Distilling Free-Form Natural Laws from Experimental Data

Michael Schmidt and Hod Lipson

Science 324, 81 (2009);
DOI: 10.1126/science.1165893

This copy is for your personal, non-commercial use only.

If you wish to distribute this article to others, you can order high-quality copies for your

Downloaded from www.sciencemag.org on September 28, 2012

colleagues, clients, or customers by clicking here.
Permission to republish or repurpose articles or portions of articles can be obtained by
following the guidelines here.

The following resources related to this article are available online at

www.sciencemag.org (this information is current as of September 28, 2012 ):

Updated information and services, including high-resolution figures, can be found in the online
version of this article at:
http://www.sciencemag.org/content/324/5923/81.full.html
Supporting Online Material can be found at:
http://www.sciencemag.org/content/suppl/2009/04/02/324.5923.81.DC1.html
A list of selected additional articles on the Science Web sites related to this article can be
found at:
http://www.sciencemag.org/content/324/5923/81.full.html#related
This article has been cited by 8 article(s) on the ISI Web of Science
This article has been cited by 7 articles hosted by HighWire Press; see:
http://www.sciencemag.org/content/324/5923/81.full.html#related-urls
This article appears in the following subject collections:
Computers, Mathematics
http://www.sciencemag.org/cgi/collection/comp_math

Science (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by the
American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. Copyright
2009 by the American Association for the Advancement of Science; all rights reserved. The title Science is a
registered trademark of AAAS.
 REPORTS

Distilling Free-Form Natural Laws section S4 in the supporting online material

(SOM)]. Unlike traditional linear and nonlinear

from Experimental Data

regression methods that fit parameters to an
equation of a given form, symbolic regression
searches both the parameters and the form of
Michael Schmidt1 and Hod Lipson2,3* equations simultaneously (see SOM section S6).
Initial expressions are formed by randomly com-
For centuries, scientists have attempted to identify and document analytical laws that underlie bining mathematical building blocks such as
physical phenomena in nature. Despite the prevalence of computing power, the process of finding algebraic operators {+, –, ÷, ×}, analytical
natural laws and their corresponding equations has resisted automation. A key challenge to finding functions (for example, sine and cosine), con-
analytic relations automatically is defining algorithmically what makes a correlation in observed stants, and state variables. New equations are
data important and insightful. We propose a principle for the identification of nontriviality. We formed by recombining previous equations and
demonstrated this approach by automatically searching motion-tracking data captured from various probabilistically varying their subexpressions.
physical systems, ranging from simple harmonic oscillators to chaotic double-pendula. Without any The algorithm retains equations that model the
prior knowledge about physics, kinematics, or geometry, the algorithm discovered Hamiltonians, experimental data better than others and aban-
Lagrangians, and other laws of geometric and momentum conservation. The discovery rate dons unpromising solutions. After equations reach
accelerated as laws found for simpler systems were used to bootstrap explanations for more a desired level of accuracy, the algorithm termi-
nates, returning a set of equations that are most

Downloaded from www.sciencemag.org on September 28, 2012

complex systems, gradually uncovering the “alphabet” used to describe those systems.
likely to correspond to the intrinsic mechanisms
underlying the observed system.
athematical symmetries and invariants no connection to the natural physics or dynamics Although symbolic regression is typically

