After eye opening, DS develops faster in PMLS than V1

DS develops in ferret V1 in the week after eye opening and matures around P37 (Li et al., 2006; Clemens et al., 2012). While we have previously shown that PMLS DS is mature at the same age (Lempel and Nielsen, 2021), the time course of maturation has not yet been determined. Here, we systematically tracked PMLS DS development and compared it directly to V1. To this end, we recorded responses to drifting gratings in both areas from P29 (around eye opening) to P52 (young adult). Recordings were performed in anesthetized animals using multichannel silicon probes, in many cases collecting data from V1 and PMLS simultaneously by recording from two probes. All data were collected with the eyes open, irrespective of the natural eye opening state of an animal. Figure 1, B and C, show example tuning functions for well isolated SU in both areas from P30 to 37. As these plots show, we were able to find neurons with reliable visual responses at all ages, which allowed us to quantify tuning properties. The remainder of the analyses in this section will be limited to these visually responsive single neurons (determined using an ANOVA across mean baseline-corrected firing rates for all conditions including the blank with a criterion of p < 0.05; see Materials and Methods). The resulting data set sizes are listed in Table 1.

Especially at the youngest ages, tuning functions were often poorly fit with standard direction (or orientation) tuning functions. We therefore quantified DS by computing two different metrics. The first was L_dir or 1-DirCircVar, a vector-based metric (Mazurek et al., 2014) that takes the entire tuning function into account (see Materials and Methods for definition). Importantly, L_dir = 1 corresponds to high selectivity (responses to a single direction only), while L_dir = 0 indicates no selectivity (equal responses to all conditions). This metric is comparable to circular variance (CircVar), commonly used to quantify orientation tuning strength (Batschelet, 1981). The second metric used to quantify DS was the DS index (DSI), which compares only the responses between the preferred direction and the one opposing it (the null direction; see Materials and Methods). Results were generally identical between the two metrics, so only the results based on L_dir (which is more robust because it is based on the entire tuning curve) are discussed further (see Fig. S2 and Table S3 for the key DSI results). Both L_dir and DSI values are indicated for each example neuron in Figure 1, B and C, and suggest a general increase in DS with age in both areas.

To capture the time course of this DS maturation and to compare it between areas, we next analyzed DS as a function of age. As a first step, we computed the mean L_dir for each individual animal by averaging across all visually responsive PMLS or V1 neurons collected in that animal (results based on the median are comparable; data not shown). This makes it possible to visualize developmental trends without having to group animals by age; it also more clearly shows variability in developmental stage across animals of the same age. Figure 1D plots these per-animal means as a function of age. The two horizontal bands correspond to the mean ± 1 standard error of the mean (SEM) for animals P40 and older. Consistent with previous data (Li et al., 2006; Clemens et al., 2012) from ferret V1, our data show V1 DS to be immature around P30. PMLS DS levels were similarly low at this age. L_dir values eventually improved to mature levels in both areas, with some of the improvements occurring very rapidly, as evident by significant changes in DS between animals with a single day age difference. Unexpectedly, this sudden developmental improvement occurred earlier in PMLS than V1—around P34 for PMLS and P36 for V1. The faster PMLS development continued, in that mature values were reached faster after the initial improvements than in V1. Judging by the overlap between the per-animal mean and the young adult range, PMLS matured around P35, but it took V1 until at least P40 to reach this point.

This difference between PMLS and V1 development was also apparent when dividing SU L_dir values for each area into three age groups used previously for ferret V1 development (collapsing data across animals). The youngest of the three groups corresponded to eye opening (P29–32), the second to the phase of rapid DS development (P33–36), and the third to animals with mature DS levels (P37 and older). Figure 1, E and F, plots the L_dir distributions for these groups for PMLS and V1. Age had a general effect on the L_dir distribution in both areas (Kruskal–Wallis test across all three groups: p < 0.001 for both areas; full statistics in Table 2, including the p values for all nonsignificant tests). For PMLS, post hoc rank-sum tests revealed that all three age groups deviated from each other (p < 0.001). In V1, on the other hand, the P29–32 and P33–35 groups were not significantly different, but both differed from the oldest group (p < 0.001).

