Calcium activity is a degraded estimate of spikes

Summary

Recording action potentials extracellularly during behavior has led to fundamental discoveries regarding neural function; hippocampal neurons respond to locations in space ¹, motor cortex neurons encode movement direction ², and dopamine neurons signal reward prediction errors ³, observations undergirding current theories of cognition ⁴, movement ⁵, and learning ⁶. Recently it has become possible to measure calcium flux, an internal cellular signal related to spiking. The ability to image calcium flux in anatomically ^7,8 or genetically ⁹ identified neurons can extend our knowledge of neural circuit function by allowing activity to be monitored in specific cell types or projections or in the same neurons across many days. However while initial studies were grounded in prior unit recording work, it has become fashionable to assume calcium is identical to spiking, even though the spike-to-fluorescence transformation is nonlinear, noisy, and unpredictable under real-world conditions ¹⁰ It remains an open question whether calcium provides a high-fidelity representation of single-unit activity in awake, behaving subjects. Here we have addressed this question by recording both signals in the lateral orbitofrontal cortex (OFC) of rats during olfactory discrimination learning. Activity in the OFC during olfactory learning has been well-studied in humans ^11-14, nonhuman primates ^15,16, and rats ^17-21, where it has been shown to signal information about both the sensory properties of odor cues and the rewards they predict. Our single-unit results replicated prior findings, whereas the calcium signal provided only a degraded estimate of the information available in the single-unit spiking, reflecting primarily reward value.

eTOC Blurb

Hart et al. use single unit recording and single-cell calcium imaging during olfactory learning to show that subtle changes in spiking that determine odor identity coding and its multiplexing with reward-predictions are lost in imaging data, thus demonstrating that the two signals do not produce the same result during a complex cognitive operation

Results and Discussion

Rats in electrophysiology and imaging groups learned novel 4-odor go/no-go discriminations within each recording session at a similar rate and with similar motivation

To test whether calcium imaging and single-unit recordings contain the same information during a complex cognitive operation, we recorded or imaged OFC neurons while rats learned novel odor sets in a go/no-go discrimination task. Rats learned a new problem during each session, responding to all four odors during the early trials and gradually learning to withhold responding to the no-go odors (Figure 1A-B). Groups did not differ in the number of trials required to meet a behavioral criterion of 90% correct in a sliding window of 20 trials, indicating they learned the underlying associations in each problem at a similar rate (Figure 1C)¹⁷. Groups also did not differ in how quickly they responded to the positive odors, indicating similar motivation for reward (Figure 1D).

Figure 1. OFC neurons signal reward predictions and odors after learning.

(A-B) Average (all sessions, rats) percentage of correctly performed trials for each of the 4 odors during the recording sessions. Rats initially responded to each of the four odors (positive odors, blue) and learned to withhold responding to the negative odors (red). The electrophysiology group is displayed in panel A, and the imaging group is shown in panel B. Shaded areas indicate standard error of the mean.

(C) The number of trials required to reach a criterion of 90% correct over the prior 20 trials during recording sessions did not differ across groups (KS test, p = 0.91).

(D) Average reaction time on responses for reward during recording sessions did not differ across groups (KS test, p = 0.19).

(E) Example (rat 2, session 1, neuron 7) reward prediction-signaling (Friedman test, go vs no-go, p = 0.009) single unit spike raster plots (top) and the mean spiking (bottom). Activity is displayed starting one second prior to odor-port-exit (odor sampling), and one second following. The positive odors are plotted in blue, and the negative odors are plotted in red. Trials are shown blocked for visualization purposes; these trials occurred pseudo-randomly.

(F) Same as in (E), but for a neuron (rat 4, session 5, neuron 39) that multiplexed reward prediction and odor identity (2-way ANOVA interaction, p < .0001).

(G) Same as in (E) and (F), but for a neuron (rat 2, session 2, neuron 7) that signaled odor identity (Rank Sum test, no-go 3 vs no-go 4, p = .047).

(H) Representative peri-odor-unpoke fluorescence response of one OFC neuron (rat 4, session 2, neuron 13) that signaled reward prediction (Friedman test, go vs no-go, p < 0.001). Two seconds of activity are displayed, starting one second prior to unpoke from the odor port (i.e., during odor sampling), and ending one second following unpoke. Individual trials are shown in the top panel, and the average fluorescence responses are shown in the second panel, where positive odors are plotted in blue, and negative odors are plotted in red.

(I) Same as in (H), however, this example neuron (rat 2, session 2, neuron 27) showed a more complex response pattern, where fluorescence responses were strongest to the positive odors and one of the negative odors.

(J-K) Proportions of neurons in the electrophysiology (I) and imaging populations (J) that signaled reward prediction, odor identity, or multiplexed reward and odor identity, which were different between the populations ( χ² = 8.90 , p = 0.01).

(L) Example multiphasic neuron odor-on responses (left) and odor-unpoke responses (right). Note the selectivity pattern differed in the two analysis periods (highlighted in gray).

