Decision timing in the face of changing sensory input: behavior and neural correlates

(1)

International School for Ad vanced Stu d ies

Decision timing in the face

of changing sensory

input: Behavior and

neural correlates

Thesis su bm itted for the d egree of

“Doctor of Philosophy”

Candidate:

Ad ina Dru m ea

Supervisor:

Prof. Mathew E. Diam ond

Febru ary 2017

Cognitive N eu roscience sector

(2)

(3)

Abstract

Most real-life situ ations requ ire organism s to extract inform ation from incom ing stim u li to p red ict fu tu re events and , from the p red iction, to p recisely tim e the ap p ropriate m otor act. In the present stu d y w e d esigned a new behavioral task that requ ires su bjects (hu m ans and rod ents) to extract tem p oral inform ation characterizing a continu ou s stream of sensory inp u t and execu te a p recisely tim ed m otor act. Fu rtherm ore, w e record ed neu rons in the p rem otor cortex of rats p erform ing this task to investigate the involvem ent of this brain area in tim ing actions as a resp onse to incom ing stim u lation.

In ou r exp erim ent rats received vibrations on their w hiskers and resp ond ed by w ithd raw ing from the nose-p oke hole, w hile hu m ans received the stim u li on their fingertip s and resp ond ed by p ressing a bu tton. The stim u li w ere form ed by m u ltip lying p ink -noise velocity valu es by an envelop e sine w ave. Resp onses m ad e arou nd the p eak of the envelop e (40% of each cycle) w ere rew ard ed . The p aram eters of the envelop e (frequ ency, am p litu d e and p hase at stim u lu s onset) changed from trial to trial to ensu re that su bjects cou ld not set an absolu te am p litu d e threshold or u se tim ing alo ne (e.g. “w ait 1 second after stim u lu s onset”) to solve the task.

Rats and hu m ans learned to tim e their resp onses to the envelop e p eak at above-chance levels across d ifferent envelop e p aram eters. Both rats and hu m ans resp ond ed in later cycles in high frequ ency and low am p litu d e stim u li, su ggesting that these stim u li w ere m ore d ifficu lt and thu s requ ired integration of m ore evid ence to su p p ort the resp onse. Fu rtherm ore, rats benefited from collecting m ore inform ation abou t the stim u lu s, as show n by better -tim ed resp onses m ad e in the second than in the first cycle of stim u lation.

As exp ected , the activity of p rem otor cortex neu rons w as p red ictive of the imminence of the animal’s action, in the time period preceding the w ithd raw al. Moreover, neu rons carried inform ation regard ing the stim u lu s, w ith a large p rop ortion cod ing for the overall stim u lu s am p litu d e. A sm all p ercentage of the record ed p rem otor cortex neu rons also show ed a correlation betw een firing rate and the stim u lu s am p litu d e at any given p oint in th e trial.

(4)

The strategy rats w ere likely to u se for solving the task em erging from these resu lts w as to u nd erstand the global am p litud e of the trial and set an am p litu d e threshold against w hich to com p are the p erceived stim u lu s.

Interestingly, the activity of prem otor cortex neu rons at d ifferent m om ents in the trial w as correlated w ith the tim e at w hich the rat w ithd rew , carrying inform ation both regard ing how m u ch tim e the rat is w illing to w ait and how m u ch tim e has p assed since the stim u lation started . We u sed an artificial neu ral netw ork (AN N ) im p lem ented in MATLAB to p red ict w ithd raw al tim e from the firing rates at d ifferent tim e bins of all the neu rons record ed sim u ltaneou sly w ithin a behavioral session, and fou nd a good netw ork p erform ance in the tim e bins p reced ing the anim al’s action. Perform ance w as better in incorrect trials, ind icating that in som e trials rats only engaged in tim ing, w hile in others they p aid attention to the stim u lu s and d id not keep track of tim e.

In su m m ary, w e d esigned a new behavioral p arad igm to investigate how the brain tim es d ecisions in resp onse to changing incom ing sensory stim u lation. Both rats and hu m ans learned to align their resp onses to the p eak of the envelop e and chose to gather m ore stim u lu s inform ation in trials characterized by low am p litu d e and high frequ ency. Finally, neu rons in the p rem otor cortex of rats p erform ing the task carried signals related to key asp ects of the task: the tim e of w ithd raw al and stim u lu s p rop erties.

(5)

1

1. Introduction

1.1. D esigning a task to measure temporal signals w ithin noise

In ou r exp erim ent su bjects received a stream of noisy sensory stim u lation based on w hose p rop erties they w ere requ ired to tim e a m otor resp onse. The stochastic natu re of the stim u lu s m eant that su bjects had to accu m u late inform ation abou t the stim u lu s in tim e.

Previou s research has been aim ed at u nd erstand ing how the brain accu m u lates evid ence abou t a noisy stream of sensory inp u t across tim e. For instance, in the rand om d ot m otion d iscrim ination (RDMD) task, su bjects have to ju d ge the d irection of m ovem ent of d ots p resented on a screen and the d ifficu lty is varied by changing the p ercentage of coherently m oving d ots (Figu re 1.1.A). Theoretical m od els, su ch as the d rift d iffu sion m od el (Sm ith, 2000) consid er that evid ence su p p orting one hyp othesis abou t the stim u lu s (e.g. its d irection of m otion) is accu m u lated in a d ecision variable that d rifts in tim e tow ard s a d ecision bou nd ary (Figure 1.1.B). The d ecision is m ad e once the bou nd ary has been reached . Accu m u lation of evid ence in tim e w as evid ent from increased p erform ance for longer stim u lu s d u ration. Moreover, w hen allow ed to collect stim u lu s inform ation at their ow n p ace, su bjects w ait ed m ore tim e before resp ond ing and w ere less accu rate low er coherence than in higher coherence trials. These effects have been revealed in hu m an (Watam aniu k & Seku ler, 1992), p rim ate (Roitm an & Shad len, 2002) and rod ent stu d ies (Dou glas, N eve, Qu ittenbau m , Alam , & Pru sky, 2006; Reinagel, Mankin, & Calhou n, 2012).

(8)

4

Figu re 1.1: A. Rand om d ot m otion d iscrim ination task (RDMD) requ ires su bjects to ju d ge the d irection of m otion of d ots p resented on the screen (Joshu a I. Gold & Ding, 2013). B. Diffu sion m od els, su ch as the rand om w alk m od el cou ld exp lain the behavior in tasks su ch as the RDMD. Evid ence is accum u lated at a rate d ep end ing on its strength (p ercentage of coherently m oving d ots), and the d ecision is m ad e once the accu m u lated evid ence reaches a threshold (Joshua I Gold & Shad len, 2007).

While the u se of noisy stim u li has been valu able in show ing how the brain m ight accu m u late evid ence to red u ce u ncertainty, the d ecision to be m ad e in su ch cases concerns the p rop erties of the stim u lu s, not tim ing. In short, a rand om d ot exp erim ent can exp lore how the brain d eterm ines “w hat” bu t not how the brain determines “when”. Our experiment originated with the idea of an u nd erlying rhythm , albeit u ncertain d u e to the noisy ch aracter. The rhythm allowed us to formulate a “when” question: when does the cyclical input reach a p eak? A few sim ilar exp erim ental p arad igm s have been p reviou sly p erform ed , only engaging hu m an su bjects. For exam p le, p articip ants w ere requ ired to pred ict the tim e of occu rrence of an au d itory stim u lu s p art of a rhythm ic tone p resentation (Arnal, Doelling, & Poep p el, 2014).

Figu re 1.2: Exam p le of a behavioral task that requ ired hu m ans to extract the tem p oral p attern of a sequ ence of tones and d ecid e w hether the last one w as d elayed w ith regard to the beat (Arnal et al., 2014).

A

(9)

5

The task w e d esigned in the cu rrent stu d y com bines the accu m u lation of evid ence in tim e to the extraction of tem p oral inform ation abou t the stim u lu s and ad d s a new d im ension: the p rep aration of a p recisely tim ed m otor act in resp onse to the incom ing stim u lu s.

