DOI: 10.1126/science.1218919, 1212 (2012);337 Science

et al.C. C. IoannouPreyPredatory Fish Select for Coordinated Collective Motion in Virtual

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

clicking here.colleagues, clients, or customers by , you can order high-quality copies for yourIf you wish to distribute this article to others

here.following the guidelines

can be obtained byPermission to republish or repurpose articles or portions of articles

): January 14, 2013 www.sciencemag.org (this information is current as of

The following resources related to this article are available online at

http://www.sciencemag.org/content/337/6099/1212.full.htmlversion of this article at:

including high-resolution figures, can be found in the onlineUpdated information and services,

http://www.sciencemag.org/content/suppl/2012/08/15/science.1218919.DC1.html can be found at: Supporting Online Material

http://www.sciencemag.org/content/337/6099/1212.full.html#relatedfound at:

can berelated to this article A list of selected additional articles on the Science Web sites

http://www.sciencemag.org/content/337/6099/1212.full.html#ref-list-1, 16 of which can be accessed free:cites 34 articlesThis article

http://www.sciencemag.org/content/337/6099/1212.full.html#related-urls1 articles hosted by HighWire Press; see:cited by This article has been

http://www.sciencemag.org/cgi/collection/evolutionEvolution

subject collections:This article appears in the following

registered trademark of AAAS. is aScience2012 by the American Association for the Advancement of Science; all rights reserved. The title

CopyrightAmerican Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by theScience

on

Janu

ary

14, 2

013

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fro

m

appreciable extent. The condensable nitrates that weobserve in the particle phase are likely second- orhigher-generation oxidation products, producedby the slower oxidation of the first-generationproducts (15, 16). Based on the measurements ofRH, and aerosol surface area and composition,we estimate that N2O5 heterogeneous loss has asmall impact on NO3 concentration (<10%), andthus NO3 variability is dominated by its sourceterm (reaction 1) and gas-phase reactivity. Figure2 also shows that the kinetics of aerosol RONO2

formation are approximately linear with PNO3,indicating that aerosol precursors are abundantand that NO3 production is rate limiting. Becausethis SOA is produced by reactions of NO3, it canbe considered anthropogenic. Although the car-bon may be of biogenic origin, without high NOx

emissions it would not be produced.The observation that VOC with high SOA

yields may suppress SOA formation is surpris-ing. To demonstrate that this is kinetically possi-ble in theNOx/VOC regime observed inBakersfield,we modeled SOA formation from NO3 oxidationof limonene (Fig. 3). We use limonene as an ex-ample VOC because of its relatively high concen-trations in Bakersfield and its high SOAyield, andbecause we have some knowledge of the kineticsof its oxidation products (16). Details of the boxmodel used are included in the materials andmethods SI. We find that because the second-generation products have SOAyields ~2.5 timesas large as those of the first-generation productsand that high concentrations of limonene inhibitthe formation of these less-volatile products, SOAproduction slows in the high-limonene regime.At the same time, given sufficient O3, increasesin NO2 always lead to more SOA owing to thehigher NO3 production rate.

Our findings suggest that SOA formation vianighttime nitrate radical chemistry in Bakersfield isa large PM source, which frequently results in thedaily maximumOA concentration during the sum-

mer. The high concentrations of NO2 and O3 atnight resulted in very high NO3 production rates[frequently greater than 1 part per billion (ppb)hour!1]. Nevertheless, concentrations of reactiveBVOCs were frequently high enough that pSANformation was inhibited, suggesting that the pSANprecursors are less reactive than the primary VOCsand have a somewhat reduced volatility. A goodcorrelation between production rates of NO3 andpSAN was observed, suggesting that the targetedreductions in NOx at this location should reduceOAmass. Although attributing sources of daytimeSOA as biogenic or anthropogenic remains chal-lenging, our results show that pSANs are a largefraction of nighttime growth and likely a resultof NO3 chemistry. That this SOA would not beproduced in the absence of NOx makes nighttimepSANs a clear tracer for anthropogenically con-trolled SOA, regardless of the carbon source.

References and Notes1. Q. Zhang et al., Geophys. Res. Lett. 34, L13801 (2007A).2. J. L. Jimenez et al., Science 326, 1525 (2009).3. C. L. Heald, D. A. Ridley, S. M. Kreidenweis, E. E. Drury,

Geophys. Res. Lett. 37, L24808 (2010).4. J. de Gouw, J. L. Jimenez, Environ. Sci. Technol. 43, 7614

(2009).5. C. R. Hoyle et al., Atmos. Chem. Phys. 11, 321 (2011).6. Q. Zhang et al., J. Geophys. Res. 112, D22306 (2007).7. U.S. Environmental Protection Agency, Our Nation’s

Air—Status and Trends through 2008 (Washington, DC,2010).

8. D. C. Carslaw et al., “Trends in NOx and NO2 emissionsand ambient measurements in the UK.” Version: July 2011.

9. B. W. LaFranchi, A. H. Goldstein, R. C. Cohen,Atmos. Chem. Phys. 11, 6945 (2011).

10. M. Hallquist et al., Atmos. Chem. Phys. 9, 5155 (2009).11. J. H. Kroll, J. H. Seinfeld, Atmos. Environ. 42, 3593 (2008).12. L. Hildebrandt et al., Geophys. Res. Lett. 37, L23801

(2010).13. A. M. Winer, R. Atkinson, J. N. Pitts Jr., Science 224,

156 (1984).14. N. L. Ng et al., Atmos. Chem. Phys. 8, 4117 (2008).15. A. W. Rollins et al., Atmos. Chem. Phys. 9, 6685 (2009).16. J. L. Fry et al., Atmos. Chem. Phys. 11, 3879 (2011).17. A. W. Rollins, J. D. Smith, K. R. Wilson, R. C. Cohen,

Environ. Sci. Technol. 44, 5540 (2010).18. L. Bianco, I. V. Djalalova, C. W. King, J. M. Wilczak,

Boundary-Layer Meteorol. 140, 491 (2011).19. B. J. Williams et al., Atmos. Chem. Phys. 10, 11577

(2010).20. A. C. Aiken et al., Atmos. Chem. Phys. 9, 6633 (2009).21. C. J. Hennigan, M. H. Bergin, A. G. Russell, A. Nenes,

R. J. Weber, Atmos. Chem. Phys. 9, 3613 (2009).22. J. D. Surratt et al., J. Phys. Chem. A 112, 8345 (2008).23. M. Hallquist, I. Wängberg, E. Ljungström, I. Barnes,

K.-H. Becker, Environ. Sci. Technol. 33, 553 (1999).24. S. S. Brown et al., Atmos. Chem. Phys. 9, 3027 (2009).

Acknowledgments: This work is supported by the CaliforniaAir Resources Board under grants CARB 08-316 and 09-328.E.C.B. was supported by NASA ESSF fellowship.

Supplementary Materialswww.sciencemag.org/cgi/content/full/337/6099/1210/DC1Materials and MethodsFigs. S1 to S4Table S1References

6 March 2012; accepted 24 July 201210.1126/science.1221520

Predatory Fish Select for CoordinatedCollective Motion in Virtual PreyC. C. Ioannou,1,2* V. Guttal,1,3 I. D. Couzin1*

Movement in animal groups is highly varied and ranges from seemingly disordered motionin swarms to coordinated aligned motion in flocks and schools. These social interactions areoften thought to reduce risk from predators, despite a lack of direct evidence. We investigatedrisk-related selection for collective motion by allowing real predators (bluegill sunfish) to huntmobile virtual prey. By fusing simulated and real animal behavior, we isolated predator effectswhile controlling for confounding factors. Prey with a tendency to be attracted toward, and toalign direction of travel with, near neighbors tended to form mobile coordinated groups andwere rarely attacked. These results demonstrate that collective motion could evolve as a responseto predation, without prey being able to detect and respond to predators.

From herding ungulates to shoaling fish,nesting birds, and swarming crickets, an-imals living in groups are generally less at

risk from predators (1). Mechanisms include the

ability of groups to detect predators sooner andfrom a greater distance (the “many eyes” effect)and cognitive confusion of the predator, caused byhaving to choose among many possible targets

Fig. 3. Simulation ofmul-tigenerational SOA for-mation from the reactionof NO3 with limonene asa function of NO2 and lim-onene at 50 ppb O3. Weassume that Bakersfield(1 to 2 hours upwind) isthemajor NOx source andtherefore show contoursthat are ppb of pSAN af-ter 2-hourmodel runs. Forlonger runs (up to5hours),the pSAN scaled approxi-mately linearly with time.Theproduction ratesofNO3corresponding to the NO2concentration are shownon the right axis. Top axis shows the total NO3 gas-phase lifetime with limonene at 34% of the total NO3loss. Red dashed box highlights the NO2 and limonene concentration range typically observed inBakersfield, showing that increases in limonene here are expected to lead to less aerosol production. pptv(ppbv), parts per trillion (billion) by volume.

7 SEPTEMBER 2012 VOL 337 SCIENCE www.sciencemag.org1212

REPORTS

on

Janu

ary

14, 2

013

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fro

m

(1). Studies have focused on the costs and ben-efits of group size and position within the group(2–4), but how predation risk varies with theresponse of individuals to their neighbors is notwell understood because of difficulties in itsmeasurement (5, 6) and manipulation (7). Onesuch behavioral response that is common in na-ture is the tendency for individuals to align theirdirection of travel with that of near neighbors,forming coordinated “polarized” groups. It is of-ten assumed that coordination between preymakesthem harder to catch by enhancing informationtransfer between individuals (5, 6) or by increas-ing the confusion effect, although little experi-mental work supports this idea (7, 8).

How animals move together has been simu-lated by agent-based models with generic behav-ioral tendencies: repulsion when neighbors aretoo close and otherwise aligning with, and/orbeing attracted to, neighbors. These models areable to recreate groupmovement such as swarms,in which attraction dominates alignment tenden-cy, and coordinated polarized flocks or schools,in which alignment tendency becomes relativelystrong (9–13). Although studies link this mech-anistic approach back to functional explanations(14–16), there are few that have explicitly dealtwith the relationship between predation and dy-namic group behavior. Models can recreate themacroscopic responses of prey when under threat(17) and the evolution of swarmlike or highlypolarized, coordinated groups as a direct result ofsimulated predator behavior (18). In many cases,however, it is unclear which properties of preybehavior are being selected for and why, and itremains to be established whether any real-lifepredators select for coordinated motion.

