ARTICLE IN PRESS

0925-2312/$ - se

doi:10.1016/j.ne

�CorrespondE-mail addr

Neurocomputing 70 (2007) 1079–1084

www.elsevier.com/locate/neucom

Letters

Data mining neural spike trains for the identification of behaviouraltriggers using evolutionary algorithms

Richard Stafford, F. Claire Rind�

School of Biology and Psychology, Ridley Building, University of Newcastle-upon-Tyne, NE1 7RU, UK

Received 30 January 2006; received in revised form 1 September 2006; accepted 1 September 2006

Communication by W.L. Dunin-Barkowski

Available online 10 November 2006

Abstract

We analysed spike trains from the descending contralateral movement detector (DCMD) neuron of locusts. The locusts either

performed jumps or did not jump in response to visual looming stimuli. An evolutionary algorithm (EA) was employed to sort spike

trains into the correct behavioural categories by optimising threshold parameters, so jump behaviour occurred if the spike-train data

exceeded the threshold parameters from the EA. A candidate behavioural trigger appeared to be prolonged high-frequency spikes at a

relatively early stage in the approach of the stimulus. This technique provides a useful precursor to a full biological analysis of the escape

jump mechanism.

r 2006 Elsevier B.V. All rights reserved.

Keywords: DCMD neuron; Locust; Behaviour; Neural coding; Evolutionary algorithm

1. Introduction

Escape or startle behaviours in some animals aretriggered by a single spike in an identified interneuron.For example, a single spike in the Mauther neuron ofteleost fish causes the C-start escape behaviour ([7]), and asingle spike in one of the crayfish’s giant interneuronscauses a tail-flip escape response ([3]). However, mostescape behaviours are not so obviously coded, relyinginstead on particular features from within complex trainsof spikes.

The descending contralateral movement detector(DCMD) neuron of locusts responds most vigorously tosmall, fast-moving stimuli that may best represent thelocust’s natural avian predators [9,10]. Of these, theDCMD neuron responds most strongly when the objectlooms, or appears to approach the eye. In flight, theDCMD neuron is thought to trigger a gliding behaviourthat may cause a locust to dive and evade capture by anapproaching predator [11,12]. DCMD neuron activity may

e front matter r 2006 Elsevier B.V. All rights reserved.

ucom.2006.09.011

ing author.

ess: claire.rind@ncl.ac.uk (F. Claire Rind).

also be involved in the locust’s evasive steering behavioursthat occur earlier during a predator’s approach [12,5].The spike-train feature of the DCMD neuron that

triggers a glide is thought to be complex. DCMD neuronspikes must occur in sufficient number above an instanta-neous threshold so as to allow them to summate and elicitaction potentials in a motor neuron that is involved in theelevation of the wings into the gliding posture [11].Furthermore, this trigger feature must occur within thecorrect phase of the wingbeat cycle so as to be summedwith ongoing flight motor activity and to allow the glide tooccur at the appropriate wingbeat phase [11].Determining the features of a neural spike train that

trigger a particular behaviour can be a complex and time-consuming process. In demonstrating the role of anidentified neuron in behaviour two approaches arecommon: Classically, necessity and sufficiency can bedemonstrated by first ‘killing’ the neuron and demonstrat-ing that behaviour no longer occurs in response to anormally effective stimulus, and secondly, exciting theneuron electrically and eliciting the behaviour artificially[7]. However, often the invasive techniques required maythemselves result in an alteration of the animal’s natural

ARTICLE IN PRESSR. Stafford, F. Claire Rind / Neurocomputing 70 (2007) 1079–10841080

behaviour, obscuring any effect of the perturbed neuronitself. An alternative method is to correlate spike-train andbehavioural features in order to indicate a possible causalrelationship [11,12]. However, obtaining neural and beha-vioural data simultaneously are highly demanding anddetermining the correct features of the behaviour and spiketrain to correlate is largely a trial and error process.

In this study, we present an objective method of dataanalysis that may help identify features of a spike trainimportant in behavioural triggering, allowing the collectionand analysis of more focussed biological data in subse-quent experiments.