1.2. The choice of stimuli

The stim u li u sed in ou r exp erim ent w ere tactile vibrations d elivered on the w hiskers and fingertip s of rats and hu m ans resp ectively.

Figu re 1.3: The vibrotactile w orking m em ory task requ ires su bjects to p erceive 2 consecu tive vibrations and com p are their strength, w hich is qu antified by the stand ard d eviation of the velocity valu es of each vibration (A). Psychom etric cu rves of rats (B) and hu m ans (C) p erform ing this task (Fassihi, Akram i, Esm aeili, & Diam ond , 2014).

Tactile stim u li are p articu larly u sefu l w hen p erform ing exp erim ents on rod ents, w ho have a highly d evelop ed sense of tou ch on w hich they rely for

A

(10)

6

su rvival. Moreover, the tactile p ercep tion of rod ents and hu m ans are highly com p arable (Diam ond , 2010). The behavior of hu m ans p erform ing the sam e tasks as rats exhibit very sim ilar resp onse p atterns, su ggesting that rod ent stu d ies are a u sefu l tool for u nd erstand ing how su ch p rocesses take p lace in the hu m an brain (Fassihi et al., 2014) (Figu re 1.3. B, C).

Recent stu d ies requ ired rats to m ake d ecisions based on the p rop erties of vibrations received p assively throu gh the w hiskers (Ad ibi, Diam ond , & Arabzad eh, 2012; Fassihi et al., 2014) (Figu re 1.3. A). Su ch stim u li allow the exp erim enter to exercise a strict control over the p aram eters of the stim u li entering the sensory system . The u se of vibrations is also ecologically relevant, as rod ents are bu rrow ing anim als and cou ld u se inform ation extracted from the earth vibrations to u nd erstand the size and d irection of m ovem ent of an ou tsid e p red ator. Fu rtherm ore, the brain m u st d ecod e the w hisker vibrations resu lting from rat’s sweeping movements over textures in order to determine the coarseness of the texture (Lottem & Azou z, 2008, 2009).

1.3. The choice of the brain area to be examined

The brain area w e focu sed on in the p resent stu d y is the rat p rem otor cortex, consid ered to be the analogu e of the p rim ate p rem otor cortex (Cond é, Maire-lep oivre, Au d inat, & Crép el, 1995; H arry M. Sinnam on, 1984; Jam es V. Corw in, 1998; Roger L. Reep , Jam es V. Corw in, Atsu taka H ashim oto, 1984)

The p rim ate p rem otor cortex has been trad itionally view ed as d ed icated to m ovem ent p rep aration (Cru tcher & Alexand er, 1990; Gentilu cci et al., 1988; Riehle & Requ in, 1989; Tanji, Tanigu chi, & Saga, 1980), bu t new fu nctions have em erged for m otor and p rem otor brain areas, su ch as stim u lu s categorization (R Rom o, Ru iz, Cresp o, Zainos, & Merchant, 1993), evid ence accu m u lation (Liu & Pleskac, 2011) and d ecision m aking (H ernánd ez, Zainos, & Rom o, 2002; Ranu lfo Rom o, H ernánd ez, & Zainos, 2004; Ranu lfo Rom o, H ernánd ez, Zainos, Lem u s, & Brod y, 2002).

The rat p rem otor cortex has been show n to contribu te to d ecision m aking by conveying significant d ecision valu e and chosen valu e signals before and after a choice w as m ad e, resp ectively. Therefore, it m ight be p art of the n eu ral system w here actions are selected and prop agated to d ow nstream m otor

(11)

7

stru ctu res for execu tion (Su l, Jo, Lee, & Ju ng, 2011). Rat p rem otor cortex has also been show n to be cru cial for orienting m ovem ents in a m em ory -gu id ed task, as neu rons resp ond selectively to contralateral or ip silateral m ovem ents, and u nilateral inactivation of this area im p airs contralateral orienting m ovem ents (Erlich, Bialek, & Brod y, 2011). Even if su ch tasks requ ire su bjects to accu m u late evid ence abou t the stim u lu s in tim e, rat prem otor cortex has been show n to not be d irectly involved in evid ence accu m u lation, bu t to represent the choice the rat w ou ld m ake based on the evid ence accu m u lated thu s far (Figu re 1.4. A) (H anks et al., 2015). Sim ilarly, the m ou se anterior lateral m otor cortex (ALM), p roposed to be the analogu e of p rim ate p rem otor cortex has been show n to be involved in p lanning licking. N eu rons in this area show resp onse p reference for contralateral or ip silateral licks, w hile ALM inactivation affects only contralateral licking m ovem ents (Gu o et al., 2014; Li, Chen, Gu o, Gerfen, & Svobod a, 2015).

Figu re 1.4: Fu nctions of rat p rem otor cortex. A. The firing rate of a p op u lation of neu rons in the p rem otor cortex ram p ed u p at a rate p rop ortional to the strength of the evid ence in a task w here rats had to d ecid e w hich of tw o sp eakers d elivered higher frequ ency au d itory stim u li (H anks et al., 2015). B. Exam p le p rem otor cortex neu ron ram p ing in the absence of any sensory stim u li. The rat resp ond ed w hen the neu ron ’s firing rate reached a certain threshold . C. Exam p le neu ron w ith transient activation p red icting how m u ch the rat w ill w ait in the nose p oke in the absence of incom ing stim u lation (Mu rakam i, Vicente, Costa, & Mainen, 2014).

Rat p rem otor cortex neu rons have also been show n to be involved in keep ing track of the tim e a rat is w aiting for an incentive. In a task w here rats cou ld give u p w aiting for a large rew ard in favor of a sm all rew ard , a p rop ortion of p rem otor cortex neu rons transiently increased or d ecreased their activity at d ifferent tim e p oints in the trial, w ith their firing rate prop ortional to the tim e the rat w aited (Figu re 1.4.C), w hile other neu rons grad u ally increased

(12)

8

or d ecreased their firing rates d u ring the trial, reaching a firing threshold ju st before the animal’s response (Figure 1.4.B). The authors suggested that the resu lts cou ld be exp lained by a neu ral integration p rocess w here the first kind of neu rons are the inp u t to the ram p ing integrator neu rons that trigger an action w hen reaching a threshold (Mu rakam i et al., 2014). N eu rons in this area have been also linked w ith tim e m easu ring in a task in w hich d ifferent stim u li ind icated the interval after w hich the rew ard w ou ld be available. As rats w aited for the rew ard , the resp onse p atterns of m otor cortex neu rons cou ld be d escribed as ram p s, p eaks and d ip s and p rovid ed su fficient inform ation to d iscrim inate the d elay d u ration (Matell, 2012).

1.4. Aim of the project

The cu rrent p roject is aim ed at u nd erstand ing the brain m echanism s resp onsible for d ecid ing the precise tim ing of one’s actions. Su bjects in this stu d y received noisy tactile vibrations based on w hich they ha d to p lan their m otor resp onse. More p recisely, the stim u li w ere m od u lated by a sinu soid al w ave and a correct resp onse (bu tton p ress for hu m ans or n ose p oke w ithd raw al for rats) w as consid ered one at the p eak of the stim u lu s. For each trial su bjects had to extract the stim u lu s p rop erties (sine am p litu d e and frequ ency) in ord er to p red ict w hen the next p eak shou ld occu r and p lan the tim ing of their m otor resp onse.

The brain area of interest w as the rat p rem otor cortex. The p rem otor cortex has been p reviou sly show n to be involved in sensory d ecision m aking as w ell as tim e p ercep tion and action p lanning, m aking it a good cand id ate for su p p orting the tim ing of resp onses to incom ing sensory stim u li.

(13)

9

2. Methods

2.1. Animal subjects

Five Wistar m ale rats (H arlan Laboratories, Italy) w ere u sed in this stu d y. At the start of the exp erim ents the anim als w ere 6-8 w eeks old . Rats w ere hou sed in p airs and m aintained on a reversed 12/ 12 hou rs d ark/ light cycle w ith ad libitum food , bu t w ater restricted d u ring the exp erim ental p eriod .

