Computational Discovery of Communicable Knowledge

1
Computational Assistance for Systems Biology of
Aging
Pat Langley School of Computing and
Informatics Arizona State University Tempe,
Arizona
Thanks to D. Bidaye, J. Difzac, J. Furber, S.
Kim, S. Racunas, N. Shah, J. Shrager, D.
Stracuzzi, and M. Verdicchio for their
contributions to this research, which was funded
in part by a grant from Science Foundation
Arizona.
2
The Complexity of Human Aging
There is mounting evidence that aging involves
many interacting mechanisms, including

• Mutation of mitochondrial DNA
• Accumulation of lipofuscin in lysosomes
• Protein crosslinking in extra-cellular matrix
• Cell senescence and cell loss

The daunting complexity of modeling these and
other processes suggests the need for
computational assistance.
3
Furbers Network Diagram of Aging
4
Challenges for a Systems Biology of Aging
Furbers diagram offers a good step toward a
systems biology of aging, but it remains

• Informal, in that the meanings of the diagrams
  nodes and links are imprecise
• Inert, in that it needs a human interpreter to
  produce model explanations or predictions and
• Static, in that the model cannot be updated or
  revised over time without considerable effort.

In this talk, I present some computational
responses to these three challenges that build on
Furbers work.
5
An Initial Modeling Environment
6
Formal Representation of Aging Processes
We want a notation for models of aging that is
precise enough for a digital computer.
Computational biology is rife with candidate
formalisms, but most are problematic

• Differential equations require functional forms
  and parameters
• Boolean networks assume discrete, not continuous,
  variables
• Bayesian networks require arbitrary probabilistic
  parameters

We need a notation that makes closer contact with
biologists ideas about aging mechanisms.
Kuipers (1986) qualitative modeling formalism
offers many of the features that we desire.
7
Qualitative Causal Models of Aging
We can state a model as a set of qualitative
elements, each of which specifies

• Places in the cell (e.g., lysosome, cytoplasm)
• Quantities that are measured in a place
• For unstable entities (e.g., ROS, Fe)
• For stable entities (e.g., oxidized proteins,
  lipofuscin)
• Causal influences between quantities
• Increase/decrease of one quantity with another
• Increase/decrease of quantitys rate of change

Each influence describes a qualitative causal
link between two quantitative variables.
8
A Formal Lysosome Model
Time



junk-protein
junk-protein

Lysosome
Cytoplasm
Fe




ROS
membrane-damage


lipofuscin
oxidized-protein

lytic-enzyme
lipofuscin


H202
H202
9
Examining the Lysosome Model

• gt load lyso.lisp
• loaded lyso.lisp
• gt show places entities
• Places
• p1 cell
• p2 lysosome in the cell
• p3 cytoplasm in the cell
• Quantities
• q0 time
• q1 junk-protein in the lysosome
• q2 junk-protein in the cytoplasm
• q3 fe in the lysosome
• q4 ros in the lysosome
• q5 oxidized-protein in the lysosome
• q6 lipofuscin in the lysosome
• q7 lipofuscin in the cytoplasm
• q8 lytic-enzyme in the lysosome
• q9 damaged-membrane in the lysosome
• q10 h2o2 in the lysosome

10
Examining the Lysosome Model
gt show claims Claims c1 junk-protein in the
lysosome decreases with time c2 junk-protein
in the cytoplasm increases with time c3
junk-protein in the lysosome increases with
junk-protein in the cytoplasm c4 fe increases
with junk-protein in the lysosome
lysosome digestion produces fe from junk
proteins c5 ros increases with fe in the
lysosome c6 oxidized-protein increases with
ros in the lysosome ros oxidizes
proteins to produce oxidized proteins c7
lipofuscin increases with oxidized-protein in the
lysosome oxidized-proteins and fe
crosslink to form lipofuscin c8 c2 decreases
with lipofuscin in the lysosome
lipofuscin reduces the rate of disassembling
junk proteins c9 lytic-enzyme decreases with
lipofuscin in the lysosome c10 ros increases
with lipofuscin in the lysosome c11
damaged-membrane increases with ros in the
lysosome c12 lipofuscin in the cytoplasm
increases with damaged-membrane in the lysosome
damaged lysosomal membranes spill
lipofuscin into cytoplasm c13 h2o2 in the
lysosome increases with h2o2 in the cytoplasm
11
Reasoning About Aging Models
A systems model of aging is lacking unless one
can relate it to observable phenomena. Such an
account should let one answer questions like

• What biological effects does the model predict?
• What observations/experiments does the model
  explain?
• What portions of the model explain a given
  phenomenon?
• How would changes to the model alter its
  predictions?