We correlate neuronal and behavioural features using anevolutionary algorithm (EA), which simultaneously opti-mises a range of parameters associated with DCMDneuron spike trains recorded from locusts either jumpingin response to a looming stimulus, or producing noresponse to it. The parameters are then used, withoutreference to the locust’s behavioural response, to divide thespike trains into ‘jump’ or ‘non-jump’ prediction groups, ina similar way to EA-based clustering methods in bioinfor-matics [4]. These predictions are then compared with thebehaviour known to have occurred during each spike train,with correct predictions increasing the ‘fitness’ of the set ofparameters used to divide the trials. An optimal set ofparameters is finally obtained that correctly dividesjumping and non-jumping spike trains based on particularresponse features. These spike train features may, there-fore, be important in triggering the jump itself.

2. Methods

2.1. Experimental methods

We used recorded data of DCMD activity and beha-vioural responses in two locusts challenged with computer-generated visual stimuli looming from their right sides(data originally presented in [2]).

Visual stimuli were displayed on a high-speed computermonitor with a refresh rate of 180Hz. The stimulus was acomputer-generated circular object of diameter 80mm,which loomed towards the locust at a constant speed of1m/s. The monitor screen was positioned 70mm from thelocust’s right eye and the simulated trajectory of objectapproach was 901 to the locust’s long axis. Initially, thelooming stimulus was simulated to be 2.07m from the eyeof the locust (including the 70mm between the position ofthe screen and the locust) and approached over 2m, finallysubtending an angle of 601 on the locust’s eye at its finalsimulated position level with the monitor screen.

Activity in the left DCMD was recorded from its axon inthe cervical connectives using implanted hook electrodes asdescribed in [12]. Electrodes were attached to 50 mmdiameter insulated copper wires and signals amplified.The copper wires acted as a loose tether. Other than thewire tether, locusts could move and jump freely. For eachtrial, DCMD activity was processed off-line in Spike2

(CED) to obtain the timing of each spike, relative to thestart of looming stimulus movement. Spike timings wererecorded in seconds to an accuracy of five decimal places.Each spike train was also classified as a jump-producingspike train or a non-jump-producing spike train based onwhether the locust jumped during or immediately after thepresentation of the looming stimulus in that trial.Jumping in response to a looming stimulus occurred in

p25% of trials; however, the frequency of jumping inresponse to a looming stimulus was far greater than thatwhen no stimulus was shown, when the locust rarelyjumped unless in response to a disturbance in thelaboratory (e.g. opening a door and entering the labora-tory). In addition, the timing of the jumps coincided withthe final phase of looming stimulus expansion, from 0.2 sbefore the end of stimulus movement until its end. Thisindicates the locust did jump in response to the loomingstimuli when jumping behaviour occurred.Due to the difficulty in collecting electrophysiological

data from locusts able to freely jump, data were obtainedfrom just two locusts. Locust 1 had four clear DCMDneuron recordings during jumps and 12 clear DCMDneuron recordings during no jumps. Locust 2 had two clearrecordings during jumps and 11 clear recordings during nojumps. Since it is possible that subtle differences may occurbetween the behavioural triggers of different animals, datawere analysed separately for the two locusts.

2.2. Parameters used to divide spike trains into jump and

non-jump groups

A total of six parameters were used by the evolutionaryalgorithm to attempt to separate spike trains recordedduring jumps from those recorded during non-jumps. Thefirst two parameters defined a time window of interest andwere (1) start time and (2) end time. These parameters weredesigned to focus on a specific window of important neuralactivity, which could, in theory, include the whole record ofthe DCMD neuron response. The limits of these para-meters, when adjusted by the EA, were between the start ofthe loom, and the last recorded spike in the DCMD neuronresponse.Three interlinked parameters were also used to deter-

mine differences between spike trains. The parameters were(3) a threshold instantaneous spike frequency (the recipro-cal of a single inter-spike interval), (4) a number ofconsecutive spikes occurring above the threshold fre-quency, (5) a time for which the instantaneous frequencyof the spike train exceeded the threshold frequency.Parameters 4 and 5 are closely correlated, but subtlydifferent. For example a time of 0.1 s above a 100 imp/sthreshold must equate to at least 10 spikes above thethreshold, but could mean more spikes than this if theinstantaneous frequency rose during this period. Unlike instatistical analysis, there is no need for the parameters to befully independent. However, correlated variables must beinterpreted with care. Parameter (6) was the maximum

ARTICLE IN PRESSR. Stafford, F. Claire Rind / Neurocomputing 70 (2007) 1079–1084 1081

instantaneous spike frequency that occurred within theregion of interest (between the start time and end timedefined above). This parameter was included in order toinvestigate whether very high-frequency spikes occurring atan optimal time could act as a jump trigger.

