Prediction errors explain mismatch signals of neurons in the medial 4 prefrontal cortex

31

stimulus arrives, such as a deviant event, there will be a failure to predict the bottom-up input and to 68 suppress the PE signal. This PE signal is used to optimize the encoding of sensory causes at higher 69 levels 8,10,11 . Thus, oddball mismatch signals are the result of the two components: RS and PE 13 . This  The MMN has interacting generators in a frontal-temporal network [15][16][17] , where the prefrontal cortex 74 (PFC) is believed to play an essential role in contextual processing and thus, in predictive coding 18 . 75 Studies in humans have shown evidence of PE-and even prediction signals in frontal cortices 76 (FCs) [19][20][21][22][23] . Some of those studies were carried out using human electrocorticography (ECoG). ECoG 77 reflects the population network dynamics because it records high-gamma activity (80-150 Hz) and 78 measures multiunit-but also synaptic activity 24 . Additionally, the clinical electrode placement is prefrontal mismatch responses to passive oddball sounds [26][27][28] . However, one study in the PFC of the 89 macaque with deep electrode recording failed to replicate the strong mismatch signals found in 90 primates 29 . Also, these animal studies lacked appropriate controls to interpret the generative system 91 of PE 13,30 . Under the oddball paradigm, two components contribute to the extraction of the standard-92 repetition rule: the RS, or attenuation of the evoked response to a specific repeated stimulus feature, 93 and more complex forms of predictive activity, such as the generation of statistical inferences. 94 Hence, the details and neuronal contribution of RS and PE at the PFC are currently unknown. In the present account, we recorded neuronal activity in the rat medial prefrontal cortex (mPFC) 97 under an oddball stimulus paradigm and adequate control sequences that allowed us to separate 98 mismatch signals into RS and PE 14 . We found robust mismatch signals that were explained by 99 maximal indices of PE across the mPFC. Our findings support the notion that mPFC neurons detect 100 unpredictable deviations from the auditory background. These cells may, therefore, represent the 101 neuronal basis of predictive activity in FCs.

105
Context-dependent responses across fields in mPFC. 106 To seek experimental evidence for predictive coding signals in the mPFC, we recorded neuronal   responses to standard, cascade, and many-standards conditions, as compared to deviant sounds of the 123 same frequency and intensity (Fig. 3a). While typically the deviant sounds evoked the largest 124 responses, in some cases other conditions also evoked robust firing rates, as the cascade descending 125 control example from the PL field ( Fig. 3a; cyan line). The latency and amplitude of mPFC 126 responses differed per multiunit. mPFC multiunits frequently presented a peak of maximum firing 127 followed by a decay (e.g., the AGm, ACC and PL examples in Fig. 3a), whereas some cells showed a 128 period of sustained firing as the IL example in Fig. 3a. For each multiunit, we tested 1-8 pairs of 129 deviant-standard combinations among the 10 frequencies comprising the control sequences. We 130 calculated the baseline-corrected spike counts per trial from 100 to 600 ms after sound onset (see 131 Methods, Fig. 1c). Overall, within the same multiunit, spike counts were larger for deviant tones, 132 either ascending or descending, than for any other condition regardless of the sound frequency ( Fig.   133 3b). Accordingly, neurons in the mPFC showed poorly tuned frequency response areas. We tested a 134 broad range of sound frequencies and sequences for each multiunit and found that mPFC cells 135 exhibit a strong and dynamic sensitivity for stimulation context rather than for spectral processing. for each multiunit. Among the 10 tones, a target tone fi (i=1-10; highlighted in color) was part of an oddball ascending or 140 descending sequence as a deviant or standard stimulus (left column). Deviant events were presented in a pseudo-random 141 fashion with a minimum separation of three standard events. Two conditions controlled for the repetition suppression 142 effect exerted over the target tone: cascade and many-standards sequences (right column). The many-standards sequence 143 comprises the random presentation of the 10 tones, and thus, the target tone is unpredictable. In the cascade sequences, 144 the 10 tones are presented in a regular succession of ascending or descending frequency, making the target tone 145 completely predictable. Responses to deviant events are compared to frequency-matching tones presented as the last 146 standard event (i.e., the standard preceding a deviant tone), the many-standards control, and the corresponding cascade 147 ascending or descending sequence. b Decomposition of the index of neuronal mismatch (iMM), under the framework of 148 the predictive coding theory. The iMM is calculated as the difference between the responses to deviant and last standard 149 8 conditions. The index of repetition suppression (iRS) is calculated as the subtraction of the standard response to one of 150 the controls. The remaining part of the mismatch signal not assigned to iRS is the index of prediction error (iPE). C