M underlie nearly all physical laws in na-

ture (1), suggesting that the search for
many natural laws is inseparably a search for con-
of the system. We introduce a principle for iden-
tifying only the useful analytical relations that are
related to the system dynamics. We then dem-
used to find explicit (12–14) and differential
equations (15), this method cannot readily find
conservation laws or invariant equations. Rather
served quantities and invariant equations (2, 3). onstrate how a search algorithm based on this than trying to model a specific signal, we are
Automated techniques for generating, collecting, principle identifies meaningful analytical links trying to detect any underlying physical law that
and storing data from scientific measurements in data captured from various physical systems the system obeys, which may or may not be
have become increasingly precise and powerful, (Fig. 1). constant (e.g., a Lagrangian).
but automated processes for distilling this data into Our goal is to find natural relations where A particular challenge is requiring the law to
knowledge in the form of analytical natural laws they exist, with minimal restrictions on their be a function of the system’s state while avoiding
have not kept pace. Thus, there is a pressing prac- analytical form (i.e., free-form). Many methods trivial or meaningless relations. For any system
tical need (4, 5) for improved forms of scientific exist for modeling scientific data: Some use over the state space x, there are infinitely many
data mining (6, 7). fixed-form parametric models derived from ex- trivial equations over x that satisfy a conserved
The most prohibitive obstacle to overcome in pert knowledge, and others use numerical models quantity, such as sin2(x1) + cos2(x1) or x1 + 4.56 –
order to search for conservation laws computa- (such as neural networks) aimed at prediction. x2x1/x2. Additionally, there are infinitely many
tionally is finding meaningful and nontrivial Still others have explored restricted model spaces arbitrarily close trivial conservations, such as
invariants. There exist an infinite number of using greedy monomial search (8, 9). Alterna- 4.56 + 1/(100 + x12). To distinguish good con-
identities that are numerically invariant but have tively, we seek the principal unconstrained servation law candidates from poor ones, we
analytical expression that explains symbolically need a more robust principle than simply invar-
1
Computational Biology, Cornell University, Ithaca, NY 14853, precise conserved relations, thus helping distill iance alone.
USA. 2School of Mechanical and Aerospace Engineering, data into scientific knowledge. The identification of nontrivial relations is a
Cornell University, Ithaca, NY 14853, USA. 3Computing and
Information Science, Cornell University, Ithaca, NY 14853, Symbolic regression (10) is an established major challenge, even for human scientists: Many
USA. method based on evolutionary computation (11) published invariant quantities have turned out to
*To whom correspondence should be addressed. E-mail: for searching the space of mathematical expres- be coincidental (16). The mere appearance of a
[email protected] sions while minimizing various error metrics [see conserved value is insufficient for a conservation

Fig. 1. Mining physical systems. We captured the angles and angular velocities these variables. Without any prior knowledge about physics or geometry, the
of a chaotic double-pendulum (A) over time using motion tracking (B), then we algorithm found the conservation law (C), which turns out to be the double
automatically searched for equations that describe a single natural law relating pendulum’s Hamiltonian. Actual pendulum, data, and results are shown.

www.sciencemag.org SCIENCE VOL 324 3 APRIL 2009 81

REPORTS
law. The key insight into identifying nontrivial We collected data from standard experimental of equations (see SOM section S8). We then
conservation laws computationally is that the systems typically used in undergraduate physics inspect this list manually to select the final equa-
candidate equations should predict connections education: an air-track oscillator and a double tion. Often the list consists of varying approx-
between dynamics of subcomponents of the sys- pendulum (Fig. 3). We used motion-tracking imations or elaborations on a particular law
tem. More precisely, the conservation equation software to record the devices’ positions over equation, but it can contain qualitatively different
should be able to predict connections among de- time. We then numerically calculated velocities equations, as discussed below.
rivatives of groups of variables over time, rela- and accelerations (see SOM section S11). All We experimented on two configurations of the
tions that we can also readily calculate from new data sets are available in SOM section S15. air track: (i) two-spring single-mass and (ii) three-
experimental data. Without any additional information, system spring double-mass. Similarly, we collected time-
One instance of such a metric is the partial models, or theoretical knowledge, the search with series data from a pendulum and a double
derivatives between pairs of variables (see SOM the partial-derivative–pairs criterion produced pendulum (Fig. 3) with the use of motion-tracking
section S1). For example, in a two-dimensional several analytical law expressions directly from (SOM section S12).
system we could measure variables x(t) and y(t) these data. For each system, the algorithm outputs The single-car air track is a harmonic os-
over time. The system’s partial derivatives esti- a short list of ~10 equations that have maximal cillator with slight damping from the air and its
mated from time-series data would then be x′/y′ ≈ accuracy found for different sizes (complexities) two springs. With only minimal noise and damp-
∆x/∆y and y′/x′ ≈ ∆y/∆x (where x′ and y′ represent
the time derivatives of x and y). Similarly, given a
candidate conservation law equation f (x,y), we

