ORIGINAL ARTICLE

On comprehensive and realistic modeling: Some

ruminations on the what, the how and the why

David Harel, PhD The Weizmann Institute of Science,

Rehovot, Israel
[email protected]

Clin Invest Med 2005; 28 (6): 334–337.

Summary cumstances, often ones that haven't yet been actually

This short paper is about comprehensive realistic tried out in the laboratory.
modeling in general. I am no expert at all on health It is obvious that comprehensive modeling, if car-
care or on modeling health-related systems. Rather, I ried out successfully, can yield very far-ranging bene-
am a computer scientist and, in recent years, have fits for biology and for science in general. However, its
spent time applying some of my work on systems and immediate benefits may be somewhat limited, since it
software engineering to the modeling of biology. is not designed to be a short term effort aimed at solv-
Indeed, the examples given in the talk are of two of ing a particular problem.
our group's biological modeling projects. The notion of realistic modeling is a key issue, and it
Nevertheless, I invited members of the audience to try is addressed throughout the lecture. To be realistic, a
to substitute "biology" for "health care" throughout model must capture not only the overall viable stochas-
the lecture. All I promised was that this experiment tic behaviour of the system as a whole, but also the
could yield interesting, perhaps thought-provoking, behaviour of the individual entities and their inter-rela-
results. Towards the end I posed a "grand challenge" tionships, their cooperation and their influence on each
for the health-care modeling community. other. In fact, it is best if the model is such that the
The lecture emphasizes the two adjectives "compre- overall emergent picture be the result of the combined
hensive" and "realistic", as applied to modeling, and behavior of the individually modeled entities. A realistic
the questions it tries to deal with include: model must be fully executable, which is more than car-
rying through a probabilistic computation of projected
• What kinds of systems should we model?
average case behaviour, or doing queuing theory analy-
• Why do we want to model?
sis of probable outcomes. Executing the system is not
• How should we model?
just producing the end results, say, in the form of the
• When are we done?
probability of some event at the end of computing a
One of the main points made is to highlight the Markov chain. Rather, we want the ability to execute
notion of comprehensive modeling – where the goal is the "program" of the system, which, just like running
to model an entire organ, an entire organism, or even any computer program, can be done on various inputs,
an entire population – and to distinguish it from more in a one-step-at-a-time debugging fashion, in ways that
conventional types of modeling, where one is interest- highlight the behaviour of individual pieces, in ways
ed in a specific aspect of a system and the modeling is that take into account the probability distribution of
aimed at getting particular results or making particu- inputs and of certain decisions made in the process, in
lar predictions. The motivation for comprehensive best and worst case fashion, and indeed in typical aver-
modeling is multi-fold. We really want to understand age cases too. Thus, model execution should be the
the system and to gain deep comprehension of how it true analogue of running a conventional computer pro-
works and of how it behaves over time, but we also gram, and model analysis is the analogue of verification,
want to predict its future behavior under varying cir- validation and complexity analysis.