(M) The proportion of neurons in the electrophysiology (gray) and imaging (green) groups that responded differently during odor sampling based on reward, positive odor identity, negative odor identity, reward + odor identity, or that exhibited multiphasic responses. The proportions of reward prediction-coding neurons ( χ² = 15.40 , p < 0.001 ), reward + odor neurons ( χ² = 4.63 , p = 0.03 ), and multiphasic neurons ( χ² = 17.90 , p < 0.001 ) were higher in the electrophysiology group than the imaging group

See also Figure S1 and Table S1.

OFC neurons signal reward predictions and odors after learning

Single-units were acquired using fine-wire microelectrodes and standard approaches for recording and isolation. Calcium signals were acquired using UCLA miniscopes. Critically, there is no generally agreed best practice for analyzing calcium data²², and different methods do not have the same output^23,24. A main difference between processing pipelines begins - or ends - with deconvolution. The deconvolution algorithms were developed in vitro^25,26 and are argued to estimate spikes in-vivo. It is proposed that temporally constraining sensor output provides a better approximation of the biological signal of interest - action potentials - and therefore should also more closely match single-unit recording. Some research groups have added deconvolution to their analysis pipelines^27-32, while others decided to forgo it and instead analyzed fluorescence traces^9,33-35. Since our goal is to evaluate how well the signal matches single-unit recording with as broad a methodology as possible, we analyzed calcium signals both with and without deconvolution and will hereafter refer to them as “fluorescence” and “deconvolved”. Our analysis pipeline (Figure S1A) was based on field-standard practice^36,37,26. Fluorescence responses to the odor cues were qualitatively like single-unit spiking (Figure S1B,S1F-S1G, top), and deconvolution temporally constrained estimated calcium flux (Figure S1C,S1F-S1G, bottom). The frequency of deconvolved calcium transients was sparser than spikes (Figure S1D), as others reported^7,31,38, even though we sampled the same approximate location and volume of tissue(Figure S1E).

As a first step toward addressing whether spiking and calcium activity contain similar information, we analyzed the trials immediately after the rats met criterion on each problem. We examined neural correlates related to odor identity and reward-prediction during odor sampling post-criterion. Consistent with prior unit recording studies^{16,17,19,20,39}, OFC spiking represented information about both the reward prediction of the current odor (Figure 1E), its identity (Figure 1G), or both (Figure 1F). Such single-unit correlates have been characterized previously during odor discrimination learning in rats^19,40 and monkeys^41,42. While it was possible to find calcium responses to the odors that were qualitatively like the single-unit correlates (Figure 1H-I), and the overall proportions were also somewhat congruent (Figure 1J-1K), fluorescence seemed less sensitive. Significantly fewer neurons met criteria for reward-signaling and multiplexing (Figure 1M). Imaging data were also simpler, containing fewer “multiphasic” neurons²⁴ - neurons which had different selectivity patterns across trial epochs, in this case during and following odor presentation (Figure 1L,1M).

OFC ensembles contain reward prediction and odor identity information, which are attenuated in imaging data

The above analyses suggest that calcium activity during odor sampling may lack the dynamic range and fidelity required to capture the information present in spike rates, particularly regarding the cues and rewards they predict. To test this, we modeled the population response to each odor with a support vector classifier built from all but one randomly selected trial of each odor. This classifier was then tested one-against-one on the left-out trials. The output of this analysis can be displayed as a confusion matrix, where the actual trial type is plotted against the classifier-identified trial type, with the proportion classified shown in false color. Exemplar patterns demonstrating idealized hypothetical outcomes are shown in Figure 2A. If the population represents only reward predictions, the classification pattern would resemble that in the right panel, where the two positive and two negative odors are confused; on the other hand, the more the population identifies the particular cues along with rewards, the more the classifier pattern would resemble the one on the left panel.

Figure 2. OFC ensembles contain reward and odor identity information, which are attenuated in imaging data.

(A) Exemplar patterns demonstrating hypothetical outcomes of classification analyses - odor identity (left) and reward prediction (right).

(B) Confusion matrices from spiking rates, fluorescence, and deconvolved fluorescence taken during odor sampling.

(C) The proportion of mis-classified trials was higher in fluorescence ( 1518/8000 ; units 625/8000 ; mean difference 95% CI: 0.10 - 0.12 ) and in deconvolved calcium (4301/8000 ; mean difference 95% CI: 0.45 - 0.47 ). * indicates the proportion is different from electrophysiology data.

(D) Correlation of binarized confusion from the matrices in panel B with exemplar patterns from panel A at different thresholds. Confusion matrices from (B) were filtered at different thresholds to extract different patterns of information. The horizontal dotted line indicates the correlation between exemplar patterns - at that level correlations cannot reliably indicate one model or the other. At very low thresholds, reward information dominated, while at a reasonable threshold (~15%, vertical dotted line), odor identity information became apparent in the electrophysiology data.

(E) Output of binary classification analyses for 10, 40, 100, and 200 neuron pseudo-ensembles. Each panel shows average decoding of reward, positive odor identity, and negative odor identity (left, middle, right, respectively). Electrophysiology, fluorescence, and deconvolved are plotted in gray, green, and cyan, respectively. The gray horizontal lines indicate theoretical chance accuracy. * indicates decoding accuracy exceeded chance, defined by mean accuracy exceeding the 95^th percentile of shuffled data.

