fpga-accelerator attractor computation of scale free gene

38
FPGA-accelerator Attractor Computation of Scale Free Gene Regulatory Networks Ricardo Ferreira, Julio Vendramini Departamento de Informática, Universidade Federal de Viçosa, Brazil [email protected] FPL 2010 20th Field Programmable Logic Conference 31 Aug 2 Sept – Milan, Italy

Upload: others

Post on 03-Feb-2022

5 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: FPGA-accelerator Attractor Computation of Scale Free Gene

FPGA-accelerator Attractor Computation of Scale Free Gene Regulatory Networks

Ricardo Ferreira, Julio Vendramini

Departamento de Informática,

Universidade Federal de Viçosa, Brazil

[email protected]

FPL 201020th Field Programmable Logic Conference31 Aug 2 Sept – Milan, Italy

Page 2: FPGA-accelerator Attractor Computation of Scale Free Gene

Contents• Basic Concepts

– Gene Regulatory Networks and Scale Free Networks

• Problem– Attractor Computation

• Contributions– FPGA-accelerator– Architecture Node Framework– Dynamic Interconnections

• Results and Conclusions

Page 3: FPGA-accelerator Attractor Computation of Scale Free Gene

Gene Regulatory Networks

• Dynamic Model of System Behaviour• Applications

– Cell differentiation and Evolution– Drug design process

• Perturbations– Robust – Adaptable

Page 4: FPGA-accelerator Attractor Computation of Scale Free Gene

Gene Regulatory Networks

• Dynamic Model of System Behaviour• Applications

– Cell differentiation and Evolution– Drug design process

• Perturbations

–Robust – Adaptable

Cell A Cell A

Page 5: FPGA-accelerator Attractor Computation of Scale Free Gene

Gene Regulatory Networks

• Dynamic Model of System Behaviour• Applications

– Cell differentiation and Evolution– Drug design process

• Perturbations– Robust

–AdaptableCell A Cell A1

Page 6: FPGA-accelerator Attractor Computation of Scale Free Gene

Gene Regulatory Networks● Models

– Ordinary Differential Equations, Bayesian Networks, Petri Nets, .....

– Boolean Networks [Kauffman, 69], “The Origins of

Orders: Self Organization and Selection of Evolution” [Kauffman, 93]

Random Graph, where each Node: Gene (1 or 0) K neighbours; Random Boolean function

Page 7: FPGA-accelerator Attractor Computation of Scale Free Gene

Network TopologyRandom Scale Free Hierarchical

Page 8: FPGA-accelerator Attractor Computation of Scale Free Gene

Network TopologyRandom Scale Free Hierarchical

Internet, citations, social networks.....•Science [Barabásie e Albert, 1999]•Nature - [Watts e Strogatz, 1998]

Kauffman

Page 9: FPGA-accelerator Attractor Computation of Scale Free Gene

Scale FreeRandom Scale Free Hierarchical

Scale Free Gene regulatory networks•[Aldana, 2003], [Irons, 2006],[Iguchi,2007]

Yeast protein Interaction Network

Page 10: FPGA-accelerator Attractor Computation of Scale Free Gene

Network Attractors ● Network State = All node states● Network Dynamics

– State Evolution;– System converges to stable cycles, called

attractors

Network

Network state

Page 11: FPGA-accelerator Attractor Computation of Scale Free Gene

Network Attractors ● Network State = All node states● Network Dynamics

– State Evolution;– System converges to stable cycles, called

attractors

Network One Step on Time

Page 12: FPGA-accelerator Attractor Computation of Scale Free Gene

Network Attractors ● Network State = All node states● Network Dynamics

– State Evolution;– System converges to stable cycles, called

attractors

Network Attractor or cycle

Page 13: FPGA-accelerator Attractor Computation of Scale Free Gene

Attractors● Cycles - Biological

– Could be observed (experiments)– Cell type

●Example:

State Diagram of Network

network

Page 14: FPGA-accelerator Attractor Computation of Scale Free Gene

Attractor Complexity● Large State Space 2gene

– 2100 ~ 10 27 states...– NP-Hard

●Example:

State Diagram of Network

network

gene

3 genes → 8 states

State transition =All nodes and edgesMust be visited

Page 15: FPGA-accelerator Attractor Computation of Scale Free Gene

Attractor Computation

●Synchronous model●Two Simulation instances: S

0 and S

1

Network State diagram

One stepTwoSteps

Page 16: FPGA-accelerator Attractor Computation of Scale Free Gene

Attractor Computation

●Synchronous model●Two Simulation instances: S

0 and S

1

Network State diagram

One stepTwoSteps

Page 17: FPGA-accelerator Attractor Computation of Scale Free Gene

Sequential Algorithm

● For each step– Visit all nodes and all edges O( N + E )

Network State diagram

Page 18: FPGA-accelerator Attractor Computation of Scale Free Gene

Sequential Algorithm

● Several steps to find an attractor....

Network State diagram

Page 19: FPGA-accelerator Attractor Computation of Scale Free Gene

Sequential Algorithm Complexity

● O( (T+C) steps ) = O( (T+C) * (N+E))– Where C = cycle size, T = transient size

Network State diagram

Page 20: FPGA-accelerator Attractor Computation of Scale Free Gene

Our Approach: Parallel Step Computation O(1)

One clock cycle to visit all Nodes and all Edges

Page 21: FPGA-accelerator Attractor Computation of Scale Free Gene

Simulation Based Model

• 100 genes, 2100 states– Impossible to visit all state space

• Generate a large number of random networks and sample

BiologicalKnowledge

models

Revise Models(topologies,

Boolean functions....)