334 Med clin exp • vol 28, n 0 6, décembre 2005

 Realistic modeling

Another aspect of the realism of the modeling has which this paragraph and the next one are adapted. The
to do with ease of comprehension – both of the model challenge is to fully model an entire multi-cellular
itself and of its dynamics during execution. We want organism. We actually have a particular organism in
the experts of the subject matter (biologists when mind, the Caenorhabditis elegans nematode worm, bet-
modeling biology, and in the present case perhaps ter known simply as C. elegans, a suggestion that is in
health care researchers, hospital officials, and decision line with the extraordinarily insightful 40-year old pro-
makers) to be able to model themselves or, at the very posal of Sydney Brenner, who chose this creature to
least, to comprehend and modify existing models. challenge biologists with the task of discovering the
Thus, heavy use of differential equations or operations entire development and neurobiology of a living crea-
research theories and techniques in the modeling has ture. (For this proposal and the tremendously influen-
the added disadvantage of being unfitting for use, or tial work that he and others did following it, Brenner
even modification by these experts, and indeed it can shared the 2002 Nobel Prize in Physiology or
easily alienate them. Medicine.)
In way of illustrating the "realistic" facet of model- This challenge – which we estimate to require many
ing, the lecture describes the general approach to mod- years of work by many research groups with diverse
eling taken by our group. It is based on viewing the backgrounds, and which might never really be
biological artifacts to be modeled as reactive systems1, achieved – is to construct a full, true-to-all-known-
and to use for their modeling and simulating visual for- facts 4-dimensional model of this worm (or of a com-
malisms.2 These are graphical, diagrammatic languages parable multi-cellular animal), which is easily
that are both intuitive and mathematically rigorous, and extendable as new facts are discovered. The front end
are supported by powerful tools that enable full model would be an anatomically correct, animated graphical
executability. They are linkable to object diagrams and rendition, tightly linked to a reactive system model of
GUIs, and other structural descriptions of the system the entire creature. The model would be fully exe-
under development and its front-end, as well as to full cutable, flexible, interactive, comprehensive and com-
animation by an idea we call reactive animation.3 At prehensible. It would enable realistic simulation of the
present, such languages and tools – often based on the worm's development and behaviour over time (the
object-oriented paradigm – are being strengthened by fourth dimension), which would help uncover gaps,
verification modules, making it possible not only to exe- correct errors, suggest new experiments and help pre-
cute and simulate the system models (test and observe) dict unobserved phenomena. It would be zoomable,
but also to verify dynamic properties thereof (prove). enabling easy switching between levels of detail
They are also linkable to tools for dealing with the sys- (reaching down at least to the cellular level, and pos-
tem's continuous aspects (e.g., Matlab) in a full hybrid sibly the molecular level at some points), and allowing
fashion. researchers to see and understand the organism and its
One of two variants of our approach is state-based, behavior in ways not otherwise possible. The underly-
encouraging an intra-object style of specification, and ing computational framework would be not only rig-
uses the language of statecharts4 to describe the sys- orous and realistic, but would be set up in such a way
tem's behaviour by objects. One powerful tool sup- that biologists would be able to enter new data them-
porting this is Rhapsody,5,6 but there are many selves as it is discovered, and even plug in varying the-
statechart tools. (Matlab has also adopted statecharts ses about aspects of behavior that are not yet known,
for its discrete aspects, in its StateFlow tool.) Another, in order to see their effects.
more recent variant is scenario-based, and inter-object In order to lend support to this outlandish idea, the
in spirit. It uses the language of live sequence charts next part of the lecture describes briefly two modeling
(LSCs),7 and allows one to play in the behaviour projects that we have been carrying out; one using the
directly from the system's GUI and to then play it out state-based intra-object approach and the other using
just as if it were an intra-object model.8 In both cases, (mainly) the scenario-based inter-object approach.
the model's objects are considered to exist as individ- The first project involves T-cell development in the
ual entities, and when executed they interact with oth- thymus,3,10 and shows thousands of cells entering the
ers in ways that are appealingly realistic. thymus, struggling and competing for the prize if
The lecture then goes on to discuss a Grand becoming fully-fledged T-calls. This model was the
Challenge that I proposed a few years ago to the com- motivation for developing reactive animation, and
puter science and systems biology community,9 from uses Flash linked with Rhapsody and its statecharts.