To explore these issues, we investigated howa predatory fish, the bluegill sunfish (Lepomismacrochirus), hunts simulated prey. Bluegills aregeneralist predators whose body plan is special-ized for hunting in complex vegetated environ-ments (19). Our system allows us to isolate thespecific selection pressure of predation risk fromthe multitude of factors influencing the evolutionof any trait, such as the associated costs of thetrait and taxonomic constraints (20). In naturethere is a wide variety of prey responses to pred-ators; some react to predators as they approach(2), whereas others only respond once an attackis made (21), and these responses are furtherinfluenced by factors such as prey group size(1, 21). Because our simulated prey could notrespond to the fish, we analyzed only the firstattack from each fish, which is analogous to sit-uations where risk is determined primarily by the

first attack (22). A simulation of animal move-ment (23) consisting of a behaviorally heteroge-neous population of 16 prey was projected ontoa translucent screen on an inner side of the testtank (Fig. 1 and fig. S1). Individual prey be-havior was encoded by three traits: the strengthof their behavioral tendencies to be attractedtoward (wai), orientate direction of travel with(woi), or ignore (wpi) near neighbors [the threetraits were normalized so that their sum was1 (12, 24, 25) (table S1)]. Depending on thesetraits, prey exhibited a range of movement be-haviors, including solitary randomwalk, formationand maintenance of aggregations, and coordi-nated polarized motion (movie S1).

The predators exerted a strong selection pres-sure on the virtual population, with some preybeing attacked often while others were never at-tacked (Fig. 2A). The risk of being targeted wasminimized for individuals with characteristics thatbalanced both attraction and orientation, resultingin an interaction between these two parameters[generalized linear model (GLM): likelihood ra-tio test (LRT1,12) = 11.43, P = 0.00072]. This sug-gests that to most effectively avoid predation,prey should both move toward and align withtheir near neighbors, behavior that generatesmov-ing groups of coordinated individuals (12).

To a large degree, the strength of attractionmediates the group size for the simulated prey;when attraction is zero, prey are most frequentlyfound alone (Fig. 2B). Although increasing theorientation parameter has only a minor effect ongroup size, it increases the straightness (i.e., de-creases the “curvedness” or tortuosity) of theprey’s path substantially when that prey is in agroup (Fig. 2C). For example, although the preytype with wai = 0.2 and woi = 0.6 is rarely sol-itary, it tends to exhibit relatively directed, low-tortuosity motion (Fig. 2C), similar to the preytype that has no social tendency (wai = 0.0 andwoi = 0.0) and is rarely found in a group. Fur-thermore, the tortuosity of a prey individual’s pathtypically scales negatively with their group’s po-larization (the directional coherence among groupmembers), so that nonsolitary individuals withlow-tortuosity paths are typically in groups withhigh polarization (fig. S2).

To explore how these behaviors mediated theeffects of both attraction and orientation on risk,we could not simply correlate the mean groupsize or tortuosity of a prey type with the numberof attacks it received, because any choice madeby the predator is constrained by the other preyphenotypes present at the time of attack. Instead,we created a null predator that chose a prey ran-domly at the same time steps as the observedattacks and compared the relationship betweenthe targets’ tortuosity and the group size expectedfrom random targeting to that actually observedfrom the real predatory fish (23). Compared torandom targeting, the fish disproportionately tar-geted prey in smaller groups, and this was stron-gest when prey were also taking less-tortuouspaths (figs. S3 and S4). Prey in groups with a

coordinated direction of motion (i.e., with highpolarization) (12) were at less risk than theircounterparts in unpolarized swarms (fig. S3).

Bluegill sunfish employ a characteristic “hover-ing” behavior during foraging (Fig. 1B) (19),allowing us to approximate the time taken tomake each targeting decision. Consistent witha confusion effect (26), this decision time in-creased with the prey target’s group size (GLM:LRT1,67 = 11.32, P = 0.00077). Although this ac-counts for the targeting of prey in smaller groups,there was no evidence that a prey’s tortuosity,either alone or as part of an interaction with groupsize, had any additional effect [LRT1,67 = 0.029,P = 0.86; and LRT1,66 = 0.22, P = 0.64, respec-tively; see also (7, 8)].

In response to the confusion effect (27), orsimply because they are nearer on average (3, 4),predators will often attack prey at the edge ofgroups. To test whether prey relatively far fromthe group center were attacked disproportion-ately, we used the null predator procedure de-scribed previously. The analysis indicates thatthis is indeed the case, but only for prey movingwith relatively low tortuosity (fig. S5A). Althougha number of the behavioral types often moved onpaths with low tortuosity (Fig. 2C), prey withhigh orientation relative to attraction tended to befound more often at the edges of groups (fig. S5,C and D). This “self-assortment” (12) should con-tribute to selection against such prey types, inaddition to their tendency to be solitary (Fig. 2B).However, the edge effect cannot explain why po-larized groups were disproportionately less at risk(fig. S3) and why prey types with no tendency

1Department of Ecology and Evolutionary Biology, PrincetonUniversity, Princeton, NJ 08544, USA. 2School of BiologicalSciences, University of Bristol, Woodland Road, Bristol BS81UG, UK. 3Centre for Ecological Sciences, Indian Institute ofScience, Bangalore 560012, India.

*To whom correspondence should be addressed. E-mail:C.C.Ioannou@bristol.ac.uk (C.C.I.); icouzin@princeton.edu(I.D.C.)

Fig. 1. The experimental system. (A) The simu-lation was projected (green arrow) onto a screen onthe opposite side of the test tank to that where thefish was released. (B) Each attack was preceded bythe fish hovering in front of the prey (white dots)(left inset) before accelerating toward a prey, open-ing the mouth and gill flaps (right inset). t, timestep in the simulation.

www.sciencemag.org SCIENCE VOL 337 7 SEPTEMBER 2012 1213

REPORTS

on

Janu

ary

14, 2

013

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fro

m

to orient with their neighbors (woi = 0) wereselected against (Fig. 2A).

In complex habitats, such as the littoral zoneof lakes where bluegill sunfish are found, bothpredators and prey often exploit boundaries (2).The boundaries of the projection presented to thepredators were periodic; i.e., when prey came incontact with a boundary, they would reappearat the opposite boundary with the same velocity.This ensured that the prey types were found withequal probability anywhere in the projected arenaand excluded possible spatial artifacts confound-ing our results, such as swarming prey being foundin the corners (23). Thus, if there is any tendencyfor prey with certain characteristics to be attackedin particular locations in the projected area, thismust be due to the fish’s behavior. We found thattargeted prey tended to be in larger groups, and tohave more tortuous paths, when they were nearerthe top and bottom of the projection (Fig. 3; thevertical axis polynomial effect: GLM: LRT2,65 =13.07, P = 0.0015; and LRT2,65 = 14.35, P =0.00077, respectively).When hunting near the edge,predators experience groups that partly, or com-pletely, cross the boundary (from the localizedperception of the predator, this is analogous toprey occlusion by a physical structure in the en-vironment). Individuals in groups with high tor-tuosity, and thus low net movement, persist forlonger in such semi-occluded states and wereparticularly at risk. In direct contrast, prey in po-larized groups were attacked less often, probablybecause their more directed movement gave lesstime for targeting to occur when they were withinthe high-risk area of the boundary. Further anal-ysis of the pattern of risk seen in Fig. 2A dem-onstrates that although individuals in swarms arevulnerable at the boundaries, further from theboundary this effect weakens (fig. S6).

To ensure that our results were not sensitive tothe distribution of prey types employed, and todemonstrate the selection of virtual prey strat-egies by real predators, we used the risk land-scape in Fig. 2A to “evolve” our prey (23). Wethen presented either this evolved population orthe original population to the fish in a secondexperiment (23). Although selection changed thefrequencies of different prey types, and hence thegroup sizes and tortuosities in the projected sim-ulation (table S2 and fig. S7), there was no evi-dence that the pattern observed in Fig. 2A changedbetween the pre- and postselection populations(population ! attraction ! orientation GLM:LRT1,18 = 0.09, P = 0.76). Selection for orien-tation and attraction thus generalizes beyond aneven distribution of behavioral parameters andappears relatively robust to frequency-dependenteffects. Neither was there any evidence that se-lection had a detrimental effect on the fish’s pred-atory behavior (table S3).

Our results show that predation risk is re-duced among prey that exhibit both attractionand orientation under the conditions of our ex-periment, through an interaction between theconfusion effect and the ability of prey to formcoherent mobile groups. This is dependent onhabitat properties that the predators exploit tofacilitate targeting opportunities, without the ne-cessity for prey individuals to react dynamical-ly to the predator’s presence, position, and/orattack (6, 18). The degree of control affordedby virtual prey populations, as developed here,could allow a closed feedback loop betweenpredator attack and prey response to explore suchproperties. This may reveal further dependen-cies and synergies between anti-predatory adap-tations, predator hunting strategies, and habitatvariables (2, 19, 27).

Fig. 2. Risk associated with attraction and orientation traits. (A) Based ontheir particular attraction and orientation tendencies (black diamonds), thenumber of attacks from a total of 70 that each prey type received (proportionalto the area of the open circles; the largest of these is 14 attacks). The coloredgradient represents the fitted values of this relationship from the fully factorialmodel, where red indicates more attacks and white fewer attacks (see color

scale, inset). (B and C) Frequency distributions of each prey type’s group sizeand tortuosity, respectively, pooled from the simulations presented to the fishin experiment 1. The scale of the histograms is shown in empty insets at topright. Group size was determined by the number of interconnected prey (23)every 250 time steps, whereas their individual tortuosity was calculated every100 time steps.