Protocols w ere in accord ance to international norm s and w ere ap p roved by the Italian H ealth Ministry and the Ethics Com m ittee of the International School for Ad vanced Stu d ies.

2.2. Apparatus

The ap p aratu s consisted of a 25×25×38 cm (H ×W×L) p lexiglass cham ber, cu stom -m ad e by the SISSA Mechatronics Lab (Fig u re 2.1A). The front w all consisted of a central head hole op ening throu gh w hich r ats cou ld access the nose-p oke hole. The nose-p oke w as a 0.7 cm d iam eter op ening w ith an infrared sensor to detect the animal’s presence (Figure 2.1B). On top of the opening a green LED light w as placed to signal to the anim al w hen it cou ld initiate a new trial. Anim als received vibrations on the vibrissa e bilaterally by vibrating p lates connected to m otors (Brü el & Kjæ r Typ e 4809 shakers). Sticky tap e w as attached to the p lates to ensu re a better ad herence of the w hiskers. Rats received w ater rew ard throu gh d rinking sp ou ts situ ated on one sid e. Anim al licking w as d etected by infrared sensors in the d rinking sp ou t and triggered the activation of a syringe p u m p d elivering the rew ard (Fig u re 2.1C). Tw o sp eakers m ou nted on the sid e w alls of the ap p aratu s d elivered au d itory cu es. One cu e w as activated w hen the nose-p oke sensor stop p ed d etecting the anim al’s nose (resp onse cu e), inform ing the rat of the tim e at w hich a resp onse w as registered . The other cu e signaled the rew ard d elivery (rew ard cu e), w orking as an au d itory reinforcem ent for the rat.

(14)

10

Figu re 2.1: Photo of the exp erim ental setu p from above (A), p oint of view of the rat (B) and p u m p for d elivery of w ater rew ard (C).

2.3. Stimuli

The stim u li w ere form ed by m u ltip lying a velocity noise vibration b y an envelop e sine w ave (Figu re 2.2A) and w e refer to them as sine-m od u lated noisy vibrations. First a noisy vibration w as obtained by choosing p robe p osition valu es from a norm al d istribu tion w ith 0 m ean and stand ard d eviation of 1 m m , u sing a sam p ling rate of 10.000 valu es p er second . The signal w as then low p ass filtered w ith a 150 H z Gau ssian filter and m u ltip lied by a low frequ ency envelop e w ith valu es betw een 0 and 1. The envelop e w ave p aram eters changed from trial to trial. The frequ encies u sed w ere 0.7 and 1 H z. The d ifference betw een the base and the p eak of the sine w ave w as constant, and qu antified by the am p litu d e d ifference ind ex ADI=0.5, w here ADI= (Peak am p litu d e-Base am p litu d e)/ (Peak am p litu d e + Base am p litu d e). Tw o envelop e am p litu d es w ere chosen, and w e d efined high am p litu d e as envelop e valu es betw een 0.33 for valley and 1 for p eak; w hile low am p litu d e envelop e w as in the range 0.11 - 0.33. Finally, tw o sine w ave p hases w ere chosen, as stim u li cou ld start at the valley or p eak.

A

B

(15)

11

Figu re 2.2: Creating the stim u lu s. A. A sine m od ulated noise vibration w as obtained by m u ltip lying a velocity noise vibration w ith a sine w ave of d ifferent am p litu d es, p hases and frequ encies. B. In early p hases of training stim u li w ere m od u lated by a step w ave, lead ing to tw o levels of am p litu d e.

2.4. Task

Rats w ere trained to d etect the p eak of the sine m od u lated noisy vibration, and resp ond by w ithd raw ing from the nose-p oke (Figure 2.3). On a given trial the id entical stim u lu s w as p resented on w hiskers on both sid es of the snou t, and continu ed w hile the nose-p oke sensor w as activated , allow ing the anim al to collect as m u ch inform ation as it chose.

Figu re 2.3: Tim eline of a trial. A trial starts as the rat enters the nose p oke. After a short d elay the stim u lu s is d elivered . Once the anim al w ithd raw s from the nose p oke the stim u lu s stop s and the rat has to tu rn to the rew ard sp ou t w here, if the w ithd raw al w as correct, the w ater rew ard is d elivered .

(16)

12

A green LED light signaled the p ossibility to start a new trial, and tu rned off at the tim e of the nose p oke. Each trial w as initiated by the rat by p lacing its snou t in the nose-p oke hole. After a short d elay (rand om ly d raw n from a Gau ssian d istribu tion w ith m ean 250 m s and stand ard d eviation of 50 m s) the stim u lu s d elivery started . The stim u lu s w as p resented continu ou sly u ntil the anim al w ithd rew from the nose-p oke, at w hich tim e the rat received an acou stic w ithd raw al cu e. One single sp ou t d elivered the rew ard after a d elay from the first lick. The d elay w as rand om ly selected from a u niform d istribu tion betw een 100 and 150 m s, and the rew ard d elivery w as accom p anied by a rew ard sou nd . The rew ard ed tim e w ind ow covered 40% of each cycle, and w as centered on the p eak. To d iscou rage very early resp onses, the rew ard for the first cycle w as 0 for stim u li starting at the p eak and 50% of the total rew ard if the stim u li started at the valley. The next trial cou ld be initiated im m ed iately (100 m s inter trial interval), bu t for som e rats a larger d elay (5 second s) w as im p os ed after incorrect trials. Anim als p erform ed on average 270 trials in each session.

Different p aram eters of the envelop e sine w ave w ere u sed to ensu re that rats p aid attention to the stim u lu s in each trial and d id not u se alternative strategies (for instance, fixed , stim u lu s-ind ep end ent w aiting tim e) for solving the task. Stim u li cou ld have tw o levels of am p litu d e ensu ring that rats d id not p erform the task by resp ond ing w hen stim u li reached a fixed am p litu d e threshold . The p eak of the low am p litu d e stim u li w as as intense as the valley of the high am p litu d e stim u li, so no u niversal threshold cou ld be ap p lied for p erform ing the task. Moreover, the u se of d ifferent frequ encies and sine p hases cau sed the rew ard ed p eriod s to occu r at d ifferent tim es in d ifferent trials, d iscou raging a strategy w here anim als cou ld tim e their actions and resp ond after a fixed tim e interval.

2.5. Rat training procedure

Before training, anim als w ere habitu ated to the exp erim enter throu gh a 30 m inu tes hand ling session p er d ay for 5 d ays. The first training step consisted of learning to activate the nose p oke for a w ater rew ard . The tim e rats w ere requ ired to sp end in the nose-p oke w as grad u ally increased , bu t there w as no cu e ind icating the m om ent at w hich a resp onse cou ld be m ad e. Stim u li w ere introd u ced w hen rats cou ld w ait for m ore than one second in the nose p oke.

(17)

13

First step -m od u lated noisy vibrations w ere d elivered (Figu re 2.2B). In this case, the velocity noise vibration w as m u ltip lied w ith a vector w ith only 2 am p litu d e levels. Rats had to w ithd raw in the high states of the stim u lu s to receive w ater. Once they p erform ed above chance on the step -m od u lated stim u li (>60% correct) the sine-m od u lated stim u li w ere introd u ced (Fig u re 2.2A).

2.6. Surgery

For chronic su rgeries Isoflu ran e (1.5-2.5%) anesthesia w as d elivered throu gh a snou t m ask. In p rep aration for the su rgery the anim al’s fu r w as shaved w ith a razor and its head w as fixed in the N arashige stereotaxic ap p aratu s. The anim al w as p laced on a heated p ad , and its tem p eratu re w a s constantly m onitored w ith a therm om eter inserted in the anal op ening. Ep igel op hthalm ic m oistu rizing ointm ent w as ap p lied to p revent d rying of the eyes, and lid ocaine gel w as u sed as a local anesthetic on the skin p reced ing the incision.