2.3. Evolutionary algorithm and fitness design

We performed data analysis using an EA described indetail in [8]. Initially, a population of individuals wasgenerated where each individual was a set of values foreach of the parameters described in the previous section. Avalue for each of these parameters was randomly deter-mined from uniform distributions with set ranges. Rangesfor parameters 1 and 2 were 0 s (the start time of the loom)and the time of the final spike in the data set. Parameter 3ranged from 0 to 500 imp/s. Parameter 4 from 1 to 30spikes. Parameter 5 from 0 to 0.35 s and parameter 6 from0 to 700 imp/s. One complete set of parameter values wasconsidered as an ‘individual’ for use in the evolutionaryalgorithm below. Initially, individuals were assigned fitness0 (see below). Fitness was a measure of how well anindividual sorted the spike trains into jump and non-jump behaviours, the higher the value the better thesolution, and the more likely the individual would be toproduce offspring in the next generation of individualsolutions.

The EA used a search strategy based on biologicalevolution and natural selection to optimise the parametervalues of the originally generated ‘individuals’ oversuccessive generations. The parameter values of eachindividual are used to separate jump from non-jump spiketrains on the basis of whether the recorded spike train valueis greater (jump) or lesser (non-jump) than that specified bythe parameter value for that individual in the EA. Duringevaluation of an individual in a generation of the EA,recorded spike train data were filtered so as only that datafalling within the time window given by parameters 1 and 2of a given individual were studied. Furthermore, if theinstantaneous spike frequencies of the data fell below thevalue in parameter 3 they were also ignored. If two of thethree parameters 4–6 exceeded the parameter value of theindividual solution from the EA, and the data were foractual jump behaviour, then the fitness for that individualwas increased by either three units for locust 1, or six unitsfor locust 2 (see below). Similarly, if two of the threeparameters 4–6 were below the threshold values for thatindividual and the data were for non-jump behaviour, thefitness for that individual solution was increased by oneunit. In all other cases, where spike trains resulting injumping behaviour were classified as non-jumps or wherespike trains not resulting in jumps were classified as jumpspike trains, fitness did not increase. In preliminaryexperiments, an EA design requiring all parameters of agiven individual to classify a spike train as jump or non-jump was tried but was found to be less effective than thedesign described here.

Different fitness increments were used for correctclassification of jumps and non-jumps because there weremore non-jump records than jump records. This preventedthe EA from finding a local optimum, where all the non-jumps were correctly classified but jumps were incorrectlyclassified. For Locust 1, correct classification of the 4 jumpspike trains resulted in a fitness of (4 jumps � 3 units perjump) 12 units, identical to the correct classification of the12 non-jump situations. For Locust 2, a total fitness of(2 jumps� 6 units per jump) 12 units would be obtained forthe jumps compared to the 11 units for non-jumps. Themaximum fitness of Locust 1 was 24 units and Locust 2 was23 units.By this method, a value of fitness was obtained for each

of the individuals in the population at a given generation,depending on how well the parameters of that individualdistinguished between jump and non-jump trials. If anindividual has parameter values that correctly classify ahigh number of jump and non-jump trials, it is awarded ahigh fitness and will be selected for by the EA. Thisindividual will then produce a number of similar indivi-duals to itself in the next generation.Recombination between different individuals in the

population, and mutation, alters the individual’s offspringin the next generation slightly in order to attempt to find asolution with increased fitness. If an individual hasparameters which poorly separate jump and non-jumptrials, it will not be selected for by the EA and theindividual will not produce any similar offspring in the nextgeneration. The EA was run for 200 generations or until‘jump’ and ‘non-jump’ data were fully separated.A fully annotated version of the Matlab encoded EA,

including details about mutation rates and recombination,is available for inspection and use at (http://www.staff.ncl.ac.uk/claire.rind/EA.htm).