Fig. 3. Preferential targeting of group size (A)and tortuosity (B) as a function of height in thevertical axis of the projection. The curves show fittedvalues from models with the (polynomial) verticalaxis effect only, because the horizontal axis had nosignificant effect. Tortuosity is expressed as a per-centage of the maximum possible value. The indi-vidual tortuosity of the target prey is shown here; asimilar trend was found for the tortuosity of thetarget’s group as a whole (GLM: LRT2,65 = 14.89, P=0.00059; group tortuosity = 1 – group polarization).

7 SEPTEMBER 2012 VOL 337 SCIENCE www.sciencemag.org1214

REPORTS

on

Janu

ary

14, 2

013

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fro

m

References and Notes1. J. Krause, G. D. Ruxton, Living in Groups (Oxford Univ.

Press, Oxford, 2002).2. W. Cresswell, J. L. Quinn, Oikos 104, 71

(2004).3. W. L. Romey, A. R. Walston, P. J. Watt, Behav. Ecol. 19,

74 (2008).4. W. D. Hamilton, J. Theor. Biol. 31, 295 (1971).5. M. Ballerini et al., Proc. Natl. Acad. Sci. U.S.A. 105,

1232 (2008).6. N. O. Handegard et al., Curr. Biol. 22, 1213

(2012).7. G. D. Ruxton, A. L. Jackson, C. R. Tosh, Behav. Ecol. 18,

590 (2007).8. K. A. Jones, A. L. Jackson, G. D. Ruxton, Behav. Ecol. 22,

831 (2011).9. R. Lukeman, Y. X. Li, L. Edelstein-Keshet, Proc. Natl.

Acad. Sci. U.S.A. 107, 12576 (2010).10. Y. Katz, K. Tunstrøm, C. C. Ioannou, C. Huepe,

I. D. Couzin, Proc. Natl. Acad. Sci. U.S.A. 108, 18720(2011).

11. J. Buhl et al., Science 312, 1402 (2006).12. I. D. Couzin, J. Krause, R. James, G. D. Ruxton,

N. R. Franks, J. Theor. Biol. 218, 1 (2002).13. J. K. Parrish, L. Edelstein-Keshet, Science 284, 99

(1999).14. S. Bazazi et al., Curr. Biol. 18, 735 (2008).

15. V. Guttal, I. D. Couzin, Proc. Natl. Acad. Sci. U.S.A. 107,16172 (2010).

16. D. Grünbaum, Evol. Ecol. 12, 503 (1998).17. R. Vabø, L. Nøttestad, Fish. Oceanogr. 6, 155 (1997).18. A. J. Wood, G. J. Ackland, Proc. Biol. Sci. 274, 1637

(2007).19. T. J. Ehlinger, D. S. Wilson, Proc. Natl. Acad. Sci. U.S.A.

85, 1878 (1988).20. H. Sayama, S. Dionne, C. Laramee, D. S. Wilson, in IEEE

Symposium on Artificial Life (IEEE Conference Proceedings,Nashville, TN, 2009), pp. 85–91.

21. J. E. Treherne, W. A. Foster, Anim. Behav. 30, 536(1982).

22. W. Cresswell, Ibis 138, 684 (1996).23. Materials and methods are available as supplementary

materials on Science Online.24. I. D. Couzin, J. Krause, N. R. Franks, S. A. Levin, Nature

433, 513 (2005).25. I. D. Couzin, J. Krause, Adv. Stud. Behav. 32,

1 (2003).26. C. C. Ioannou, C. R. Tosh, L. Neville, J. Krause, Behav. Ecol.

19, 126 (2008).27. C. R. Tosh, J. Theor. Biol. 281, 24 (2011).

Acknowledgments: We thank A. Hundal, M. Jiang, andM. Singh for assistance and A. S. I. Wade, the Couzin

lab, and three anonymous reviewers for comments on themanuscript. Funded by Office of Naval Research awardN00014-09-1-1074, NSF award PHY-0848755, SearleScholar award 08-SPP-201, and Army Research Officegrant W911NG-11-1-0385 (I.D.C.); Defense AdvancedResearch Projects Agency grant HR0011-05-1-0057(Princeton University); Leverhulme Trust Early CareerFellowship (C.C.I.); and a Ramalingaswami Fellowshipfrom the Department of Biotechnology, Government of India(V.G.). All experiments were conducted in accordance withfederal and state regulations and were approved by thePrinceton University Institutional Animal Care and UseCommittee. Data and code are freely available athttp://icouzin.princeton.edu.

Supplementary Materialswww.sciencemag.org/cgi/content/full/science.1218919/DC1Materials and MethodsFigs. S1 to S7Tables S1 to S3References (28–40)Movie S1

9 January 2012; accepted 26 July 2012Published online 16 August 2012;10.1126/science.1218919

Molecular Mechanics ofCardiac Myosin-Binding Protein Cin Native Thick FilamentsM. J. Previs,1 S. Beck Previs,1 J. Gulick,2 J. Robbins,2 D. M. Warshaw1*

The heart’s pumping capacity results from highly regulated interactions of actomyosinmolecular motors. Mutations in the gene for a potential regulator of these motors, cardiacmyosin-binding protein C (cMyBP-C), cause hypertrophic cardiomyopathy. However, cMyBP-C’sability to modulate cardiac contractility is not well understood. Using single-particlefluorescence imaging techniques, transgenic protein expression, proteomics, and modeling,we found that cMyBP-C slowed actomyosin motion generation in native cardiac thick filaments.This mechanical effect was localized to where cMyBP-C resides within the thick filament(i.e., the C-zones) and was modulated by phosphorylation and site-specific proteolyticdegradation. These results provide molecular insight into why cMyBP-C should be considereda member of a tripartite complex with actin and myosin that allows fine tuning of cardiacmuscle contraction.

Cardiac muscle’s pumping capacity isproduced by the sarcomere (Fig. 1A), aparallel array of proteins assembled into

thick filaments, composed of myosin molecu-lar motors that cyclically interact with actin-containing thin filaments, generating force thatpropels the thin filaments past the thick filaments.These actomyosin interactions can be modulatedon a beat-to-beat basis by cardiac myosin-bindingprotein C (cMyBP-C) (Fig. 1B), a 140-kD immuno-globulin (Ig) protein superfamily member (Fig.1C) that is confined to two distinct regions (i.e.,

C-zones) of the thick filament (1, 2) (Fig. 1A).Mutations in the MYBPC3 gene are a leadingcause of familial hypertrophic cardiomyopathy(FHC) (1, 2). Proposed mechanisms of cMyBP-C’s function assume that several Ig-like domainsand their linkers (C0–C2) (Fig. 1C) extend awayfrom the thick filament backbone (Fig. 1B) (3)and reversibly bind to myosin’s motor domainsand/or actin filaments (1, 2), with this bindingtunable by phosphorylation of four serines (S273,S282, S302, and S307) in the motif linker betweendomains C1 and C2 (Fig. 1C) (4, 5). Insight intocMyBP-C’s function and regulation by phos-phorylation has benefited from intact heart andmuscle fiber studies, but these complex prepa-rations make molecular-level interpretations dif-ficult. Isolated protein studies, although simpler,lack the sarcomere’s spatial relation between thethin and thick filaments. Here, we developed an

in vitro sarcomere model system in which singleactin filaments could be visualized moving overnative cardiac thick filaments with and withoutcMyBP-C.

Cardiac thick filaments that retained their na-tive length (~1.6 µm), bipolar structure, and cen-tral bare zone devoid of myosin heads (Fig. 1D)were isolated from mouse hearts by fine-tissuedissection and limited enzyme-induced proteindegradation (0.2 U/µl µ-calpain) (6). Quantita-tive liquid chromatography–mass spectrometry(LC-MS) (6) showed that filaments contained thenormal complement of cMyBP-C (fig. S1). Be-cause cMyBP-C is a target for calpain-mediatedprotein degradation (7), protein immunoblottingwith domain specific antibodies was used to showthat 79 T 4% (SD, n = 3) of the cMyBP-C mol-ecules were intact (fig. S2), and in combinationwith LC-MS analyses, we determined that theremainder of the molecules had 29 kD of theirN terminus removed (i.e., C0–C1 plus 17 aminoacids of the motif, C0C1f) (Fig. 1C) by cleavagebetween amino acids R266 and R271 (fig. S3).

To assess cMyBP-C’s mechanical impacton actin filament sliding, native cardiac thick fila-ments were adhered to a microscope cover slip.Fluorescently labeled actin filaments were thenintroduced onto the cover slip (25 mM KCl,100 µM ATP, 22ºC), and their sliding along thethick filaments tracked (6) with high time (8.3-ms)and spatial (30-nm) resolution (Fig. 2A). The useof short [250 T 9 nm (SEM)] actin filaments (fig.S4) prevented these filaments from spanning thebare zone (fig. S5), which allowed us to probeone-half of the thick filament where the acto-myosin interactions were oriented as in the sar-comere (Fig. 1B). To ensure that a given actinfilament traversed regions of the thick filamentwith and without cMyBP-C, we only analyzedtrajectories greater than the C-zone length [i.e.,~350 nm (8)], averaging 658 T 8.7 nm (SEM)

1Department of Molecular Physiology and Biophysics, Uni-versity of Vermont, Burlington, VT 05405, USA. 2Departmentof Pediatrics and the Heart Institute, Cincinnati Children’sHospital Medical Center, Cincinnati, OH 45229, USA.

*To whom correspondence should be addressed. E-mail:david.warshaw@uvm.edu

www.sciencemag.org SCIENCE VOL 337 7 SEPTEMBER 2012 1215

REPORTS

on

Janu

ary

14, 2

013

ww

w.s

cien

cem

ag.o

rgD

ownl

oade

d fro

m

www.sciencemag.org/cgi/content/full/science.1218919/DC1

Supplementary Materials for

Predatory Fish Select for Coordinated Collective Motion in Virtual Prey C. C. Ioannou,* V. Guttal, I. D. Couzin*

*To whom correspondence should be addressed. E-mail: C.C.Ioannou@bristol.ac.uk (C.C.I.);

icouzin@princeton.edu (I.D.C.)