Model complexity can make these difficult to do
manually, but we can provide computational
support for such reasoning.
12
Encoding Predictions and Observations
Before one can relate a models predictions to
observations, we must represent them both. Our
environment uses a notation similar to model
claims

• A quantity increases with another quantity
• E.g., lipofuscin in the lysosome increases with
  time
• A quantity decreases with another quantity
• E.g., lytic-enzyme decreases with ROS in the
  lysosome
• A quantity does not change with another quantity
• E.g., H2O2 does not vary with ROS in the
  lysosome

Note These describe phenomena that the model
does or should predict they are not part of the
model themselves.
13
Examining Facts, Queries, and Predictions
gt show facts Facts f1 lipofuscin in the
lysosome increases with time f2 lipofuscin in
the cytoplasm increases with time f3
lytic-enzyme decreases with ros in the lysosome
f4 h2o2 does-not-change with ros in the lysosome
14
Examining Facts, Queries, and Predictions
gt show facts Facts f1 lipofuscin in the
lysosome increases with time f2 lipofuscin in
the cytoplasm increases with time f3
lytic-enzyme decreases with ros in the lysosome
f4 h2o2 does-not-change with ros in the
lysosome gt does lytic-enzyme change with ros in
the lysosome ? p1 lytic-enzyme decreases with
ros in the lysosome gt does oxidized-protein
change with h2o2 in the lysosome ? p2
oxidized-protein does-not-change with h2o2 in the
lysosome
15
Examining Facts, Queries, and Predictions
gt show facts Facts f1 lipofuscin in the
lysosome increases with time f2 lipofuscin in
the cytoplasm increases with time f3
lytic-enzyme decreases with ros in the lysosome
f4 h2o2 does-not-change with ros in the
lysosome gt does lytic-enzyme change with ros in
the lysosome ? p1 lytic-enzyme decreases with
ros in the lysosome gt does oxidized-protein
change with h2o2 in the lysosome ? p2
oxidized-protein does-not-change with h2o2 in the
lysosome gt predict f1 f2 f3 f4 f1 lipofuscin
in the lysosome increases with time the
model makes ambiguous predictions. f2
lipofuscin in the cytoplasm increases with time
the model makes ambiguous predictions. f3
lytic-enzyme decreases with ros in the lysosome
p3 the model predicts the same relation. f4
h2o2 does-not-change with ros in the lysosome
p4 the model predicts the same relation.
16
Generating Model Predictions
Given a qualitative query about the relation
between two model quantities, the environment

• Chains backward from dependent term D to find all
  paths that connect it with independent term I.
• If no paths exist, then it predicts no empirical
  relationship.
• If a path has an even number of decreases links,
  it predicts D increases with I, else that D
  decreases with I.
• Provides a definite relationship if all path
  predictions agree.
• Notes an ambiguous relationship if path
  predictions disagree.

The same form of qualitative reasoning predicts
both changes over time and responses to
experimental manipulation.
17
Conflicting Model Pathways
Time



junk-protein
junk-protein

Lysosome
Cytoplasm
Fe




ROS
membrane-damage


lipofuscin
oxidized-protein

lytic-enzyme
lipofuscin


H202
H202
18
Altering a Qualitative Model of Aging
Biologists should also be able to extend and
revise models of aging easily and efficiently.
Our environment lets users alter the current
model interactively in five basic ways

• Adding a new place, quantity, claim, or fact
• Adding a note (e.g., a URL) to a claim or fact
• Removing a place, quantity, claim, or fact
• Disabling or enabling a claim or fact
• Saving revisions to a file that can be loaded
  later

