1. Neuroscience
Download icon

Switch-like and persistent memory formation in individual Drosophila larvae

  1. Amanda Lesar
  2. Javan Tahir
  3. Jason Wolk
  4. Marc Gershow  Is a corresponding author
  1. Department of Physics, New York University, United States
  2. Center for Neural Science, New York University, United States
  3. NYU Neuroscience Institute, New York University Langone Medical Center, United States
Research Article
Cite this article as: eLife 2021;10:e70317 doi: 10.7554/eLife.70317
4 figures, 5 tables and 2 additional files

Figures

Figure 1 with 3 supplements
Y-maze assay to quantify innate and learned preference.

(A) Image sequence of a larva making two consecutive decisions in the Y-maze assay. White arrows indicate direction of air flow; red arrow shows direction of larva’s head. (B) Probability of choosing channel containing CO2 without any training. (C) Schematic representation of experiments in (D,E,F). All larvae were tested in the Y-maze for 1 hr to determine initial preference and again following manipulation to determine a final preference. The manipulations were: Paired Training - reward in concert with CO2 presentation, 15 s intervals, 20 repetitions; Offset After - reward presentation 7.5 s after CO2 onset, 15 s intervals, 20 repetitions; Reverse-Paired Training - reward opposite CO2 presentation, 15 s intervals, 20 repetitions; Offset Before - reward presentation 7.5 s before CO2 onset, 15 s intervals, 20 repetitions; DAN Activation Without CO2 - CO2 is never presented, while reward is presented at 15 s intervals, 20 repetitions; no training - no manipulation between two testing periods; Forward Paired (extended spacing) - 15 s reward follows 15 s CO2 presentation, followed by 60 s of air, 20 repetitions; Backwards Paired (extended spacing) - 15 s reward prior to 15 s CO2 presentation, followed by 60 s of air, 20 repetitions; Reward Between CO2 (extended spacing) - 15 s reward presentation between two 15 s CO2 presentations, followed by 45 s of air, 20 repetitions. (D) Probability of choosing CO2 containing channel before and after manipulation. All animals were fed ATR supplemented food, except those marked ATR-. (E) Probability of choosing CO2 containing channel before and after training as a function of reward timing, in training protocols with extended air spacings. All animals were DANi1>CsChrimson and fed ATR. (F) Probability of choosing CO2 containing channel before and after 20 cycles of paired training, as a function of CO2 concentration, used both during training and testing. All animals were DANi1>CsChrimson and fed ATR. * p<0.05, ** p<0.01, *** p<0.001.

Figure 1—source data 1

Spreadsheet containing each individual animal’s decisions in temporal sequence.

https://cdn.elifesciences.org/articles/70317/elife-70317-fig1-data1-v1.xlsx
Figure 1—video 1
Recording of a larva making 2 decisions within the Y-maze.

The direction of airflow and the larva’s decisions are noted. Video was recorded at 20 frames per second; the playback speed of 25 fps represents 1.25x real time.

Figure 1—video 2
Recording of a larva before training, showing a sequence of decisions made at the Y-maze juncture.

Recordings show 12.5 s before and 12.5 s after each decision. Video was recorded at 20 frames per second; the playback speed of 100 fps represents 5x real time.

Figure 1—video 3
Recording of a larva after training, showing a sequence of decisions made at the Y-maze juncture.

Recordings show 12.5 s before and 12.5 s after each decision. Video was recorded at 20 frames per second; the playback speed of 100 fps represents 5x real time.

Dose dependence of learning DANi1>CsChrimson were given varying cycles of paired training (as in Figure 1C).

(A) Probability of choosing CO2 containing channel before and after training, as a function of amount of training. *** p<0.001. (B) Histograms of individual larva preferences after training, grouped by number of training cycles. (C) Histogram of individual larva preference after training for all larvae. (D) Population average probability of choosing CO2 following training vs. dose. (E) Fraction of larvae untrained vs. number of training cycles. Teal: fit parameters and error ranges from quantized model, purple lines, prediction and error ranges from memoryless model. Note logarithmic y-axis on insert. (C–E) Orange: graded model prediction - post-training preference is represented by a single Gaussian distribution whose mean and variance depend on amount of training; Teal: quantized model prediction - post-training preference is represented by two fixed Gaussian distributions and the fraction of larvae in each population depends on the amount of training; Purple: all-or-none model prediction - post-training preference is represented by two fixed Gaussian distributions and the effect of a single training cycle is to train a fixed fraction of the remaining untrained larvae.