Published 16 August 2012 on Science Express DOI: 10.1126/science.1218919

This PDF file includes:

Materials and Methods Figs. S1 to S7 Tables S1 to S3 Caption for Movie S1 References (28–40)

Other Supplementary Material for this manuscript includes the following: (available at www.sciencemag.org/cgi/content/full/science.1218919/DC1)

Movies S1A and S1B

2

Materials and Methods 24 25

Predator fish 26 Juvenile bluegill sunfish, Lepomis macrochirus (46.5±4.9mm mean±SD standard 27

body length) were caught using a seine net from the littoral zone of Lake Carnegie, 28 Princeton NJ, and kept in 50 × 29 × 24cm (length × height × width) glass aquaria for at 29 least 6 months before testing. The number of fish per tank ranged from 11 to 16, and fish 30 were approximately size assorted between tanks. Fish were fed defrosted chironomid 31 larvae once per day ad libitum. Fish were not fed 24hr prior to testing. 32

33 Prey population simulation and projection 34

We developed a simple simulation framework that allowed prey to exhibit a wide 35 range of behaviors with a minimal set of parameters. The simulation was based on the 36 models of Couzin et al. (12, 24, 28). We describe it below in detail. In our mathematical 37 notation, all scalars are italicized whereas vectors are italicized with bold fonts. A hat ^ 38 over  a  vector  means  that  it’s  length  has  been  normalized  to 1. 39

We denoted the position of prey i at time t by ci(t), its direction of heading by v̂i(t), 40 and its speed by s0. The index i can take values from 1 to population size n. We assumed 41 that prey have a circular body shape, with a diameter zb, moving in a two dimensional 42 arena of size Lx and Ly in x and y directions, respectively. Prey maintain a personal space, 43 with highest priority, by moving away from others within a zone of repulsion, zr (>zb). 44 The direction of movement due to repulsion d̂’i(t+dt) given by 45

46 d'i(t+dt)= −∑ 𝒄𝒋( ) 𝒄𝒊( )

|𝒄𝒋( ) 𝒄𝒊( )|𝐣 𝐢    . 47

48 In the absence of any prey within this repulsion zone, and depending on their 49

‘behavioral  traits’,  individuals  may  locally  interact  with  their  ‘nearby  neighbors’.  Nearby  50 neighbors are defined as those within a zone of socialization (zs). Here, we considered 51 three behavioral traits corresponding to how individuals interact with their nearby 52 neighbors. These are the capacity to be attracted towards (denoted by ai for individual 53 i), to orientate direction of travel with (oi), or to ignore (pi), others. We normalized 54 three scalar values such that ai + oi + pi =1, representing the balance of these social 55 tendencies. For example, if pi =1, then individual i ignores neighbors. If ai =1, 56 individual i would be maximally attracted towards neighbors. If oi =1, individual i 57 would not be attracted to neighbors, but would orientate direction of travel with 58 neighbors. For intermediate regions of these parameters, individuals balance these 59 tendencies to varying degrees. For example if ai =0.9 and pi =0.1, individual i is 60 attracted to neighbors, but also biases its preferred direction of travel based on its 61 previous direction of travel. In our model, individuals employ their social tendencies 62 (ai, oi and pi,) only when no other individual is within their zone of repulsion. 63 Therefore, even an individual with maximal ignoring tendency (pi =1) will have 64 repulsion tendencies at short distances. 65

As a consequence of these social interactions, the desired direction of travel at the 66 next time step (d̂i(t+t)) is given by 67

68

3

di(t+t) = pi*v̂i(t) + ai*d̂ai(t)+ oi*d̂oi(t) 69 70

where d̂ai(t) is the direction towards local center of mass and d̂oi(t) is the local direction of 71 heading given by the respective equations 72

73 dai(t)= ∑ 𝒄𝒋( ) 𝒄𝒊( )

|𝒄𝒋( ) 𝒄𝒊( )|𝒋 𝒊   and doi(t)= ∑ 𝒗𝒋( )  |𝒗𝒋( )  |𝒋 𝒊   . 74

75 Individuals are subject to a maximum turning rate max; if the angle between current 76

velocity ( v̂ i(t)) and the desired direction of travel is less than max t, then the new 77 direction of travel will be d̂’i(t+t) = d̂i(t+t); else they turn max t towards it. Finally, 78 to   account   for   errors   in   individuals’   perception   and   motion,   we   rotated   this   direction  79 vector by a small but random angle, a Gaussian distributed noise with zero mean and 80 variance m, leading to velocity vector v̂i(t+t). The new position vector is then given by: 81 ci(t+t)= ci(t) + siv̂i(t +t). A list of all parameters and their values are given in Table 82 S1. 83

We began simulations with random orientations and positions at time step 0 and ran 84 simulations for a minimum of 5,000 time steps before the fish were allowed to observe 85 them on the screen. We employed periodic boundary conditions (see main text). 86

A population of 16 simulated prey was back-projected onto a white translucent film 87 (Rosco gel #118), at an angle to reduce the line of sight to the projector bulb and to 88 remove reflections (Fig. S1). The projection was adjusted using the keystone to correct 89 any distortion from angling the projector in this way. Prey could move over a projected 90 area of 35×17.5cm (Table S1). Projections were made using a Dell M409WX Projector at 91 60Hz (beyond the flicker-fusion frequency of bluegill sunfish (29)) and at a resolution of 92 1200×800. Prey were circular and projected as approximately 2mm in diameter. Prey size 93 and shape was approximately modeled on a common prey of sunfish (Daphnia spp.) and 94 were readily attacked without training. Daphnia actively form swarms, including in 95 response to predator presence (30), and can form vortices with some degree of alignment 96 between individuals at local scales (31, 32). 97

We limited the population size (n) of prey to this relatively small number for two 98 reasons. Firstly, if prey number, and thus density, were too high, prey may appear 99 aggregated without any actual attraction or orientation behavior occurring between them. 100 A larger simulated space to lower the density of prey would have reduced the spatial 101 resolution of the video recordings and hence made it difficult to see (and locate) clearly 102 each prey item. Secondly, the greater the number of prey, the more trials (and fish) we 103 would need to gain a representative view of differential risk between different prey 104 behaviors. At the other extreme, a very small population size would not yield diverse 105 movement strategies needed to investigate selection pressure by the predators. The 106 population size was thus a trade-off between these limitations and being able to sample 107 adequately the attraction-orientation-persistence (aioiandpi) parameter space. Our 108 choice of a modestly sized population of 16 heterogeneous individuals exhibits a range of 109 individual and collective behaviors including solitary correlated random walks, 110 unpolarized and polarized groups of different sizes, and a fission-fusion process among 111 them as exhibited by many real organisms (25, 33). 112

113

4

Trials procedure 114 To prepare fish for the trials we familiarized them with the experimental tank by 115

netting all fish from a single aquarium into the main test area the evening before being 116 tested (Fig. S1). The following day, all fish were moved into the holding chamber, which 117 had partial visual and olfactory contact with the main test area (see Fig. S1). They were 118 then given 10 minutes here to habituate. The simulation was projected and run to 5,000 119 time steps, which was an adequate amount of time for the simulated prey behavior to 120 reach a steady state. Video recording of the tank wall used for the projection was then 121 started (60fps progressive scan, at a resolution of 1280×720 using a Sony PMWEX1R 122 camcorder), followed by a single fish being netted from the holding chamber to the start 123 position (X in Fig. S1). Observations of the fish were made remotely, the trials being 124 ended once the fish had made their first attack on the prey. If no attack was made, trials 125 were limited to 28 minutes in experiment 1 (the maximum recording time of the 8Gb 126 memory cards) and 20 minutes in experiment 2. The fish was returned to its aquarium. 127 The simulation was restarted with a new random seed (random seeds were recorded so 128 the exact simulations could be later reconstructed for analysis), allowed to reach 5,000 129 time steps, and the rest of the procedure repeated. Each fish was tested only once per 130 experiment.  All  work  was  carried  out  under  Princeton  University’s  IACUC  (Protocol  No.  131 1736). 132

The frame of the first attack by each fish was identified. Sunfish use a distinctive 133 suction-feeding method (34, 35), which allowed the moment of the attack to the nearest 134 1/60th second to be determined. This includes the mouth being opened, expansion of the 135 operculum (to increase the volume of the buccal cavity), and, related to this movement of 136 the operculum, to the shift of the eyes upwards and forward relative to the body axis. The 137 positions   of   the   fish’s  mouth,   the   nearest   prey,   and   all   other   prey  were   recorded  using  138 ImageJ (36). These coordinates were then plotted and compared to the coordinates 139 simulated for prey at that time step. This process was used only to identify the targeted 140 prey, rather than in the calculations for group size and tortuosity (see below), which used 141 data directly from the simulations; thus, any possible image distortion and error caused 142 by the recording would not have impacted the results. The prey nearest to the attack was 143 recorded as the attacked prey. This distance was not more than 82% of the distance to the 144 next nearest prey, although in most trials this percentage was much lower (the 3rd 145 quartile was at only 24%, i.e. the next nearest prey, on an average, was approximately 146 four times further away). 147

Bluegill sunfish adopt a hover behavior in which they approach a prey item and 148 pause before either making an attack or swimming away from the prey (Movie S1). We 149 recorded   the   length  of   this   ‘decision   time’   for  each  attack  as  a  measure  of  difficulty   in  150 targeting the prey, calculated as the time step difference between the approach to the prey 151 and the attack. Prey behavioral types were not known until these measurements were 152 completed, thus analysis was conducted blind. 153 154 Individual risk of prey types and the post-selection population 155

In the first experiment, each prey had a unique set of behavioral parameters 156 corresponding to their respective combination of behavioral tendencies ai, oi and pi 157 (see above). In order to maximally detect selection we distributed these ai, oi and pi 158

5

evenly in parameter space. 100 trials were conducted, 70 of which resulted in an attack 159 within the 28 minutes. 160

Based on the number of attacks for each prey type in this population (shown in Fig. 161 2A),  we  generated  a  ‘post-selection’  population  of  prey.  To  do so, we first assigned a risk 162 (ri) that corresponded to the number of attacks (si) for each prey type i as ri= si/(1+ si). 163 This implies that risk increases linearly as a function of number of attacks for small 164 values of si and it saturates to a constant value for a larger number of attacks. We have 165 also chosen other forms of risk functions, such as ri = si, and the qualitative nature of our 166 results continue to hold. The fitness of an individual is then assumed to be a negative 167 function of the risk, i.e. Fi =R - ri where R is the maximum value of the risk among all 168 prey. Thus, the fitness is always a non-negative number such that the prey with maximum 169 risk has the least (zero) fitness while the prey type with the least number of attacks has 170 the maximum fitness. 171