© 2005 CIM Clin Invest Med • Vol 28, no 6, December 2005 335
Harel et al

The second project involves vulval cell fate determina- what is needed for comprehensive modeling to be
tion in the C. elegans nematode,11,12 and its key play- complete. The Popperian twist comes from the fact
ers are six vulval precursor cells who have to decide that once such a model passes the test, it will inevitably
which of them gets the honour of working with a spe- change over time as science develops and we learn
cial anchor cell to form the worm's vulva, which is its more about the system we are modeling – all this in
egg-laying venue. This model was built mainly from the good spirit of Popper's philosophy of science.
LSCs using the Play-Engine, but we have also done
some verification work of cell mechanistic behavior Bibliography
against lab observations, using LSCs and statecharts. 1. Harel D, Pnueli A. On the Development of Reactive
At this point, I propose a Grand Challenge for this Systems. In: Apt KR, editor. Logics and Models of
community. The challenge – in full analogy with the Concurrent Systems.New York: Springer-Verlag; 1985.
challenge for modeling biology9 – is to model a com- p. 477-98.
plete health care system, fully and realistically. This 2. Harel D. On Visual Formalisms. Communications of
could be "merely" an entire hospital, but my feeling is the ACM 1988;31:514-30.
that it should be larger: perhaps the complete hospital 3. Efroni S, Harel D, Cohen IR. Reactive animation:
system for a region or a state. It could, and possibly Realistic modeling of complex dynamic systems.
should, also include (or at least solidly interface with) Computer 2005 Jan;38:38-47.
other relevant entities, such as governmental health 4. Harel D. Statecharts: A visual formalism for complex
offices, medical schools, health insurance companies, systems. Sci Comput Program 1987;8:231-74.
etc. This kind of challenge – again, in full analogy with (Preliminary version: Technical Report CS84-05, The
modeling a biological organism – is very long term Weizmann Institute of Science, Rehovot, Israel,
and incredibly complex and might never be achieved. February 1984.)
However, it also enjoys the same potential benefits, 5. Harel D, Gery E. Executable object modeling with
i.e., providing an unparalleled understanding of a vast statecharts. IEEE Computer 1997 ;30:31-42.
system of relevance. If achieved, such a challenge will 6. I-Logix web site. http://www.ilogix.com
no doubt result in new ideas, predictions, and recom- 7. Damm W, Harel D. LSCs: Breathing life into message
mendations, that could help improve the overall qual- sequence charts. Formal Methods in System Design
ity of health care. Interestingly, truly grand challenges 2001;19(1):45-80. (Preliminary version in Proc. 3rd
often yield significant advances even if they are not IFIP Int. Conf. on Formal Methods for Open Object-
successful, simply by the massive amounts of work that Based Distributed Systems (FMOODS'99 ), (P.
come from the talent, energy, money and dedication Ciancarini, A. Fantechi and R. Gorrieri, eds.), Kluwer
concentrated around them. Academic Publishers, 1999, pp. 293-312.)
The final part of the lecture addresses the particu- 8. Harel D, Marelly R. Come, Let's Play: Scenario-Based
larly interesting question of how we know when we Programming Using LSCs and the Play-Engine.
are done. Or, in other words, when is a comprehen- Springer-Verlag; 2003.
sive, realistic model deemed complete, or valid? Here 9. Harel D. A grand challenge for computing: Towards
I propose a sort of Turing test, but with a Popperian full reactive modeling of a multi-cellular animal.
twist: a model of an entire biological system is com- Bulletin of the EATCS 2003;81:226-35. (Reprinted in
plete and valid if a team of professionals cannot tell the Current Trends in Theoretical Computer Science: The
difference between the model and the real thing.13 Challenge of the New Century, Algorithms and
There are many issues that have to be addressed for Complexity, Vol I, Paun, Rozenberg and Salomaa, eds.,
such a test to be even conceivable, such as the World Scientific, pp. 559-68, 2004.)
"buffer" that has to be set up to prevent the interro- 10. Efroni S, Harel D, Cohen LR. Toward rigorous com-
gating team from knowing the difference simply by prehension of biological complexity: Modeling, execu-
peripheral things like sight and smell or the time dif- tion, and visualization of thymic T-cell maturation.
ference between a computerized model answering a Genome Research 2003;13:2485-97.
query and a lab experiment set up to do the same. 11. Kam N, Harel D, Kugler H, Marelly R, Pnueli A,
Of course, this test is perhaps too wild and far- Hubbard E.J.A, Stern M.J. “Formal Modeling of C.
fetched, almost imaginary, but it deserves discussion elegans Development: A Scenario-Based Approach”.
because it does try, just like Turing's original test for Proc. 1st. Int. Workshop on Computational Methods
computerized intelligence14 to put an upper bound on in Systems Biology (ICMSB 2003), Lecture Notes in

336 Med clin exp • vol 28, n 0 6, décembre 2005

 Realistic modeling

Computer Science, Vol. 2602, Springer-Verlag, pp.

4–20, Feb. 2003. (Revised version in Modeling in
Molecular Biology (G. Ciobanu and G. Rozenberg,
eds.), Springer, Berlin, 2004, pp. 151–173.)
12. Fisher J, Piterman N, Hubbard EJA, Stern MJ, Harel
D. Computational insights into Caenorhabditis elegans
vulval development. Proceedings of the National
Academy of Sciences of the United States of America
2005 Feb 8;102(6):1951-6.
13.Harel D. A Turing-like test for biological modeling.
Nature Biotechnology 2005;23:495-6.
14. Turing AM. Computing Machinery and Intelligence.
Mind 1950;59:433-60.

© 2005 CIM Clin Invest Med • Vol 28, no 6, December 2005 337