It should be noted that the oldest age group here only includes “young adults” up to P52 because we focused on the rapid initial phase of DS development. It is possible that DS levels in either area exhibit further changes over much longer timescales. The mean L_dir in PMLS in the oldest group here was 0.59 ± 0.08 [95% confidence interval (CI)]. We have previously observed a mean L_dir of 0.56 ± 0.05 for animals aged P57–497 (Lempel and Nielsen, 2019), which suggests that at least PMLS L_dir values remain stable after the early developmental phase studied here. In V1, the mean L_dir in the oldest group equaled 0.46 ± 0.08, with a median value of 0.49. In our previous study in animals aged P57–497, the mean L_dir equaled 0.29 ± 0.06 in V1 (Lempel and Nielsen, 2019). This would suggest that V1 L_dir levels are transiently higher around P50 than in fully mature animals. Indeed, V1 DS levels have been shown to trend down from P60 to 90 (although that trend did not reach significance), with the mean DSI falling from ∼0.7 to 0.5 from P60 to 90 (Popovic et al., 2018). This previous P60 DSI value is in close agreement with our data here (mean DSI for the oldest group here, 0.71 ± 0.08).

Since development is obviously dynamic and averaging across age groups cannot capture all trends, Figure 1G uses an alternate analysis approach to show the relative DS development in V1 and PMLS in more detail. Here, L_dir values between P29 and 38 were binned into much smaller age windows of 2 d, combining data across animals. We also added a final bin from P40 to 52 (corresponding to the mature levels indicated in Fig. 1D). Since in adult animals DS is stronger in PMLS than V1, the comparison of the developmental trajectories between the areas is complicated by the fact that they reach different endpoints. To account for these endpoint differences, we chose to normalize L_dir values so that the average L_dir in the 2 d window around eye opening (P29–30) was set to 0, and the average L_dir values in the P40–52 group was set to 1 (see Materials and Methods). V1 and PMLS data were normalized independently, and the normalized L_dir values were averaged in each age bin. These normalized time courses again show that PMLS DS development occurred earlier than in V1. PMLS L_dir values rapidly increased starting with the P33–34 window, while V1 L_dir values remained at the eye opening level until P35–36. A two-way ANOVA with factors area and age confirmed these observations. For the ANOVA, the age windows used for normalization were excluded (since they are forced to be 0 and 1, respectively), reducing the data to P31–38. For these data, the two-factor ANOVA then resulted in a significant effect for age (p < 0.001), consistent with the general age-dependent change in DS in both areas. Importantly, a significant effect of area (p < 0.001), as well as an interaction between age and area trending toward significance (p = 0.01), supports differences in DS levels between areas during this time period, which because of the normalization cannot be explained by overall higher DS levels in PMLS but reflect developmental differences. Further post hoc rank-sum tests indeed revealed significantly higher normalized L_dir values in PMLS than in V1 at P33–34 (p < 0.001) and P34–35 (p = 0.008), confirming the earlier development of DS in PMLS (see Table 3 for full test statistics). It is worth noting that the time courses in Figure 1G also suggest that V1 DS development does not just start later; it also takes longer to complete (note the shallower slope of the V1 time course).

As a final comparison of the developmental trajectories between areas, we took advantage of the fact that L_dir values increased rapidly in both areas once maturation set in, but were relatively stable before and after, to determine the onset of maturation in PMLS versus V1. To this end, we binned the (non-normalized) SU data from P29–38 into the same 2 d windows used for the analysis above and computed the mean L_dir per age bin per area. The oldest group reflecting the more steady-state adult data was excluded here since the focus was on the rapid initial transition in DS levels. Data were binned to increase the sample size per bin, but an analysis without binning yielded similar results. We then used MATLAB's findchangepts function to divide the average L_dir time course per area into two segments with locally stable means. Specifically, the algorithm determines the changepoint as the sample (here, age bin) that divides the dataset into two segments such that the residual error between each segment's data and the segment mean is minimized. A 95% CI was determined for every changepoint using bootstrapping (see Materials and Methods). For PMLS L_dir, this approach indicated a change at P33 (CI: P33–35), while V1 L_dir did not change until P35 (CI: P35–37), further confirming the lag in DS development in V1 relative to PMLS.

