1. Aligning Transcribed Sequences to the Human and Mouse Genomes
Yongchang Gan, Jonathan Crabtree, Chris Stoeckert
Computational Biology and Informatics Laboratory (CBIL)
Center for Bioinformatics
University of Pennsylvania

2. The Genomes: Human Recent events Current public draft sequence
June 2000: “working drafts” announced
Feb. 2001: first analyses published
Feb. 2002: UCSC exits assembly business
Current public draft sequence
July, 2002: NCBI Build #30
June 28, 2002 freeze of GenBank data
87% finished seq., est % coverage

3. The Genomes: Mouse Recent events (public sequence)
Late 2000: shotgun sequencing begun
Late 2001: first assemblies created
April 2002: Arachne chosen over Phusion
Current public draft sequence
April, 2002: MGSCv3
February, 2002 freeze of ~7X shotgun
Estimated 90-95% coverage

4. The Transcribed Sequences
dbEST expressed sequence tags (ESTs)
~4 million human
~2.5 million mouse
Highly variable quality
GenBank mRNAs and RefSeqs
Many are “full length”, high quality
Includes RIKEN cDNAs
Did not include GenBank HTC division

5. DoTS: Database of Transcribed Sequences
Cluster ESTs & mRNAs by similarity
Assemble the clusters with CAP4
Annotate resulting consensus seqs.
Predict protein sequences
Run BLAST searches
Predict GO function
Link to RH maps, gene trap cell lines, expression data, MGI, GeneCards, etc.
Results at

6. A Sample DoTS Assembly

7. Is EST assembly still relevant?
Not every organism has genome project
EST sequencing is still a relatively cheap way to survey a transcriptome
Though array-based approaches are also very powerful if the sequence is known
Not every EST will necessarily align to the draft genome
Annotation component is useful, regardless of assembly method

8. Aligning transcripts with DNA
5’ UTR CDS ’ UTR
Transcribed
sequences
(e.g., mRNA)
Genome
(i.e., DNA)

9. Aligning transcripts with DNA
5’ UTR CDS ’ UTR
Transcribed
sequences
(e.g., mRNA)
Genome
(i.e., DNA)
exon exon exon 3
*** DRAMATIZATION ***

10. What are the goals? Find genes & delineate their boundaries
Investigate alternative splicing
Validate DoTS assemblies
Gain insight into sources of error
Assess whether we gain anything by assembling ESTs before aligning them

11. Potential “unsplicing” tools
BLAST
Good general-purpose local alignment tool
But not well-suited to this specific task
Special-purpose alignment tools
e.g., est2genome (Birney, Durbin), est_genome (Mott), sim4 (Florea et al.)
Do a good job but are very slow

12. Unsplicing: our first attempt
BLAST-sim4 heuristic algorithm
Employs a two-step approach
BLASTN - find candidate locations
sim4 – perform precise alignments
Much faster than sim4 alone
But still slow for whole-genome analysis
Similar to Spidey (Wheelan et al.)

13. Unsplicing: BLAT BLAT: BLAST-Like Alignment Tool
Written by Jim UCSC
Indexes target db, not query sequence
Takes advantage of additional constraints
Adjusts exon boundaries using splice signals
Attempts to locate small exons
500x speedup with no loss of sensitivity

14. Overview of alignment process
BLAT searches (vs. human and mouse)
RefSeqs and DoTS consensus sequences
Load alignments into database
Compute summary information
Including alignment “quality”
Merge selected alignments into “genes”

15. Introducing GUS plugin LoadBLATAlignments
Process raw BLAT output
Perl modules BLAT::Alignment, BLAT::PSL
Load alignments into GUS
BLATAlignment table (not Similarity)
10% minimum length cutoff applied
Compute and store summary info.
Alignment quality (requires target seq.)
Poly(A) detection (requires query seq.)
max_query_gap, unaligned_3p_bases, etc.

