HUG @ NGCLE@e-Novia 15.11.2017

1. HUGenomics
Hardware for hUman Genomics
e-Novia
15/11/2017
Lorenzo Di Tucci, Giulia Guidi, Chiara Crippa, Davide Sampietro,
Alessandro Comodi, Davide Conficconi, Sara Notargiacomo, Marco D. Santambrogio

2. 2
Personalized medicine
Responds to a standard dose
Responds to a higher dose
Responds to a lower dose
Responds to an alternative
medication
• Thousands of DNA variants are associated with different diseases
• Possibility to develop a more precise and personalized medicine for
each individual patient

3. 3
Challenges
a. Development of new architectures to efficiently process the huge
amount of data generated in the years
b. Development of technologies to efficiently and clearly visualize
the computed data

4. 4
Proposed solution
HUG is an integrated hardware and software system that aims
at:
• becoming an advanced support for bioinformatics
• facilitating the identification of correlations, making less
complicated the identification of biological meanings

5. 5
Implementation

6. 5
Implementation

7. 5
Implementation
hardware accelerator

8. 5
Implementation
hardware accelerator

9. 5
Implementation
hardware accelerator

10. 6
Applications

11. 6
Applications
FPGA-based implementation of a Smith-Waterman
algorithm

12. 6
Applications
FPGA-based implementation of a Smith-Waterman
algorithm
FPGA-based implementation of an Ab Initio Protein
Folding algorithm

13. 7
Applications
HARDWARE ACCELERATION
OF GENETIC VARIANT
CALLER

14. 7
Applications
About 80% of rare diseases
are of genetic origin

15. 7
Applications
About 80% of rare diseases
are of genetic origin
By identifying gene changes
specific treatments
may be recommended

16. 7
Applications
About 80% of rare diseases
are of genetic origin
By identifying gene changes
specific treatments
may be recommended
Finding the changes implies
a comparison against
huge databases
of known variants

17. 7
Applications
About 80% of rare diseases
are of genetic origin
By identifying gene changes
specific treatments
may be recommended
Finding the changes implies
a comparison against
huge databases
of known variants
Even modestly complex
algorithms would require a
prohibitively long time to
complete

18. 8
https://macarthurlab.org
1
2
DNA is extracted from a blood sample of the individual and
loaded on to a sequencing machine
GATK Best
Practice
• Data
Pre-processing
• Variant
Discovery
• Callset
Refinement
FASTQ VCF
2
Variant Discovery

19. 9
Workflows are specific to the variant type:
• Germline SNPs & Indels
• Somatic SNVs & Indels
Workflows
https://software.broadinstitute.org/gatk/

20. 9
Workflows are specific to the variant type:
• Germline SNPs & Indels
• Somatic SNVs & Indels
Workflows
https://software.broadinstitute.org/gatk/

21. 9
Workflows are specific to the variant type:
• Germline SNPs & Indels
• Somatic SNVs & Indels
Workflows
https://software.broadinstitute.org/gatk/
PairHMM

22. 10
https://software.broadinstitute.org/gatk/
Pair-HMM
Stage Time Runtime %
Assembly 2,598s 13%
Pair-HMM 14,225s 70%
Transversal + Genotyping 3,379s 17%
Biological sample - NA12878 80xWGS

23. 10
https://software.broadinstitute.org/gatk/
Pair-HMM
Stage Time Runtime %
Assembly 2,598s 13%
Pair-HMM 14,225s 70%
Transversal + Genotyping 3,379s 17%
Biological sample - NA12878 80xWGS

24. 11
Pairwise alignment of each read
against each candidate haplotype
The likelihood of each
read-haplotype pair is calculated
Pair-HMM
Every possible alignment is taken
into account for cumulative
probability

25. 12
Parallelization
read
haplotype
Store only 3 diagonals
Diagonal-wise computation
Hardware implementation

26. 13
Overall system

27. 13
Overall system
C/C++
Libraries

28. 13
Overall system
C/C++
Libraries

29. 14
Applications
Tiled Regular eXpression
matching architecture

30. 14
Applications
Genomic data efficiently
analyzed through regular
expressions

31. 14
Applications
Genomic data efficiently
analyzed through regular
expressions
RE as instructions
the need for a
customized processor

32. 14
Applications
Genomic data efficiently
analyzed through regular
expressions
RE as instructions
the need for a
customized processor
Reconfigurable Instruction
Set Architecture

33. Architecture Implementation 15

34. 16Architecture Implementation
Fetch & Decode Unit
Produces three kind of
information:
• reference data
• valid reference number
• opCode

35. Multiple clusters of comparators
to increase the throughput
17Architecture Implementation
Execution Unit

36. 18Architecture Implementation
Control Path
Handles the execution of the
whole pipeline

37. Results: Single Core
VC-707 Board Slice LUTs Slice Reg. F7 Muxes F8 Muxes
Available 303600 607200 75900 75900
TiReX
1921
[0. 63%]
1775
[0.29%]
261
[0.17%]
114
[0.15%]
19
Area Utilization
Performance *
Regular Expression
Time Flex
(intel i7
2.80Ghz)
Time TiReX
(100 MHz)
Speedup
ACCGTGGA 271 µs 39,9 µs 6x
(TTTT)+CT 121 µs 81,35 µs 1.48x
(CAGT)|(GGGG)|(TTGG)TGCA(C|G)
+
263 µs 173,835 µs 1.5x
* Dimension of the test file: 16 KB

38. From single core to multicore 20
Processor Data

39. From single core to multicore 20
Processor Data

40. From single core to multicore 20
Processor Data

41. From single core to multicore 20
Processor Data

42. From single core to multicore 21
Processor Data

43. • Tiled Architecture
• Two ways to operate:
– Single Instruction Multiple
Data (SIMD)
– Multiple Instructions Single
Data (MISD)
22
From single to multicore

44. 23Results: Quad-Core implementation
Performance *
Regular Expression
Time Flex
(intel i7
2.80Ghz)
Time TiReX
(Quad-Core100 MHz)
Speedup
ACCGTGGA 271 us 10 us 27.1x
(TTTT)+CT 121 us 6.23 us 19.42x
(CAGT)|(GGGG)|(TTGG)TGCA(C|
G)+
263 us 24 us 10.9x
* Dimension of the test file: 16 KB

45. 24
Work in progress

46. 25
What’s NECST?

47. 25
What’s NECST?
Algorithms integration
Development of tools to
efficient visualize data

48. Thanks for your attention
e-Novia
15/11/2017
Lorenzo Di Tucci, Giulia Guidi, Chiara Crippa, Davide Sampietro,
Alessandro Comodi, Davide Conficconi, Sara Notargiacomo, Marco D. Santambrogio
{giulia.guidi, chiara2.crippa, davide.sampietro, alessandro.comodi, davide.conficconi}@mail.polimi.it
{lorenzo.ditucci, sara.notargiacomo, marco.santambrogio}@polimi.it