Leveraging Succinct Data Structures for DNA Sequence Mapping on FPGA

27th IEEE Reconfigurable Architecture Workshop, May 18th 2020

POLITECNICO DI MILANO

MANUALE DI CORPORATE IDENTITY

MILANO, MAGGIO 2015

Guido Walter Di Donato guidowalter.didonato@mail.polimi.itAlberto Zeni

Lorenzo Di TucciMarco D. Santambrogio

alberto.zeni@mail.polimi.itlorenzo.ditucci@polimi.itmarco.santambrogio@polimi.it

Dipartimento di Elettronica, Informazione e Bioingegneria (DEIB) - Politecnico di Milano

2

DNA Sample

Sequencing

Genome

Assembly

Genome

Analysis

Drug

Creation

Computational Bottleneck

DNA Analysis for Precision Medicine

Sequence Mapping

3

Sequence MappingSequenced reads

Sequence mapping is a computational intensive step involved in genome assembly and

genomic analysis pipelines, leading to long execution times and high computational costs.[1]

[1] Beretta, S.. Algorithms for strings and sequences: Pairwise alignment. Encyclopaedia of Bioinformatics and Computational Biology. Oxford: Academic Press, 2019.

Sequence Mapping

4

Sequence MappingSequenced reads

Sequence mapping is a computational intensive step involved in genome assembly and

genomic analysis pipelines, leading to long execution times and high computational costs.[1]

[1] Beretta, S.. Algorithms for strings and sequences: Pairwise alignment. Encyclopaedia of Bioinformatics and Computational Biology. Oxford: Academic Press, 2019.

Sequence mapping is the computational bottleneck of

many genome assembly and genome analysis pipelines:

Long execution time

High computational cost

5

Burrows-Wheeler Transform

Reversible string permutation that can be searched directly and has long strings

of repeated characters (great for compression).

TACGACGTCGACT$

1 TACGACGTCGACT$

2 ACGACGTCGACT$T

3 CGACGTCGACT$TA

… …

13 T$TACGACGTCGAC

14 $TACGACGTCGACT Sort each row by their prefixes in

lexicographic order 14 $TACGACGTCGACT

2 ACGACGTCGACT$T

5 ACGTCGACT$TACG

… …

1 TACGACGTCGACT$

8 TCGACT$TACGACG

TTGGATAACCCC$G

14-2-5-11-3-9-6-12-4-10-7-13-1-8

BWT output

Suffix array

Burrows, M. and Wheeler, D. J.: A block-sorting lossless data compression algorithm. 1994.

BWaveR Data Structure

6

A flexible combination of succinct data structures, namely Wavelet Tree and RRR

Sequences, able to store long sequences in an amount of space "close" to the theoretic

lower bound, while still allowing for efficient rank query operations.

S., Beller, T., Moffat, A., and Petri, M.: From theory to practice: Plug and play with succinct data structures, 2014

RRR Sequences

Parameters: block size & superblock factor

Backward Search on BWaveR Data Structure

start = C(X) + 1

end = C(X+1)

INITIALIZATION ITERATIONS

start = C(X) + rankWT(X, start-1) +1

end = C(X) + rankWT(X, end)end < pos($)

start = C(X) + rankWT(X, start-1) +1

end = C(X) + rankWT(X, end-1)start < pos($) ≤ end

start = C(X) + rankWT(X, start-2) +1

end = C(X) + rankWT(X, end-1)start ≥ pos($)C(X)

$ A C G T

0 1 4 8 11

TTGGATAACCCC$G

7

Backward Search on BWaveR Data Structure

start = C(X) + 1

end = C(X+1)

INITIALIZATION ITERATIONS

start = C(X) + rankWT(X, start-1) +1

end = C(X) + rankWT(X, end)end < pos($)

start = C(X) + rankWT(X, start-1) +1

end = C(X) + rankWT(X, end-1)start < pos($) ≤ end

start = C(X) + rankWT(X, start-2) +1

end = C(X) + rankWT(X, end-1)start ≥ pos($)C(X)

$ A C G T

0 1 4 8 11

TTGGATAACCCC$G

8

A pattern of length p is found in O(p) time,

independently of the reference size!

Custom Sequence Mapping Architecture

FPGA-tailored optimizations:

1. Reference structure stored in BRAM to reduce memory access time

2. Dimensioned query sequences structures to exploit memory burst

3. Parallel mapping of each query sequence and its reverse complement

4. Implemented rank queries using a reduction strategy9

Load reference structureHost DRAM FPGA

Load query sequence

Send mapping result

Compute reverse complement

Mapping via backward search

Initialization

Loop

10

Results - Memory Footprint

Data

stru

ct s

ize [b

ytes

]

1M

5M

10M

14M

18M

Superblock factor2 7 17 32 50 100

3 7 11 15

Data

stru

ct s

ize [b

ytes

]

10M

43M

75M

108M

140M

Superblock factor2 7 17 32 50 100

Block size

Memory reduction up to

63.7% w.r.t BWT sequence

98.1% w.r.t naive Occ matrix

E.Coli Genome Human Chr. 21

Memory reduction up to

68.3% w.r.t BWT sequence

98.4% w.r.t naive Occ matrix

11

Results - Mapping Time (HW)

Map

pin

g T

ime

[ms]

1

10

100

1000

10000

100000

1 million reads 10 millions reads 100 millions reads

BWaveR - FPGA BWaveR - CPU Bowtie2 - 1 threadBowtie2 - 8 threads Bowtie2 - 16 threads

Time for mapping 100 bp reads against Human Chr. 21 (b = 15, sf =50)

14x7,8x

1,4x0,7x

62x 42x

7,6x4x

70x 51x

9,5x4,9x

12

En

ergy

Co

nsu

mp

tio

n [J

]

1

10

100

1000

10000

1 million reads 10 millions reads 100 millions reads

BWaveR - FPGA BWaveR - CPU Bowtie2 - 1 threadBowtie2 - 8 threads Bowtie2 - 16 threads

Energy for mapping 100 bp reads against Human Chr. 21 (b = 15, sf =50)

Results - Energy Consumption (HW)

74x42x

7,7x4x

337x225x

41x21x

380x274x

51x27x

13

Conclusions

This paper presents BWaveR, a fast and memory efficient sequence mapper

leveraging succinct data structures and HW acceleration on FPGA:

➡ Flexibility of usage

➡ Great compression capabilities

➡ Time independence from reference size

➡ Low energy consumption

➡ Low execution time

For questions regarding this work, email:

Guido Walter Di Donato: guidowalter.didonato@mail.polimi.it