First, the skin on top of the anim al’s sku ll w as cu t, and the connective tissu e w as rem oved . N ext, 3 screw s w ere inserted in the bone, in contact w ith d u ra m ater. These screw s have a d ou ble role of fixing the im p lant and connecting to the reference electrod e. A craniotom y w as d rilled accord ing to know n coord inates of p rem otor cortex (center +2AP, ±1.3ML from Bregm a) (Erlich et al., 2011). After rem oving the skull covering the craniotom y, d u ra m ater w as also rem oved u sing a bent need le. Once exp osed , the brain w as constantly w ashed w ith p hosp hate bu ffer solu tion (PBS).

In ord er to avoid d im p ling of the brain, a sm all d rop of Vaseline -based ointm ent w as p laced in the m id d le of the op ening, and bio com p atible glu e w as ap p lied on the ed ges of the craniotom y. Electrod es w ere slow ly low ered in the brain, w hile grad u ally w rap p ing the grou nd / reference w ires arou nd the screw s fixed in the sku ll. The p resence of neu rons w as m onitored online u sing a TDT record ing system . Once the d esired d ep th w as reached (800-1200 m m ), the craniotom y w as covered w ith silicone, and d ental cem ent (Secu re Starter Kit, Su n Med ical) w as u sed to cover the area w here the sku ll w as exp osed .

Rym ad il (5m g/ kg) analgesic w as injected intram u scu larly one hou r after the anesthesia onset, and at the end of su r gery. An antibiotic (Baytril, 5 m g/ kg)

(18)

14

w as injected su bcu taneou sly before the anim al w as aw oken, and also d elivered throu gh the w ater bottle in the 48 hou rs follow ing the su rgery.

After the su rgery rats had a w eek of recovery tim e d u ring w hich w ater and soft food w ere available ad libitum.

At the end of the exp erim ent rats w ere sed ated w ith Urethane (1.5 m g/ kg) and transcard ially p erfu sed w ith 0.1% p hosp hate bu ffer solu tion follow ed by 4% p araform ald ehyd e. The brain w as then rem oved and p laced in p araform ald ehyd e at 4oC for 24 to 48 hou rs, and then transferred to a su crose solu tion (15% to 30%). Finally, the brain w as sliced w ith a 25 µm thickness u sing a m icrotom e and stained w ith N issl solu tion.

2.7. Electrophysiological recordings

Rats w ere im p lanted w ith Tu cker-Davis Zif-Clip based 32 m icrow ire arrays. Record ed d igital signals w ere p assed throu gh a Tu cker -Davis ACO-32 com m u tator to prevent w ires from tangling w hen the rat tu rned arou nd in the cage, and then a PZ-4 connection m anifold . N ext, the signal w as transm itted throu gh optical cables to a RZ2 BioAm p Processor.

A cu stom m ad e Op enEx circu it w as u sed to m onitor the neu ral activity online, and save raw d ata for fu rther processing. Together w ith the neu ral d ata, the behavioral ep ochs w ere also saved (nose p oke, w ithd raw al, lick and m otor trigger tim es).

2.8. Human psychophysics

H u m an p articip ants w ere tested on a m od ified version of the task u sed for rats. They felt the stim u li on the fingertip of their left ind ex finger, and resp ond ed by p ressing a bu tton p laced in their right hand . Each su bject p erform ed 10 sessions of 48 trials each. For each session the p articip ant w as instru cted to resp ond at the p eak or at the valley of the stim u lu s (5 p eak and 5 valley sessions, p resented p seu d orand om ly).

The stim u li w ere created in the sam e m anner as the stim u li u sed for rats bu t w ith d ifferent p aram eters inasm u ch as p erform ance w ou ld be nearly p erfect w ith the rat p aram eters. We tested 3 envelop e frequ encies: 0.35, 0.7 and

(19)

15

1.4 H z and tw o levels of am p litu d e. ADI w as fixed at 0.5 and all trials started at the valley of the stim ulu s. Once the p articip ants m ad e a resp onse, a front p anel LED tu rned green or orange, ind icating a correct or incorrect resp onse.

2.9. D ata analysis

All d ata analysis w as p erform ed u sing Matlab (Mathw orks) scrip ts. Sp ike sorting w as p erform ed offline u sing the UltraMegaSort 2000 algorhythm (Daniel N . H ill, Mehta, & Kleinfeld , 2011) im p lem ented in Matlab. Tw o typ es of trials w ere exclu d ed from analysis of neu ral activity: first, trials in w hich the rat retu rned to the nose p oke w ithin 50 m s from w ithd raw al, in w hich case w e consid ered the registered resp onse to be d u e to the anim al’s shaking, and not to the intention to w ithd raw ; second , trials in w hich the anim al d id not start licking the d rinking sp ou t w ithin 4 second s from w ithd raw al.

To calcu late the correlation betw een stim u lu s am p litu d e and firing rate at every tim e p oint in the trial w e d ivid ed each trial into 200 m s tim e bins and calcu lated the Sp earm an correlation coefficient betw een firing rate and average envelop e am p litu d e in all bins. In ord er to avoid any bias cau sed by neu rons changing firing rate preced ing the w ithd ra w al, from the firing rate of each tim e bin the average firing rate of the neu ron for all trials in that tim e bin w as su btracted . Only tim e bins w here at least 10 trials w ere record ed and therefore u sed for averaging w ere consid ered for analysis. Sim ilarly, w hen calcu lating the firing rate over the w hole trial, w e su btracted the average firing rate of all trials.

Phase coherence ind ex w as com p u ted in ord er to check if the firing of p rem otor cortex neu rons w as p hase locked to the envelop e sine w ave. The p hase coherence ind ex w as calcu lated as follow s:

𝑃𝐶𝐼 =∑ 𝑒

𝑖𝑃𝑗 𝑁 𝑗

𝑁

Where PCI=p hase coherence ind ex, i=im aginary u nit, N =nu m ber of sp ikes over all trials, j is the ind ex of sp ikes p ooled across all trials, and Pj=p hase of firing of sp ike j (w here the p hase is the angle of the H ilbert transform of the sine w ave). To select neu rons w ith significant p hase coherence (p <0.01), the obtained p hase correlation ind ices w ere com p ared to the nu ll

(20)

16

d istribu tion of p hase coherence ind ices given the nu m ber o f trials record ed for each neu ron.

Artificial neu ral netw ork (AN N ) analysis w as p erform ed u sing the N eu ral N etw ork toolbox in Matlab. In all com p u tations, a netw ork w ith zero hid d en layers w as u sed . The inp u t of the netw ork w as alw ays the firing rate of neu rons record ed in the sam e session at a sp ecific tim e p oint, and the target cou ld be the w ithd raw al tim e, or the tim e bin to w hich the firing rate belonged . Before ap p lying AN N the inp u ts and targets w ere z scored .

In ord er to pred ict the w ithd raw al tim e from the neu ronal firing rate the fitting tool w as u sed . The netw ork p erform ance w as calcu lated as the m ean squ ared error betw een the target and the netw ork ou tp u t. In ord er to avoid a low netw ork p erform ance cau sed by a low ratio of nu m ber of inp u ts to n u m ber of trials, for each session and tim e bin only the 5 neu rons w ith highest correlation coefficient betw een firing rate and w ithd raw al tim e w ere consid ered .

The fitting tool w as also u sed for p red icting w hich tim e bin the firing rate belonged to. The inp u t of the netw ork w as the firing rate in consecu tive 300 m s tim e bins, either starting w ith the start of the stim u lation or lead ing to the w ithd raw al, and the ou tp u t w as the bin nu m ber 1, 2, or 3.

Finally, to classify if a firing rate belonged to the tim e bin ju st before the start of the stim u lation or another tim e bin d u ring the trial the p attern recognition tool w as u sed . The netw ork w as trained on the classification betw een the tim e bin before stim u lu s start and the tim e bin p reced ing the w ithd raw al and then tested on the p reviou s 2 tim e bins before w ithd raw al. The sam e analysis w as p erform ed training the netw ork to d iscrim inate betw een the tim e bin before start and the tim e bin ju st after start, and then tested on the 2 su bsequ ent bins.