Downloaded from www.sciencemag.org on September 28, 2012

can derive the same values through differentia-
tion: (df/dy)/(df/dx) ≈ dx/dy and (df/dx)/(df/dy) ≈
dy/dx. We can now compare ∆x/∆y values from
the experimental data with dx/dy values from a
candidate conservation expression f (x,y) to mea-
sure how well it predicts intrinsic relations in the
system. In higher-dimensional systems, multiple
variable pairings and higher-order derivatives
yield a plethora of criteria to use. See SOM
sections S2 and S3 for generalization to higher-
dimensional systems. Using the partial-derivative
pairs, we define a new type of search criteria for
measuring how well a candidate analytical ex-
pression represents a nontrivial invariance over
the experimental data.
An important consequence of the partial-
derivative–pair measure is that it can also identify
relations that represent other nontrivial identities
of the system beyond invariants and conservation
laws. For example, if the system is confined to a
manifold, the manifold equation can also derive
accurate partial-derivative pairs. Similarly, the
partial-derivative pair can identify equations such
as Lagrangian equations, the energy equivalent to
the equation of motion in classical mechanics,
which summarize the systems dynamics but are
not invariant.
One can control, to an extent, the type of law
that the system might find by choosing what
variables to provide to the algorithm. For
example, if we only provide position coor-
dinates, the algorithm is forced to converge on
a manifold equation of the system’s state space.
If we provide velocities, the algorithm is biased
to find energy laws. If we additionally supply
accelerations, the algorithm is biased to find force
identities and equations of motion. However,
given these or other types of variables, other or
previously unknown analytical laws may exist.
We used an algorithm (Fig. 2) to search for Fig. 2. Computational approach for detecting conservation laws from experimentally collected data. (A)
analytical laws in data captured from several First, calculate partial derivatives between variables from the data, then search for equations that may
synthetic and physical systems using various describe a physical invariance. To measure how well an equation describes an invariance, derive the same
sets of system variables. We present results for partial derivatives symbolically to compare with the data. (B) The representation of a symbolic equation in
a number of physical experimental systems (see computer memory is a list of successive mathematical operations (see SOM section S6). (C) This list
SOM section S7 for a study of synthetic systems, representation corresponds to a graph, where nodes represent mathematical building blocks and leaves
geometric symmetries, and manifolds). A video represent parameters and system variables. Both (B) and (C) correspond to the same equation. The
is available online (see SOM section S14). algorithm varies these structures to search the space of equations.

82 3 APRIL 2009 VOL 324 SCIENCE www.sciencemag.org

 REPORTS
ing, it was the simplest physical system that we data for the pendulum, the algorithm converged ergy terms. By combining the chaotic data with
examined. The double-mass air track consisted of on the equation of a circle, indicating that the low-velocity in-phase oscillation data, the algo-
two coupled harmonic oscillators of different pendulum is confined to a circle. The algorithm rithm converged onto the precise energy laws
masses. There was considerable noise in this data also produced several inexact expressions after several hours of computation.
set as a result of compression of the middle through small-angle approximations—for exam- In the absence of appropriate building blocks,
spring. The pendulum is a nonlinear oscillator ple, using x in place of sin(x) and 1 – x2 in place the algorithm developed approximations. For ex-
that is masked by small-angle approximations. of cos(x) in the pendulum and double-pendulum ample, eliminating the sine and cosine operations
The double pendulum is a coupled nonlinear systems. from the set of equation building blocks caused
oscillator system that exhibits rich dynamics (17) An interesting approximate law for the double the pendulum invariant to be expressed as w2 +
and chaos at certain energies (18), making it pendulum that emerged was conservation of an- k1q2 – k2q4 (where q is the pendulum’s angle, w is
challenging to model (19, 20). Additionally, there gular momentum. Given only data measured the angular velocity, and k1 and k2 are constants),
is higher measurement noise and dampening while the pendulum was chaotic (at high energy), thereby exploiting the 4th-order Taylor series ex-
errors due to higher velocities. the algorithm fixated on this law. The conserva- pansion of the cosine function. Eliminating cosine
Given position and velocity data over time, tion of momentum equation is simpler than other but not sine drove the algorithm to converge on
the algorithm converged on the energy laws of valid laws and is approximately correct for high the equality cos(q) = sin(q + p/2) or more com-
each system (Hamiltonian and Lagrangian equa- velocities where gravity is negligible, as with the plex equivalences (see SOM section S13).
tions). Given acceleration data also, it produced high-energy chaotic data set. Useful scientific theory is both predictive and
the differential equation of motion corresponding Similarly, given only data from low-velocity parsimonious. Similarly, some equations may be