MapNetwork

Initial statesimulate

Page 22: FPGA-accelerator Attractor Computation of Scale Free Gene

Simulation Based Model

• 100 genes, 2100 states– Impossible to visit all state space

• Generate a large number of random networks and sample

BiologicalKnowledge

models

Revise Models(topologies,

Boolean functions....)

MapNetwork

Initial statesimulate

GenerateNetworks

10 000 Randomnetworks

1000Initialstates

Page 23: FPGA-accelerator Attractor Computation of Scale Free Gene

Previous Approaches on FPGA

NetworkGeneration

SíntesePlace & Route

para FPGA

SynthesisPlace & Route

FPGA

SíntesePlace & Route

para FPGAFPGA

Configuration

SíntesePlace & Route

para FPGAExecution

TimeConsuming > Configuration Execution

Time Time.

[Zerarka, 2004], [Pournara, 2005]Synthesis Time is not reported

Minutes, hours μs

Map NetworkOn FPGA

Page 24: FPGA-accelerator Attractor Computation of Scale Free Gene

Our Approach

NetworkGeneration

Size N

SíntesePlace & Route

para FPGA

SynthesisPlace & Route

FPGA

FPGAConfiguration

ArchitectureFramework

Map GenericFree ScaleNetwork

interconnections

NodeVhdl generator

Dynamic Reconfiguration

To generate severalnetworks

Page 25: FPGA-accelerator Attractor Computation of Scale Free Gene

Our Approach

NetworkGeneration

Size N

SíntesePlace & Route

para FPGA

SynthesisPlace & Route

FPGA

FPGAConfiguration

ArchitectureFramework

Map GenericFree ScaleNetwork

interconnections

SynthesisOnce !

Page 26: FPGA-accelerator Attractor Computation of Scale Free Gene

Proposed Architecture Framework

ArchitectureFramework

interconnections

Each node isMapped on

Process Element (PE)Which implement a

FSM

S0

S1

FSM

Receive data fromThe neighbour

Compute new state

Page 27: FPGA-accelerator Attractor Computation of Scale Free Gene

We propose Dynamic InterconnectionsMultistage Interconnection

O( N log2 N)

Edges are mapped

networkmappednetwork

Page 28: FPGA-accelerator Attractor Computation of Scale Free Gene

Change configuration

New Network remapped

network

Generate a new randomNetwork

reconfigure Multistage

Page 29: FPGA-accelerator Attractor Computation of Scale Free Gene

Generation and Simulation

Newnetwork

DynamicInterconnection

Network

V1

V2

Vn

Several Networks are generated

FPGA

NewConfig.

returnAttractor

size

InitialStates

Page 30: FPGA-accelerator Attractor Computation of Scale Free Gene

Scale Free Networks and Architecture

Page 31: FPGA-accelerator Attractor Computation of Scale Free Gene

Scale Free Networks and Architecture

Page 32: FPGA-accelerator Attractor Computation of Scale Free Gene

Scale Free PE architecture

Network PE Size Distribuition

Size 1 2 4 8 16 32

100 49 28 11 12

200 98 57 25 15 5

300 176 44 30 30 15 5

400 248 80 30 30 8 5

Large numberOf poorly connected

Few number ofStrongly connected

Generic PE distribution

Few number ofStrongly connected

Each random Scale Free

Has different PEdistribution

Page 33: FPGA-accelerator Attractor Computation of Scale Free Gene

Scalable Architecture

Network PE Size Distribuition FPGA

Size 1 2 4 8 16 32 Occupancy

100 49 28 11 12 4.4%

200 98 57 25 15 5 9.5%

300 176 44 30 30 15 5 10.9%

400 248 80 30 30 8 5 12.2%

Double Size

Page 34: FPGA-accelerator Attractor Computation of Scale Free Gene

CPU and FPGA Execution Time

Network sub- CPU FPGA

size edges steps (ms) (μs) speed-up

100 933 5 8.12 13 706

200 2331 7 55.9 43 1300

300 4285 14 104.4 98 1128

400 4816 19 69.0 70 976

3 order of magnitude

10.000 networks + 100 Initial states → 27 hours 10.000 networks + 100

Initial states → 2 minutes

Page 35: FPGA-accelerator Attractor Computation of Scale Free Gene

Dynamic Interconnection Size and configuration bits

Multistage FPGA memories

Size LUTs occupancy (max. 416)

128 1015 0.6% 14

256 1777 1.2% 32

512 4065 2.3% 72

1024 8447 5.6% 160

O(n log2 n)

ConfigurationBits

Up to 512 differentconfigurations

Page 36: FPGA-accelerator Attractor Computation of Scale Free Gene

Conclusions

• FPGA acellerators• Bioinformatics – Large amount of Data

and parallelism• Proposed Implementation

– Speed up : 2-3 Order of Magnitude– Dynamic Interconnection Reconfiguration – Real World: 100-2000 nodes

• Suitable for FPGA Technology

Page 37: FPGA-accelerator Attractor Computation of Scale Free Gene

Conclusions

• Model Scale Free Network on FPGA• FPGA Embedded Memories

– Reduce space of reconfiguration bits

• Multistage Interconnection– Dynamic

– O(n Log2 n)

• Generic Architecture Framework– FSM computation → nodes– Multistage → edges

Page 38: FPGA-accelerator Attractor Computation of Scale Free Gene

Future Works

• Exploration of Gene Regulatory Networks– Boolean Functions– Topologies– Probabilistic Models

• Generic Architecture– Nodes + Dynamic Edges– Model others Cellular Automata Problems