Some changes can alter the models implications
and its ability to explain observed results.
19
Examining the Effects of Model Revision
gt predict f1 lipofuscin in the lysosome
increases with time the model makes
ambiguous predictions. f2 lipofuscin in the
cytoplasm increases with time the model
makes ambiguous predictions. f3 lytic-enzyme
decreases with ros in the lysosome p3 the
model predicts the same relation. f4 h2o2
does-not-change with ros in the lysosome p4
the model predicts the same relation.
20
Examining the Effects of Model Revision
gt predict f1 lipofuscin in the lysosome
increases with time the model makes
ambiguous predictions. f2 lipofuscin in the
cytoplasm increases with time the model
makes ambiguous predictions. f3 lytic-enzyme
decreases with ros in the lysosome p3 the
model predicts the same relation. f4 h2o2
does-not-change with ros in the lysosome p4
the model predicts the same relation. gt disable
c1 junk-protein in the lysosome decreases with
time
21
Examining the Effects of Model Revision
gt predict f1 lipofuscin in the lysosome
increases with time the model makes
ambiguous predictions. f2 lipofuscin in the
cytoplasm increases with time the model
makes ambiguous predictions. f3 lytic-enzyme
decreases with ros in the lysosome p3 the
model predicts the same relation. f4 h2o2
does-not-change with ros in the lysosome p4
the model predicts the same relation. gt disable
c1 junk-protein in the lysosome decreases with
time gt predict f1 lipofuscin in the lysosome
increases with time p5 the model predicts the
same relation. f2 lipofuscin in the cytoplasm
increases with time p6 the model predicts the
same relation. f3 lytic-enzyme decreases with
ros in the lysosome p7 the model predicts the
same relation. f4 h2o2 does-not-change with
ros in the lysosome p8 the model predicts the
same relation.
22
Status of the Interactive System
Our computational assistant is still under
development, but the system already lets users

• Load, examine, and alter a qualitative model of
  aging
• Answer queries about how one quantity affects
  another
• Compare the models predictions to observed facts
  
• Determine effects of model revisions on
  predictions

Also, we have an initial encoding for the
lysosomal portion of Furbers network diagram.
We have much work ahead of us, but we also have
the basic machinery and a partial knowledge base
in place.
23
Support for the Aging Research Community
Initial users of our interactive modeling system
are likely to be individual biologists or
laboratories. However, a Web-based version would
benefit the distributed aging research community,
which could use it to

• Store shared beliefs about aging phenomena and
  processes
• Exchange new observational facts and propose
  hypotheses
• Update the community model of aging incrementally

The system would serve as a graphical wiki that
organizes knowledge, directs discussion, and
grows over time. However, content would consist
of formally stated models and phenomena, rather
than documents.
24
Intellectual Influences
Our approach to computational biological aides
incorporates ideas from many traditions

• formalizations of biological knowledge (e.g.,
  EcoCyc, 2003)
• qualitative reasoning and simulation (e.g.,
  Kuipers, 1986)
• languages for scientific simulation (e.g.,
  STELLA, PROMETHEUS)
• Web-supported tools for biological visualization
  (e.g., KEGG)
• Web-based tools for biological processing (e.g.,
  BioBike, 2007)

However, it combines these ideas in novel ways to
assist in the construction of system-level
models of aging.
25
Directions for Future Research
Our effort is still in its early stages, and we
need further work to

• expand our encoding of aging events and processes
• formalize more content from Furbers network
  diagram
• provide a graphical interface for viewing and
  using models
• make the system accessible remotely using the Web
• support community-based development of models
• evaluate the softwares actual usability for
  biologists

Together, these changes should make our
interactive system a powerful computational aid
for aging researchers.
26
Key Contributions
In summary, we are developing an interactive
computational aid for systems biology of aging
that supports

• Formal models of aging that specify clear
  relationships among known biological quantities
• Interpretable models that let a computer predict
  and explain known phenomena and
• Revisable models that users can evaluate, update,
  and revise with little training and effort.

The system is still immature, but a more advanced
version would offer many benefits to the aging
community.
27
End of Presentation