Figure 2—source data 1

Spreadsheet containing each individual animal’s decisions in temporal sequence.

https://cdn.elifesciences.org/articles/70317/elife-70317-fig2-data1-v1.xlsx
Figure 3 with 3 supplements
Memory extinction (A) Testing and training protocols for B,C.

Training + Extinction: larvae were exposed to 18 cycles of alternating CO2 and air following training. Habituation + Training: larvae were exposed to 18 cycles of alternating CO2 and air prior to training. (B) Probability of choosing CO2 containing channel (top) and fraction of larvae in trained group according to double Gaussian model fit (bottom) before and after training scheme. (C) Histograms of individual larva preference after training, for all larva and for larva trained with 2–4 training cycles. * p<0.05, ** p<0.01, *** p<0.001.

Figure 3—source data 1

Spreadsheet containing each individual animal’s decisions in temporal sequence.

https://cdn.elifesciences.org/articles/70317/elife-70317-fig3-data1-v1.xlsx
Figure 3—figure supplement 1
After extinction, larvae can be trained again.

(A) Testing and training protocol for B,C. Larvae were trained with three cycles of paired training, followed by 18 extinction cycles of alternating CO2 and air with no reward presented. After extinction, larvae were presented with three additional paired training cycles before testing. (B) Probability of choosing CO2 containing channel before and after training scheme. (C) Fraction of larvae in trained group according to double Gaussian model fit before and after training scheme. All larvae were DAN-i1>CsChrimson and raised on ATR+ food. *** p<0.001.

Figure 3—figure supplement 2
Larvae population average response following training.

(A) Larvae population average response in the first ten minutes following training (0–10 min), compared to the latter fifty minutes of testing (10–60 min), for larvae that had been given 2, 5, or 20 cycles of paired training. (B) Larvae population average response for the first five choices made by the larvae following training, compared to the remaining choices, for larvae that had been given 2, 5, or 20 cycles of paired training and made at least 10 decisions following training. (C,D,E) Larvae population average response over 15-min segments following training, for larvae trained with (C) 2 cycles, (D) 5 cycles, or (E) 20 cycles of training. All larvae were DAN-i1>CsChrimson and raised on ATR+ food.

Figure 3—figure supplement 3
Larvae given additional training between testing periods.

(A) Testing and training protocols for experiments in B. All larvae are tested in the Y-maze for one hour to determine initial preference. Larvae were then trained with 10 cycles of paired training, followed by a 15-min test period. The 10 cycle train/15 min test was repeated four times. (B) Probability of choosing CO2-containing channel before training, and during each of the four test periods. All larvae were DAN-i1>CsChrimson and raised on ATR+ food. *** p <0.001.

Memory retention overnight.

(A) Testing and training protocols. Except where indicated, larvae were tested, trained immediately after testing, tested again, then placed on food overnight and tested the following day. For extinction experiments, larvae were trained three times, and then exposed to 18 cycles of alternating CO2 and air either immediately following training or prior to testing the next day. (B,C,D) Probability of choosing CO2 containing channel (top) and fraction of larvae in trained group according to double Gaussian model fit (bottom) prior to training, immediately following training, and the next day. When the center bar is missing, larvae were not tested immediately following training but instead removed immediately to food. M Nx = massed training, N repetitions, S 10x = spaced training 10 total pairings, RP = reverse paired (see Figure 1C), No Train = no training. Larvae in (B,C) were DANi1>CsChrimson. Larvae in (D) were DANi1>hs-dCREB2-b;CsChrimson. Larvae were raised on food containing ATR, except for ATR+/CXM-, ATR+/CXM+ larvae who were fed ATR supplemented yeast paste (without/with cycloheximide) for 4 hr prior to initial testing. For reverse-paired (RP) and no training schemes, see Figure 1B. * p<0.05, ** p<0.01, *** p<0.001.

