three fragments that include some of the alternative splicing from the gene models and two events found in
              the ESTs. It also predicts an exon skipping event that is not evident either in the ESTs or in the RNA-
              Seq data. (EST=expressed sequence tag; fl-cDNA=full-length complementary DNA; SR=serine-arginine,
              RNA-Seq=RNA sequences produced by next-generation sequencing)
                                                                   19
                                                     Gene Model for AT1G55310




                                                     ESTs/fl-cDNAs for AT1G55310




                                          Cufflinks AT1G55310 Prediction (76nt reads)




                                                        TopHat Spliced Reads


                                                                            27                        57
                            56                  82                  11                    20          42   83




             250                                        Bowtie Read Coverage
             200
Read Depth




             150
             100
              50
               0           20630500           20631000                   20631500              20632000         20632500

                                                       Alt. 3'              False-Positive Junction
                                                       Alt. 3'/SE           IR

              Supplementary Figure 17: Comparison between the gene models, EST/fl-cDNA alignments and the Cufflinks
              prediction for the SR gene AT1G55310 in A. thaliana. Given 76nt reads and gene models, Cufflinks predicts
              only the gene models and none of the events found in the ESTs. This is consistent with our other results which
              suggest that Cufflinks may not generalize well when provided gene model annotations. (EST=expressed
              sequence tag; fl-cDNA=full-length complementary DNA, SR=serine-arginine)

                                                                           20

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:4
posted:6/13/2012
language:
pages:20