GPU Acceleration of DNA Alignment of Long Reads for DNA Assembly

110 Pages·2017·5.55 MB·English

GPU Acceleration of
DNA Alignment of Long Reads
for DNA Assembly
Tong Dong Qiu CE-MS-2018-32
y
g
o
ol
n Abstract
h
c
e
T
f
o Third generation sequencing machines produce reads with tens of thousands of base
y
t pairs. To perform de novo assembly, all reads must be compared with every other read
si
er to find overlaps. Finding overlaps with the optimal Smith-Waterman is not feasible,
v
ni since the complexity of Smith-Waterman is quadratic with the length of the reads.
U
Heuristics are designed be faster, but are not guaranteed to give the optimal solution.
t
f
el Two heuristic DNA aligners are Daligner and Darwin. Daligner uses an edit graph based
D
algorithm that has an O(ND) complexity, where N is the read length, and D the number
of differences between the two aligned reads. Darwin creates overlapping tiles to search
promising areas of the Smith-Waterman matrix, and is empirically shown to be optimal.
This work implements these algorithms on a GPU, and compares the two with respect
to sensitivity and specificity. Daligner is not suitable for GPU acceleration, but Darwin
hasshownspeedupof109xvs8CPUthreads, usingaTeslaK40. Thespeedupincreases
to 148x when the Smith-Waterman scores are not calculated. Despite large speedups
for Darwin, Daligner is 2-6x faster than Darwin, and slightly more sensitive and specific.
An advantage of Darwin is that is produces generally longer overlaps, calculates the
Smith-Waterman score, and is able to report the aligned sequences, where Daligner only
reports the start and end of the overlap.

GPU Acceleration of DNA Alignment of Long
Reads for DNA Assembly
THESIS
submitted in partial fulfillment of the
requirements for the degree of
MASTER OF SCIENCE
in
COMPUTER ENGINEERING
by
Tong Dong Qiu
born in Ede (Gld), The Netherlands
Quantum & Computer Engineering
Department of Electrical Engineering
Faculty of Electrical Engineering, Mathematics and Computer Science
Delft University of Technology

GPU Acceleration of DNA Alignment of Long
Reads for DNA Assembly
by Tong Dong Qiu
Abstract
Third generation sequencing machines produce reads with tens of thousands of base pairs.
Toperformdenovoassembly,allreadsmustbecomparedwitheveryotherreadtofindoverlaps.
FindingoverlapswiththeoptimalSmith-Watermanisnotfeasible,sincethecomplexityofSmith-
Watermanisquadraticwiththelengthofthereads. Heuristicsaredesignedbefaster,butarenot
guaranteed to give the optimal solution. Two heuristic DNA aligners are Daligner and Darwin.
Daligner uses an edit graph based algorithm that has an O(ND) complexity, where N is the
read length, and D the number of differences between the two aligned reads. Darwin creates
overlapping tiles to search promising areas of the Smith-Waterman matrix, and is empirically
shown to be optimal. This work implements these algorithms on a GPU, and compares the two
with respect to sensitivity and specificity. Daligner is not suitable for GPU acceleration, but
Darwinhasshownspeedupof109xvs8CPUthreads, usingaTeslaK40. Thespeedupincreases
to148xwhentheSmith-Watermanscoresarenotcalculated. DespitelargespeedupsforDarwin,
Daligner is 2-6x faster than Darwin, and slightly more sensitive and specific. An advantage of
Darwin is that is produces generally longer overlaps, calculates the Smith-Waterman score, and
is able to report the aligned sequences, where Daligner only reports the start and end of the
overlap.
Laboratory : Quantum & Computer Engineering
Codenumber : CE-MS-2018-32
Committee Members :
Advisor: Nauman Ahmed, QCE, TU Delft
Chairperson: Zaid Al-Ars, QCE, TU Delft
Member: Arjan van Genderen, QCE, TU Delft
Member: Matthias M¨oller, AM, TU Delft
i
ii
Dedicated to my family and friends
iii
iv
Contents
Acknowledgements vii
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Thesis goal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Background 5
2.1 Biology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 DNA sequencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Sanger sequencing . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.2 Next Generation Sequencing . . . . . . . . . . . . . . . . . . . . . 10
2.2.3 Third Generation Sequencing . . . . . . . . . . . . . . . . . . . . . 15
2.3 DNA alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.1 Dynamic programming . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.2 Global alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3.3 Local alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.4 Semi-global methods . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3.5 Heuristic algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4 DNA assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4.1 De novo assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.2 Reference based assembly . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 GPU processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5.1 CUDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.5.2 Memory hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.5.3 Streams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
2.5.4 Bioinformatics on GPU . . . . . . . . . . . . . . . . . . . . . . . . 37
3 Concept 39
3.1 Pacific Biosciences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2 Daligner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.2.1 Seeding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2.2 Local alignment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3 Darwin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3.1 D-SOFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3.2 GACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
v
4 Specification 49
4.1 Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1.1 Daligner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1.2 Darwin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2.1 Daligner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2.2 Darwin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3 Acceleration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.1 Daligner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.2 Darwin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5 Results 61
5.1 Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.2 Daligner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.2.1 Runtimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.2.2 Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.3 Darwin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3.1 Runtimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.3.2 Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.4 Sensitivity and specificity . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6 Conclusion 77
6.1 Research questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6.2 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Bibliography 98
vi