At the most fundamental level, high DS requires strong responses to the preferred direction, and weak responses to the opposite (null) direction. To probe some of the mechanisms underlying DS development, we compared the baseline-corrected firing rates evoked by preferred and null direction at different ages. To capture the overall trends, data were divided into the same three age groups used for Figure 1, E and F. In both areas, the responses to the preferred stimulus became more robust with age (Fig. 1H,I; see Table 2 for statistics). Kruskal–Wallis tests comparing the three age groups were significant for both areas (both p < 0.001), with significant differences between the two youngest groups and the oldest group in both areas (rank-sum tests, all p < 0.001). The increase in the V1 preferred direction response is consistent with previous reports (Clemens et al., 2012); our data now show that a similar increase occurs in PMLS.

Responses to the null stimulus changed with development as well, but with differences across areas (Fig. 1J,K). In V1, null direction responses temporarily increased in the P33–36 group, a change that reached significance in comparison with the oldest group, and trended toward significance relative to the P29–32 group (Kruskal–Wallis test, p = 0.008; rank-sum tests P33–36 vs P37+, p = 0.006; P33–36 vs P29–32, p = 0.01). The difference between the two older groups parallel a previous observation of a drop in V1 null responses after P35 (Clemens et al., 2012). In PMLS, on the other hand, null direction responses decreased with age, so that the oldest group had the lowest null responses (Kruskal–Wallis and post hoc test between both younger and the oldest age group; p < 0.001; Table 2). This weaker response to the null direction, especially because it was paired with an increase in the preferred direction response, would support better discriminability of the two directions and points to a possible role of null direction inhibition especially in the development of PMLS DS (see also Discussion).

Preferred and null responses can also be used to begin to address another question regarding the data. Even in adult animals, the transformation of information from V1 to PMLS results in enhanced DS in PMLS. It is possible that the same mechanisms—like integration across V1 inputs and threshold nonlinearities (Rust et al., 2006; Lempel and Nielsen, 2021)—amplify subtle direction biases present in V1 early in life and that these biases increase in V1 during the phase of rapid PMLS DS development ahead of the major changes in DS in V1. Ultimately, measurements of the V1 input to PMLS neurons are required to fully answer this question. The data shown in Figure 1, however, argue against this explanation: at a per-animal level, animals with mean L_dir levels exceeding the levels found around eye opening can be found first at P33 in PMLS but only at P36 in V1 (Fig. 1D). Figure 1G similarly shows that the first bin in which the normalized L_dir values deviate from 0 is P33–34 for PMLS and P35–36 for V1. As long as the entire tuning function is considered, our data therefore show stable tuning levels in V1, while PMLS levels are rapidly changing, inconsistent with the notion of small V1 shifts supporting the PMLS development. However, subtle shifts may be hard to detect at the level of the entire tuning function. We therefore also computed the ratio of null to preferred responses for every neuron, as a way to capture the relative strength of the null direction response independent of changes in overall firing rates (the DSI represents an alternate way to quantify the relative null direction response and yields the same results). Comparing ratios between V1 and PMLS then provides a sense of the amplification of V1 direction differences present in the PMLS output. If there were subtle changes in direction biases in V1 after eye opening, these would be reflected in decreasing ratios during this time period. Figure S2G–I plots the response ratios for V1 and PMLS divided into the same three age groups used above, as well as into 2 d windows equivalent to Figure 1G. Note that in this case, we did not normalize the data since the difference between areas is of interest. Both types of analyses show that V1 null responses do not change relative to preferred responses up to P36. At the same time, PMLS null responses decrease relative to the preferred responses starting around P33. This again strongly argues against small changes in V1 responses that are then amplified in PMLS.

In summary, our data show earlier and faster DS development in PMLS than V1, in a process that involves area-specific decreases in null direction responses as well as increases in preferred responses in both areas.