3. Results and discussion

The EA was able to fully sort the jump and non-jumpdata for Locust 1 into the correct behavioural groups (maxfitness obtained ¼ 24 units). For Locust 2, one non-jumprecord was misclassified as a jump (max fitnessobtained ¼ 22 units).The EA was run five times for each locust in order to

give an indication of the distribution of the thresholdparameters required to sort the data (Table 1). Theprecision on any one parameter value between replicateruns of the EA indicates the importance of that particularparameter in distinguishing between jump and non-jumpbehaviours on the basis of the DCMD response.For both locusts, a precise value for the end time of the

time window of interest was found in consecutive runs ofthe EA. The mean value of this parameter was 1.8 s afterthe start of the loom for both locusts. The high precision ofthis parameter, combined with the low precision of thestart of time window parameter, is likely to indicate thatimportant features of the spike train, that can be used to

ARTICLE IN PRESS

Table 1

The parameters obtained from the EA that best sort the DCMD responses recorded during jump and non-jump trials into the correct behavioural

categories

Parameter Replicate 1 2 3 4 5 Mean S.D. S.D. as

% of

mean.

Locust 1

Start time (s after start of loom) 0.779 0.302 0.615 0.9163 0.851 0.693 0.246 35.4

End time (s after start of loom) 1.793 1.910 1.788 1.823 1.825 1.83 0.0489 2.67

Instantaneous spike threshold (imp/s) 86.38 123.1 83.68 79.10 94.80 93.4 17.6 18.8

Number of spikes above threshold 6.52 5.80 7.46 7.60 6.75 6.83 0.734 10.7

Time above threshold (s) 0.2601 0.1928 0.1455 0.201 0.306 0.221 0.0625 28.3

Maximum instantaneous frequency (imp/s) 234.3 286.5 7.20 58.0 400.3 197.3 162.8 82.6

Locust 2

Start time (s after start of loom) 1.498 0.128 0.4367 1.5786 1.395 1.01 0.674 66.8

End time (s after start of loom) 1.874 1.715 1.8053 1.9409 1.893 1.84 0.0878 4.76

Instantaneous spike threshold (imp/s) 53.1 20.98 82.16 111.94 96.54 72.9 36.3 49.6

Number of spikes above threshold 17.12 13.06 3.68 12.01 9.64 11.1 4.95 44.6

Time above threshold (s) 0.0814 0.0253 0.2385 0.215 0.196 0.151 0.0928 61.3

Maximum instantaneous frequency (imp/s) 274.1 461.2 279.7 142.6 274.6 286.4 113.5 39.6

Parameters for Locust 1 fully sort the data into the correct categories. Parameters for Locust 2 result in one non-jump spike train being categorised as a

jump. Data for each locust is provided for five replicate optimisations by the EA as is the mean, S.D. and S.D. as a percentage of the mean of the five

replicate trials for each parameter.

R. Stafford, F. Claire Rind / Neurocomputing 70 (2007) 1079–10841082

distinguish jump and non-jump trials, occur immediatelyprior to 1.8 s after the start of the loom. Since the loomingobject would, in theory, collide with the locust 2.07 s afterthe start time, this equates to 0.27 s before the calculatedtime of collision.

For DCMD neuron spikes to result in the triggering ofbehaviour, such as jumping, they are likely to have to sumin a postsynaptic neuron such as a motor neuron and resultin the threshold of this neuron being exceeded and either asingle-action potential or several action potentials beingproduced [1,6]. This is because EPSPs mediated by theDCMD neuron in leg motor neurons, such as the fastextensor of the tibiae, are small (o1mV) and one EPSP onits own is not big enough to cause a spike [1].Instantaneous spike frequency is, therefore, an importantresponse feature in determining the potential responses ofsuch postsynaptic neurons.