Downloaded from www.sciencemag.org on September 28, 2012

to Newton’s second law for the harmonic oscil- in-phase oscillations, the algorithm fixated on more accurate but overfit the data, whereas others
lator and pendulum systems. Given only position small-angle approximations and uncoupled en- may be more parsimonious but oversimplify

Fig. 3. Summary of laws inferred from experimental data collected from the algorithm detects position manifolds; given velocities, the algorithm detects
physical systems. Depending on the types of variables provided to the energy laws; given accelerations, it detects equations of motion and sum of
algorithm, it detects different types of laws. Given solely position information, forces laws (q, angle; w, angular velocity; a, angular acceleration).

www.sciencemag.org SCIENCE VOL 324 3 APRIL 2009 83

REPORTS
(21, 22); the right balance is difficult to specify these factors (see SOM section S9). For compar- Computational systems such as this could
in advance. Instead of producing a single result, ison, the simulated double-mass air-track and play a role in modeling high-dimensional and
the algorithm produces a small set of final can- simulated double-pendulum data sets (where complex phenomena (23, 24) that currently stress
didate analytical expressions on the accuracy- measurements are noiseless) took ~1/10th of the the reach of expert-driven research. A key chal-
parsimony Pareto front (see SOM section S8), computational effort to analyze. A summary of lenge is scaling to higher complexity. To accom-
which represents the optimal solutions as they performance versus noise level is provided in plish this, scientists leverage knowledge from
vary over equation complexity and the maximum SOM section S9. simpler systems to explain more complex sys-
predictive ability. Parsimony is measured as the Though the algorithm can present equations tems. Can an algorithm do this as well?
inverse of the number of terms in the expression, corresponding to physical laws in their mathe- One method to use prior knowledge is seed-
whereas the predictive accuracy is measured as matical form, we are still faced with the challenge ing the equation search by initializing the
error on withheld experimental data used only for of justifying and giving words to their meaning. algorithm’s initial set of candidate expressions
validation. One difficulty is that we cannot know with cer- with terms from equations from simpler systems.
The Pareto front for the double pendulum tainty the units of bulk constants in the law ex- For example, the single-pendulum and the double-
(Fig. 4A) reveals a few particularly simple pressions (for example, combinations of masses, harmonic oscillator equations provide clues to the
equations that predict the partial-derivative pairs lengths, etc. embodied in the system). Second, laws governing the more complex double pendu-
accurately. Predictive accuracy was measured by the equation may model something that is in- lum. We shuffled terms of the simpler systems
cross-validation with the partial-derivative–pairs herently difficult to observe directly, such as total (for example, exchanging velocity symbols with
criterion (see SOM section S2). The Pareto front energy. Requiring equations to maintain consist- double-pendulum velocity variables) and ran-

