Learning-related population dynamics in the auditory thalamus

Learning to associate sensory stimuli with a chosen action involves a dynamic interplay between cortical and thalamic circuits. While the cortex has been widely studied in this respect, how the thalamus encodes learning-related information is still largely unknown. We studied learning-related activity in the medial geniculate body (MGB; Auditory thalamus), targeting mainly the dorsal and medial regions. Using fiber photometry, we continuously imaged population calcium dynamics as mice learned a go/no-go auditory discrimination task. The MGB was tuned to frequency and responded to cognitive features like the choice of the mouse within several hundred milliseconds. Encoding of choice in the MGB increased with learning, and was highly correlated with the learning curves of the mice. MGB also encoded motor parameters of the mouse during the task. These results provide evidence that the MGB encodes task- motor- and learning-related information.


30
The thalamus was long considered a passive relay station of sensory information to cortex. 31 However, recent evidence suggests that thalamic nuclei may also be involved in high-order 32 processing of cognitive functions such as attention, working memory and learning (Acsády, (Jones, 1985), the role of specific thalamic nuclei in cognitive processing 38 remains unclear. In this study we focused on the neural correlates of higher-order thalamic 39 nuclei of the medial geniculate body (MGB; auditory thalamus) during auditory learning. 40 The MGB is the thalamic relay center of the auditory pathway, predominantly 41 receiving direct input from the inferior colliculus (Calford and Aitkin, 1983;Peruzzi et al., 42 1997), but also from cortex (Winer et al., 2001) and other sources (Crabtree, 1998;Lee, 43 2015; Winer, 1992). Its projections target the cerebral cortex and numerous other brain  (Rouiller et al., 1989;Smith et al., 2012). This pathway is called the lemniscal 50 pathway, and is considered to be the main auditory processing pathway. In contrast, the 51 dorsal and medial parts of the MGB project to higher-order cortical areas (Huang and 52 Winer, 2000a; Lee, 2015) and receive cortical feedback, among others, from layer 5 of the 53 auditory cortex (Bartlett et al., 2000a;Lee, 2015;Llano and Sherman, 2008a). Thus, the 54 medial and dorsal parts of the MGB, considered to be part of the non-lemniscal pathway, 55 are well positioned to encode higher-order information of sensory, motor and associative 56 nature. 57 Learning, the process of acquiring new knowledge through experience, is 58 traditionally thought to involve the neocortex. Learning to discriminate between different 59 stimuli leads to changes in the respective primary sensory areas (Blake et al., 2002;Chen 60 Results 83 Calcium imaging from the MGB along learning 84 To study learning-related changes in the MGB we first injected AAV-GCaMP6f into the 85 MGB of C57BL/6 mice and implanted a 400 µm optical fiber directly above the injection 86 site. After a week of handling and habituation to head-fixation, we trained mice on a go/no-87 go auditory discrimination task. Each trial started with a visual start cue (orange LED; 88 duration 0.1 s; 2 seconds before stimulus onset) followed by an auditory stimulus, either a  The learning curves with respect to the go and no-go trials shows that mice varied in 105 performance and strategy (Fig S1). Learning the task took different forms: some mice 106 increased their CR rate, others had a steep increase in hit rate, whereas others gradually 107 increased both hit and CR rates. Thus, mice learned the task using a range of behavioral 108 strategies.
significantly higher as compared to the no-go frequency ( Fig. 1G To evaluate how sounds and other task attributes were represented in the MGB across 139 learning, we plotted responses to the go and no-go sounds as mice learned the task. Figure   140 2A shows responses to the go and no-go sounds of one representative mouse ( Fig. 2A).

141
The most evident change in MGB responses across learning was during the late part of the 142 trial (0.6-1 seconds after stimulus onset), and particularly so for go trials ( Fig. 2A).