151
Population firing rate as normalized spike-density functions (mean ± SEM), n = 83 multiunits in the mPFC) for 8 152 consecutively presented tones within a cascade sequence (green trace), many-standards sequence (yellow trace) and 153 oddball sequence with the last standard tone (blue trace) before and after a deviant event (red trace). Tones are illustrated 154 as gray horizontal lines and lasted 75 ms with an inter-stimulus interval of 500 ms. Baseline spontaneous firing rates 155 were computed in the time window from 0-50 ms after stimulus onset (gray dotted lines). Analysis windows to compute

172
Ascending and descending denote whether the target tone's frequency was lower or higher than the previous tone,      (Table 1). Therefore, in the following, we will report only analyses obtained 203 under the cascade control because this sequence not only controls for the presentation rate, as the 204 many-standards condition does, but also for the refractoriness of the response and RS effects due to denoted as n.s. (non-significant) and *** (p < 0.001), within-field multiple comparisons Friedman test (see Table 1).

217 218 219
We computed the index of neuronal mismatch (iMM = deviantstandard), which is the classical decrease from the control condition to the standard condition, which is due to repetition effects 31 .

225
The index of prediction error (iPE = deviantcontrol) denotes the relative response increase from 226 the control condition to the deviant condition, which represents the error signal generated when a 227 prediction is inaccurate 13,30 . The iMM revealed values close to the index maximum and was 228 significantly larger than zero in all the mPFC fields (within-field multiple comparisons Friedman 229 test, Table 1, Fig. 4b). More remarkably, all these robust mismatch signals were due to strong and 230 highly significant PE signals, because RS signals were absent and not significant in any field; iRS of  The temporal dynamics of mismatch and prediction error signals coincide in time. 251 To identify the overall response pattern of each mPFC field, we computed the population temporal 252 dynamics of the average firing rate as normalized spike-density functions (SDF; Fig. 5a). The 253 neuronal firing rate was very similar between cascade and many-standards controls in all the fields at   Fast time course of the repetition suppression effect to predictable auditory input. 310 In order to explore the dynamics of adaptation or RS effect to repetitive stimuli over time, we 311 averaged the responses to deviant, standard, and cascade stimuli across recordings for every trial 312 number within the sequence. Thus, we averaged firing rates at their absolute position within the 313 sequence and generated the time course of responses from the beginning of the sequence within each 314 field (Fig. 6a). The variability of deviant, many-standards and cascade responses in all fields was 315 minimally explained with the tested models (linear, exponential, double exponential, inverse 316 polynomial, and power-law models; adjusted r 2 < 0.15). Within these conditions, two tones were 317 never repeated and did not undergo the same influence of the repetitive effect as standard events. By  The models demonstrated a comparable rate of RS among fields (b parameter [with 95% confidence it only needed a second stimulus repetition to generate a response decay less than 50% of the initial 346 response (AGm 39.4%; ACC 46.6%, PL 37.6%, IL 28.6%; Fig. 6b, arrows). A third repetition 347 further attenuates the response, to levels comparable to the steady-state, where the firing rate remains 348 constant until the end of the sequence (Fig. 6b, note the overlap between mean ± standard error of the  Strong responses to unpredictable sounds under a regular context of silence. 356 We studied the effect of yet another type of regular stimulation context on the PE signals generated In this study, we recorded neuronal activity while presenting an oddball stimulus paradigm across all 375 mPFC fields in the anesthetized rat. Neurons in mPFC showed robust neuronal mismatch signals to 376 unpredictable events. We performed a quantitative separation of neuronal mismatch signals into PE 377 and RS components with the cascade and many-standard control sequences as we did previously at 378 lower levels of the auditory hierarchy 14 . We found that maximal iMMs are almost exclusively due to 379 iPEs. We also verified that this PE mechanism generalizes to multiple frequency sounds within the 380 same neurons. At the population level, we observed a delay period of sustained activity beyond the 381 presentation of deviant events, which may represent a time window that eliminates PEs 382 progressively. Importantly, we found that RS over repeated stimuli was extremely fast, such that 383 only two repetitions of the standard tone suffice to encode a regularity representation. The fact that 384 this generative system of PE signals is present in urethane-anesthetized rats, in the absence of 385 behavioral relevance or wakefulness, suggests the mechanism of hierarchical inference as a 386 fundamental process of the rat mPFC.