Downloaded from www.sciencemag.org on September 28, 2012

tends to contain a cliff where predictive ability ent physical units still leaves room for ambiguity. domized parameters to generate many inexact
jumps rapidly at some minimum complexity. A more systematic approach to parsing the initial expressions. This seeding approach does
Predictive ability then improves only marginally coefficients is to analyze multiple data sets from not constrain the equation search, but it biases the
with more complex equations (Fig. 4A). The the same systems, albeit with different config- search to reuse terms from previous laws.
conservation of angular momentum equation lies urations and parameters. To demonstrate this Bootstrapping the double-pendulum search in
on the Pareto front, though it is inexact. The approach, we used several virtual double pendula this fashion reduced the search time by nearly an
double pendulum’s Hamiltonian lies at the point with randomly chosen masses and lengths to order of magnitude, from 30 to 40 hours of com-
representing the simplest equation with the larg- generate several synthetic data sets. We fit the putation to 7 to 8 hours (Fig. 4B). On the basis of
est increase in predictive ability. In all of our free coefficients of the automatically discovered this result, we conjecture that bootstrapping may
experiments, the solution at this point has been an model to each data set and then invoked the be critical for detecting laws in higher-order
exact theoretical law (see SOM section S7 for equation search algorithm again to seek a relation systems that are veiled in complexity.
additional systems). between the coefficients and parameter sets. A statistical analysis of the subexpression
In the worst case, the time to converge on the After arbitrarily defining k1 = 1, the algorithm frequency and complexity across populations of
law equations depends exponentially on the identified that k2 = m2L22/(m1L12 + m2L12), k3 = various physical systems revealed that the terms
complexity of the law expression itself and 2m2L2/(m1L1 + m2L1), k3 = 19.6/L1, and k4 = that are both frequently used and complex tend to
roughly quadratically on the system dimension- 19.6m2L2/(m2L12 + m1L12), where 19.6 is the be more physically meaningful, such as trigono-
ality (the number of variable pairings) (Fig. 4B). only absolute constant (over all parameter metric terms representing potential energies,
The algorithm’s search is readily parallelizable, variations) whose units are necessarily meters squared velocities representing kinetic energies,
as many candidate functions need to be evaluated per square seconds (see SOM section S5). In the or linear force combinations (see SOM section
simultaneously. In a 32-core implementation, the above expressions, m is mass and L is length. A S10). These terms may make up an “emergent
time required ranged from a few minutes for the similar approach can be used to identify coef- alphabet” for describing a range of systems,
harmonic oscillator to 30 hours for the double ficients that vary slowly over time (for example, which could accelerate their modeling and sim-
pendulum. The impact of noise also couples with because of damping, creeping, or ecological drift). plify their conceptual understanding.

Fig. 4. Parsimony versus accuracy and computation time. (A) Pareto front time required to detect different physical laws for several systems. The
(solid black curve) for physical laws of the double pendulum and the computation time increases with the dimensionality, law equation complexity,
frequency of sampling during the law equation search (grayscale). The and noise. A notable exception is the bootstrapped double pendulum, where
equation at the cliff corresponds to the exact energy conservation law of reuse of terms from simpler systems helped reduce computational cost by
the double pendulum (highlighted in the figure). A second momentum almost an order of magnitude, suggesting a mechanism for scaling higher
conservation law that we encountered is also highlighted. (B) Computation complexities.