We then normalized fitness to obtain relative fitness (fi) such that ∑ 𝑓 = 1. This 172 can be done by defining 173

𝑓 = 𝐹∑ 𝐹

Such a normalization helps us interpret fi as the probability of reproduction of prey type i. 174 We use these probabilities in a Roulette wheel selection algorithm to generate 10,000 175 sample post-selection populations. Although these selected populations exhibit clear 176 trends with an increase in mean values of both attraction and orientation traits, there was 177 large variability in the distribution of traits among different samples due to the small prey 178 population size. Therefore, we implemented the following deterministic alternative. We 179 denoted nri to be the number of offspring for the prey type i in a population generated by 180 the Roulette wheel algorithm. We note that ni =N*fi represents the expected number of 181 offspring for the prey type i. The Roulette wheel selected population that minimized the 182 sum of squared difference between the expected number offspring and realized number of 183 offspring was then chosen as the post-selection population and was employed in the 184 second experiment (Table S2). 185

In experiment 2, the behavior of the fish was compared when faced either with the 186 pre- and post-selection populations as the experimental treatment. 45 trials were carried 187 out for each population, with the presentation of populations being ordered using a 188 complete random block. As in experiment 1, videos were analyzed blind with the 189 population treatment (pre- or post-selection) only being known to the experimenter after 190 the analysis was complete. 191 192 Statistical analysis 193

In experiment 1, a Poisson-distributed Generalized Linear Model (GLM) was used 194 to analyze the effects of attraction (ai) and orientation (oi) as continuous variables on 195 the number of attacks per prey type. The trials were then split between the targeted prey 196 that were in the 50% of targets closer to the nearest horizontal boundary and the 50% 197 further away (i.e. in the horizontal middle of the projection). In other words, targets with 198 a vertical position in the upper or lower quartiles were in one subset, and targets with a 199 vertical position in the two middle quartiles were in the other. Poisson GLMs with the 200 same explanatory factors as previously was then applied to each of these data subsets. 201

6

The next step was to determine which prey behaviors visible to the fish, generated 202 by   the   prey’s   attraction   and   orientation   parameters,   the   fish   were   selecting   for.   We  203 initially investigate two characteristics, the group size and the tortuosity of prey. A group 204 is a set of individuals where each individual is in the social zone of at least one other 205 individual of the same set. Such group membership was calculated using an extension of 206 the calculation of equivalence classes (37), where the criterion of interest is the presence 207 of at least one individual within the zone of social interaction zs, i.e., within a distance of 208 0.12 (this algorithm finds all interconnected individuals) (15, 37). Tortuosity is a measure 209 of curvedness, or departure from straightness, in the path of an individual i. It is obtained 210 by 211

212 Individual tortuosity (i) = 1-   ( ) 213

214 The numerator of the second term d12(i) represents displacement of individual i in a 215 certain time interval starting from t1 to t2 whereas the denominator dmax represents 216 maximum possible displacement in the same duration. Since individuals in our model 217 move at a constant speed s0, dmax = s0(t2-t1) and dmax>d12(i). Therefore, the second term is 218 always less than or equal to 1 and consequently, the tortuosity has a range of 0 to 1. For a 219 prey travelling in a straight line, the tortuosity is nonexistent (i.e. equal to 0). As the path 220 becomes increasingly curved, the actual displacement from their start to end points (the 221 distance between positions at times t1 and t2) is much smaller than distance travelled by a 222 prey moving on a straight path leading to higher tortuosity. Individual tortuosity can 223 reach a maximum value of 1 when there is no net displacement over the duration of 224 interest. 225

As there was variation in behavioral type between each prey in the first experiment, 226 individual   behavioral   ‘phenotype’   was   the   primary   focus   of   interest.  We   now   discuss 227 individual and group level properties and how they relate to one another. By definition, 228 all members of a group have the same group size, i.e., the total number of individuals in 229 that  group  (although  each  individual’s  position  relative  to   the  group  center of mass, i.e. 230 how relatively far they are from the edge of the group, is an individual property; see Fig. 231 S5). On the other hand, tortuosity is defined at the level of individuals and therefore will 232 be different even for members of a same group. To relate individual-level characteristics 233 to group properties, we also defined a group-level tortuosity as the tortuosity of the group 234 centroid 235

236 Group tortuosity = 1-   237

238 where D12 represents displacement of the group centroid in a certain time interval starting 239 from t1 to t2 whereas the denominator Dmax (= s0(t2-t1)) represents maximum possible 240 displacement of centroid in the same duration. Like individual tortuosity, this quantity 241 also takes values between 0 and 1. It is worth recalling that polarization, i.e. the degree of 242 directional alignment among group members, also acts as a measure of directedness of 243 movement, and is often defined at the level of groups, and even populations. Group 244 polarization, denoted by Pgj, is defined as (12, 38) 245

7

𝑃 = | 1𝑡 − 𝑡

1𝑛   𝑣 (𝑡)|

246 where ng is the size of the group g, and vertical side bars indicate we take the modulus or 247 the length of the final resulting vector. 248

These definitions raise a question however; how are individual tortuosity, group 249 tortuosity and group polarization related to each other? First, it can be mathematically 250 shown that Group tortuosity = 1 – Group polarization. Second, our simulations show that 251 individual tortuosity and the tortuosity of the group to which an individual belongs are 252 highly correlated with each other, and that the tortuosity of prey attacked by the fish is 253 representative  of  the  group’s  tortuosity  (Fig  S2).  In  other  words,  individual  tortuosity  is  a  254 good representation of group tortuosity, and hence group polarization as well. 255

Analogous to real animals, the group size and tortuosity of each prey individual 256 changed   throughout   due   to   the   simulation’s   fission-fusion dynamics; thus although the 257 underlying traits (ai, oi, pi) for a prey type were constant, their behavior visible to the 258 fish (including their group size and tortuosity) varied (see Fig. 2B and C). Additionally, 259 of the prey behavior at the moment of the attacks, many were mutually exclusive (as the 260 population was limited to 16), while others interdependent (e.g. for a prey to have a group 261 size of 5 must mean another 4 individuals also have this group size value). This lack of 262 independence ruled out a direct comparison of phenotype distributions and the number of 263 attacks received per prey type (i.e. a direct comparison of Fig. 2A to 2B and 2C). It is 264 also not clear that summary statistics such as the mean would be representative of the 265 prey   type’s   behavior   given   some   of   the   skewed   distributions   in   Fig.   2B   and   C.   The  266 predatory behavior of the fish was thus compared to a computer-generated  ‘null’  predator  267 that chose a prey individual randomly per trial at the same time step as the observed 268 predation events. For the attacked prey, the relationship between tortuosity and group size 269 was fitted with a polynomial negative binomial GLM fit (n=70). The quadratic 270 coefficient and constant term parameters were extracted. This was then repeated 1,000 271 times but for the null predator that chose an individual randomly at each trial (n=70 in 272 each of the 1,000 runs). The probability (P) that the observed constant term and quadratic 273 coefficient were greater than (or equal to) expected from random was then calculated. P 274 values of <0.05 indicate that the observed value was significantly lower than expected by 275 chance. 276

Although the distance between prey within which they interact (when pi <1) was 277 set to 0.12 (the zone of socialization, zs), this was not necessarily the scale of importance 278 to the predator. The randomization analysis was thus repeated across a range of distances 279 between prey that defined their group membership, hence varying their group size value. 280

To confirm that any non-random targeting by the predator regarding the tortuosity of 281 the   target’s   path   could   be   extrapolated   to   the   motion   of   the   target   prey’s   group   as   a  282 whole, the null predator analysis was repeated using the relationship between group size 283 and group, rather than individual, tortuosity. The slope and intercept of the linear 284 relationship between group size and group tortuosity observed in the actual attacks was 285 compared to the corresponding relationship from 1,000 null predator iterations. 286

The randomized null predator procedure was also employed to test whether the path 287 tortuosity  of  the  attacked  individual  was  representative  of  their  group’s  overall  tortuosity  288

8

(and hence group polarization which is 1 – group tortuosity; see above). The null 289 predators were constrained to attack only individuals within the same groups as those 290 actually attacked by the fish, including solitary prey. A single prey was randomly chosen 291 from each of the same groups as in the trials (n = 70) and summary statistics, such as the 292 mean, were calculated. This was repeated 1,000 times to generate a distribution of these 293 summary statistics expected if the behavior of the fish was no different from random, and 294 the distribution was compared to the equivalent statistic from the 70 observed attacks (see 295 Fig. S2 for further details). 296

A   negative   binomial   GLM   was   used   to   test   the   effect   of   the   attacked   prey’s  297 tortuosity and group size on the time taken to make the attack after being approached (i.e. 298 the decision time). To examine whether the fish were disproportionately attacking prey 299 on the edges or centers of groups, we repeated the randomized null predator analyses 300 above for the relationship between relative position in the group (i.e. relative distance 301 from   the  group’s   center)   and   tortuosity   (Fig.  S5).  As   there   is   little   variation   in   relative  302 position in groups of fewer than 4 individuals, this analysis was limited to the 22 trials 303 where attacks were carried out on prey in groups of >3 (similarly the random predator 304 could also only attack prey in groups of >3). A linear relationship between relative 305 position and tortuosity was found in the observed data (Fig. S5), and the observed 306 regression line was compared to the equivalent from 1,000 randomized null predator 307 attacks. We then repeated this analysis using the tortuosity of the whole group, rather 308 than the tortuosity of individual prey. 309

