Can we Verify an Elephant?
David HarelThe Weizmann Institute of Science
Surprisingly many parts of this were influenced by
Amir Pnueli
In recent years he became very interested in biological modeling, and actively participated in some of the
projects
Here are some static computerized elephants
Computer science is really the science of the
dynamic
As are certain parts of mathematics
So here are some dynamic computerized elephants
And now for a really dynamic one
Just to get us in the mood….
Why do we do computerized modeling ?
What and how should we model?
What makes models “valid”, “complete”, and how do we verify this?
Such questions become especially acute when we try to model Nature
Biological artifacts are really reactive systems (Harel & Pnueli, 1986) on all levels: the molecular and the cellular, and all
the way up to organs and full organisms
Biology as Reactivity
Biological systems can be modeled and analyzed as reactive systems,
using languages/tools developed for constructing computerized systems
A thesis follows:
Put simply: Let’s reverse-engineer an elephant rather than engineer an F-15…
What to model?
Be comprehensiveThat is, do the whole thing ...
But what is the whole thing? (horizontal delineation)
• An entire cell
• An entire organ or organism
• An entire population?
On (or up to) what level of detail? (vertical delineation)
• Inter-cellular
• Intra-cellular (inter-molecular)
• Probably also genomic/proteomic
• Maybe biochemistry & even physics (particles, quantum mechanics, string theory…)??
Crucial point:
Comprehensive modeling entails capturing everything that is known
about the system, and doing everything else any which way…
To construct a “full”, true-to-all-known-facts, 4-dimensional model of
a multi-cellular organism
WOP: Whole Organism Project A Grand Challenge for Comprehensive
Modeling (H, 2003)
Which animal would be a good
choice? Later (but it’s not an
elephant…)
Another crucial point(otherwise we’re wasting our time):
The model should make researchers excited, enabling them to observe,
analyze and understand the organism ― development and
behavior ― in ways not otherwise possible; e.g., to predict
• Help uncover gaps, correct errors, form theories and explanations
• Suggest new experiments, and help predict unobserved phenomena
• Help discover emergent properties
• Verify biological theories against laboratory observations
• Pave the way for in silico experimentation, and possibly synthesis, drug construction,…
Additional potential gains are enormous
How to model?
Be realisticThat is, make it look good…
Project I (thymus)
(with S. Efroni and I. Cohen, ‘03 )
• T-cell (thymocyte) behavior in the thymus.
• Many cells, complex internal behavior, interaction and geometric movement.
• Enormous amount of biological knowledge assimilated and modeled (~ 400 papers).
The front end
Statechart outline for a single T-cell
Migration
Interaction
Receptors
Cell phase
Receptorsdecisions
Entry to thymus
Straight run
Interaction, etc.
The model reveals emergent properties (with Efroni and Cohen, ‘07)
Competition change:
Project II (pancreas) (with Y. Setty, Y. Dor and I. Cohen; 2007)
• Embryonic development of the pancreas (very different characteristics).
• Here we use 3D animation and are interested in organ formation.
Cell count results:
Normal growth:
Wild “playing” yielded insights into the role of blood vessel density
into organ development
Experimental confirmation in progress!
Project III (C. elegans)(with N. Kam, M. Stern, J. Hubbard, J. Fisher, H. Kugler, A. Pnueli;
2001−7)
• Vulval precursor cell (VPC) fate determination in the C. elegans nematode
• Few cells, lateral and inductive signaling with subtle timing; many mutation-driven variants.
C. elegans:
Development
Behavior
Meet the Grand Challenge by modeling the C. elegans
nematode
Proposal:
Or some comparable creature
P6.p
P7.p
ayIs4;e1282;lin-15(n309)
P5.pP4.p
anchor cell
P5 ablated wildtype vulva fate
Carry out multi-level modeling, with different abstraction levels
modeled with different languages and methods
Then combine all to yield a smoothly zoomable & executable model
Central CS problem to be solved:Vertical linkage
(hierarchy, abstraction and levels )
A modest step forward: Biocharts
(with H. Kugler and A. Larjo, 2009)
• A compound, fully executable 2-tier language for modeling biology
• Upper level captured using Statecharts
• Lower level captures networks, pathways, etc.; e.g., with semantics-rich biological diagrams.
When are we done?
Aha! The $64m question…
But,… comprehensive modeling is about understanding a whole thing
You really and truly understand a thing when you can build an interactive
simulation that does exactly what the original thing does on its own.
Q: How do you tell when you’ve managed to achieve that ?
A: We want prediction-making taken to the utmost limit; the key to this is to fool
an expert.
Hence, for comprehensive modeling, I propose a Turing-like test, but with a
Popperian twist
We are done when a team of biologists, “well versed” in the relevant
field, won’t be able to tell the difference between the model and the
real thing
A Turing-like test for modeling (H’ 2005)
This is not a test for the weak-hearted, or for the impatient…
And it’s probably not realizable at all…
But as the ultimate mechanism for prediction-confirming, it can serve as a lofty, end-of-the-day,
goal for the WOP Grand Challenge
Thank you for listening