388
In the rat mPFC, we found maximal iMMs, which were disentangled into maximal iPEs and non-389 significant iRSs using two controls, the many-standards and cascade sequences 30,35 . These indices Our findings under urethane-anesthesia, in the absence of attention or reward-related behavior, 431 suggest that the mechanism of hierarchical inference is a fundamental process of the rat mPFC.

432
Using the same methodological approach, we recently showed slightly higher iPEs in the awake state  Fig. 7) 14 . In other words, responses to 440 our predictable events will be "filtered out" at lower levels of the hierarchy such as ACs to prevent 441 their arrival at a higher cortical processing level, i.e., mPFC. By contrast, deviations from 442 expectations that cannot be explained in lower levels will be forwarded to the mPFC and registered 443 as robust iPEs (Fig. 7) 10,32 . Hence, the rat mPFC acts as a higher level in the hierarchy and PE are   accommodated in a stereotaxic frame with hollow specula to facilitate direct sound delivery to the 528 ears. Rectal temperature was maintained at ~37 ºC with a homeothermic blanket system (Cibertec). 529 We surgically exposed bregma by making an incision in the scalp at the midline and retracting the 530 periosteum. A craniotomy of ~3 mm in diameter was performed above the left mPFC and the dura 531 was removed. 532 533 Neurophysiological recordings. 534 We recorded neuronal activity to look for evidence of predictive coding signals under acoustic dorsoventrally. Therefore, we covered the four fields of the mPFC and various cortical layers (II-VI). 541 We performed extracellular neurophysiological recordings with glass-coated tungsten  Histological procedures and verification of recording sites. 553 The neuroanatomical location of the recording tracts was marked with electrolytic lesions. Post-554 mortem brains were fixed with 4% paraformaldehyde in phosphate-buffered saline and cryoprotected 555 in 30% sucrose. 40 µm sections were cut in the coronal plane with a freezing microtome and Nissl-556 stained with 0.1% cresyl violet to visualize cytoarchitectural landmarks. Histological assessment of 557 the electrolytic lesions to any of the fields of the mPFC was processed blindly to each animal history.

558
Multiunits locations were assigned to AGm, AC, PL or IL within a rat brain atlas, accordingly with 559 the histological verification and the stereotaxic coordinates in the three axes of recording tracts 68 .  Trains of white noise bursts of 75 ms duration with 5 ms onset and offset ramps were presented to 569 search for neurophysiological responses to acoustic stimuli. While searching, sound presentation rate 570 and intensity were modified online to prevent strong response adaptation. All experimental 571 paradigms consisted of pure tones of 75 ms duration with 5 ms onset and offset ramps at a 572 presentation rate of 2 Hz. To identify neurons suitable for recording, we computed their frequency-573 response area or receptive field, which consisted of tones of various frequency and intensity 574 combinations that ranged from 1 to 44 kHz (in 4-6 frequency steps/octave) and 0 to 70 dBs (10 dB 575 steps) and were presented randomly with 1-3 repetitions per tone. We found sound-driven 576 multiunits, but neurons in the mPFC did not show clear auditory receptive fields for pure tones.