16. BLATAlignmentQuality
Very good (formerly “consistent”)
>= 95% identity (average)
max_query_gap <= 5
both ends consistent
no more than 10bp mismatch unless polyA
not polyA on both ends

17. BLATAlignmentQuality II
Very good with gaps
same as very good but internal and end mismatches allowed if there is a sufficiently large genomic sequence gap (within 10X mismatch length for ends)
Good
same as very good, but with max_query_gap <= 15 (allow large internal gaps if there is a sufficiently large genomic sequence gap), and inconsistent ends allowed if unaligned_bases <= 50
Not so good
everything else

18. Why “very good with gaps” and how we arrived at it?
Align Refseqs to hChr22 and mChr5
Compare consistent (very good) alignments to annotations at UCSC
False positives: close to 0
False negatives: ~18% and ~35%
With new quality filter, false negatives reduced to ~15% and ~13%

19. Why “good” and how we arrived at it?
If a Refseq has very good (consistent) alignment, would the Refseq-containing assembly too?
hChr22: 98/255 (38%) did not
mChr5: 109/271 (40%) did not
Mostly due to minor problems at end(s)
With new filter, false negatives reduced to 25/255 (~10%) and 33/271 (~12%)

20. Some alignment statistics
hDoTS (08/02) vs hGenome (GP 06/02)
Total DoTS sequences: 859,545
Alignments loaded: 5,544,300 / 8,975,529
Quality
#Align. (non-singlton)
#Seq. (non-singleton)
#Align. / #Seq. 1
343,819 (119,301)
303,555 (107,818)
1.13 (1.11) 2
3,554 ( 1,108)
3,305 ( 1,069)
1.08 (1.04) 3
262,885 ( 78,228)
195,248 ( 60,596)
1.35 (1.29)
1,2,3
610,258 (198,637)
494,023 (166,960)
1.24 (1.19) 4
4,934,042 (809,600)
292,320 (79,155)
16.9 (10.23)

21. Some alignment statistics II
mDoTS (07/02) vs mGenome (GP 02/02)
Total DoTS sequences: 579,906
Alignments loaded: 3,208,572/4,663,903
Quality
#Align. (non-singlton)
#Seq. (non-singlton)
#Align. / #Seq. 1
163,270 (57,993)
155,444 (56,035)
1.05 (1.03) 2
64,062 (23,556)
25,470 ( 8,524)
2.52 (2.76) 3
140,542 (40,063)
101,565 (32,932)
1.38 (1.22)
1,2,3
367,874 (121,612)
271,476 (93,695)
1.36 (1.30) 4
2,840,698 (883,619)
300,595 (52,493)
9.45 (16.8)

23. “Gene” creation algorithm
Select BLAT alignments
Parameters: min quality, genomic region
Merge overlapping alignments
Merge nearby alignments with at least one EST sequence in each assembly from common clone
Parameter: max distance (default 20kb)
Merge nearby alignments
Parameter: max distance (default off)

24. “Gene” creation algorithm II
~/dots2gene]$ ./dotsAlignment2Gene.pl -h
Usage: dotsAlignment2Gene.pl --sp sp --chr chr --start start --end end
--qf qf --xs xs --am am --cm cm --lm lm --mis mis
--of out [--test] [--debug debug] [--help], where...
sp: scientific/common name of species, e.g. human, Mus musculus
chr: chromosome of interest, e.g. 5, 22, X, 3_random
start: start genomic position, default to 1
end: end genomic position, default to chromosome length
qf: quality filter to select blat alignments for gene creation
xs: exclude blat alignments of singletons
am: merge by genomic alignment overlap
cm: merge by shared clone info between gene seeds within specified distance
lm: merge by alignment proximity (within specified distance)
mis: min intron size for a gene to be kept
out: output format, one of s[ummary], v[erbose] or gff
test: test case using DeiGeorge region
debug: specify level of debug output, can be 1, 2, 3, and 4+

25. Initial Algorithm Calibration
Human chr22q (~34Mb) as test case
Sanger annotation release 2.3: 832 genes (341 gene, 118 gene_segment, 112 related, 109 predicted, 152 pseudogenes)
Focus on DiGeorge Region
DGCR6 to ZNF74 (~ 1.6Mb)
Contains genes based on literature (Sanger: 44 genes with 33 known)
* Used DoTS 02/02 release vs Golden Path 12/01 release, and old BlatAlignment table (limited quality classes).