(F) 3D projection of the first three LDA components (LD) from PCA-reduced unit-spiking data. The first LD perfectly separated the odors by reward (λ = .0001, p < 0.001), and the second by identity (λ = .01, p < 0.001).

(G) 3D projection of the first three LDA components from PCA-reduced fluorescence data. The first LD separated the odors by reward (λ = .0009, p < 0.001), while the second did not perform as well (λ = .05, p = 0.02). The positive (KS test, p < 0.0001) and negative odors (KS test, p < 0.0001) were more separated (better represented) in the electrophysiology data.

See also Figure S2.

As is apparent in the confusion matrix in Figure 2B (left), single-unit spiking during odor sampling signaled odor identity nearly perfectly, consistent with known single unit correlates, which frequently showed an influence of reward prediction as well as odor. In the fluorescence, however, classification accuracy was significantly worse than for spikes, and classifier errors tended to occur within positive or negative exemplar pairs (Figure 2B, middle). Odor 1 tended to misclassify as odor 2, and vice versa, indicating that calcium flux – unlike single unit activity - did not distinguish well between the odor cues within each reward class. Notably, contrary to assumptions that deconvolution recapitulates spiking, the deconvolved signal showed an even greater tendency for within-exemplar confusion (Figure 2B-C).

Critically, these differences did not reflect a lack of reward-related coding in the unit spiking, but rather the addition of information about the specific cues associated with reward or non-reward. This can be revealed by filtering and binarizing the classification results from Figure 2B at different thresholds. We set thresholds, 5% for example, and all values in the confusion matrices above that threshold as equal to 100. At low thresholds, all three confusion matrices from panel 2B correlate best with the reward prediction pattern in Figure 2A. At higher thresholds ( ~15%), odor identity dominated spiking data (Figure 2D, right), while fluorescence data remained more similar to the reward prediction exemplar pattern (Figure 2D, middle). Odor identity information never emerged in the deconvolved calcium (Figure 2D, right).

We next tested whether we could separately decode reward prediction, positive odor identity, and negative odor identity. This served as a control for the clear effect of reward on OFC activity and allowed us to determine whether classifier errors were simply due to the “best match” approach when all four odors were considered simultaneously. We randomly sampled pseudoensembles of 10-200 neurons, trained a support vector classifier on half the data, and tested on the other half. Very few neurons were required to decode reward prediction; classification accuracy in the electrophysiology data was above chance at the smallest population size (Figure 2E, left). However, fluorescence required one order of magnitude more neurons to decode reward above chance, and double that for the deconvolved calcium (Figure 2E). A similar effect was observed for odor identity, though larger samples were generally required. Using pseudoensembles of 200 neurons, reward, positive odor, and negative odor identity information were present in the spike data. However, even at this threshold, odor identity decoding was not above chance in the fluorescence and deconvolved data (Figure 2E, right), leaving reward prediction as the sole “correlate” in the calcium signal.

Neural representations are high dimensional, and a small number of components retain meaningful properties of the full data set⁴³. We therefore reduced both populations to the first 20 principal components (PCs) explaining approximately 80% of the variance in activity. We utilized the fluorescence data to provide a “best-case scenario” comparison, since fluorescence performed better than the deconvolved data. We took the PCs from spiking and fluorescence and performed linear discriminant analysis to find linear combinations of features that separated the data points⁴⁴, in this case trials, of each odor. These data were projected onto three-dimensional linear discriminant (LD) subspace (Figure 2F-G). In visualizing the underlying structure of this space, the electrophysiology data form four distinct clusters, each representing one of the four odors, where the first LD separates the points by valence, and subsequent components by odor identity. While the overall pattern is similar in the fluorescence data, there is a marked compression in the overall activity space, particularly for LD2, which separated the odors by identity.

Although the above analyses were performed on equally sized populations thus ruling out this potential confound, we also performed analyses to rule out potential effects of sample size and bias within groups. These data are outlined in Figure S2, where we show that individual ensembles did not differ in size across groups (Figure S2A), the effects that were apparent in population analyses replicated in the individual “real-time” ensembles, as well as across rats within each group, and subsampling the highest frequency calcium activity did not replicate spiking data (Figure S2H-J).

Imaging captures degraded spike activity

OFC spiking contained rich multiplexed representations of rewards and odors; imaging did not capture the subtler changes in neural activity that determined odor identity coding and its multiplexing with information about reward. However, we did not perform extracellular recording and imaging concurrently in individual rats. While this was intentional to avoid confounding the information gathered by one approach with tissue damage or other factors introduced by the other approach, and the recording location, task, odors, learning rates and motivational measures were similar across the two groups, it is possible that differences between the two signals could arise because they came from different subjects. To address this, we utilized a proven spikes-to-fluorescence model²⁴ to transform odor-sampling spike times into synthetic fluorescence (Figure 3A). We then used synthetic fluorescence for identical analyses as in Figure 2. We sampled 200-neuron pseudoensembles, which was the threshold where spikes reliably signaled rewards and odors but fluorescence did not. Synthetic fluorescence perfectly recapitulated real fluorescence data; classifier confusion was due to reward, and we failed to decode odor identity above chance, even though reward information was still apparent (Figure 3B-C).