N etw ork test p erform ance w as alw ays calcu lated by u sing the leave-one-ou t cross valid ation m ethod . For each session AN N w as ap p lied a nu m ber of tim es equ al to the nu m ber of trials in that session. Each tim e the netw ork w as trained w ith all trials bu t one, and tested on the left ou t trial. Test p erform ance w as com p u ted by calcu lating the p erform ance (m ean squ are error or cross entrop y) given the targets and the ou tp u ts of the test trials. The exp ected netw ork p erform ance in the absence of any inform ation from neu ronal fir ing rate w as 1, w hich is the stand ard d eviation of the z scored inp u t.

(21)

17

When netw ork p erform ance w as calcu lated for correct and incorrect trials sep arately the sam e nu m ber of trials w as consid ered for the 2 grou p s. To d o so, for every session a nu m ber of trials equ al to the nu m ber of incorrect trials w ere selected rand om ly from the correct trials.

(22)

18

3. Results

3.1. Rat behavior

Five rats w ere trained to d etect the p eak of sine m od u lated noisy vibrations received throu gh their w hiskers, and fou r of them achieved above chance p erform ance for 10 or m ore consecu tive sessions. The follow ing behavioral analyses are p erform ed on 10 sessions record ed from each rat once it reached stable p erform ance (Figu re 3.1). Trained rats learned to align their w ithd raw als to the p eak of the stim u lu s, and resp ond ed m ostly in the first or second envelop e cycle (Figu re 3.2, Figu re 3.4A). The m ean p erform ance (p ercentage rew ard ed resp onses) of each rat w as 54.5, 57.1, 70.7 and 50.6 for the 4 rats AD1, AD2, AD4 and AD5, all significan tly higher than chance. Theoretical chance level w as 40%, corresp ond ing to the p ercentage of each cycle that w as rew ard ed . Chance levels w ere also com p u ted for each rat sep arately, by shuffling the animal’s response times with respect to the stimulus p aram eters. The valu es of these calcu lated chance levels w ere 40.1, 41.2, 40.6 and resp ectively 41.4 for the 4 rats.

Figu re 3.1: Perform ance (% rew ard ed resp onses) in all sessions of all rats. Each d ot rep resents one session, the central m ark of the box p lo t ind icates the m ed ian p erform ance valu e, the ed ges are the 25th

and the 75th

p ercentile, and the w hiskers extend to the m ost extrem e valu es. The grey d otted line corresp ond s to the theoretical chance level.

(23)

19

Figu re 3.2: Exam p les of w ithd raw als for one rat (AD4) to all trials of frequ ency 0.7 H z (A) and 1 H z (B) starting at the valley. The green bars rep resent the intervals in w hich w ithd raw al w ou ld be rew ard ed .

Rat behavior d ep end ed on the p aram eters of the envelop e sine w ave characterizing each trial. Most im p ortantly, rats show ed better p erform ance in trials w here the higher am p litu d e envelope w as d elivered (48.5% vs 62.9%, p <0.01, Welch t-test on ranks) (Figu re 3.3), w hile the envelop e frequ ency and p hase d id not influ ence p erform ance.

Stim u lu s p aram eters also had an effect on how long rats w ere w illing to w ait, and therefore on the cycle in w hich they m ad e their resp onses. Rats resp ond ed in later cycles in trials characterized by low am p litu d e and high frequ ency (p <0.01, Welch t-test on ranks) (Figu re 3.4A), su ggesting that these trials requ ired them to accu m u late m ore evid ence before resp ond ing. H ow ever, there w as no significant d ifference betw een the absolu te w aiting tim es in 0.7 versu s 1 H z trials (Figu re 3.4B). The absence of effect of envelop e frequ ency and p hase on p erform ance, bu t their effect on the nu m ber of the resp onse cycle ind icates that rats ad ju sted their w aiting tim es to achieve a satisfactory p erform ance.

B

A

(24)

20

Figu re 3.3: Perform ance of all rats in trials of d ifferent am p litu d es and frequ encies. Each d ot rep resents one session. The grey d otted line corresp ond s to the theoretical chance level (40%).

Figu re 3.4: Stim u lu s p aram eters effects on w ithd raw al tim e. A. Mean nu m ber of the cycle in w hich the w ithd raw al w as m ad e. B. The average w ithd raw al tim e. Data from the 4 rats are show n, each d ot rep resenting one session.

To test w hether resp ond ing in later cycles benefit ted p erform ance, w e com p ared w ithd raw als m ad e in the first and second cycles in all trials starting at the valley. Ou r resu lts show that w ithd raw als m ad e in the second cycle w ere closer to the p eak than those m ad e in the first cycle (Figu re 3.5). The m ean w ithd raw al tim e, w here 0 is the tim e of the nearest p eak, w as 165.16 for the first cycle and -30.876 m s for the second cycle for 1 H z stim u li and 224.99 and resp ectively -202.04 for 0.7 H z stim u li. Moreover, for the 1 H z stim u li, the stand ard d eviation of resp onse tim e to the second cycle w as significantly low er

(25)

21

(222.5 vs 241.2 m s, p <0.01, t-test on bootstrap p ed stand ard d eviations), show ing that w ithd raw als w ere better aligned to the p eak of the stim u lu s in trials w here rats resp ond ed in the second cycle.

Figu re 3.5: H istogram of w ithd raw als m ad e in the first (blu e) and second (red ) cycle for all rats in trials starting at the valley. 0 on the x axis rep resents the tim e of the envelop e sine w ave p eak. (A) frequ ency 1 H z. (B) frequ ency 0.7 H z.

3.2. Human psychophysics

H u m an p articip ants (total 14: 6F, 8M) p erform ed the behavioral task and achieved p erform ances better than chance both in sessions w here they w ere instru cted to p ress the bu tton at the p eak and in sessions w hen they had to resp ond at the valley.

Overall p erform ance w as not significantly d ifferent for p eak and valley sessions (m ean p erform an ce 87.0 on p eak vs 84.7 on valley sessions, p = 0.1678, Welch t-test on ranks). H ow ever, p articip ants resp ond ed in later cycles in valley trials (m ean cycle of resp onse 5.78 vs 5.06; p = 0.0016), ind icating that these trials requ ired m ore evid ence accu m u lation for m aking a d ecision. Sim ilar to rat behavior, in both peak and valley trials hu m ans p erform ed better and resp ond ed in earlier cycles in high am p litu d e trials (Figu re 3.7). Moreover, hu m ans w aited for m ore cycles and had low er p erform ance in higher fre qu ency trials, show ing that it w as m ore d ifficu lt to tim e their d ecisions w hen the stim u lu s am p litu d e w as changing at a fast rate.

(26)

22

Figu re 3.6: Behavior of one hu m an su bject (S1) to all trials w here the instru ction w as to p ress the bu tton at the p eak (A) or valley (B) of the stim ulu s in all trials w here the envelop e frequ ency w as 0.35 H z. The green bars rep resent the intervals in w hich w ithd raw al w ou ld be rew ard ed .

Figu re 3.7: (A) Mean perform ance in all sessions w hen hu m ans w ere instru cted to resp ond at the p eak of the stim u lu s. For visu alization p u rp oses, each p erform ance p oint w as shu ffled by ad d ing a valu e d raw n rand om ly from a u niform d istribu tion betw een -5 and 5. (B) Mean cycle of bu tton p ress for all hu m ans.

A

B

A

BUTTON PRESSES BUTTON PRESSES

BUTTON PRESS HISTOGRAM BUTTON PRESS HISTOGRAM STIMULUS STIMULUS

(27)

23

3.4. N europhysiology

A total of 214 neu rons w ere record ed in 34 sessions from p rem otor area (Figu re 3.8) of 2 rats (AD2 and AD4) p erform ing the task.

Figu re 3.8: Coronal section of the brain of one of the record ed rats (AD2) at 4.2 m m anterior from Bregm a. The location of the electrod e tip s are show n w ith red d ots.