143
A particularly informative comparison was between hit and miss trials, where the 144 stimulus is identical but the choice of the mouse is different (either lick or no-lick). Thus, 145 differences between hits and misses represent encoding of choice. In expert mice (defined 146 as the last 500 trials), MGB responses were higher for hit as compared to miss trials, and 147 more so than in the novice mice (defined as the first 500 trials; Fig. 2B; compare blue to 148 light blue traces; For comparison, CR and FA trials are plotted in gray). Higher responses 149 in hit versus miss trials were evident in all (6/6) expert mice and in 50% (3/6) of novice 150 mice ( Fig. 2C; p<0.05; Wilcoxon rank sum test for each mouse separately). Choice 151 responses (i.e. hit minus miss) increased gradually after stimulus onset, and were stronger 152 in expert mice (Fig. 2D). These data provide evidence that the MGB encodes more than 153 only sounds. MGB encodes the choice of the mouse, which implies that the auditory 154 thalamus is involved in higher-level sensory-motor processing or cognitive attributes of the 155 task.

195
To quantify this effect across mice, we defined the latency for discrimination as the 196 time it takes to reach significance for each AUC measure in the expert case (arrows in Fig.   197 3B). The latency of discrimination for 'stim-AUC' was significantly lower than that of the were quite weak to the no-go sound, resulting in a low discrimination value. In summary, 213 these data indicate that mostly choice encoding changes with learning and that this 214 information develops late -several hundred milliseconds after stimulus onset.  The strong effects of movement suggest that the abovementioned MGB 'choice' 304 signal may simply represent body movements rather than choice, per se. Indeed, some 305 MGB responses and body movement covaried strongly whereas others did not (Fig. 5E).  Specifically, we found higher responses in hit compared to miss trials (Fig. 5F). Notably, 312 the effect size was smaller and levels of statistical significance were weaker ( Fig. 5G; 313 compare to Fig. 2C 'Expert'). Importantly, when mice were still novice, movement-free 314 MGB responses did not encode choice, indicating that choice encoding develops with 315 learning (Fig. S6). To rule out movement-related artifacts, we also imaged mice during the 316 task while exciting the MGB with a wavelength that does not excite the calcium indicator 317 (565 nm). We found no movement related artifacts (data not shown). In summary, body 318 movements during the task are evident in the calcium signal and are learning-related. 319 Nevertheless, MGB responses still maintains choice information that is separate from the 320 motor parameters and, critically, develops with learning. These data strengthen the claim 321 the MGB encodes high-level information.

323
Learning related changes in MGB differ in different frequency bands 324 As noted above, MGB responses were tuned to frequency (Fig. 1F-I). We next tested 325 whether the changes we observed in MGB are general or specific to the response properties 326 at the frequencies we used. To this end, we trained three additional mice on the task. In along the trial and was evident only in expert mice (Fig. 6C). Here again, MGB response 337 curves were sigmoid-like (Fig. 6D) and strongly correlated with learning (Fig. 6E). A more 338 detailed description of these effects is shown in Figure S7. Taken together, we infer that

429
We find that motor parameters, such as body movements, are also learning-related.

430
As mice learn to associate between the go sound and a reward, they increase their body   imaging, mice were anesthetized with 2% isoflurane (in pure O2) and body temperature 465 was maintained at 37°C. We applied local anesthesia to the area of surgery (lidocaine 1%),  Auditory discrimination task. Mice were trained on a go/no-go auditory discrimination 477 task (Fig. 1A). Each trial started with a visual start cue (orange LED placed in front of the 478 mouse; duration 0.1 s; 2 seconds before stimulus onset) followed by an auditory stimulus, were not punished or reinforced. Licking in response to the no-go sound were false alarm 486 trials (FA) that were followed by a mild punishment of white noise (duration of 3 seconds).

487
Withholding licking for the go sound were counted as misses, and were not punished. The 488 licking detector remained in a fixed and reachable position throughout the entire trial and 489 mice were free to lick at any time. Licking before the response cue was allowed and did 490 not lead to punishment or early reward. Note that the visual cue merely signals the start of 491 the trial, but had no predictive power with respect to go or no-go condition.

492
Training and performance. Nine mice were trained on the task. Mice were first 493 handled and accustomed to head fixation before starting the schedule of water restriction.