Figure 4—source data 1

Spreadsheet containing each individual animal’s decisions in temporal sequence.

https://cdn.elifesciences.org/articles/70317/elife-70317-fig4-data1-v1.xlsx

Tables

Key resources table
Reagent type (species) or resourceDesignationSource or referenceIdentifiersAdditional information
Genetic reagent (D. melanogaster)w[1118]; P{y[+t7.7]w[+mC]=20XUAS-IVS-CsChrimson.mVenus}attP2 (w;;UAS-CsChrimson)Bloomington Stock CenterRRID:BDSC_55136
Genetic reagent (D. melanogaster)SS00864 split-Gal4 (DAN-i1-Gal4)Saumweber et al., 2018Gift of Marta Zlatic, Janelia Research Campus
Genetic reagent (D. melanogaster)w[*]; Gr63a[1]Bloomington Stock CenterRRID:BDSC_9941
Genetic reagent (D. melanogaster)w[1118]; P{y[+t7.7] w[+mC]=GMR58E02-GAL4}attP2 (GMR58E02-Gal4)Bloomington Stock CenterRRID:BDSC_41347
Genetic reagent (D. melanogaster)w;hs-dCREB2-b 17–2Yin et al., 1995FlyBase_ FBti0038019Gift of Jerry Chi-Ping Yin, University of Wisconsin, Madison
Genetic reagent (D. melanogaster)w[*]; P{w[+mW.hs]=GawB}ey[OK107]/In(4)ci[D], ci[D] pan[ciD] sv[spa-pol] (OK107-Gal4)Bloomington Stock CenterRRID:BDSC_854
Genetic reagent (D. melanogaster)w[*]; P{w[+mC]=UAS-Hsap\KCNJ2.EGFP}7 (UAS-kir2.1)Bloomington Stock CenterRRID:BDSC_6595
Genetic reagent (D. melanogaster)w[*]; P{w[+mC]=Gr21a-GAL4.C}133t52.1 (Gr21a-Gal4)Bloomington Stock CenterRRID:BDSC_23890
Software, algorithmlivetrackergithub.com/GershowLab/TrainingChamber (copy archived at URL swh:1:rev:e2a7ccc4e8d845e6cac59d3b2f344cca826c4727, Lesar, 2021)This work
Table 1
Crosses used to generate larvae for experiments throughout this work.

For strain information, see key resource table.

FigureDesignationFemale parentMale parent
1Gr63a1w;Gr63a1
1OK107>Kir2.1UAS-Kir2.1-GFPOK107-Gal4
1Gr21a>Kir2.1UAS-Kir2.1-GFPGr21a-Gal4
1-4DANi1>CsChrimsonw;;UAS-CsChrimsonSS00864
1Driver ctrlSS00864
1Effector ctrlw;;UAS-CsChrimson
158E02>CsChrimsonw;;UAS-CsChrimson58E02-Gal4
4hs-dcreb2-b;DANi1>CsChrimsonw;hs-dcreb2-b;UAS-CsChrimsonSS00864
Table 2
Data for experiments in Figure 1, Figure 2, Figure 3, and Figure 4.

# Larva: number of individual larvae tested for experiment type; # Approach Pre-Train: total number of times all larvae chose the channel containing air with CO2 prior to training; # Avoid Pre-Train: total number of times all larvae chose the channel containing pure air prior to training; # Approach Post-Train: total number of times all larvae chose the channel containing air with CO2 after the indicated training scheme; # Avoid Post-Train: total number of times all larvae chose the channel containing pure air after the indicated training scheme; # Approach Next Day: total number of times all larvae chose the channel containing air with CO2 during testing approximately 20 hr after training; # Avoid Next: total number of times all larvae chose the channel containing pure air during testing approximately 20 hr after training. All tests were 1 hr (for each larva).