In these analyses of predator behavior, when the target prey individual was in a 310 group split over a boundary of the projected simulated space, we used the group size 311 ‘apparent’  to  the  predator  as  this  is  more  likely  to  be  relevant  to,  and  affect,  the  predator’s  312 behavior (rather than including the rest of the group near the facing boundary which is 313 likely to be outside of their  visual  field).  In  contrast,  the  ‘true’  group  size,  which  includes  314 all   individuals   irrespective   of   the   boundary,   is   more   representative   of   each   prey’s  315 underlying behavior (their values for ai, oi and pi) and hence potential selection on 316 that prey type (as shown in the phenotype frequency distributions in Fig. 2B and C). The 317 true group size was used to test whether there was a relationship between the position of 318 the targeted prey (horizontal and vertical coordinates of the target as explanatory factors) 319 and  the  target’s  group  size  using  a  negative  binomial  GLM.  Visual  inspection  of  the  data  320 suggested a polynomial fit for the Y variable (Fig. 3), so this was included for both X and 321 Y explanatory variables. The analysis was repeated with the tortuosity of the targeted 322 prey  as  the  response,  and  then  again  using  the  tortuosity  of  the  target’s  group  rather  than  323 its individual tortuosity. In all cases, there were no significant interaction terms, nor did 324 the X variable have an effect. Neither did removing the polynomial from the X variable 325 and repeating the analysis have any qualitative effect on the results (P>0.05 in these 326 cases). 327

For each of the 70 prey that were attacked, the difference between their true and 328 apparent group size was calculated to measure of how split the group was when attacked. 329 A Poisson GLM was used to test whether this splitting was significantly affected by the 330 targeted  prey’s   tortuosity,   controlling   for   (true)  group   size  as   an  additional   explanatory  331 variable. 332

To compare differences in predator behavior between the two different populations 333 of prey (experiment 2), survival analyses (Cox Proportional Hazards) were used to 334

9

compare the total time taken for the trial, the time taken for the first approach towards a 335 prey individual, and the time taken between the first approach and the end of the trial 336 (Table S3). Additionally, the total number of approaches in a trial, their mean duration 337 and the decision time (i.e. the duration of the approach immediately preceding an attack) 338 were compared between the two populations using negative binomial Generalized Linear 339 Models. A binomial GLM compared the probability of attack between the two 340 populations. Finally, a negative binomial GLM was used to analyze the effect of 341 population (pre- or post-) and the attraction and orientation parameters on the per capita 342 risk of each prey type, to examine whether the pattern of selection generated in the first 343 experiment was likely to change after selection operated. 344

In cases where more than one explanatory variable is included in a model, models 345 began fully-factorial, with higher-order terms removed where non-significant. GLMs 346 using Poisson, negative binomial or binomial distributed errors met the assumption of the 347 dispersion parameter being approximately equal to 1  (i.e.  0.5  to  2).  α  was  set  at  P=0.05.  348 All statistical tests were carried out in R 2.10.1 (39). 349

10

Fig. S1. Top and side views of the experimental set up (to scale, except projector and 350 camera). 351

352 The prey simulation was projected onto the wall opposite to the holding chamber and 353 allowed to reach 5,000 time steps before the focal predator was netted to the position 354 labeled by the X. The projection and surrounding area were filmed. The yet-to-be-tested 355 fish were kept in the holding chamber; this had olfactory contact with the main area 356 through the mesh walls of the upper half (grey area) and visual contact with the test fish 357 in only a part of the tank (the dashed lines). There was no direct line of sight between the 358 holding chamber and the prey projection (double line represents a white opaque barrier). 359 Water level is represented by the dotted line. The base of both tanks was covered with 360 yellow gravel. Outer walls of the main test area were covered with white cloth to 361 minimize disturbance, while the inner walls were lined with the same material used as the 362 projection screen to diffuse any internal reflections of predator and prey. The projection 363 was at least 5 cm from the nearest perpendicular wall and it covered less than a third of 364 the total glass wall area. We note that due to the nature of the water-glass-air interface, 365 even without the internal light diffuser, it would not have been possible for reflections of 366 the prey to have been seen in the side walls, from the perspective of the targeting 367 predator. These measures ensure that secondary reflections from other walls of the tank 368 do not result in formation of new prey images within the visual field of the fish. 369

holding chamber

20

.3 cm

26

cm

52

.7 c

m

122.5 cm

46

cm

40.6 cm

24

.7 cm

35

cm1

7.5

cm

Top view

Side view

x

main test area

main test areaholding chamber

11

Fig. S2. Individual tortuosity is representative of group tortuosity and hence group 370 polarization. 371

372

373

0.0

0.2

0.4

0.6

0.8

1.0A

1 0.99 0.98 0.97

0.0

0.2

0.4

0.6

0.8

1.0

0.96 0.92 0.83 0.84

0.0

0.2

0.4

0.6

0.8

1.0

0.81 0.79 0.71 0.66

0.0 0.4 0.8

0.0

0.2

0.4

0.6

0.8

1.0

0.530.0 0.4 0.8

0.410.0 0.4 0.8

0.480.0 0.4 0.8

0.12

Tortuosity (of individual prey)

Gro

up to

rtuos

ity

rel_tor_mean

Freq

uenc

y

0.90 0.94 0.98

0

100

200

300

Mean

Freq

uenc

y

B

rel_tor_var

Freq

uenc

y

0.02 0.04Variance

C

gradient1

Freq

uenc

y

-0.06 -0.04 -0.02Slope

(group size)

D

gradient2

Freq

uenc

y

-0.3 -0.1Slope

(group tortuosity)

E

12

In experiment 1, each prey type had a different set of parameters determining their 374 interactions with their neighbours (ai, oi, pi); we thus quantified the toruosity of each 375 individual’s  path,  rather  than  for  each  group  as  a  whole  (i.e.  the  tortuosity  of  the  group’s  376 centroid). To be able to extrapolate from analyses of individual tortuosity to the overall 377 group polarisation (which is 1 – group tortuosity), here we analyse the relationship 378 between the two variables. A, As  expected,  the  tortuosity  of  an  individual’s  path  is  highly  379 positively correlated with the tortuosity of the group it is in (data is pooled from all prey 380 in the simulations shown to the fish during the first experiment, as in Fig. 2B, 2C and 381 S5). When prey are solitary, individual tortuosity equals group tortuosity (top left panel; 382 Spearman’s  rank  correlation  coefficient  (rs) is given in the bottom right corner). As group 383 size increases (left to right, then top to bottom), individual tortuosity steadily becomes 384 less correlated as each individual contributes proportionally less to the overall group 385 motion, although remains highly correlated until only the largest group sizes. This 386 positive relationship between individual and group tortuosities was also found at the 387 moment at which the fish attacked the prey (the correlation coefficient rs was 0.94 both 388 when groups split by the periodic boundary conditions were considered as two separate 389 groups,  i.e.  the  ‘perspective’  of  the  fish,  or  as  a  single  group,  i.e.  the  prey’s  perspective).  390 B to E, Although strong correlations clearly exist between the tortuosity of an 391 individual’s   path   and   the   path   of   the   group   they   are   in,   there   may remain enough 392 variation, particularly at larger group sizes and tortuosities (A), for the fish to attack 393 individuals  with  paths  unrepresentative  of  the  group’s  tortuosity  (so  that  the  tortuosity  of  394 the   attacked   prey’s   path   does   not   extrapolate   to   the   group’s   overall  motion).   For   each  395 prey we calculated their relative tortuosity as the toruosity of their path divided by the 396 tortuosity  of  their  group’s  path.  We  then  compared  the  observed  mean  (B) and variance 397 (C) in the relative tortuosities of the 70 attacked prey to a distribution of the 398 corresponding values from 1,000 null predator iterations where a prey individual is 399 randomly selected from each of the same groups, and at the same time steps, as attacked 400 by the fish (each null mean and variance is calculated across the 70 random attacks in 401 each iteration). The mean and variance of relative tortuosities for prey attacked by the 402 fish were not significantly different than expected by chance (the observed statistic being 403 greater than expected: P=0.36 and P=0.17, respectively; the frequency plots show the 404 distribution of means and variances for the random predation with the red line indicating 405 the value from the observed attacks). It is possible however that non-random targeting by 406 the fish was only observed in larger groups, or groups with more tortuous paths (A). We 407 thus regressed relative tortuosity of individuals attacked in the trials against the size of 408 their group, and separately, against the toruosity of their group. The slope of the fit was 409 then extracted (shown in D and E, respectively, indicated by the red line). We repeated 410 this process for the 1,000 null predator iterations as previously, each of which generated a 411 value for the gradient between relative tortuosity and group size and another for the 412 gradient between relative tortuosity and group tortuosity (the histograms in D and E, 413 respectively). In neither case were the gradients observed different than expected by 414 chance (the observed gradient being more positive than expected: P=0.39 for group size 415 and P=0.50 for group tortuosity), i.e. the fish were not targeting prey with less or more 416 tortuous paths relative to the group as group size and tortuosity increased. 417 418

13

Fig. S3. Preference of predatory fish for tortuosity and group size prey traits. 419 420

421

422 423 A, The polynomial fit between these two characteristics for all attacked prey (black 424 curve, n=70) was a significantly better fit than the linear equivalent (GLM: 425 LRT67,68=12.67,   P=0.00037).   A   target’s   group   size   is   determined   by   the   number   of  426 interconnected prey (within 0.12 units) at the moment of the attack, whilst their 427 individual  tortuosity  is  calculated  over  the  100  time  steps  before  the  attack  (or  the  fish’s  428 decision time if >100 time steps). To detect any significant active preference by the 429 predator, we compared this polynomial fit to the equivalent from 1,000 computer-430 generated null predators that chose a prey randomly at the same time steps (the red and 431 light red curves represent the mean and 95% intervals, respectively, of the randomly-432 generated fitted values). Across the range of tortuosities, attacked prey were in smaller 433 groups than expected by chance (as the constant term of the observed fit was significantly 434 lower than the randomly generated terms: P<0.001). Moreover, the quadratic coefficient 435