577
Thus, for each multiunit, we selected tones within the audible range to generate control sequences 578 ( Supplementary Fig. 1). Control sequences consisted of 10 tones evenly spaced by 0.5 octaves 579 delivered at a fixed sound intensity for the same multiunit and varied among multiunits. Among 580 those 10 tones, we selected pairs of consecutive frequencies to generate oddball sequences. Both 581 control and oddball sequences lasted 400 tones with a 500 ms interstimulus interval (2 Hz). Each 582 target or studied tone of a specific frequency was presented as a deviant or standard condition within 583 an oddball paradigm, as well as part of the many-standards and cascade control sequences. This 584 approach allowed comparing the same physical stimuli within various stimulations contexts. We 585 presented the sequences in a random fashion, with periods of ~12 mins of silence between sequences 586 to minimize long-term habituation effects 69 .

588
We used the oddball paradigm to find neuronal evidence of predictive coding signals to the violation 589 of regularity rules. Oddball sequences consisted of frequently repeating stimuli (standard tones) 590 presented with a 90% probability, which were pseudo-randomly interleaved with rare events (deviant 591 tones) occurring with a 10% probability (Fig. 1a). Sequences started with a minimum of 10 592 repetitions of standard stimuli, and a minimum of 3 standard events separated deviant tones. We 593 classified deviant sounds as ascending or descending depending on whether the frequency of the 594 standard tone was lower or higher, respectively (Fig. 1a).

596
We generated the many-standards and cascade controls as sequences with 10 tones in a different 597 presentation order to account for the contribution of RS effect of mismatch signals exerted to the 598 target tone. Thus, we can identify the remaining part of the mismatch signal that is not explained by 599 the RS effect as the PE signal (Fig. 1b). Both sequences control for the presentation rate because they 600 were delivered at the same rate than the oddball paradigm. The many-standards control is the 601 consecutive presentation of blocks of 10 tones randomly ordered within the block (Fig. 1a) 35 . The 602 target tone was among those ten tones and presented at the same ratio as the deviant condition. The 603 frequency separation (in octaves) between the tones in the many-standards sequence was equal to the 604 separation between deviant and standard in the oddball sequence. The target tone is unpredictable in 605 both oddball and many-standards sequences. The many-standards control cannot establish a 606 prediction or internal rule as the oddball paradigm because an unpredictable sequence of tones 607 replaces the repetition of standard tones. Therefore, the many-standards control cannot account 608 precisely for the undergone RS effect during the oddball sequence.

610
We also used two cascade control sequences, which consist of the regular presentation of 10 tones 611 (the same as in the many-standards sequence) in ascending or descending frequency succession (Fig.   612 1a) 30 . PE signals to the target tone are minimized in the cascade sequence because it is embedded in 613 a predictable context, which conforms the internal representation rule and undergoes a RS effect.

614
Ascending and descending deviant targets were compared with the corresponding ascending and 615 descending cascade paradigms. Thereby, we controlled for the state of refractoriness and pitch 616 gliding effects that the preceding tone could exert over the target stimulus. In sum, the cascade 617 sequence makes a more rigorous control (for review 13 ).

618
After controlling for RS effects on target tones, we studied the effect of yet another type of regular showed a response latency of ~150 ms with a sustained firing spanning up to the next tone (Fig. 1c). 640 We calculated the baseline spontaneous firing rate as the mean firing rate from 0-50 ms during the 641 tone presentation. We measured baseline-corrected spike counts as the spiking activity that exceeded 642 the firing rate of the baseline. This baseline-corrected spike count is the area above the baseline and 643 below the SDF in the period of 100-600 ms after stimulus onset. To avoid overlap of consecutive 644 tone responses, the response analysis window preserved the interstimulus interval of 500 ms and was 645 delayed 100 ms (Fig. 1c). Although the many-standards and cascade sequences controlled for RS effects to target tones, we can 659 separate the neuronal mismatch signal into RS and PE with just one control (Fig. 1b). Both In order to discern the response latencies among fields of the mPFC, we measured latencies as the where iPE is the distribution of iPE values for our sample of frequency tested tones per field (Table   714 1). Note that our sample sizes were ~14 times larger than the required sample due to the 715 homogeneity of neuronal responses of the mPFC. As mentioned, these recordings were also diverse 716 in anatomical localization and therefore, we consider the current sample size as a representative 717 sample of each field.