Figure 3. Imaging captures degraded spike activity.

(A) Example single-unit spiking around odor-unpoke (left) and normalized synthetic fluorescence (right). The spikes to fluorescence transformation attenuated odor identity signaling.

(B) Confusion matrix displaying results of synthetic fluorescence classification.

(C) Results of binary synthetic fluorescence classification. We failed to decode odor identity above chance. * indicates reward decoding accuracy exceeded chance.

(D) The trials utilized for population classification in Figure 2 were reduced to 20 principal components, and these trials are shown projected onto principal component space. Note the first PC perfectly separates the trials by reward.

(E) Confusion matrix displaying results from utilizing just the first PC for classification.

(F) Results of binary first-PC classification. We failed to decode odor identity above chance. * indicates reward decoding accuracy exceeded chance.

(G) Histogram of average spiking rates after downsampling to match the deconvolved calcium population. The vertical gray line indicates the mean activity rate.

(H) Confusion matrix displaying results from downsampled spikes for classification.

(I) Results of binary downsampled spikes classification. We failed to decode odor identity above chance. * indicates reward decoding accuracy exceeded chance.

See also Figure S3.

Of course, this does not address the specific transformation occurring when calcium flux is used as a proxy for spikes. In a motor planning task, the signals had different underlying modes of population activity, or PCs. Variance in the first PC in anterolateral motor cortex was due to temporal dynamics, in this case movement²⁴. Incidentally, the most convergence across electrophysiology, imaging and processing pipelines also occurred during movement. Does imaging primarily capture the first PC in spiking data? The largest determinant of variance in OFC spiking is reward value in most settings^45-50. To test whether imaging simply captured the first PC in our single-unit spiking, we performed a principal component analysis (20 PCs, ~80% variance) of our single unit population. These data are displayed in Figure 3D, where odor sampling trials are shown projected onto PC space, and the first PC separates the trials by reward. Using just PC1 for classification analyses also replicated results from imaging (Figure 3E-F).

It is generally accepted that single spikes are frequently missed in imaging data¹⁰, while high frequency bursts are supralinearly amplified⁵¹, yielding a sparser signal. Could the difference in the results between the two signals be accounted for by this data loss? To test this, we downsampled our spikes by a factor reflecting the median difference between average spiking and calcium transient rates (Figure S1D). This produced a distribution of spiking rates with a median frequency roughly equal to the deconvolved calcium activity (Figure 3G). As before, we randomly sampled 200-neuron pseudo-populations and performed identical analyses. Like the synthetic fluorescence and PC1, downsampled spikes produced a result that resembled imaging data (Figure 3H-I). Taken together, these results indicate that calcium activity in this setting is essentially a degraded estimate of the information contained in neural activity, where fluorescence captures primarily the information conveyed by the strongest changes in unit spiking.

We performed additional analyses designed to rule out the possibility that the difference in signal resulted because our unit separation was better or more accurate than the separation of neurons in the calcium signal. Figure S3 shows that unsorted “multi-unit activity” yields worse encoding of information about the odors than calcium signal, and that the information in the calcium signal was further degraded when we analyzed fluorescence data without demixing single neurons. Together, these results indicate that differences in information between the signals are unlikely to be caused by poorer single neuron “isolation” in the imaging data.

OFC electrophysiological ensembles acquire reward prediction representations prior to imaging

The analyses presented thus far have utilized activity following learning, when rats had already learned the meaning of the cues and were performing well. But might even more significant differences be evident in neural correlates of higher-order cognitive functions? Odor-evoked responses in OFC change during learning^17,34, and the OFC is critical to adapting such learning about individual problems to facilitate rapid acquisition of like-problems, a function evident in single-unit spiking in OFC^52-54. Serial learning of novel odor discrimination problems, as in the current study, could be performed more simply by learning that odors 1 and 3 share the same contrasting rules as odors 2 and 4; indeed, reward prediction is confounded since cues with different reward value necessarily had different sensory identities¹². This is typically addressed by training a classifier on one set of exemplars and testing it on the other^13,55. Such learning has been described as “abstract representation” or “deploying a schema”, and the ability to generalize rules across novel stimuli and contexts is thought to facilitate learning and other flexible behaviors^45,56-58. If rats acquired sensory-independent neural representations of reward prediction formatted in this manner then cross-cue decoding should improve with learning.

To test whether and how rapidly unit activity and calcium signal developed such correlates, we analyzed the trials at the beginning of each session. Sessions were separated into 12-trial blocks, three of each odor, totaling 96 trials. We then trained and tested a support vector classifier on all possible permutations of pairs of go/no-go exemplars. Sensory-independent reward information was absent during the initial trial blocks (Figure 4A, left), before rats were responding accurately. The development of these correlates in single-unit spiking coincided with the increase in behavioral performance (Figure 4A, right). In the fluorescence data, on the other hand, these correlates lagged several blocks behind the single-unit spiking (Figure 4B) and never reached the high accuracy shown by spiking data, despite the increase in behavioral performance (Figure 4B, right) and the acquisition of simple go/no-go based activity (Figures 1-2).