0.0 0.2 0.4 0.6 0.8 1.0

1

2

5

10

Tortuosity

Gro

up s

ize

A

0.0 0.2 0.4 0.6 0.8 1.0

Tortuosity

B

0.0 0.2 0.4 0.6 0.8 1.0

1

2

5

10

Group tortuosity

Gro

up s

ize

C

0.0 0.2 0.4 0.6 0.8 1.0

Group tortuosity

D

14

(how curved the relationship is) was also significantly different to that expected from 436 random (P<0.001). This occurred because disproportionately targeting small groups 437 changed  with  how  tortuous  a  target’s  path  was,  being  greatest  at  low  values  of  tortuosity. 438 Shown here are results using a distance of 0.12 units between prey to define group 439 membership, although the trends were significant across a wider range (Fig. S4). B, This 440 pattern held when group size was redefined to include all interacting members of a group 441 irrespective of the splitting effect of the boundary, unlike in A that uses the size of the 442 group  ‘apparent’  to  the  predator  (as  in  A, black and red curves represent the observed and 443 random fits, respectively). The splitting had little effect on the randomly-generated 444 relationship between the two characteristics, as the red curves are very similar across the 445 range of tortuosities, indicating that prey being in a split group was not common at the 446 moment of the attacks. In marked contrast, the relationships in the observed attacks were 447 similar at low tortuosities but diverged as the target prey took a more tortuous path (the 448 observed relationship from A is given in grey for comparison). In other words, the extent 449 to   which   a   target’s   group   was   split   across   the   boundary   increased   with   the   target’s  450 tortuosity (GLM: LRT1,67=28.03, P=1.2×10-7). C, The analysis in A was repeated using 451 group, rather than individual, tortuosity to confirm that disproportionately attacking 452 smaller groups was most pronounced   when   the   target’s   group   had   low   tortuousity   (as  453 well as the target prey itself having a low tortuosity, as in A and B). Here a linear fit was 454 adequate to describe the relationship between tortuosity and group size, as the fit from the 455 polynomial was not significantly different (GLM: LRT67,68=1.78, P=0.18). Again, smaller 456 groups were attacked more often than expected as the intercept of the observed 457 relationship was significantly less than expected from random targeting (P<0.001). Also 458 in agreement with A and B, this difference was greatest at low tortuosities (the slope of 459 the line was significantly greater than expected: P<0.001). D, This pattern remained 460 unaffected   by   whether   groups   were   considered   split   by   the   boundary   (the   predator’s  461 perspective, as in C),  or  whether  they  are  not  split  (the  prey’s  perspective,  D). Again the 462 slope and intercept were significantly different to random targeting (P<0.001 for both 463 variables). Note, however, that unlike individual tortuosity, the splitting effect of the 464 boundary also affects the calculation of group tortuosity, as well as group size. For C and 465 D, line and dot labeling as in A and B. 466

467

15

Fig. S4. Spatial scale effect on non-random targeting of prey characteristics. 468

469 When the constant term (red line) of the observed polynomial fit between group size and 470 tortuosity is significantly less than expected from random behavior (Fig. S3A), it 471 indicates  that  the  fish  disproportionately  targeted  smaller  groups  regardless  of  the  prey’s  472 tortuosity. This was found to be the case across a large range of distances between 473 individuals within which they were defined as belonging to the same group (the x axis). 474 Across a smaller, but still substantial range, the quadratic coefficient of the fit (black 475 line),   i.e.   how   ‘peaked’   or   ‘curved’   the   curve   is   in   Fig.   S3A,   was   significantly   more  476 negative than expected by chance. This effect was greatest at a distance of 0.12 units, the 477 actual distance within which prey would attract and orientate with one another (the zone 478 of socialization, zs). The grey line indicates P=0.05. 479

0.0 0.5 1.0 1.5

0.0

0.1

0.2

0.3

0.4

0.5

Distance between prey defining group membership

Pro

babi

lity

obse

rved

is ra

ndom

16

Fig. S5. Targeting of prey nearer the group edge and relative position in groups as a 480 function of attraction and orientation parameters. 481

482

483 As  individuals  are  equidistant  to  the  group’s  center  of  mass  (COM)  in  groups  of 2, and 484 usually in groups of 3, we restricted these analyses to groups of >3 only. Thus there are 485 only 22 targeted prey in A compared to 70 in Fig. S3, and the sample size is unequal 486 between prey types in the histograms as they are found in groups of >3 individuals with 487 different frequencies (Fig. 2B). A, The randomly-choosing null predator procedure in 488 Fig. S2 and S3 was employed to detect any non-random targeting by the real predatory 489 fish regarding attacks on prey nearer the edges or centers of groups. The black line 490 represents   the   fitted  observed   relationship  between   the   tortuosity   of   the   targeted  prey’s  491 path  and  their  distance  to  the  group’s  center  of  mass  (relative  to  the  mean  value  for  their  492 group). As in Fig. S3, the red and light red curves represent the mean and 95% intervals, 493 respectively, of the fitted values from equivalent fits but where the random predator 494

0.0 0.2 0.4 0.6 0.8 1.0

0.5

1.0

1.5

2.0

Tortuosity

Rel

ativ

e di

stan

ce to

CO

MA

0.0 0.2 0.4 0.6 0.8 1.0

Group tortuosity

B

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Attraction ( ai)

Orie

ntat

ion

(o i

)

C

0 1Relative rank

0

5000

Freq

uenc

y

0.0 0.2 0.4 0.6 0.8 1.0Attraction ( ai)

D

0 3Relative distance

0

2000

Freq

uenc

y

17

chose a target randomly from all grouped prey (>3 individuals) at the same time steps the 495 real predators attacked a prey in a group. 1,000 random predator iterations were carried 496 out. The gradient of the observed relationship is significantly more negative than 497 expected from random targeting (P=0.016); the fish disproportionately targeted prey 498 relatively further from the group center but only when  the  prey’s  tortuosity  was  low  (i.e.  499 when they were in polarized groups). B, This analysis was repeated using group 500 tortuosity rather than individual tortuosity. Although the slope was only marginally more 501 negative compared to random targeting (P=0.099), the qualitative pattern remained the 502 same as in A and the intercept was significantly greater than random targeting (P=0.003). 503 C, and D, The distributions for each behavioral prey type over the course of the 504 simulations shown to the fish in experiment 1 as in Fig. 2B and C (sampled every 100 505 time steps). The relative rank (C)  is  the  rank  proximity  of  the  prey  from  its  group’s  center  506 of mass divided by the group size (hence 1 is the furthest prey). In a similar manner, the 507 relative distance (D) is the distance of the prey from the center of mass divided by the 508 mean distance from the center of mass for all prey in the group (values of 1 indicate the 509 prey has an average distance to the group center). It is this measure of relative position 510 that we use for the random predator analysis (A and B). 511

512

18

Fig. S6. Risk for prey types closer (A) and further (B) from the boundaries of the 513 projection. 514

515 Data  on  the  targeted  prey  were  split  into  two  halves  according  to  the  target’s  distance  to  516 the nearest horizontal boundary: the 50% of targeted prey closer to the nearest horizontal 517 boundary (A) and the 50% further away (i.e. in the horizontal middle of the projection; 518 B). The effect of the attraction and orientation parameters were then analyzed using 519 Poisson distributed GLMs. As in Fig. 2A, the bubble area is proportional to the number 520 of attacks each prey type received, the largest of these being 10 attacks. The color 521 gradient represents the fitted values of the relationship for the fully factorial model in A 522 (GLM: LRT1,12=10.98, P=0.00092) and the main-effects model in B (attraction: 523 LRT1,13=15.087, P=0.00010; orientation: LRT1,13=5.54, P=0.019; the interaction term 524 was not significant: LRT1,12=1.75, P=0.19). Splitting the data in this way was chosen to 525 equalize the sample size (and hence test power) between the two statistical models (n=35 526 attacks  for  each).  Although  it  is  unlikely  to  capture  exactly  what  is  ‘close’  or  ‘far’  from  527 the perspective of the fish, it does illustrate that selection for the attraction × orientation 528 synergy is greater closer to the boundary, and that the effect of attraction is relatively 529 strong throughout the simulated space. 530 531

0.0 0.2 0.4 0.6 0.8 1.0Attraction ( ai)

0.0

0.2

0.4

0.6

0.8

1.0

Orie

ntat

ion

(o i

)A

-----------------------------------------------------------------------------------------------------

0

1

2

3

4

Fitte

d n

atta

cks

0.0 0.2 0.4 0.6 0.8 1.0Attraction ( ai)

B

-----------------------------------------------------------------------------------------------------

0

2

4

6

Fitte

d n

atta

cks

19

Fig. S7. Group size and tortuosity distributions pre- and post- selection. 532

533

534 Distributions are pooled over all prey types and trials from experiment 2, split by the 535 experimental treatment (left panels: pre-selection (A and C) and right panels: post-536 selection  (B  and  D)).  Group  size   is   the  ‘true’  group  size,   i.e.   the  number  of  prey  in   the  537 group irrespective of the splitting effect of the boundary. 538

0.0

0.1

0.2A

Pro

babi

lity

Group size0 5 10 15

B

Group size0 5 10 15

0.0

0.1

0.2

0.3C

Pro

babi

lity

Tortuosity0.0 0.2 0.4 0.6 0.8 1.0

D

Tortuosity0.0 0.2 0.4 0.6 0.8 1.0

20

Table S1. Model parameters and their values. 539 The size, shape and speed of the simulated prey as projected is similar to Daphnia, a 540 common prey of juvenile bluegill sunfish (40). 541 542

Quantity Description Values Dimension in projection

Lx Size of arena in X dimension 4.0 units 35cm Ly Size of arena in Y dimension 2.0 units 17.5cm zb Body size (diameter) of prey 0.025 units 2.19 mm zr Zone of repulsion 2.56* zb = 0.064 5.6 mm zs Zone of social interactions (i.e.

socialization) 4.8* zb = 0.12 10.5 mm

s0 Speed 0.7* zb = 0.0175/unit time

1.53 mm/sec

dt Time step 0.1 of a unit time 1 frame of video = 1/60th of a sec; 1 unit time = 10 time steps = 1/6th of a second

m Noise in angular motion 5.7 degrees (per time step)

NA

max Maximum turning rate 11.4 degrees per time step

11.4 degrees per frame

Size of the prey population 16 NA 543

544

21

Table S2. Prey type behavioral parameters and their frequencies in pre- and post-545 selection populations. 546 All other parameters were as in Table S1. 547 548

