High performance computing in biology: challenges and perspectives

1
High performance computing in biology challenges
and perspectives
PNNL-SA-61363

• SciDAC 2008
• Christopher Oehmen
• Pacific Northwest National Laboratory

2
Computational Biology computing at many scales

• Biological processes occur simultaneously at many
  temporal and spatial scales
• Molecules form working units and control systems
• Cells encapsulate information and function
• Tissues, communities can work in concertemerging
  behaviors linked to survival

3
Genes

• Challenge is to characterize cells by their
  genetic signals
• Genes turned on indicate processes that are
  activated
• Dynamic, complex signals result from and
  contribute to many combined effects

4
Proteins and proteomics

• Proteins working molecules in cells
• Understand what processes a cell is using by
  taking a snapshot of molecular activity
• Mass spectrometry a good way to characterize all
  proteins at once
• The key is effectively identifying proteins or
  fragments that give rise to MS peaks.

5
Modeling and simulation

• Using theory from chemistry and physics to
  predict behavior of biomolecular systems.
• insight into complex or unmeasurable behaviors
• parameters, boundary conditions, etc... for
  higher level simulations
• Can help predict protein structure from sequence

6
Molecular interactions and relationships

• Graphs used in many different ways in biology
• Can we use graph theory to produce meaningful
  visual metaphors and analysis techniques?
• Biological graphs can be very complex,
  simplifying approximations often obscure the
  underlying biology

7
Multicell tissues and communities

• Understanding emerging behavior requires
  integration across many levels
• Control systems
• Signals
• Coordinated Responses
• ...
• Needs computing at terascale, petascale and
  beyond

8
HPC for genome analysis

• Joint Genome Institute (JGI)
• National Center for Biotechnology Information
  (NCBI)
• The Institute For Genomic Research (TIGR-JCVI)
• ...

Online tools
Data resources
9
Genome sequencing
consortium
Cost/genome Base pairs/day genomes
M 100K 1K-1M 100M-1B Few thousands
institute
single-investigator project
1995
2001
today
10
More genomes more analysis!
11
But what I really want to do with my genomes is...

• Human health
• Drug design
• Biomarkers
• Energy
• Engineered systems for renewable energy
• Carbon management
• Defense
• Rapid identification
• Forensic analysis

How can we enable users to develop, refine
hypotheses in real time?
12
Conventional interface to HPC
Quality of solution related to theory used,
reliability of input and solution domain,
accuracy and precision of mathematical method.
Calculation driven by theoretical formulation and
observations. Generally the hypothesis being
evaluated maps well to the output of the
calculation.
13
Biology needs a different interface to HPC
DATA MATH
Calculations often driven by data and a mix of
statistical methods and chemistry, physics, other
equations. Driving question is often a high-level
hypothesis that is not easily mapped to the
output.
14
Spectrum of HPC coupling
High degree of coupling
Low degree of coupling
Remote HPC services
Local HPC integrated workflow
Local HPC services

• Website/web services interface to external,
  shared large-scale facilities may require
• Manual data entry
• Password/authentication
• Limit on priority for large-scale tasks

• Website/web services interface to internal,
  dedicated facilities
• User has more control over resource allocation

• Application-level access to dedicated HPC
• User doesnt even have to know HPC is being used

15
Taking advantage of local HPC resources

• Web services, websites for launching parallel jobs

Local dedicated cluster
User or applications can integrate output into
workflow
Polygraph and ScalaBLAST both implemented using
this model at PNNL
16
More tightly couple HPC resources to biological
workflows

• Case study multiple whole-genome analysis
  workflow

Visualization
High performance hardware, software
Genome 1
Genome 2
Genome 3
Post-processing

Genome n
17
More tightly couple HPC resources to biological
workflows

• Case study proteomics workflow

3. Public tools
Instrument output
Bioinformatics Resource Manager (BRM)
2. PQuad
1. HPC Polygraph
18
Challenges for HPC users in biology

• Biological computing is different than most HPC
• Integer mathematics more important than floating
  point arithmetic
• Memory or memory-latency bound
• Data-driven, data-intensive
• We need different kinds of hardware
• Mathematical challenges
• Biological data is growing exponentially
• 1 false positive rate is unacceptable when you
  have 1 billion items
• Often want to understand space of good solutions
  instead of 1 optimal solution
• Scalable applications must continually evolve
  with new mathematical theory

19
More challenges...

• Policy/permissions
• Most HPC systems operate using multiuser, batch
  model
• Tightly coupling HPC into biological workflows
  means applications will need more immediate
  access to compute cycles, NOT batch mode
  operation
• Local HPC systems not normally available for
  anonymous users

20
Summary

• Biology sciences are just scratching the surface
  of how high performance computing might be used.
• HPC solutions in biology will need to accommodate
  exponentially growing datasets with evolving
  mathematics
• Biology will likely continue to provide science
  drivers for novel architectures that prioritize
  integer mathematics and memory bandwidth/capacity
• HPC can maximize its value to biology by more
  tightly coupling with analytical pipelines and
  workflowsbut this will require different access
  and user models

21
Acknowledgments

• Support provided by the Data Intensive Computing
  For Complex Biological Systems funded by the
  Office of Advanced Scientific Computing Research,
  and under the LDRD Program at Pacific Northwest
  National Laboratory.