Mean values for the instantaneous spike thresholdparameter of 93.4 and 72.9 imp/s were obtained for Locusts1 and 2, respectively, although the precision of this estimatewas low for Locust 2 (Table 1). Clearly, there are potentialinteractions between this term and the number of spikes, orthe time, above the threshold frequency parameter. Highvalues of instantaneous frequency often coincided withshorter time periods or lower spike numbers above thethreshold frequency (e.g. Locust 1, replicate 2 in Table 1),whereas low values often resulted in longer times or highernumbers of spikes above the threshold frequency (e.g.Locust 2, replicates 1 and 2). The values of these threeparameters are, as expected, tightly correlated. However,discrepancies can arise between spike number above thethreshold frequency and time above the threshold fre-quency. These discrepancies may arise due to the fitness

design requiring only fulfilment of two of the parameters4–6 in order to correctly classify a spike train as occurringduring a jump or non-jump behaviour.Values for the maximum instantaneous frequency

parameter were much more precise between replicateoptimisations for Locust 2 than for Locust 1. For Locust1, the parameter played little role in classifying spiketrains as occurring during jumps or non-jumps, often beinglower than the instantaneous spike threshold. For Locust2, the EA results indicate values of around 200 imp/sor greater may be important in triggering a jump.Although not all values exceed 200 imp/s (e.g. Locust 2,replicate 4), this may also have occurred because the fitnessdesign only required correct classification of 2 of the 3parameters.The optimal parameters obtained from the EA indicate

areas where differences between jump and non-jump spiketrains occurred. Since collecting electrophysiological datafrom freely moving animals is difficult, our data sets weresmall. Furthermore, since the frequency of jumping inresponse to a purely visual stimulus is low, the amount ofjump data collected was particularly low. Therefore, we donot have sufficient data with which to test the optimalparameters generated by the EA. However, the purpose ofthis study was to develop a method of objectivelyidentifying key aspects of a spike train that could beimportant in triggering behaviour.It is unlikely that the key features of the spike train,

which are identified as important by the EA, would beobtained by traditional methods such as investigating peri-stimulus plots (Fig. 1). In both jump and non-jump plots,the maximum spike frequency is far higher at the end of thestimulus approach and immediately after the stimulus stops

ARTICLE IN PRESS

Fig. 1. Peri-stimulus plots of DCMD neuron spike rate for jump and non-

jump behaviours against time from onset of visual stimulus movement.

Spike rate is instantaneous frequency averaged over 25ms time window in

order to smooth the plot. Grey box indicates the time period 1.7–1.8 s

from the start of movement of the object. Horizontal dotted line indicates

a frequency of 80 imp/s. Vertical dotted line indicates the end of stimulus

approach, where the stimulus subtends an angle of 601 on the locust’s eye,

70mm from the theoretical point of collision. DCMD spike activity

remains higher during the time indicated by the grey area in jump rather

than non-jump behaviour.

R. Stafford, F. Claire Rind / Neurocomputing 70 (2007) 1079–1084 1083

than around 1.8 s after the start of the stimulus, defined asa point of interest by the EA. Although differences can beseen between jump and non-jump trials within the timewindow before 1.8 s after the onset of stimulus movement,identifying this time period as important would be difficultdirectly from the figures due to the similarity of what mayappear more important features of the spike train else-where. Furthermore, the spike frequency of non-jumpspike trains can often exceed 80 imp/s during this periodprior to 1.8 s after the onset of stimulus movement, but theduration of this high spike frequency tends to be lowerthan for jump behaviours. Thus, separating jump andnon-jump behaviours using peri-stimulus plots may bedifficult.

A potential limitation of the technique is the identifica-tion of features of the spike trains that occur throughchance differences. This is particularly important ifthe number of spike trains to be analysed is small.Where possible, large numbers of spike trains shouldbe used to identify differences in behaviour. In thisstudy, although the number of spike trains analysed wassmall, the similarities between the two different locusts ofsome key parameters, such as high-frequency spikes480 imp/s occurring 0.27 s before the point of impact,indicates these are likely to be important parameters indetermining whether a locust jumps to a looming stimulusor not.

The information obtained from the EA can be used infurther studies to try and identify key interneurons ormotor neurons that synapse with the DCMD and respond,

producing action potentials, to the types of inputsdesignated as important from the results of the EA.