84 3 APRIL 2009 VOL 324 SCIENCE www.sciencemag.org

 REPORTS
We have demonstrated the discovery of 3. J. Hanc, S. Tuleja, M. Hancova, Am. J. Phys. 72, 428 19. Y. Liang, B. Feeny, Nonlinear Dyn. 52, 181 (2008).
physical laws, from scratch, directly from ex- (2004). 20. M. Mor, A. Wolf, O. Gottlieb, in Proceedings of the 21st
4. D. Clery, D. Voss, Science 308, 809 (2005). ASME Biennial Conference on Mechanical Vibration and
perimentally captured data with the use of a 5. A. Szalay, J. Gray, Nature 440, 413 (2006). Noise (ASME Press, Las Vegas, NV, 2007), pp. 1–8.
computational search. We used the presented 6. R. E. Valdés-Pérez, Commun. Assoc. Comput. Mach. 42, 21. P. Gregory, R. Denis, F. Cyril, in Evolution Artificielle,
approach to detect nonlinear energy conservation 37 (1999). 6th International Conference, vol. 2936, L. Pierre,
laws, Newtonian force laws, geometric invari- 7. R. D. King et al., Nature 427, 247 (2004). C. Pierre, F. Cyril, L. Evelyne, S. Marc, Eds. (Springer,
8. P. Langley, Cogn. Sci. 5, 31 (1981). Marseilles, France, 2003), pp. 267–277.
ants, and system manifolds in various synthetic 9. R. M. Jones, P. Langley, Comput. Intell. 21, 480 22. E. D. De Jong, J. B. Pollack, in Genetic Programming and
and physically implemented systems without (2005). Evolvable Machines, vol. 4 (Springer, Berlin, 2003),
prior knowledge about physics, kinematics, or 10. J. R. Koza, Genetic Programming: On the Programming of pp. 211–233.
geometry. The concise analytical expressions that Computers by Means of Natural Selection. (MIT Press, 23. S. H. Strogatz, Nature 410, 268 (2001).
Cambridge, MA, 1992). 24. P. A. Marquet, Nature 418, 723 (2002).
we found are amenable to human interpretation 11. S. Forrest, Science 261, 872 (1993). 25. This research was supported in part by Integrative Graduate
and help to reveal the physics underlying the 12. J. Duffy, J. Engle-Warnick, Evolutionary Computation in Education and Research Traineeship program in nonlinear
observed phenomenon. Many applications exist Economics and Finance 100, 61 (2002). systems, a U.S. NSF graduate research fellowship, and NSF
for this approach, in fields ranging from systems 13. F. Cyril, B. Alberto, in 2007 IEEE Congress on Creative-IT grant 0757478 and CAREER grant 0547376.
Evolutionary Computation, S. Dipti, W. Lipo, Eds. We thank M. Kurman for editorial consultation and
biology to cosmology, where theoretical gaps
(IEEE Press, Singapore, 2007), pp. 23–30. substantive editing of the manuscript.
exist despite abundance in data. 14. B. Elena, B. Andrei, L. Henri, in Seventh International
Might this process diminish the role of future Symposium on Symbolic and Numeric Algorithms for Supporting Online Material
Scientific Computing (SYNASC '05) (IEEE Press, 2005), www.sciencemag.org/cgi/content/full/324/5923/81/DC1
scientists? Quite the contrary: Scientists may use

Downloaded from www.sciencemag.org on September 28, 2012

Materials and Methods
processes such as this to help focus on interesting pp. 321–324.
15. J. Bongard, H. Lipson, Proc. Natl. Acad. Sci. U.S.A. 104, SOM Text
phenomena more rapidly and to interpret their 9943 (2007). Figs. S1 to S7
meaning. Tables S1 to S3
16. S. Nee, N. Colegrave, S. A. West, A. Grafen, Science 309,
1236 (2005). References
Movie S1
References and Notes 17. P. Jäckel, T. Mullin, Proc. R. Soc. London Ser. A 454,
Data Sets S1 to S15
1. P. W. Anderson, Science 177, 393 (1972). 3257 (1998).
2. E. Noether, Nachr. d. König Gesellsch. d. Wiss. zu 18. T. Shinbrot, C. Grebogi, J. Wisdom, J. A. Yorke, 15 September 2008; accepted 19 February 2009
Göttingen, Math-Phys. Klasse 235 (1918). Am. J. Phys. 60, 491 (1992). 10.1126/science.1165893