Figure 4. OFC electrophysiological ensembles acquire reward prediction representations prior to imaging.

(A) (left) Average cross-cue decoding performance across blocks of 12 trials in the electrophysiology data. Electrophysiology ensembles showed significant sensory-independent reward correlates starting during the second 12-trial block. Shaded areas indicate standard deviation from the mean. (right) Response accuracy averaged across all four odors during learning across 12-trial blocks. Shaded areas indicate standard error from the mean.

(B) (left) Average cross-cue decoding performance across blocks of 12 trials in the fluorescence data. Performance did not exceed chance until the fifth trial-block. Shaded areas indicate standard deviation from the mean. (right) Response accuracy averaged across all four odors during learning across 12-trial blocks. Shaded areas indicate standard error from the mean.

Here we have shown that traditional single-unit recording and single-photon calcium imaging produced different results during a 4-odor discrimination learning task. Calcium signals primarily represented whether odor cues signaled reward/non-reward, whereas spiking data, consistent with decades of unit-recording studies, contained rich, multiplexed representations of odor identity and reward. Spiking data rapidly supported generalization, while this information was slower to appear and attenuated in calcium imaging data. These data show that calcium activity is a degraded representation of unit spiking in cortical areas of an awake, freely behaving rat.

Several aspects of these results are worth discussing. First, although we used a 1-photon scope, we believe the findings apply also to 2-photon approaches. Transforming spike data into fluorescence, with a model that was developed using 2-photon imaging²⁴, replicated our 1-photon findings, findings essentially identical to what others reported using a similar discrimination task and 2-photon imaging in OFC³⁴. Second, we believe the information loss is due to the inherent noisiness and unpredictability in the spikes-to-fluorescence transformation. Indeed, even under ideal conditions – single cell recording and 2-photon imaging – calcium responses to 1-5 spike events had considerable overlap²⁴. Given one level of calcium sensor output, this makes inferring the precise number of spikes impossible. Of course, it remains an empirical question whether one-to-one concordance between spikes and calcium flux could be achieved under real-world conditions – outside proof-of-concept - with much improved sensors and supra-kilohertz imaging. However with current and even near-future approaches, it seems to us that better analytical approaches to capture the spike-to-fluorescence transformation are unlikely to fully recover the information lost. Consistent with this, deconvolution, which is thought to improve the match between spiking and calcium signal, only made the informational loss worse, and spikes were sensitive to any form of information degradation, all of which produced results that were similar to imaging. It therefore seems prudent to interpret calcium flux as signaling whichever variable/s drive the strongest, clearest changes in neural activity, with the understanding that information conveyed by subtler changes in spiking and multiplexing may be lost.

We do not intend to suggest that calcium imaging is not useful. The advantages conferred by the abilities to track neurons longitudinally^7,30, label them genetically or anatomically⁵⁹, or to image at a large scale⁶⁰, have increased our understanding of neural function. Calcium imaging is necessary if having such advantages is central to test a given hypothesis. However, in each of these lines of work, the underlying ground truth was tied to prior single-unit recording studies. Calcium imaging was used to extend our knowledge of a confirmed neural correlate. By contrast it is sometimes used as a stand-alone proxy for neural activity – without single-unit evidence and sometimes even contrary to it. The current data provide a striking illustration of this – even in a relatively simple odor discrimination setting, calcium activity lacked the sensitivity to resolve finer features of the neural activity it is thought to closely estimate. The absence of such features - sensory encoding, multiplexing - is not trivial since this is a result that is incompatible with current theories of OFC function^12,16,61-67.

Despite the known biases inherent to single unit recording⁶⁸, its early successes provided information that described fundamental roles of visual⁶⁹, hippocampal¹, motor², and dopaminergic areas³. These early findings helped spawn current theories of vision⁷⁰, cognition⁴, movement⁵, and learning⁶. We therefore contend that extracellular electrophysiology provides a clearer readout of neural signaling, whereas calcium activity provides an approximation. This makes sense when one considers that neurons communicate via electrical signals⁷¹, and calcium activity is several steps removed from these. We think the resultant limitations are critical to keep in mind when using this approach to characterize neural correlates when prior unit correlates do not exist or when information in the calcium signal does not match that in prior recording studies.

STAR Methods

Lead Contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact Dr. Geoffrey Schoenbaum (Geoffrey.schoenbaum@nih.gov).

Materials availability

This study did not generate unique reagents.

Data and code availability

All data reported in this paper are archived at DOI:10.17632/vwtgfj5vzg.1 and are publicly available as of the date of publication. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. This paper does not report original code.

Experimental model and subject details

Animals

Eight adult male, experimentally naive (n=8, n=4 electrophysiology, n=4 imaging) wild type Long-Evans rats were acquired from the NIDA-IRP breeding colony. Male rats were used since experiments were conducted during the SARS-CoV-2 pandemic; we were not able to order rats from an external vendor and so were required to utilize male rats, which were available from the breeding colony. Rats were three to four months of age and weighed 300-350 g during experiments. Rats were individually housed and given ad libitum food during behavioral training and testing. One day prior to behavioral training rats began water restriction, under which they received 10 m of free access to water, following each session, in addition to the water they consumed during testing. Rats were given free access to water on days they were not tested. Rats were maintained on a 12hr/12hr light/dark cycle and tested during the light phase between 7:00 am and 2:00 pm 5-7 days per week. Experiments were performed at the National Institute on Drug Abuse Intramural Research Program, in accordance with NIH guidelines.

