Dynamic coding of choice targets in the perirhinal cortex

Abstract Cortical neurons flexibly respond to multiple task-events characterized by distinct computational aspects. There are two possible explanatory scales, dynamic population code realizing computational flexibility and single-neuron code associating relevant information across task events. Since these interpretations were independently developed, it is unclear how neurons contribute to population dynamics with their individual ability to integrate relevant information. Here we show systematic changes in single-neuron representations of choice targets shape population dynamics. We analyzed neural activities in the perirhinal cortex while rats performed a two-alternative target-choice task. A time-resolved pattern analysis indicated dynamic population code of choice targets with two time-stable states during the cue and reward periods. The population patterns during these periods were substantially reversed and therefore allowed us to decode task-periods as well as choice targets. Our results clarify the contribution of single-neuron representations to population dynamics, suggesting its potential advantages such as reconciling different computational demands.

Since these interpretations were independently developed, it is unclear how 18 neurons contribute to population dynamics with their individual ability to 19 integrate relevant information. Here we show systematic changes in 20 single-neuron representations of choice targets shape population dynamics. 21 We analyzed neural activities in the perirhinal cortex while rats performed a 22 two-alternative target-choice task. A time-resolved pattern analysis 23 indicated dynamic population code of choice targets with two time-stable 24 states during the cue and reward periods. The population patterns during 25 these periods were substantially reversed and therefore allowed us to decode 26 task-periods as well as choice targets. Our results clarify the contribution of 27 single-neuron representations to population dynamics, suggesting its 28 potential advantages such as reconciling different computational demands. Individual neurons across cortical areas show temporally flexible responses 32 to multiple task-events characterized by different computational aspects, 33 such as cue, action and reward (e.g., Sakurai et al, 2017;Sakurai, 1990a,b). 34 Recent studies indicate that such complex single-neuron responses are only 35 interpretable in terms of their contribution to population dynamics which 36 flexibly perform different computations, particularly in association and 37 motor cortices (Elsayed et al., 2016;Lara et al., 2018;Mante et al., 2013;38 Raposo et al., 2014). On the other hand, in some cortical areas, it has been 39 shown that individual neurons respond to multiple task-events while question how such neurons contribute to population dynamics which would 47 be suitable for differentiating task events. 48 In the present study, we explored neural activities in the perirhinal cortex 49 (PRC) which has been implicated in association memory of objects (Ahn and 50 Lee, 2017Lee, , 2015Naya et al., 2003aNaya et al., , 2003bNaya et al., , 1996Sakai and Miyashita, 1991). 51 This region receives sensory inputs from almost all modalities, behavioral 52 context information from the prefrontal cortex and reward-related signals 53 from the amygdala (Burwell, 2001;Burwell and Amaral, 1998;Furtak et al., 54 2007;Tomás Pereira et al., 2016), creating integrated object representations 55 (Ohyama et al., 2012;Qu et al., 2016;Taylor et al., 2006). Some studies show 56 that some PRC neurons respond to the same visual cue across time contexts 57 but in different degrees of strength (Naya et al., 2017;Naya and Suzuki, 58 2011). Therefore, it is possible that the PRC is able to represent the same 59 target even across explicitly different task-events such as cue, action and 60 reward, while differentiating those task-events. 61 To determine this possibility and elucidate underlying mechanisms, we 62 analyzed single-unit activities recorded from the PRC while rats performed a 63 two-alternative target-choice task. We used randomly-interleaved visual and 64 olfactory cues to investigate choice-target representations. We found that 65 . CC-BY 4.0 International license It is made available under a (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/514612 doi: bioRxiv preprint first posted online Jan. 9, 2019; 3 dynamic population code for choice targets evolved through the task. 66 Systematic shift of the choice-target code represented task-periods and was 67 meditated by neurons characterized by activation across the task-periods. 68 Our results clarify the contribution of single-neuron responses to population 69 dynamics, suggesting potential advantages of flexible but structured 70 single-neuron responses as encoding strategy. We trained four rats to perform a two-alternative target-choice task where 76 they chose a target port (left/right) associated with a presented cue to obtain 77 reward (Figure 1a-b). The task performance was similar level regardless of 78 the cue modality (mean correct rate in visual trials = 94.9±5.8 %; olfactory 79 trials = 93.3±3.8 %). We recorded spiking activities from the PRC (n = 182 80 neurons) during the task performance. 81 The PRC neurons frequently encoded choice targets during the task 82 performance. We used ROC analysis to generate a target-preference index 83 which quantifies preferential responses of a neuron to a choice target. We CC-BY 4.0 International license It is made available under a (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/514612 doi: bioRxiv preprint first posted online Jan. 9, 2019; 4 was sustained during the presentation of the cue and was followed by 102 transient response during movement toward a target port. Soon after the 103 rats chose a target port, the target code settled again into a time-stable state 104 ( Figure 2a). We computed mean performance of the classifier during the cue 105 and reward periods and compared them with the baseline (−400 to 0 ms 106 before cue onset). As shown in Figure 2b, we successfully decoded choice 107 targets in both task-periods (P < 0.01), indicating that the PRC tracks target 108 information across task-periods. Furthermore, in Figure 2c, we found 109 decreased target-code when we decoded choice targets using correct visual 110 trials but erroneous olfactory trials (n = 89 neurons). To quantify this effect, 111 we computed the mean classification performance in correct and erroneous 112 trials. As shown in Figure 2d, the mean performance in both task-periods 113 decreased to chance level in erroneous trials despite the fact that animals 114 chose the same target-port as correct trials (cue period: P ≈ 0.10 ; reward 115 period: P ≈ 0.24). These results indicate that the target code was relevant to 116 successful task performance. 