The Automation of Science experiments can be executed, each individual ex-

periment cannot be designed to test a hypothesis
about a model. Robot scientists have the potential
Ross D. King,1* Jem Rowland,1 Stephen G. Oliver,2 Michael Young,3 Wayne Aubrey,1 to overcome this fundamental limitation.
Emma Byrne,1 Maria Liakata,1 Magdalena Markham,1 Pınar Pir,2 Larisa N. Soldatova,1 The complexity of biological systems neces-
Andrew Sparkes,1 Kenneth E. Whelan,1 Amanda Clare1 sitates the recording of experimental metadata in
as much detail as possible. Acquiring these meta-
The basis of science is the hypothetico-deductive method and the recording of experiments in data has often proved problematic. With robot
sufficient detail to enable reproducibility. We report the development of Robot Scientist “Adam,” scientists, comprehensive metadata are produced
which advances the automation of both. Adam has autonomously generated functional genomics as a natural by-product of the way they work.
hypotheses about the yeast Saccharomyces cerevisiae and experimentally tested these hypotheses Because the experiments are conceived and ex-
by using laboratory automation. We have confirmed Adam’s conclusions through manual ecuted automatically by computer, it is possible
experiments. To describe Adam’s research, we have developed an ontology and logical language. to completely capture and digitally curate all as-
The resulting formalization involves over 10,000 different research units in a nested treelike pects of the scientific process (11, 12).
structure, 10 levels deep, that relates the 6.6 million biomass measurements to their logical To demonstrate that the robot scientist meth-
description. This formalization describes how a machine contributed to scientific knowledge. odology can be both automated and be made
effective enough to contribute to scientific knowl-
omputers are playing an ever-greater role advantages of logic, most scientific knowledge is edge, we have developed Robot Scientist “Adam”

C in the scientific process (1). Their use to

control the execution of experiments con-
tributes to a vast expansion in the production of
expressed only in natural languages. This is now
changing through developments such as the
Semantic Web (4) and ontologies (5).
(13) (Fig. 1). Adam’s hardware is fully automated
such that it only requires a technician to period-
ically add laboratory consumables and to remove
scientific data (2). This growth in scientific data, A natural extension of the trend to ever-greater waste. It is designed to automate the high-
in turn, requires the increased use of computers computer involvement in science is the concept of throughput execution of individually designed
for analysis and modeling. The use of computers a robot scientist (6). This is a physically imple- microbial batch growth experiments in micro-
is also changing the way that science is described mented laboratory automation system that exploits titer plates (14). Adam measures growth curves
and reported. Scientific knowledge is best ex- techniques from the field of artificial intelligence (phenotypes) of selected microbial strains (geno-
pressed in formal logical languages (3). Only (7–9) to execute cycles of scientific experimenta- types) growing in defined media (environments).
formal languages provide sufficient semantic tion. A robot scientist automatically originates Growth of cell cultures can be easily measured in
clarity to ensure reproducibility and the free hypotheses to explain observations, devises exper- high-throughput, and growth curves are sensitive
exchange of scientific knowledge. Despite the iments to test these hypotheses, physically runs the to changes in genotype and environment.
experiments by using laboratory robotics, inter- We applied Adam to the identification of
1
Department of Computer Science, Aberystwyth University, prets the results, and then repeats the cycle. genes encoding orphan enzymes in Saccharomy-
SY23 3DB, UK. 2Cambridge Systems Biology Centre, Depart- High-throughput laboratory automation is trans- ces cerevisiae: enzymes catalyzing biochemical
ment of Biochemistry, University of Cambridge, Sanger forming biology and revealing vast amounts of reactions thought to occur in yeast, but for which
Building, 80 Tennis Court Road, Cambridge CB2 1GA, UK.
3
Institute of Biological, Environmental and Rural Sciences, new scientific knowledge (10). Nevertheless, ex- the encoding gene(s) are not known (15). To set
Aberystwyth University, SY23 3DD, UK. isting high-throughput methods are currently in- up Adam for this application required (i) a
*To whom correspondence should be addressed. E-mail: adequate for areas such as systems biology. This comprehensive logical model encoding knowl-
[email protected] is because, even though very large numbers of edge of S. cerevisiae metabolism [~1200 open

www.sciencemag.org SCIENCE VOL 324 3 APRIL 2009 85