1. Spliced Transcripts Alignment & Reconstruction
STAR
Alexander Dobin, Philippe Batut, Sudipto Chakrabortty, Carrie Davis, Delphine Fagegaltier, Sonali Jha, Wei Lin,
Felix Schlesinger, Chenghai Xue, Christopher Zaleski,
Thomas Gingeras
CSHL

2. STAR: spliced transcript alignment and reconstruction
'Ab initio' detection of splice junctions
un-annotated, non-canonical, distal exons, chimeric ...
Any read length, any number of SJs per read
Any (reasonable) number of mismatches and indels
Unique and all multiple mappers
Alignment scoring utilizing reads quality scores
"Auto" trimming of poor quality ends
Non-templated poly-A tails detection
Very Fast: human 75-mer reads: 60 Million read per hour
Memory: RAM~9*(Genome length) bytes: 25GB for human
II. Algorithm

3. Maximum mappable length
Typical short read aligner: does the read map entirely, i.e. at full length?
What is the maximum mappable length?
can detect many mismatches
can precisely "trim" poor quality tails
can detect splice junctions
With suffix arrays we find maximum mappable length in no extra time
Map
Extend
Map
Map
Map again
II. Algorithm

4. Scoring with quality scores
Similar to local alignment scoring,
but penalties have probabilistic meaning
Illumina quality score:
+QS for matches; -QS for mismatches
Penalty for gap opening:
Total score
A more elaborate iterative penalty system is being developed
gap penalty is calculated from mapped gap length distribution
mismatch penalties vs QS scores are re-calibrated after mapping
Choose the alignment(s) with highest score
II. Algorithm

5. STAR alignment algorithm
Split each read into "good" pieces by quality scores
Map good pieces using suffix arrays
Stitch and extend mapped pieces
Score and select the best alignment

6. Splitting the reads
Split the read at poor quality bases (QS<15), 'N'
Map each good piece separately
Recover mismatches caused by poor SNR
Avoid erroneous mapping caused by sequencing errors:
just 1 SNP can cause mis-mapping from paralog to paralog

7. Suffix array based search
For each good piece
find maximum exactly mappable length (could be a multiple mapper)
if a long portion of the good piece is still unmapped - repeat
repeat this procedure backwards (from 3' to 5' of a good piece)

8. Stitch and extend mapped pieces
Each uniquely mapped piece originates an alignment window (cluster)
Collect all mapped pieces within an alignment window (e.g. 200kb)
Consider all collinear combinations of mapped pieces
Choose the combination with the highest score for each cluster
Choose the alignment cluster with the highest score
Stitch
Extend
Extend

9. Comparison with exhaustive search
Fly embryo 76mer RNA seq
1 Illumina lane: 8,930,945 total reads, good quality
Exhaustively mapped
Only in STAR
Missed by STAR
Exact
5,125,614
2,425
1MM
1,353,709 94
3,217
2MM
417,225 23
4,172
Multiple mappers by exhaustive search,
<0.002% of all reads
STAR maps 99.8% of all exhaustively mapped reads
poor quality reads which did not have a single unique "anchor"
III. Application

10. with exhaustive search
Reads mapped by STAR
1.5% multi-mappers
8.5% STAR splice junctions
1.8% not mapped by STAR
0.2% STAR InDels
gap < 20b
11% STAR
>2MM or shorter length
77% STAR overlap
with exhaustive search
III. Application

11. STAR alignments
~1,000,000 alignments found by STAR and not by exhaustive search
Distribution of mapped lengths
mean length = 72
Distribution of mismatches
spliced portions
poor quality
tails
III. Application

12. Benchmarks BLAT Bowtie STAR Fly 13 19 91 Human 1 58
Single thread benchmarks
75-mer reads
Bowtie (-v2 -k1) only reports non-spliced alignments
with 0-2 MM, 1 or 2 alignments per read
BLAT and STAR report >2MM and spliced alignments,
and all the multiple alignments
Million of reads aligned per hour
BLAT
Bowtie
STAR
Fly 13 19 91
Human 1 58
III. Application

13. % mapped: unique+multiple
Human K562/GM: 2x75
Lane
All reads
% mapped: unique
% mapped: unique+multiple
GM 1/1
16,730,063 75 83
GM 1/2
16,721,853
GM 1/3
54,477,453 35 38
GM 2/1
23,817,621 42 45
GM 2/2
25,536,631 39
K562 1/1
12,200,529 79 86
K562 1/2
12,845,645
K562 1/3
47,382,765 47 50
K562 2/1
25,597,881
K562 2/2
25,996,379 36

14. Splice junctions Total # of Gencode junctions: 284k Canonical
Annotated
Number of junctions
Canonical
Un-Annotated
Non-Canonical
Un-Annotated
Minimum number of reads per junction

15. Transcript assembly algorithm
Use contigs and splice junctions only
Find all possible collinear maximally extended transcripts
by following all possible paths

16. Examples of transcripts
STAR
transcripts

17. Examples of transcripts
STAR
transcripts

18. Summary STAR: ab initio splice junction detection
Maximum mappable length search with suffix arrays
Alignment scoring uses quality scores of the reads
Very fast: 60M/hour for 75-mer reads in human, requires large amount of RAM (~25GB for human)
The code will be beta-released in November '09

19. Examples of transcripts
STAR
transcripts

20. Another Mapped Cluster
Chimeric stitching
READ
Best Mapped Cluster
Another Mapped Cluster
chr1
chr2
If the Best Mapped Cluster leaves enough un-mapped read space,
try to stitch other clusters that cover the unmapped space
II. Algorithm