494
Before imaging began mice were conditioned to lick for reward after the go sound 495 (presented within a similar trial structure as the task itself). Imaging began only after mice 496 reliably licked for the presented sound (typically after the 1 st day; 200-400 trials). On the 497 first day of imaging, mice were presented with the go sound for 50 consecutive trials, after 498 which the no-go sound was gradually introduced (starting from 10% and increasing by 499 10% approximately every 50 trials). By the end of the 1 st day, the no-go sound reached 500 50% probability (Guo et al., 2014). During the 2 nd day, most mice continuously licked for 501 both sounds. Thus, after roughly 100 trials, we increased no-go probability to 80% and 502 waited for mice to perform three consecutive CR trials before returning to 50% probability.

503
This was done for several times until mice increased their performance, specifically to learn 504 to withhold licking for the no-go sound. In mice that still continued to lick for both sounds 505 we also repeated the no-go sound several times until the mouse performed correctly. In all 506 mice, a 50-50% protocol was reached typically on the 1 st or 2 nd day. Most mice learned the 507 task within 3-7 days corresponding to roughly 1200-2000 trials (Fig. 1D). An effort was 508 made to maintain a constant position of the mouse, speaker and cameras across imaging 509 days in order to maintain similar stimulation and imaging conditions.

526
Fiber tracks and GCaMP6f expression were detected in all mice and were mainly localized 527 to the MGB ( Fig. 1C and Supplementary Fig. 2).

529
Body tracking. In addition to fiber photometry, we tracked body movements of the mouse 530 during the task (Fig. 1B and 5A). The mouse was illuminated with a 940-nm infra-red LED.

531
A body camera monitored the movements of the mouse at 30 Hz (using a DMK 23UV024 532 camera). We used movements of both forelimbs and the head/neck region to assess body 533 movements ( Fig. 1A; see Data Analysis below). Mice performed the task in the dark.  Next, we divide trials based either on stimuli (i.e. go or no-go) or on choice (i.e. 547 lick or no-lick). MGB ΔF/F signals were plotted in 2 dimensional temporal spaces where 548 the x-axis is the trial temporal structure and the y-axis is the learning profile across trials 549 and days ( Fig. 2A). From this 2D temporal space we averaged across trials during the 550 novice and expert phases (defined as the first and last 500 trials respectively; Fig. 2B).

551
Alternatively, we averaged across time frames within the trial structure, to obtain an MGB 552 response curve across learning ( Fig. 4A; see below).  Calculating body movements. We used a body camera to detect general movements of the 591 mouse (30 Hz frame rate; Fig. 1A and 5A). For each imaging day, we first outlined the 592 forelimbs and the neck areas (one area of interest for each), which were reliable areas to 593 detect general movements. Next, we calculated the body movement (1 minus frame-to-594 frame correlation) within these areas as a function of time for each trial. Thresholding at 3 595 s.d. (across time frames before stimulus cue) above baseline resulted in a binary movement 596 vector (either 'moving' or 'quiet') for each trial (Gilad et al., 2018;Gilad and Helmchen, 597 2019). This was done for each trial to achieve a 2D space of movement probability within 598 the trial temporal structure (x-axis) versus the learning process (i.e. trial number; y-axis; 599 Fig. 5B). To obtain MGB signals that are 'movement-free', i.e. do not contain direct effect 600 of body movements, we detected the first movement onset for each trial, defined as 0.2 601 seconds before crossing the movement threshold (Fig. 5E). Next, MGB signals were 602 truncated from movement onset and onwards in a single trial manner. This analysis resulted 603 in MGB trials, each with a different length, that did not contain body movements (Fig.   604 5F,G). Similar results were obtained using other body parts such as the nose and mouth 605 (including licking).

606
Statistical analysis. In general, non-parametric two-tailed statistical tests were used. The

607
Whitney rank sum test was used to compare between two medians from two populations 608 and the Wilcoxon signed rank test was used to compare a population's median to zero (or 609 between two paired populations). Multiple group correction was used when comparing 610 between more than two groups. Significance was set at p=0.05.