26. Choosing initial parameters
# DiGeorge Chromosome Region (DGCR6 - ZNF74, 1.6Mb) #
CBIL Gene Param Num CBIL Num Sanger Num Overlap Avg %overlap
qf=4, am, cm=10k 27/ / vs 71.3
qf=4, am, cm=20k 24/ / vs 75.5
qf=4, am, cm=50k 20/ / vs 77.6
qf=6, am, cm=10k 26/ / vs 75.9
qf=6, am, cm=20k 25/ / vs 80.5
qf=6, am, cm=50k 17/ / vs 87.4
# Chr22 (Chr22q ~34M)
qf=4, am, cm=20k / / vs 72.4
qf=6, am, cm=20k / / vs 81.0
* Sanger annotation in different coordinate system, did approximate translation

27. Initial parameters Derived from old alignments
qf=4: “spliced” and “consistent” alignments
xs=off: not exclude alignments by singleton assemblies
am=on: merge by alignment overlap
cm=20K: merge by shared clone
lm=off: merge by genomic location proximity
mis=15: filter putative genes by max “intron” size
Adjusted for new alignment quality categorization
qf=7: “spliced” alignment of quality id 1, 2 or 3

28. Preliminary results
Applied algorithm to new alignments of hChr22 and mChr5 [4-6 hours each]
Displayed as custom tracks at UCSC genome browser
DiGeorge region CBIL and Sanger genes
Human chr22 CBIL and Sanger genes
Mouse chr5 CBIL genes

32. Preliminary full genome runs
Excluded singletons
Tried lm parameter at off, 30, 50
~24 hours per run, with some stress on database server
Per chromosome statistics for mouse
See next slide

33. Lm off, qf7/am/xs/cm20k/mis15
Lm qf7/am/xs/cm20k/mis15 (* publicly visible DoTS only) Lm 50
mChr
#assm.
#gene
ratio
Ratio 1
2599
1705
1.52
2246
1449
1.55
1448 2
3525
2097
1.68
3024
1777
1.70
3025
1774 3
2162
1367
1.58
1860
1159
1.60 4
2665
1621
1.64
2279
1390
1389 5
2843
1747
1.63
2429
1487
2430
1485 6
2245
1409
1.59
1954
1200 7
3325
1997
1.66
2864
1696
1.69
1695 8
2216
1364
1.62
1912
1150
1148
1.67 9
2478
1503
1.65
2118
1287
1286 10
2063
1274
1784
1095 11
3723
2188
3160
1821
1.74 12
1541
958
1.61
1358
822
1359 13
1501
1008
1.49
1315
878
1.50
1316
877 14
1548
982
1355
836 15
1907
1129
1652
893
1.85
891 16
1489
927
1322
758
757
1.75 17
2194
1356
1910
1054
1.81
1053 18
1236
731
1044
595
593
1.76 19
1544
921
769
1.71 X
1320
867
1108
753
1.47 Un
1874
2306
0.81
1713
1665
1.03
1659
Total
<44124
27151
<38009
22869
<38013
22851
128615
90829
1.42

34. Directions
Assessment of results in selected 14Mb on mouse chr5 (Maja Bucan lab)

35. Directions II Quantitative evaluation of results
Correlation coefficients – next slide
(Science Kapranov et al. 296 (5569): 916.)

37. Directions III Fine tune/revise algorithm parameters
qf: recruit more alignments
cm: widen from 20K to 500K?
lm: turn on (lm<75bp or 0<lm<75bp)?
mis: intron size distribution model?
Evidence suggesting new parameters
See next slide
Conservatively assume UCSC ref genes uniformly samples 1/3 of all genes on chr22

38. hChr21 RefGenes hChr22 RefGenes hChr22 gene bounds Genomic Sizes Max+ -
Genomic
Sizes # 88 95
195
182
370
380
Min
189
408
675
309
190
343
Med
34619
20084
16903
17225
10744
9875
Avg
53134
59300
31290
35909
31999
27700
Max
259519
834228
288888
647340
715044
647564
<0
Distances
0-10
>10 11 22 19 14
-47831
-17229
-19026
-20601
-46312
-33082
-60639
-26611
-29580
-57429
-76554
-7955
-5807
-1752
-476
-3179
-8554 1 6 76 83
172
162
354
366
1138
223
216
108
109
104203
144877
62721
67898
27971
34914
372763
379780
161200
174314
70366
68781
861971
889255

39. Directions IV Problem fixes Error rate estimations
Handle assemblies on the wrong strand - see next slide
Error rate estimations
Simulate effects of sequence/assembly errors on BLAT

41. Ongoing work Combine alignments with other sequence signals (Artemis)
Detailed examination of regions of interest on mouse chr. 5 (Maja)
Incorporate alignments into DoTS assembly process

42. Acknowledgements Alignments Database of Transcribed Sequences
Yongchang Gan (see poster!)
Database of Transcribed Sequences
Brian Brunk
Steve Fischer
Deborah Pinney
Manual annotation
Joan Mazzarelli
Kolchanov group