ExperimentGenotype# Larva# Approach Pre-Train# Avoid Pre-Train# Approach Post-Train# Avoid Post-Train# Approach Next Day# Avoid Next Day
Figure 1B
Gr63a1Gr63a144831745----
DANi1> CsChrimson, ATR+DANi1> CsChrimson15917144978----
DANi1> CsChrimson, ATR-DANi1> CsChrimson16256614----
Figure 1D
PairedDANi1> CsChrimson645611760936868--
Offset AfterDANi1> CsChrimson20288757316305--
Reverse PairedDANi1> CsChrimson293151022154530--
Offset BeforeDANi1> CsChrimson19218512136315--
Paired, ATR-DANi1> CsChrimson16256614127307--
No TrainingDANi1> CsChrimson5057815994791295--
DAN w/o CO2DANi1> CsChrimson16260597161354--
Driver ctrlSS0086417110289158358--
Effector ctrlUAS-CsChrimson18214516114294--
58E02> CsChrimson58E02> CsChrimson21380912493501--
Figure 1EDANi1> CsChrimson
Forward Paired22181496350337--
Backwards Paired18181438124320--
Btw CO223272652165283--
Figure 1FDANi1> CsChrimson
6.5%19361568319290--
8%27256567295255--
15%19170368249233--
18%645611760936868--
Figure 2ADANi1> CsChrimson
0 Cycles5057815994791295--
1 Cycles35218606317495--
2 Cycles87840255210811292--
3 Cycles31310930686686--
4 Cycles32245712493511--
5 Cycles638632491975993--
10 Cycles14100287154144--
20 Cycles645611760936868--
Figure 3BDANi1> CsChrimson
2 Cycles, Training87840255210811292--
2 Cycles, Habituation + Training303851127422554--
2 Cycles, Training + Extinction30336946375793--
3 Cycles, Training30308924675679--
3 Cycles, Habituation + Training18222591260294--
3 Cycles, Training + Extinction26279695195416--
4 Cycles, Training30225659490502--
4 Cycles, Habituation + Training18239701372352--
4 Cycles, Training + Extinction273841074394475--
5 Cycles, Training638632491975993--
6 Cycles, Habituation + Training19266758367324--
6 Cycles, Training + Extinction18253687309317--
10 Cycles, Training14100287154144--
10 Cycles, Habituation + Training304061193607503--
10 Cycles, Training + Extinction304261180401386--
Figure 4BDANi1> CsChrimson
20x283801172509499459409
20x (Only Test Next Day)14224768--296250
5x294721427488480404461
2x425141537594693201548
2x (Only Test Next Day)22209696--213283
No Train20316889187544104337
RP 20x21282905121430109361
Ext Post-Train23181477--158365
Ext Pre-Test314171002--385429
Figure 4CDANi1> CsChrimson
M 20x (CXM+/ATR+)20110282252237237272
M 20x (CXM-/ATR+)17159419271236228235
S 20x (CXM+/ATR+)23191486--150316
S 20x (CXM-/ATR+)20197511--254264
M 10x (CXM+/ATR+)23136345--331344
M 10x (CXM-/ATR+)20175454--419375
Figure 4DDANi1> hs-dCREB2-b;CsChrimson
M 10x HS21175434392370253246
M 10x No HS22248656367353451490
S 10x HS2417242068156153339
S 10x No HS22294736212184335352
Table 3
p-Values for experiments in Figure 1, Figure 2, Figure 3, and Figure 4.

P-values for experiments were calculated: Bootstrap - p-values calculated as explained in Materials and methods; Fisher - p-values calculated using Fisher’s exact test; U-test - p-values calculated using two-sided Mann–Whitney U test. Unless otherwise noted, p-values are calculated between pre-train and post-train data. A shaded row indicates not all tests reach the same significance level (out of ns, p <0.05, p <0.01, p <0.001).