Prem otor cortical neu rons exhibited heterogeneou s resp onse p atterns, w ith firing rates changing at d ifferent tim es in the trial, su ch as the nose p oke tim e (Figu re 3.9A), at w ithd raw al (Figu re 3.9B), or ju st after w ithd raw al (Figu re 3.9C). A large p rop ortion (40.2%) of the record ed neu rons increased their firing rate after the first lick in correct versu s incorrect trials, show ing p articip ation in the rew ard -related netw ork (Figu re 3.10).

Figu re 3.9: Raster plot (u p p er p lots) and p eristim lu s tim e histogram (PSTH , low er p lots) of sp ikes record ed from exam p le neu rons in the p rem otor cortex. 0 on the x axis corresp ond s to the nose p oke (A) or w ithd raw al tim e (B and C).

(28)

24

Figu re 3.10: Firing rate of an exam p le neu ron w ith rew ard related activity. Trials are d ivid ed in correct (green) and incorrect (red ). 0 on the x axis corresp ond s to the tim e of the first lick.

D o premotor neurons reflect the stimulus properties?

First w e investigated w hether p rem otor neu rons carried inform ation abou t the trial am p litu d e (high versu s low am p litu d e). We calcu lated the Sp earm an correlation betw een the average firing rate d u ring the w hole trial and the trial am p litu d e and fou nd that 23.5% of all neu rons exhibited significant correlation betw een w hole trial firing rate and envelope am p litu d e, 54% of w hich w ere p ositive (Figu re 3.11A). Figu re 3.11B show s the firing rates of an exam p le neu ron w ith p ositive correlation betw een firing rate and trial am p litu d e.

Fu rtherm ore, w e calcu lated the p ercentage of neu rons w hose firing rate d ep end ed on the stim u lu s am p litu d e in correct and incorrect trials, to investigate w hether the am p litu d e inform ation carried by p rem otor cortex neu rons is necessary for su ccessfu lly solving the task . 13% of neu rons in correct trials and 11.6% of neu rons in incorrect trials show ed significant correlation betw een the average firing rate and the overall stim u lu s am p litu d e. The am p litu d e of the trial is sim ilarly d ecod ed in correct and incorrect trials, ind icating that error trials are not cau sed by incorrectly id entifying the trial am p litu d e.

N ext w e checked w hether the firing rate of p rem otor neu rons increase d in resp onse to increased stim u lu s velocity, follow ing the p hases of the envelop e, sim ilar to barrel cortex neu rons (Arabzad eh, Petersen, & Diam ond ,

(29)

25

2003; Antopolskiy et al. (in p rep aration )). For this analysis w e d ivid ed each trial in 200 m s tim e bins and com p u ted the Sp earm an correlation coefficient betw een the firing rate and the stim u lu s envelop e am p litu d e for each bin, w hich p rod u ced tim e bins characterized by d ifferent average am p litu d es for each frequ ency-p hase-am p litu d e com bination. 31.2% of neu rons show ed significant correlation betw een firing rate and stim u lu s am p litu d e (p =0.01), 55.2% of w hich w ere p ositively correlated . To exclu d e th e p ossibility that the observed correlation w as cau sed by neu rons resp ond ing d ifferent ly to the overall trial am p litu d e, w e com p u ted the sam e p oint -by-p oint correlation d ivid ing trials by the envelop e am p litud e. Calcu lating the correlation for high am p litu d e trials w e fou nd that 13.5% of all neu rons had significant correlation (62.1% p ositive correlation). When only low am p litu d e trials w ere analyzed only 4.65% of all neu rons show ed significant correlation, 80% of w hich w ere p ositively correlated . Better am p litu d e cod ing in high am p litu d e trials cou ld be related to the observed better am p litu d e in high am p litu d e trials.

Figu re 3.11: H istogram of Sp earm an correlation coefficients betw een average firing rate on each trial and am plitu d e id entity (high/ low ) (A). Firing rates of an exam p le neu ron in trials w ith high (d ark brow n) and low (light brow n) am p litu d e (B).

(30)

26

Figu re 3.12: H istogram of Sp earm an correlation coefficients betw een firing rate and envelop e am plitu d e for each neu ron in high am p litu d e (A) a nd low am p litu d e (B) trials. Bins colored in p u rp le are neu rons w ith significant correlation (p <0.01).

In ord er to investigate if these resu lts w ere d u e to neu rons cod ing for the local am p litu d e or the p hase of the stim u lu s w e analyzed w hether firing of p rem otor cortex neu rons w as locked to the p hase of the envelope sine w ave. Ou r resu lts show ed that 8.5% of all record ed neu rons show ed significant p hase ind ex. Firing of p rem otor cortex neu rons w as not only locked to the p eak or the valley of the stim u lu s; instead it sp anned the w hole length of the sine w ave, m ore strongly arou nd the valley (Figure 3.13).

Figu re 3.13: Phase cod ing of p rem otor cortex neurons. A. Polar p lot of all neu rons w ith significant p hase coherence (grey) and average vector of all significant neu rons (red ). B. Phase locking of all significant neu rons to the stim u lu s envelop e (red triangles). The grey line rep resents the envelop e am p litu d e (all stim u lu s am p litu d es, frequ encies and p hases collap sed ).

A

B

(31)

27

Therefore, the firing of p rem otor cortex neu rons d ep end ed on the incom ing sensory inp u t at m u ltip le levels. Many neu rons carried inform ation regard ing the overall stim u lu s am p litu d e of the trial, p ossibly help ing rats set a low or high am p litu d e threshold . A sm aller, bu t significan t p rop ortion of neu rons reflected the p erceived stim u lu s am p litu d e at every tim e p oint in the trial. This cou ld rep resent the valu e the rat com p ares to the threshold for d ecid ing w hen to act. Fu rtherm ore, som e neu rons w ere p hase locked to the envelop e sine w ave, m ore in the d ecreasing and increasing p ortions of the stim u lu s arou nd the valley. This m igh t be ind icative of rats u sing the slop e of the stim u lu s to d ecid e w hen to act.

D o premotor neurons represent the passage of time?

N ext w e exp lored w hether th e firing rate of p rem otor cortex neurons p red icted the w ithd raw al tim e in the trial. We aligned the neu ral activity by the start of the stimulus or by the animal’s withdrawal and calculated the Sp earm an correlation betw een firing rate and w ithd raw al tim e (calcu lated as the tim e p assed from the start of the stim u lation) in 300 m s tim e bins. Figu re 3.14 show s the p ercentage of neu rons w ith significant correlation (p <0.01) for each tim e bin. The nu m ber of significant neu rons is higher than the chance -exp ectation of 1% in all tim e bins, even before the start of the stim u lation. More neu rons show significant correlation after the stim u lu s starts, and as m any as 26.4% of all neu rons show significant correlation betw een 600 and 300 m s before w ithd raw al.

Therefore, p rem otor cortex neu rons carry inform ation abou t how m u ch tim e the rat is p lanning to w ait, evid ent from neu rons correlated w ith w ithd raw al tim e w hen sp ikes w ere aligned w ith the start of the stim u lation, even in the tim e p eriod p reced ing the stim u lu s init iation. The firing rate of a great p ercentage of neu rons w as correlated w ith w ithd raw al tim e also w hen neuronal activity was aligned with the anim al’s action, therefore representing how m u ch tim e has p assed since the start of the stim u lation . More neu rons carried tim ing inform ation once the stim u lu s start ed , ind icating that the tim e at w hich the rat is intend ing to resp ond is u p d ated by the p ercep tion of the stim u lu s.

(32)

28

Figu re 3.14: Percentage of neu rons w ith significant Sp earm an correlation betw een firing rate and w aiting tim e in 300 m s tim e bins aligned by stim u lu s start (A) or animal’s action (B).

Figu re 3.15: PSTH of exam p le neu rons w ith significant correlation betw een firing rate and w ithd raw al tim e. Trials w ere d ivid ed accord in g to the w aiting tim e into 5 equ ip op u lated grou p s. The PSTH for each grou p is show n in d ifferent colors, from yellow for the shortest to d ark red for the longest average w ithd raw al tim e.