METHOD DETAILS

Apparatus

Behavioral training and testing took place in custom aluminum boxes (18 in per side) equipped with a central odor delivery port and a well for fluid delivery. Task events were timestamped and controlled by custom software written in C++ and a system of relays and solenoid valves. Odor port and fluid well responses were detected by infrared sensors, and all timestamps were sent to the electrophysiological recording or imaging system.

Surgical procedures

Rats were anesthetized with isoflurane (3% induction, 1-2% maintenance in 2 L/m O₂) and placed in a standard stereotaxic device (Kopf Instruments, Tujunga, CA). In the electrophysiology group, two electrode bundles, each consisting of 16 individual 25 μm formvar insulated nichrome wires (A-M systems, Carlsborg, WA), were implanted. The bundles were each contained in 27 gauge cannulae and mounted in a custom-built 3d printed microdrive⁷². The centers of the bundles were separated by 800 μm and aligned on the M-L axis. Prior to surgery, electrode bundles were trimmed to 1-2 mm past the cannula tips and spread to allow at least 25 μm space between each wire. Arrays were centered unilaterally above right OFC: AP +3.0 mm; ML +3.2 mm and lowered 4.0 mm from brain surface. Headcaps were secured with 0-80 1/8” machine screws and dental acrylic and encased in a custom 3d printed protective cover. Rats received 5 mg/kg carprofen s.c. on the day of surgery and four days following, and 60 mg/kg p.o. cephalexin was administered for ten days following surgery to prevent infection. Electrodes were slowly advanced to final recording coordinates (~4.8 mm from skull, matched to GRIN lens implants) during surgical recovery.

All general surgical procedures were identical in the imaging group. These rats were infused with 0.5 μL of AAV9-GCaMP7f (Addgene 104488-AAV9) ⁷³ at the following coordinates relative to bregma: AP +3.0 mm; ML +3.2 mm; DV −5.6, −5.2, −4.8 mm. During this same surgery, a 0.5 mm diameter, 8.4 mm long GRIN lens (Inscopix 1050-002183) was implanted at the same AP and ML coordinates as the electrodes, and lowered 4.8 mm from the skull, referenced at the implant site. GRIN lenses were secured with cyanoacrylate glue and bone cement. Three weeks following infusion/implant surgery, rats were anesthetized and placed in the stereotaxic frame. Tissue under the GRIN lens was visualized with a Miniscope, and the focal plane was adjusted until healthy tissue, blood vessels, and fluorescent cells were in focus. The Miniscope was then chronically attached with dental acrylic and encased in a 3D printed protective cap.

Behavioral training

The general structure of each trial followed that of previously published work ¹⁷. During each recording session, rats learned to perform go/no-go responses to four novel odor cues. Two of the odors signaled rats to make a go-response to receive reward, and two signaled the rats to make a no-go-response to avoid a prolonged timeout. The onset of each trial was signaled by illumination of two house lights located on the wall above the odor port. Trials were initiated if rats entered the odor port within 5 s following light onset. Rats were required to remain in the odor port for 500 ms, after which odors were delivered. Rats were required to sample the odor for at least 500 ms. If rats did not remain in the odor port for at least 1 s total (500 ms following port entry, 500 ms following odor-on), the trial was aborted, the house lights were extinguished, and rats received a 6 s timeout. Upon odor port exit during go-trials, rats were required to respond at the fluid well within 2 s. A correct go-response led to delivery of 50 μL 5% sucrose in water, following a 400 to 1,500 ms variable delay, post fluid well entry. On no-go trials, rats were required to withhold responding at the fluid well for 2 s following odor port exit. Upon fluid well exit on go-trials, or the 2 s delay on no-go trials, the house lights extinguished, indicating the end of the trial. Correctly performed trials were followed by a 4 s intertrial interval (ITI), and error trials were followed by an ITI of 6 - 10 s. During each recording session, odors were presented pseudo-randomly such that the same odor could not occur more than three times consecutively, and odors would be sampled equally over a period of 240 trials. Rats were considered to have learned the odor-go/no-go associations once they reached a criterion of 90% accuracy in a moving block of 20 consecutive trials ^17,74.

Each trial was segmented into six epochs (Figure 1) - “light”, “poke”, “odor-on”, “odor-unpoke”, “choice”, and “outcome”. “Choice” was defined as the time when rats either entered the fluid well (when they responded within 2 s) or the time marking the end of the 2 s response window (when they did not respond). “Outcome” was defined as the time of fluid delivery (on rewarded trials) or the time 700 ms following the 2 s response window (on non-rewarded trials).

Histology

Rats were euthanized by carbon dioxide, and their brains were removed and fixed in 4% formaldehyde in PBS overnight. 0.05 mm sections of OFC were taken, mounted, cover slipped with DAPI mounting medium, and visualized using a Keyence BZ-X microscope.