ExperimentGenotypeHierarchical BootstrapBootstrap Animal OnlyFisherU-test
Figure 1B
Gr63a1/DANi1> CsChrimson, ATR+<10−4<10−4<10−4<10−4
Gr63a1/DANi1> CsChrimson, ATR-<10−4<10−4<10−4<10−4
Figure 1D
PairedDANi1> CsChrimson<10−4<10−4<10−4<10−4
Offset AfterDANi1> CsChrimson<10−4<10−4<10−4<10−4
Reverse PairedDANi1> CsChrimson0.34290.26890.61660.9379
Offset BeforeDANi1> CsChrimson0.44790.43730.94790.9770
Paired, ATR-DANi1> CsChrimson0.47620.43151.0000.2658
No TrainingDANi1> CsChrimson0.40660.36640.77260.9835
DAN w/o CO2DANi1> CsChrimson0.39350.31020.71730.4852
Driver ctrlSS008640.31060.03130.34110.3977
Effector ctrlUAS-CsChrimson0.33830.23610.63360.8366
58E02> CsChrimson58E02> CsChrimson<10−4<10−4<10−4<10−4
Figure 1CDANi1> CsChrimson
Forward Paired<10−4<10−4<10−4<10−4
Backwards Paired0,33680.1630.68010.1939
Btw CO20.01070.00010.0065430.0003257
Figure 1DDANi1> CsChrimson
6.5%<10−4<10−4<10−4<10−4
8%<10−4<10−4<10−4<10−4
15%<10−4<10−4<10−4<10−4
18%<10−4<10−4<10−4<10−4
Figure 2ADANi1> CsChrimson
0 Cycles0.41320.36470.77260.9835
1 Cycles0.0003<10−4<10−40.0591
2 Cycles<10−4<10−4<10−4<10−4
3 Cycles<10−4<10−4<10−4<10−4
4 Cycles<10−4<10−4<10−4<10−4
5 Cycles<10−4<10−4<10−4<10−4
10 Cycles<10−4<10−4<10−4<10−4
20 Cycles<10−4<10−4<10−4<10−4
Figure 3BDANi1> CsChrimson
2 Cycles, Training<10−4<10−4<10−4<10−4
2 Cycles, Habituation + Training<10−4<10−4<10−4<10−4
2 Cycles, Training + Extinction0.01170.00200.0013390.04743
3 Cycles, Training<10−4<10−4<10−4<10−4
3 Cycles, Habituation + Training<10−4<10−40.0007459<10−4
3 Cycles, Training + Extinction0.11330.01760.17630.03069
4 Cycles, Training<10−4<10−4<10−4<10−4
4 Cycles, Habituation + Training<10−4<10−4<10−4<10−4
4 Cycles, Training + Extinction<10−4<10−4<10−4<10−4
5 Cycles, Training<10−4<10−4<10−4<10−4
6 Cycles, Habituation + Training<10−4<10−4<10−4<10−4
6 Cycles, Training + Extinction<10−4<10−4<10−4<10−4
10 Cycles, Training<10−4<10−4<10−4<10−4
10 Cycles, Habituation + Training<10−4<10−4<10−4<10−4
10 Cycles, Training + Extinction<10−4<10−4<10−4<10−4
Figure 4BDANi1> CsChrimson
20x Pre-Test/Post-Test<10−4<10−4<10−4<10−4
20x Pre-Test/Next Day<10−4<10−4<10−4<10−4
20x (Only Test Next Day) Pre-Test/Next Day<10−4<10−4<10−4<10−4
5x Pre-Test/Post-Test<10−4<10−4<10−4<10−4
5x Pre-Test/Next Day<10−4<10−4<10−4<10−4
2x Pre-Test/Post-Test<10−4<10−4<10−4<10−4
2x Pre-Test/Next Day0.20860.05010.35240.07216
2x (Only Test Next Day) Pre-Test/Next Day<10−4<10−4<10−4<10−4
No Train Pre-Test/Post-Test0.40350.33190.78930.2003
No Train Pre-Test/Next Day0.15830.05300.30710.8884
RP 20x Pre-Test/Post-Test0.26770.15070.42760.7396
RP 20x Pre-Test/Next Day0.42050.34810.84740.3765
Ext Post-Train Pre-Test/Next Day0.18010.01460.33150.01336
Ext Pre-Test Pre-Test/Next Day<10−4<10−4<10−4<10−4
Figure 4CDANi1> CsChrimson
M 20x (CXM+/ATR+) Pre-Test/Post-Test<10−4<10−4<10−4<10−4
M 20x (CXM+/ATR+) Pre-Test/Next Day<10−4<10−4<10−4<10−4
M 20x (CXM-/ATR+) Pre-Test/Post-Test<10−4<10−4<10−4<10−4
M 20x (CXM-/ATR+) Pre-Test/Next Day<10−4<10−4<10−4<10−4
S 10x (CXM+/ATR+) Pre-Test/Next Day0.10990.0140.16710.02985
S 10x (CXM-/ATR+) Pre-Test/Next Day<10−4<10−4<10−4<10−4
M 10x (CXM+/ATR+) Pre-Test/Next Day<10−4<10−4<10−4<10−4
M 10x (CXM-/ATR+) Pre-Test/Next Day<10−4<10−4<10−4<10−4
Figure 4DDANi1> hs-dCREB2-b;CsChrimson
M 10x, HS Pre-Test/Post-Test<10−4<10−4<10−4<10−4
M 10x, HS Pre-Test/Next Day<10−4<10−4<10−4<10−4
M 10x, No HS Pre-Test/Post-Test<10−4<10−4<10−4<10−4
M 10x, No HS Pre-Test/Next Day<10−4<10−4<10−4<10−4
S 10x, HS Pre-Test/Post-Test0.38040.28300.73100.2750
S 10x, HS Pre-Test/Next Day0.26450.088600.46500.3802
S 10x, No HS Pre-Test/Post-Test<10−4<10−4<10−4<10−4
S 10x, No HS Pre-Test/Next Day<10−4<10−4<10−4<10−4
Table 4
Model fits to data in Figure 2.