Prey type

Attraction ai)

Orientation oi)

Persistence (pi)

Pre-selection frequency

Post-selection frequency

1 0 0 1 1 0 2 0 0.2 0.8 1 1 3 0 0.4 0.6 1 0 4 0 0.6 0.4 1 0 5 0 0.8 0.2 1 1 6 0.2 0 0.8 1 1 7 0.2 0.2 0.6 1 1 8 0.2 0.4 0.4 1 2 9 0.2 0.6 0.2 1 2

10 0.4 0 0.6 1 0 11 0.4 0.2 0.4 1 1 12 0.4 0.4 0.2 1 4 13 0.6 0 0.4 1 0 14 0.6 0.2 0.2 1 2 15 0.8 0 0.2 1 0 16 1 0 0 1 1

549 550

22

Table S3. The effect of prey type frequencies (i.e. pre- versus post- selection prey 551 populations) on predatory behavior in experiment 2. 552 T denotes the number of time steps. 553 554

Predator behavior Statistical test Test statistic N P value total T taken Survival analysis Z = -0.0079 90 0.99 T taken for 1st approach to prey Survival analysis Z = 0.044 90 0.96 T taken from 1st approach to attack

Survival analysis Z = -0.55 74 0.58

number of approaches Neg. Bin. GLM LRT = 1.47 90 0.22 mean decision T per trial Neg. Bin. GLM LRT = 0.011 74 0.92 decision T of attacks Neg. Bin. GLM LRT = 0.053 58 0.82 probability of attack Binomial

GLM LRT = 0.194 90 0.66

555 556

23

Movie S1. Examples of predator behavior leading to selection for coordinated 557 collective motion in virtual prey. 558 In the first clip, an isolated prey is targeted and attacked by the fish. In the second, the 559 fish attacks a prey with high tortuosity in a group (i.e. a prey in a swarm) split by the 560 periodic boundary conditions. Shown at the bottom is the time step of the simulation 561 (left) and the random seed unique to each trial (right). Clips are taken from pilot trials and 562 have been down-sampled to reduce file size (to 15 frames per second and a lower picture 563 quality). Two versions are provided, one as the video was captured (A), and the other 564 with the contrast adjusted to facilitate the visibility of the fish and prey (B). 565

References and Notes 1. J. Krause, G. D. Ruxton, Living in Groups (Oxford Univ. Press, Oxford, 2002).

2. W. Cresswell, J. L. Quinn, Faced with a choice, sparrowhawks more often attack the more vulnerable prey group. Oikos 104, 71 (2004). doi:10.1111/j.0030-1299.2004.12814.x

3. W. L. Romey, A. R. Walston, P. J. Watt, Do 3-D predators attack the margins of 2-D selfish herds? Behav. Ecol. 19, 74 (2008). doi:10.1093/beheco/arm105

4. W. D. Hamilton, Geometry for the selfish herd. J. Theor. Biol. 31, 295 (1971). doi:10.1016/0022-5193(71)90189-5 Medline

5. M. Ballerini et al., Interaction ruling animal collective behavior depends on topological rather than metric distance: Evidence from a field study. Proc. Natl. Acad. Sci. U.S.A. 105, 1232 (2008). doi:10.1073/pnas.0711437105 Medline

6. N. O. Handegard et al., The dynamics of coordinated group hunting and collective information transfer among schooling prey. Curr. Biol. 22, 1213 (2012). doi:10.1016/j.cub.2012.04.050 Medline

7. G. D. Ruxton, A. L. Jackson, C. R. Tosh, Confusion of predators does not rely on specialist coordinated behavior. Behav. Ecol. 18, 590 (2007). doi:10.1093/beheco/arm009

8. K. A. Jones, A. L. Jackson, G. D. Ruxton, Prey jitters; protean behaviour in grouped prey. Behav. Ecol. 22, 831 (2011). doi:10.1093/beheco/arr062

9. R. Lukeman, Y. X. Li, L. Edelstein-Keshet, Inferring individual rules from collective behavior. Proc. Natl. Acad. Sci. U.S.A. 107, 12576 (2010). doi:10.1073/pnas.1001763107 Medline

10. Y. Katz, K. Tunstrøm, C. C. Ioannou, C. Huepe, I. D. Couzin, Inferring the structure and dynamics of interactions in schooling fish. Proc. Natl. Acad. Sci. U.S.A. 108, 18720 (2011). doi:10.1073/pnas.1107583108 Medline

11. J. Buhl et al., From disorder to order in marching locusts. Science 312, 1402 (2006). doi:10.1126/science.1125142 Medline

12. I. D. Couzin, J. Krause, R. James, G. D. Ruxton, N. R. Franks, Collective memory and spatial sorting in animal groups. J. Theor. Biol. 218, 1 (2002). doi:10.1006/jtbi.2002.3065 Medline

13. J. K. Parrish, L. Edelstein-Keshet, Complexity, pattern, and evolutionary trade-offs in animal aggregation. Science 284, 99 (1999). doi:10.1126/science.284.5411.99 Medline

14. S. Bazazi et al., Collective motion and cannibalism in locust migratory bands. Curr. Biol. 18, 735 (2008). doi:10.1016/j.cub.2008.04.035 Medline

15. V. Guttal, I. D. Couzin, Social interactions, information use, and the evolution of collective migration. Proc. Natl. Acad. Sci. U.S.A. 107, 16172 (2010). doi:10.1073/pnas.1006874107 Medline

16. D. Grünbaum, Evol. Ecol. 12, 503 (1998). doi:10.1023/A:1006574607845

17. R. Vabø, L. Nøttestad, An individual based model of fish school reactions: predicting antipredator behaviour as observed in nature. Fish. Oceanogr. 6, 155 (1997). doi:10.1046/j.1365-2419.1997.00037.x

18. A. J. Wood, G. J. Ackland, Evolving the selfish herd: Emergence of distinct aggregating strategies in an individual-based model. Proc. Biol. Sci. 274, 1637 (2007). doi:10.1098/rspb.2007.0306 Medline

19. T. J. Ehlinger, D. S. Wilson, Complex foraging polymorphism in bluegill sunfish. Proc. Natl. Acad. Sci. U.S.A. 85, 1878 (1988). doi:10.1073/pnas.85.6.1878 Medline

20. H. Sayama, S. Dionne, C. Laramee, D. S. Wilson, in IEEE Symposium on Artificial Life (2009), pp. 85-91.

21. J. E. Treherne, W. A. Foster, Group size and anti-predator strategies in a marine insect. Anim. Behav. 30, 536 (1982). doi:10.1016/S0003-3472(82)80066-3

22. W. Cresswell, Ibis 138, 684 (1996).

23. Materials and methods are available as supplementary materials on Science Online.

24. I. D. Couzin, J. Krause, N. R. Franks, S. A. Levin, Effective leadership and decision-making in animal groups on the move. Nature 433, 513 (2005). doi:10.1038/nature03236 Medline

25. I. D. Couzin, J. Krause, Self-organization and collective behavior in vertebrates. Adv. Stud. Behav. 32, 1 (2003). doi:10.1016/S0065-3454(03)01001-5

26. C. C. Ioannou, C. R. Tosh, L. Neville, J. Krause, The confusion effect from neural networks to reduced predation risk. Behav. Ecol. 19, 126 (2008). doi:10.1093/beheco/arm109

27. C. R. Tosh, Which conditions promote negative density dependent selection on prey aggregations? J. Theor. Biol. 281, 24 (2011). doi:10.1016/j.jtbi.2011.04.014 Medline

28. I. D. Couzin et al., Uninformed individuals promote democratic consensus in animal groups. Science 334, 1578 (2011). doi:10.1126/science.1210280 Medline

29. E. Wolf, G. Zerrahn-Wolf, Threshold intensity of illumination and flicker frequency for the eye of the sun-fish. J. Gen. Physiol. 19, 495 (1936). doi:10.1085/jgp.19.3.495 Medline

30. J. Pijanowska, A. Kowalczewski, Predators can induce swarming behaviour and locomotory responses in Daphnia. Freshw. Biol. 37, 649 (1997). doi:10.1046/j.1365-2427.1997.00192.x

31. R. Mach, F. Schweitzer, Modeling vortex swarming in Daphnia. Bull. Math. Biol. 69, 539 (2007). doi:10.1007/s11538-006-9135-3 Medline

32. A. Ordemann, G. Balazsi, F. Moss, Pattern formation and stochastic motion of the zooplankton Daphnia in a light field. Physica A 325, 260 (2003). doi:10.1016/S0378-4371(03)00204-8

33. I. D. Couzin, Behavioral ecology: Social organization in fission-fusion societies. Curr. Biol. 16, R169 (2006). doi:10.1016/j.cub.2006.02.042 Medline

34. S. W. Day, T. E. Higham, A. Y. Cheer, P. C. Wainwright, Spatial and temporal patterns of water flow generated by suction-feeding bluegill sunfish Lepomis macrochirus resolved by Particle Image Velocimetry. J. Exp. Biol. 208, 2661 (2005). doi:10.1242/jeb.01708 Medline

35. T. E. Higham, S. W. Day, P. C. Wainwright, Multidimensional analysis of suction feeding performance in fishes: Fluid speed, acceleration, strike accuracy and the ingested volume of water. J. Exp. Biol. 209, 2713 (2006). doi:10.1242/jeb.02315 Medline

36. W. S. Rasband (National Institutes of Health, Bethesda, 2004).

37. W. H. Press, S. A. Teukolsky, W. T. Vetterling, B. P. Flannery, Numerical Recipes in C++ (Cambridge Univ. Press, Cambridge, 1992).

38. T. Vicsek, A. Czirók, E. Ben-Jacob, I. Cohen, O. Shochet, Novel type of phase transition in a system of self-driven particles. Phys. Rev. Lett. 75, 1226 (1995). doi:10.1103/PhysRevLett.75.1226 Medline

39. R Development Core Team (R Foundation for Statistical Computing, Vienna, 2011).

40. K. Jensen, P. Jakobsen, O. Kleiven, Hydrobiologia 368, 123 (1998). doi:10.1023/A:1003233728870