4. Conclusions

The technique presented in this study is useful inindicating important parameters within the spike trains ofa single neuron that may trigger or partly triggerbehavioural responses. This study cannot directly concludeany important characteristics of spike trains that may causejumping behaviour in locusts. It can, however, greatly aidfurther investigations into the potential trigger for anescape jump. It is likely that the DCMD neuron responsethat may fully or partially trigger a jump occurs around0.27 s before collision for a 1m/s looming visual stimulus.Analysis of the spike train around this time period injumping locusts suggests that a burst of spikes, possibly�80 imp/s in frequency lasting for around 8–10 spikes, orpossibly more spikes of a lower frequency or fewer spikesof a higher frequency, may be important in jumptriggering. Also, as part of the burst of spikes, a highinstantaneous frequency 4200 imp/s may be important intriggering jumps. DCMD neuron spike train characteristicsof locusts during non-jumping trials are likely to differ inthese respects from DCMD neuron spike trains thatproduce jumping behaviour.The identification of these characteristics of DCMD

spike trains in further records of jump and non-jumpbehaviour will be important in assessing the generality ofthis result. Furthermore, the identification of suitablepostsynaptic neurons to the DCMD, which can summateEPSPs at the predicted frequencies and produce actionpotentials as a result, will result in a complete picture of theneural escape jump pathway of locusts.

Acknowledgements

Funded by the EU (LOCUST – IST-2002-38097) and theBBSRC.

References

[1] M. Burrows, C.H.F. Rowell, Connections between descending visual

interneurons and metathoracic motoneurons in the locust, J. Comp.

Physiol. A 85 (1973) 221–234.

[2] E.W. Childs, The roles of the locust DCMD in collision detection,

Ph.D. Thesis, University of Newcastle-upon-Tyne, 1999.

[3] D.H. Edwards, W.J. Heitler, F.B. Krasne, Fifty years of a command

neuron: the neurobiology of escape behavior in the crayfish, Trends

Neurosci. 22 (1999) 153–161.

[4] E. Faulkenauer, A. Marchand, Clustering microarray data with

evolutionary algorithms, in: G.B. Fogel, D.W. Corne (Eds.),

Evolutionary Computation in Bioinformatics, Morgan Kaufmann,

San Francisco, 2003.

[5] J.R. Gray, J.K. Lee, R.M. Robertson, Activity of descending

contralateral movement detector neurons and collision avoidance

behaviour in response to head-on visual stimuli in locusts, J. Comp.

Physiol. A 187 (2001) 115–129.

ARTICLE IN PRESSR. Stafford, F. Claire Rind / Neuroc1084

[6] I.C. Gynther, K.G. Pearson, An evaluation of the role of identified

interneurons in triggering kicks and jumps in the locust, J.

Neurophysiol. 61 (1989) 45–57.

[7] H. Korn, D.S. Faber, The Mauthner cell half a century later: a

neurobiological model for decision-making?, Neuron 47 (2005)

13–28.

[8] LOCUST, Project IST-2001-38097: LOCUST, Deliverable D1,.

Model Parameterization and Optimization for Robust Operation,

2006. /www.imse.cnm.es/locustS[9] F.C. Rind, R.D. Santer, Collision avoidance and a looming sensitive

neuron: size matters but biggest is not necessarily best, Proc. Roy.

Soc. Lond. B 271 (2004) S27–S29.

[10] F.C. Rind, P.J. Simmons, Orthopteran DCMD neuron: a reevalua-

tion of responses to moving objects. I. Selective responses to

approaching objects, J. Neurophysiol. 68 (1992) 1654–1666.

[11] R.D. Santer, F.C. Rind, R. Stafford, P.J. Simmons, Role of an

identified looming-sensitive neuron in triggering a flying locust’s

escape, J. Neurophysiol. 95 (2006) 3391–3400.

[12] R.D. Santer, P.J. Simmons, F.C. Rind, Gliding behaviour elicited by

lateral looming stimuli in flying locusts, J. Comp. Physiol. A 191

(2005) 61–73.

Richard Stafford is a researcher in computational

biology and ecology. His work has included

simulations of ecosystems, populations, indivi-

duals, neural networks and evolutionary pro-

cesses, as well as the use of computational

techniques to analyse biological data.

omputing 70 (2007) 1079–1084

Claire Rind is a reader in invertebrate neurobiol-

ogy at the University of Newcastle-upon-Tyne.

Her research interests are the neural mechanisms

used by insects to detect motion, and the

resultant behaviours triggered by motion detec-

tion. She also applies her knowledge in these

areas to the development of artificial visual

systems.