Shifting Mean and σ~, shifting fraction, and exponential fraction models are presented in Figure 2. Model name: name of the model. Formula: expression for the probability of the data given the model and its parameters. # params: number of free parameters in the model. Δlog(P) logarithm of the probability of the data given best fit to this model minus logarithm of the probability of the data given the best fit model overall. A higher (less negative) value means the model better fits the data without regard to the number of parameters. ΔAIC, ΔBIC - Aikake and Bayes Information Criterion minus the lowest values over the models tested. Lower numbers indicate model is favored. According to both criterion, the exponential fraction model is strongly favored over the shifting fraction model, and the shifting fraction model is strongly favored over all models except the exponential fractional model.

Model nameFormula# paramsΔlog(P)ΔAICΔBIC
Shifting Mean (fixed σ~)Pnc(0,1,2,3,4,5,10,20)j𝒩(p(nc,j),μ(nc),σ~μ(nc)*(1-μ(nc))n(nc,j))9−42.784.86104.45
Shifting Mean and σ
(Graded learning)
Pnc(0,1,2,3,4,5,10,20)j𝒩(p(nc,j),μ(nc),σ(nc)n(nc,j))16−12.939.386.3
Shifting Fraction
(Quantized learning)
Pnc(0,1,2,3,4,5,10,20)jfu(nc)𝒩(p(nc,j),μu,σ~μu*(1-μu)n(nc,j))+
(1-fu(nc))𝒩(p(nc,j),μt,σ~μt*(1-μt)n(nc,j))
11−3.9311.338.7
Shifting Fraction
(3 clusters)
Pnc(0,1,2,3,4,5,10,20)jf1(nc)𝒩(p(nc,j),μ1,σ~μ1*(1-μ1)n(nc,j))+
f2(nc)𝒩(p(nc,j),μ2,σ~μ2(1μ2)n(nc,j))+
(1-f1(nc)-f2(nc))𝒩(p(nc,j),μ3,σ~μ3*(1-μ3)n(nc,j))
20021.484.1
Exponential Fraction
(All-or-none)
Pnc(0,1,2,3,4,5,10,20)jλnc𝒩(p(nc,j),μu,σ~μu*(1-μu)n(nc,j))+
(1-λnc)𝒩(p(nc,j),μt,σ~μt*(1-μt)n(nc,j))
4−5.300
SymbolDefinitionSymbolDefinition
ncnumber of training cyclesp(nc,j)fraction of times jth larva chose CO2 after nc cycles
μ(nc)mean probability of choosing CO2 after nc training cyclesn(nc,j)# choices made by jth larva after nc training cycles
σ~global adjustment to binomial standard deviationσ(nc)training dependent standard deviation
μuprobability of larva in untrained group choosing CO2μtprobability of larva in trained group choosing CO2
fu(nc)fraction of larvae in untrained group after nc cyclesμ1,μ2,μ3probability of larva in group 1,2,3 choosing CO2
f1(nc),f2(nc)fraction of larvae in groups 1,2 after nc cyclesλfraction of larvae not trained after one cycle
𝒩(x,μ,σ)normal cdf: 12πσ2e-(x-μ)22σ2Δlog(P)relative log probability of data given model
AICAikake Information Criterion: 2k-2log(P), k = # paramsΔAICAIC - lowest AIC
BICBayes Information Criterion: klog(nA)-2log(P), k = # params, nA = # animalsΔBICBIC - lowest BIC

Additional files

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)