Quantification and statistical analysis

Behavior

Behavioral data were collected using custom software written in C++. Raw data were processed and analyzed in Matlab 2020b (Mathworks, Natick, MA). The number of trials to reach criterion during each session was compared across groups by two-sample Kolmogorov-Smirnov test test. Each rat’s average reaction time across all sessions was compared across groups by two-sample Kolmogorov-Smirnov test. When accuracy was analyzed during learning in figure 4, we averaged accuracy across all four odors during each block of 12 trials.

Electrophysiology

Electrophysiological signals were recorded with a Plexon OmniPlex (Plexon, Dallas, TX) system. Neural signals were digitized, amplified, and bandpass filtered (300 – 8,000 Multiunit activity. We used scripts modified from the Signals on each electrode channel from each session were median referenced prior to spike detection. We used only the same channels from which we isolated single units. For spike detection, the wideband signals were bandpass filtered between 300Hz and 3000 Hz (elliptic filter). Spike times were first detected using a 4th order filter, and then the corresponding waveforms were extracted from 2nd order filtered data. A negative spike detection threshold of −5σ, where σ = median(∣x∣/0.6745), and x = the filtered signal (calculated for each 5 minute recording period). A maximum threshold of −50 σ was used to exclude possible noise. Putative spikes were extracted as 64 point (1.6 ms) snippets of data, with the signal peak aligned to data point 20 (0.5 ms). The signal peak was defined as the minima up to 0.5 ms post threshold, and refined using a cubic spline-interpolated waveform with 320 samples (downsampled back to 64 points after alignment). The minimum inter-spike interval was limited using a detection "dead-time" of 0.2 ms.

Next, the extracted spikes were automatically clustered using default parameters with Wave_clus 3^75,76 [https://github.com/csn-le/wave_clus]. Clustering was performed on 20,000 spikes (randomly sampled across the session), and then the remaining spikes were included into the clusters using template matching. Finally, clusters were manually inspected and categorized as noise or multi-unit activity (i.e. combining any isolated or multi-unit clusters).

Calcium imaging

Imaging procedures closely followed those in previously published work ²⁷. Images (480 x 752 pixels, 30 Hz) were collected with in-house produced miniaturized fluorescence microscopes (V3, Miniscope.org) via a CMOS sensor (Labmaker.org) interfaced to a custom data acquisition board via a 1.5 mm coaxial cable and custom-made commutator. Raw data were written to .avi files, which were spatially downsampled by a factor of 3 (160 x 251) and motion corrected in Matlab with NoRMCorre³⁶. Signal extraction, demixing, denoising, and deconvolution were performed using constrained non-negative matrix factorization and fast nonnegative deconvolution ^26,37,77.

Contour-free analysis.

We performed a “contour-free” imaging analysis, which blindly segments imaging data into a grid where each tile approximates the area of a single neuron. This analysis is agnostic to individual cells’ contributions to the overall fluorescence of each tile, and so can provide the imaging equivalent of “multiunit” activity⁷⁸. Motion corrected .avi files ( 160 x 251 pixels ) were imported into Matlab (Matlab importdata) and segmented into a 6 x 9 grid of 27 x 27 pixel tiles (729 pixels²). We chose this tile area since it closely matched the mean footprint size across the 463-neuron imaging population (757.80 pixels²). The rightmost tiles were 8 pixels in width and did not contain any active neurons and so were not analyzed. The bottom-most row consisted of 24 x 27 pixel tiles. Tiles that contained no active neurons or spillover fluorescence from adjacent tiles were not analyzed (Figure S3C). The median fluorescence of each tile was calculated for each imaging frame and concatenated into an n-dimension x time activity space. Each of these n-vectors (totaling 663 tiles) was analyzed identically to the single neuron fluorescence activity for the analyses in Supplemental Figure 3.

Neural analysis

Most neural analyses utilized average activity during the 594 ms (18 imaging frames or 600 ms for simplicity) immediately preceding odor port exit (i.e., odor sampling). In the multiphasic analysis described below we used 528 ms of activity (16 imaging frames); again, we say 500 ms for simplicity. Spikes were sorted into 33 ms bins, equal to a single imaging frame. Activity trains were smoothed with a gaussian kernel, SD=1. Colormaps were adopted from the viridis package ⁷⁹ to maximize mono-, di-, and trichromatic linear perception and uniformity.

Single neuron analyses.

All trial numbers were chosen to allow sufficient sample sizes and statistical power ^17,19,20. Neurons were deemed reward modulated if average activity during the 16 positive versus 16 negative trials immediately following criterion was significantly different as indicated by Friedman test where positive/negative odor identity was treated as a nuisance variable. Neurons were considered odor identity modulated if average activity during the two pairs (eight trials each) of positive or negative exemplars was significantly different as indicated by Rank Sum test. The non-parametric Friedman test does not test for interactions ⁸⁰, so we utilized a two-way ANOVA to test for reward+odor identity signaling, where neurons were classified as such if the p-value for the interaction between reward and odor identity was below a stricter criterion (p < 0.025). Neurons were considered “multiphasic” if they met the following three criteria: 1. Average activity to each of the 4 odors during the 500 ms following odor-on was not the same, as indicated by Kruskal-Wallis test;2. Average activity on each of the 4 trial types during the 500 ms following odor-port-unpoke was not the same, as indicated by Kruskal-Wallis test; 3. The pattern of pair-wise comparisons was not the same during these two periods, using Tukey’s correction for multiple comparisons ⁸¹. Thus, neurons were considered multiphasic only if they showed significant differences in activity during both odor sampling and following odor port exit, and if the selectivity pattern during these epochs was different. These criteria are like those used previously ²⁴, but with additional constraints to accommodate the complexity of our task.