117 Importantly, population patterns during these two task-periods were 118 substantially reversed (white arrows in Figure 2a), suggesting that the PRC 119 distinguished different task-periods by systematic shift of the target code. 120 We computed mean classification performance across the two task-periods 121 and found significant shift of the target code (P < 0.01; Figure  The shift of target code is possibly signature of the encoding strategy 134 which allows the PRC to integrate the different information, choice targets 135 and task periods. If it is the essential nature of the PRC rather than noise, 136 we should be able to decode richer information from neurons contributing to 137 . CC-BY 4.0 International license It is made available under a (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/514612 doi: bioRxiv preprint first posted online Jan. 9, 2019; 5 the shift than the others. We therefore repeated the pattern analysis for each 138 of the following types (classified in Figure 1f): neurons with shifting response, 139 sustained response, and specific response (for either task-period). As In this study, we showed that the PRC dynamically encoded choice targets 156 during task performance. We demonstrated that the target code was highly 157 consistent across cue modality and reduced in erroneous trials. These results 158 suggest that neural responses during the cue period might reflect retrieved 159 target as shown by electrophysiological studies, with visual tasks, both in 160 rodents (Ahn and Lee, 2017) and non-human primates (Naya et al., 2001;161 Naya et al., 1996). Our results thus provide substantial evidence for 162 integration across sensory modalities which might support object memory. 163 We also found that the PRC reversed its target code between cue and 164 reward periods. This allowed us to decode task periods as well as choice CC-BY 4.0 International license It is made available under a (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/514612 doi: bioRxiv preprint first posted online Jan. 9, 2019; 6 encoding strategy which reconciles different demands, keeping target 174 information and differentiating task-events. This suggests that functional 175 flexibility could be achieved by explicitly structured single-neuron responses. 176 Moreover, this structured code has additional benefits. First, it can be easily We trained the rats step-by-step to perform the task described above. The 230 training period typically lasted 4 to 7 weeks. First, rats were trained to poke 231 into the central port and then collect the water reward (0.02 ml) from the left 232 or right port. We gradually extended the duration of the poke by delaying the 233 go sound up to 1 s after the poke onset. Next, the rats were trained to 234 discriminate visual cues based on the same contingencies as recording 235 sessions. A variable delay (200-600 ms) was inserted before the cue onset. 236 After the rats became able to successfully discriminate visual cues (> 80%), 237 they were also trained to discriminate odors based on the same contingencies 238 as recording sessions (> 80%). Finally, we interleaved visual and olfactory 239 trials within a session and trained the animals to perform the task to a 240 training performance criterion (> 80%). CC-BY 4.0 International license It is made available under a (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/514612 doi: bioRxiv preprint first posted online Jan. 9, 2019; CC-BY 4.0 International license It is made available under a (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/514612 doi: bioRxiv preprint first posted online Jan. 9, 2019; recording sessions to maximize the number of neural responses included in 318 the decoding analysis. Here we refer to the pooled pseudo-population of the 319 PRC neurons (n = 182) as full population. The instantaneous firing rate of 320 each neuron was estimated by spike counts in a 150 ms sliding window (10 321 ms increment). We converted the instantaneous firing rates for each neuron 322 into the target-preference index independently calculated for visual and 323 olfactory trials. In this manner, we generated two independent population 324 vectors for the full population (cell×time vector each for the cue modalities). 325 We obtained a pattern similarity index by calculating the task-periods that were main focus of our study, the cue and the reward 337 periods. We repeated this process 100 times to obtain a distribution of 100 338 different measurements of the pattern classification. To investigate neural 339 responses during the task periods, we averaged the performance within the 340 cue period and reward period (e.g. Figure 2b). To obtain a baseline 341 performance, we averaged the classification performance during the baseline 342 period, −400 to 0 ms before the cue onset. We also quantified the temporal 343 change of neural responses between these task-periods by averaging the 344 following two pattern classifications: pattern classification for cue-period 345 responses in visual trials and reward-period responses in olfactory trials, 346 and cue-period responses in olfactory trials and reward-period responses in 347 visual trials (e.g. Figure 2e). We calculated the 95th percentile range for each 348 of distribution. If the 95th range of 100 different pattern classification 349 measurements within a task period did not overlap with the 95th range of 350 baseline distribution, we considered the choice targets to be decodable from 351 the neural population response during the task event. We considered zero to 352 be chance level instead of the baseline distribution when we checked 353 . CC-BY 4.0 International license It is made available under a (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.
The copyright holder for this preprint . http://dx.doi.org/10.1101/514612 doi: bioRxiv preprint first posted online Jan. 9, 2019; 11 statistical significance of the classification across the task periods. CC-BY 4.0 International license It is made available under a (which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.