A

B

A

(33)

29

We fu rther ap p lied an artificial neu ral netw ork (AN N ) to p red ict th e w ithd raw al tim e. Inp u ts to the AN N w ere the firing rates of neu rons record ed sim u ltaneou sly. For each record ed session w e consid ered only the neu rons w ith highest correlation coefficient betw een firing rate and w ithd raw al tim e, and trials w ith w ithd raw al tim e longer than 0.6 second s. We aligned the neu ral activity by the start of the stim u lu s or w ithd raw al tim e and ap p lied AN N having as inp u ts the firing rate from fou r 300 m s tim e bins, tw o before the alignm ent p oint and tw o after.

N etw ork p erform ance w as calcu lated as m ean squ ared error betw een the actu al w aiting tim e and the AN N outp u t for each test trial. Since the AN N inp u ts w ere z scored , the exp ected p erform ance in the absence of any firing rate inform ation w as 1.

Figu re 3.16: AN N p erform ance in all sessions in 4 tim e bins w hen neu ral activity w as aligned by start of the stimulus (A) or animal’s w ithdraw al (B). ANN performance w as obtained taking the 5 neu rons w ith highest Sp earm an correlation betw een w ithd raw al tim e and firing rate for each bin. Trials w ere fu rther d ivid ed into correct (green) and incorrect (red ).

When consid ering the netw ork p erform ance in all sessions record ed from the 2 rats w e observed that w hen sp ikes w ere aligned w ith the stim u lu s start, the average netw ork p erform an ce w as close to 1, althou gh there are som e sessions w ith p erform ance low er than 1 even before the stim u lu s starts (Figure 3.16A). Consistent w ith ou r find ing that a large nu m ber of p rem otor cortex neu rons w ere correlated w ith w ithd raw al tim e w hen sp ikes w e re aligned w ith the animal’s action, network performance was also significantly below 1, more

(34)

30

noticeably in the 2 tim e bins p reced ing the w ithd raw al. Interestingly, netw ork p erform ance w as better in incorrect than in correct trials (Figu re 3.16B). This show s that in incorrect trials rats relied on tim ing, w hile in correct trials anim als relied on the stim u lu s and d id not have to rem em ber how m u ch tim e had p assed since the start of the stim u lation.

Finally, w e investigated w hether the firing rate of p rem oto r cortex neu rons p red icted the tim e p oint in the trial. In ord er to d o so, w e first d ivid ed each trial into 300 m s tim e bins and calcu lated the Sp earm an correlation coefficient betw een firing rate and the bin nu m ber for all trials together. When aligning neu ral activity w ith the start of the stim u lation w e obtained that 61.4% of neu rons had significant correlation (p <0.01 Figu re 3.17A) betw een firing rate and bin nu m ber, 54.55% of w hich had p ositive correlation. Likew ise, 76.28% of all neu rons had significant correlation betw een firing rate and bin nu m ber when neural activity was aligned by the animal’s action (Figure 3.17B), 49.39% of w hom had p ositive correlation. N o d ifference w as observed w hen com p aring correct and incorrect trials.

Figu re 3.17: H istogram of Sp earm an correlation coefficients betw een tim e bin and firing rate for each neu ron w hen neu ral activity w as aligned by stim u lus start (A) or animal’s withdraw al (B). In purple the neurons with significant correlation (p=0.01) are show n.

N ext w e u sed the AN N to p red ict the cu rrent m om ent in the trial from neu ral firing rate. We chose trials w ith w aiting tim e larger than 900 m s, aligned them by start of the stim u lu s, and d ivid ed them into 300 m s bins. The inp u t of the netw ork w as the neu ronal firing rate in each bin, and the ou tp u t w as the

(35)

31

nu m ber of the bin the firing rates belonged to (1, 2 or 3). Figu re 3.18A show s the d istribu tion of netw ork p erform ances for each session of record ings in the 2 rats. In m ost sessions netw ork p erform ance w as lo w er than the exp ected valu e of 1 (p <0.01, one w ay t-test), show ing that firing rate carried inform ation abou t how m u ch tim e has p assed since the start of the stim u lu s. The sam e analysis w as p erform ed by aligning neu ral activity w ith the w ithd raw al tim e and considering the 3 time bins preceding the animal’s action, again obtaining p erform ance below 1 for m ost sessions (Figu re 3.18B). These resu lts show that the firing rate of p rem otor neu rons is also inform ative of how m u ch tim e w ill pass until the animal’s response.

Figu re 3.18: H istogram of netw ork p erform ance in p red icting the tim e bin each firing rate belonged to w hen neu ral activity w as aligned w ith start of the stim u lu s (A) or animal’s w ithdraw al (B)

Therefore the firing rate of neu rons in t he p rem otor cortex throughou t the trial is inform ative of both tim e p assed since stim u lu s start and tim e m issing u ntil w ithd raw al.

D o premotor cortex neurons represent the w ithdraw al?

To check w hether p rem otor cortex neu rons changed their activity in p red iction of the w ithd raw al w e com p ared for each neu ron the firing rates in the two 300 ms time bins preceding the animal’s action, by computing the Welch t-test on ranks. Ou r resu lts show that 35.7% of all neu rons had a significant d ifference betw een the 2 tim e bins p reced ing the w ithd raw al (57.9% increased firing rate). This change in neu ral activity ju st before w ithd raw al

B

A

(36)

32

cou ld be d u e to action p red iction or stim u lu s cod ing, given that m ost resp onses rats m ad e w ere d u ring high am p litu d e stim u lation. To d isen tangle betw een these 2 p ossibilities, w e com p ared the firing rate ju st before w ithd raw al in correct and incorrect trials. Of all neu rons w ith d ifferent firing rate in the last 2 tim e bins p reced ing the w ithd raw al, only 1.3% had a significant d ifference betw een correct and incorrect trials in the last bin, show ing that the observed effect is not d ep end ent on the stim u lu s am p litu d e. Figu re 3.19 show s the average firing rate and the average stim u lu s am p litu d e in correct and incorrect trials of tw o exam p le neu rons.

Figu re 3.19: Average firing rate and the correspond ing stim u lu s am plitu d e in correct versu s incorrect trials, w hen activity is aligned w ith the anim al’s w ithd raw al. Tw o exam p le neu rons are show n: one w ith transient activation (A) and one w hose firing rate d ecreases before w ithd raw al (B).

Finally, w e u sed the AN N to assess how the netw orks of p rem otor neurons predicted the imminence of the animal’s action in the time bin p reced ing the w ithd raw al. We trained the AN N to d istingu ish betw een the 300 m s tim e bin p reced ing w ithd raw al and the 300 m s tim e bin ju st before the stim u lu s onset, u sing firing rate as inp u t (Figu re 3.20B). We fu rther tested the netw ork on d iscrim inating betw een the tim e bin before stim u lu s start and the tim e bin betw een 600 and 300 m s before w ithd raw al, and the tim e bin betw een 900 and 600 m s before w ithd raw al (Figu re 3.20B). The p ercentage of tim e bins correctly classified w as very high (91.0%) w hen d iscrim inating betw een the tim e ju st before start and ju st before w ithd raw al and d ecreased grad u ally to 82.8 and then 76.6% as bins fu rther from the w ithd raw al w ere consid ered

(37)

33

(Figu re 3.20D). This resu lt show s that as the rat ap p roached the tim e of the action, neu rons in the p rem otor cortex also ap p roached a state m ost d istingu ishable from the state before the stim u lu s started .

Resu lts w ere less conclu sive w hen the netw ork w as trained to classify betw een firing rates belonging to the tim e bin before start and the tim e bin after start of the stim u lu s, and later tested on the tw o su bsequ ent tim e bins after stim u lu s start (Figu re 3.20A). The p ercentage of trials correctly classified w as 77.8, 80.8 and 79.9% for the 3 tim e bins after w ithd raw al (Figu re 3.20C).