Classification analyses.

Classification analyses all utilized average odor-sampling activity as described above and were performed using a support vector classifier (Matlab fitcecoc). When classification accuracy was compared across groups, we used a two-sample Kolmogorov-Smirnov test. In population analyses where all four odors were considered, we trained the model on all but one randomly selected trial of each odor and tested on the left-out trials. We performed 2,000 trial-order-shuffled simulations, and accuracy was compared to label-shuffled data. To compare the total number of mis-classified trials across the signals, we calculated the difference in the number of misclassifications for each simulation and computed a 2,000-sample bootstrap confidence interval on the mean of this difference. For the PCA-reduced data, we counted z-scored average odor sampling firing rates as observations and neurons as variables, and we used the singular value decomposition algorithm and the first 20 PCs explaining 78.38% variance (units) and 78.31% variance (fluorescence); we say 80% variance for simplicity. Pairwise distances between trial-type pairs in reduced data were compared by two-sample Kolmogorov-Smirnov test.

In binary classification analyses, classifiers were trained on one randomly selected half of the data and tested on the other. We performed 2,000 trial-order-shuffled simulations. For the population subsampling analyses we randomly sampled 10-200 neurons for each simulation 10,000 times. Decoding accuracy was considered above-chance if the average accuracy exceeded the average of the 95^th percentile of label-shuffled simulations.

Cross-cue classification was performed using average activity during odor sampling as mentioned above. Rather than leave-out, validation was performed by training classifiers on one data set (odor 1 / odor 3 trials) and testing on separate data (odor 2 / odor 4 trials). The order of training and testing utilized all possible permutations: 1/3-2/4 , 1/4-2/3 , 2/4-1/3 , 2/3-1/4. The trial order was shuffled in the training set for each iteration, and this process was repeated 10,000 times for each permutation. Accuracy was compared to label-shuffled data as before and was considered above chance if average accuracy exceeded the 95^th percentile of shuffled data.

Spikes-to-fluorescence.

Spike times during odor sampling time t_k were convolved with a double exponential kernel: c(t)=Σ_t>tkexp(− t−t_k / τ_d)[1−exp(−t−t_k / τ_r)]+n_i(t). c(t) was then converted to synthetic fluorescence through a sigmoidal function: F_m/(1+exp[−k(c(t)−c_1/2)]) + n_e(t). The constants τ_r, τ_d, k, c_1/2, and F_m represent sensor rise and decay times, nonlinearity sharpness parameter, half-activation parameter, and maximum possible fluorescence change. Since these parameters are unknown for single-photon data, we used published rise and decay times of GCaMP7f⁷³, k = 5, F_m = 9, with all other parameters set to default according to previously published procedures²⁴. For each trial, this produced a vector of 33 ms bins (totaling ~594 ms) of synthetic fluorescence, similar to real fluorescence data or to spikes that were sorted into 33 ms bins. Like the other analyses, this ~594 ms vector was averaged to yield average “activity” during odor sampling.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
	This paper	DOI:10.17632/vwtgfj5vzg.1
Bacterial and virus strains
GCaMP7f viral vector	Addgene	https://www.addgene.org/104488/
Experimental models: Organisms/strains
Wild type Long-Evans Rat	NIDA IRP Breeding colony	n/a
Software and algorithms
MATLAB 2020b	Mathworks	RRID: SCR_001622
Plex Control software	Plexon	RRID:SCR_014803
Other
3d printed headcap	Custom made in house	https://github.com/SchoenRats/3D-Parts/blob/main/STL%20CAD%20files/Oval_Base_v2.stl
3d printed microdrive	Custom made in house	https://github.com/evan-hart/3D-Parts
Microwire	A-M systems	Catalog No. 761500
Plexon Omniplex	Plexon	https://plexon.com/plexon-systems/omniplex-neural-recording-system/
CMOS sensory	Labmaker	https://www.labmaker.org/collections/miniscope-v3-2/products/cmos-pcb-miniscope
GRIN lens	Edmund Optics	https://www.edmundoptics.com/p/18mm-dia-670nm-dwl-00mm-wd-uncoated-grin-lens/19231/
GRIN lens	Inscopix	Part 1050-002183
Miniscope	Custom made	Miniscope.org

Highlights.

• Orbitofrontal spiking multiplexed reward predictions with odors

• Orbitofrontal calcium signaled reward predictions

• Calcium activity represented degraded spike activity during learning

• Spiking acquired sensory-independent reward prediction information prior to calcium