Figu re 3.20: AN N w as used to classify to w hich tim e bin the fir ing rate belonged . A, C AN N p attern recognition tool w as u sed to classify betw een the 300 m s tim e bin before stim u lu s start and the tim e bin ju st after the stim u lu s started and then tested on the consequ ent 2 tim e bins. B, D AN N w as u sed to classify betw een the tim e bin before stim u lu s start and the bin before w ithd raw al and tested on the 2 p reviou s tim e bins. AN N w as ap p lied on every record ed session of the 2 rats and the average p ercentage of bins correctly classified w as calcu lated for each tim e bin.

A

C

D

(38)

34

4. Discussion

We d evelop ed a new behavioral task that requ ired su bjects to integrate an inp u t stream of sensory stim u lation in ord er to d ecid e for the op tim al m om ent to initiate a m otor act. This p arad igm is u nu su al, as only a few stu d ies have investigated how u p com ing stim u lu s tim ing can be p red icted based on the p attern of p reviou s stim u li (Arnal et al., 2014; Bengtsson et al., 2009; Saleh, Reim er, Penn, Ojakangas, & H atsop ou los, 2010), bu t none incorp orated the p rep aration of a w ell-tim ed resp onse. Ou r stu d y is innovative becau se su bjects are requ ired to tim e their m otor resp onse p recisely d ep end ing on the incom ing stim u lu s. Moreover, w e im p lem ented this p arad igm in hu m ans and rats, w hich allow ed u s to com p are the behavioral strategies in the tw o sp ecies.

4.1. Rats and humans employ similar behavioral strategies

In ou r task rats and hu m ans w ere p resented w ith noisy vibrations (Fassihi, Akram i, Esm aeili, & Diam ond , 2014, Antop olsky et al. (in p rep aration)), m od u lated by an envelope sine w ave that m ad e them periodically increase and decrease in amplitude. The subjects’ task was to resp ond at the p eak of the stim u lu s. By changing the am p litu d e, frequ ency and stim u lu s onset p hase of the envelop e sine w ave the su bjects w ere d iscou raged from u sing alternative strategies to solve the task, su ch as ap p lying an am p litu d e threshold or tim ing their resp onses relative to trial onset.

Rats learned to solve the task by w ithd raw ing at the p eak am p litu d e, and d id so better than chance. Rats w ithd rew at d ifferent tim es accord ing to stim u lu s p aram eters, ind icating that they d id not em p loy a rigid tim ing strategy for solving the task. This claim is fu rther su p p orted by the fact that chance levels calcu lated by shu ffling the w aiting tim es w ith resp ect to the stim u lu s p aram eters is ap p roaching the theoretical chance level of 40%, corresp ond ing to how m any trials w ou ld be rew ard ed if rat w ithd rew rand om ly throu ghou t the trial. Moreover, p erform ance w as better than chance for both levels of am p litu d e, ind icating that rats d id not solve the task by sim p ly ap p lying an am p litu d e threshold irresp ective of the p erceived stim u lu s.

(39)

35

H u m ans sp ent m ore tim e than rats id entifying the stim u lu s p eak and reached a p erform ance that w as consid erably high er. This w as p robably d u e to d ifferences in p atience and goals in the tw o sp ecies - rats w ere trying to get the largest rew ard in the shortest tim e, w hile hu m ans w ere trying to p erform the task correctly. H ow ever, there w ere m any sim ilarities in the behavioral strategies em p loyed by rats and hu m ans, su ch as resp ond ing in later cycles in low am p litu d e trials as w ell as having low er p erform ance in those trials. This ind icates that these trials w ere m ore d ifficu lt and necessitated m ore evid ence integration. Moreover, althou gh hu m an p erform ance w as sim ilar in trials w here they w ere instru cted to bu tton p ress at the p eak and trials w here they w ere requ ired to respond at the valley of the stim u lu s, they resp ond ed in later cycles for valley trials, ind icating that tim ing resp onses to the valley of the stim u lu s w as m ore d ifficu lt.

At the beginning of each trial, su bjects w ere not aw are of the cu rrent stim u lu s p aram eters, so they w ere requ ired to accu m u late stim u lu s evid ence in ord er to d ecid e w hen to initiate the m ot or act. Rats have p reviou sly p roven to be cap able of accu m u lating evid ence in tim e (Dou glas et al., 2006; Reinagel et al., 2012). A valid strategy for solving the task w ou ld be to id entify the am p litu d e and the frequ ency of the envelop e and w ithd raw as soon as enou gh evid ence has been accu m u lated . A m inim u m necessary for a high p erform ance w ou ld be to gather evid ence in the first cycle of the stim u la tion and w ithd raw d u ring the second cycle. Withd raw als rats m ad e in the second cycle w ere better aligned to the stim u lu s p eak, ind icating that gathering stim u lu s inform ation d u ring the first cycle im p roved p red iction of the p eak in the second cycle. Also, both rats and hu m ans resp ond ed in later cycles in low am p litu d e and high frequ ency trials, ind icating that these trials requ ired m ore evid ence accu m u lation before m aking a d ecision.

4.2. Rat premotor cortex neurons carry task-relevant signals

We su bsequ ently investigated if neu rons in the p rem otor cortex carried inform ation abou t the stim u lu s or w ithd raw al tim e. Ou r choice of the brain area w as based on p reviou s stu d ies d em onstrating the role of p rim ate and rat p rem otor cortex in d ecision m aking (J I Gold , Shad len, J.I., & M.N ., 2000), tim e

(40)

36

p rocessing (Rao, Mayer, & H arrington, 2001), and stim u lu s p ercep tion (R Rom o et al., 1993). Finally, hu m an frontal areas have been show n to be involved in p red icting the m om ent at w hich a stim u lu s w ill occu r based on the p attern of p reviou s incom ing stim u li (Arnal et al., 2014; Saleh et al., 2010; Schu botz & von Cram on, 2002).

First, w e investigated if the rat p rem otor cortex n eu rons carried inform ation regard ing the incom ing stim u lu s for each trial.

A high p ercentage of p rem otor cortex neu rons show ed correlation betw een firing rate over the w hole trial and the am p litu d e id entity (high/ low ) for each trial. Am p litu d e inform ation cou ld be help ing rats solve the task by id entifying the m axim u m am p litu d e in each trial and w ithd raw ing w hen the m axim u m am p litu d e w as p erceived again.

Inform ation abou t the valu e of the stim u lu s at each p oint d uring the trial w as rather low in p rem otor cortex neu rons, in contrast to barrel cortex neu rons that follow the cou rse of the stim u lu s, firing m ore w hen the stim u lu s am p litu d e increases (Antop olsky et al. (in p rep aration)). N evertheless, inform ation abou t the stim u lu s intensity w as p resent in the p r em otor cortex, p articu larly in high am p litu d e trials, w hich cou ld exp lain a better p erform ance in these trials. Moreover, neu rons show ed little p hase coherence w ith the envelop e sine w ave, and w ere m ostly p hase locked to the increasing and d ecreasing am p litu d e su b-segm ents arou nd the valley of the envelop e sine w ave, p ossibly ind icating a strategy w here rats take into accou nt the slop e of the stim u lu s for d ecid ing w hen to act.

Interestingly, w e fou nd a correlation betw een the neu ronal firing rate and the w aiting tim e at d ifferent tim es in the trial, even before the start of the stimulation, indicating that neurons carried a signal related to the rats’ w illingness to w ait. A larger than exp ected p ercentage of neu rons had firing rates correlated w ith the w ith d raw al tim e both w hen neu ral activity w as aligned w ith the start of the stim u lu s, ind icating how m u ch tim e w ill p ass u ntil the rat m akes its response, and w hen sp ikes w ere aligned w ith the anim al’s w ithd raw al, w hich is associated w ith the am ou nt of tim e having p assed since the start of the stim u lu s. These neu rons resem ble the transiently active neu rons fou nd by Mu rakam i et al (Mu rakam i et al., 2014). The inform ation regard ing the w ithd raw al tim e increased at the start of the stim u lation and w as highest in the tim e bins p reced ing the w ithd raw al, p ossibly show ing that the w aiting tim e

Decision timing in the face of changing sensory input: behavior and neural correlates

International School for Ad vanced Stu d ies