Early PMLS development appears specific to DS

Since both V1 and PMLS are tuned for orientation in addition to direction, the development of OS provides the opportunity to assess the areas’ relative development along a second tuning dimension, a crucial test for determining whether PMLS develops early in general or whether the observations made so far are specific for DS. OS was quantified using two metrics similar to those for DS, a vector-based metric (Batschelet, 1981; Mazurek et al., 2014) called L_ori or 1-CircVar and an OS index (OSI). Again, analyses using either metric showed the same results, and only L_ori data will be shown here (for OSI results, see Fig. S2 and Table S3).

Intriguingly, the relative development of OS in PMLS versus V1 followed a different pattern than that of DS. In V1, consistent with previous reports (Chapman and Stryker, 1993; Chapman et al., 1996), L_ori values were low up to P30. Starting with P31, much higher L_ori values could be observed, even if there was significant variability across animals (Fig. 2A). Around P35, V1 L_ori values reached the same level as in the young adult group (P40–52, mean L_ori 0.53 ± 0.07, median 0.51), which is similar to previous studies in young adults and even in older animals reporting median V1 L_ori values of 0.56 (Lempel and Nielsen, 2019), ∼0.5 (Clemens et al., 2012), and ∼0.7 (Popovic et al., 2018). As a consequence of the rapid rise in L_ori right around eye opening, dividing the data into the same three age groups used for L_dir before (P29–32, P33–36, P37+; Fig. 2C) revealed significant differences between the P29–32 group and both other groups (Kruskal–Wallis test and both post hoc rank-sum tests, p < 0.01; Table 4), but not between the two older groups.

In PMLS, OS levels were as low as those found in V1 around P30 (Fig. 2A), with the exception of two outliers (one animal contributing two neurons, the other four). Since in all other respects (e.g., their L_dir values) these animals were not outliers, their data were included in the OS dataset as well. Other than the two outliers, OS development in PMLS was delayed relative to V1, with consistent improvements in PMLS L_ori values starting only at P34, later than the increase in V1 L_ori values. At P34, however, PMLS L_ori values immediately assumed the same levels as in the young adult group in a very rapid development. After this point, OS levels appear stable, as the values for the young adult group here (mean L_ori 0.5 ± 0.06, median 0.49) are within the range we have previously reported for older animals (mean 0.45 ± 0.04; median 0.45; Lempel and Nielsen, 2019). OS maturation was therefore complete at the same time point in both V1 and PMLS. Indeed, dividing the PMLS dataset into the same three age groups as above (Fig. 2B) revealed differences only between the P29 and 32 age groups and both other age groups, comparable to the results for V1 (Kruskal–Wallis test and both post hoc rank-sum tests, p < 0.001; see also Table 4).

Given the rapid OS development especially in PMLS, we again turned to binning L_ori data into 2 d age windows, normalizing the data to assume 0 and 1 for eye opening and mature levels, respectively, as before (Fig. 2D). Note that the two animals with high L_ori values around P30 caused the normalized L_ori time course for PMLS to initially decrease at P31–32 before rapidly increasing in the next age group. A two-factor ANOVA with age and area (excluding the endpoints used for normalization) was again used to compare development across the areas. It resulted in a significant main effect of age (p < 0.001; Table 5), consistent with the increase in OS with age. At same time, there was a significant interaction between area and age (p = 0.001), due to significantly higher normalized L_ori values in V1 in the P31–32 group (post hoc rank-sum test, p < 0.001). To rule out possible effects of including the two PMLS outliers in the normalization, we confirmed that a direct comparison of the raw (non-normalized) L_ori values for animals aged P31–32 similarly resulted in a significant difference between V1 and PMLS (rank-sum test, p = 0.003).

Finally, we also computed changepoints for the non-normalized (but binned) L_ori trajectories for each area. In V1, this analysis indicated a change around P31 (95% CI: P31–35; the larger interval reflects the fact that determining the changepoint for V1 L_ori is noisier because V1 L_ori levels are rising much more gradually with development). The changepoint for PMLS L_ori instead occurred at P33 (CI: P33–35), i.e., with a delay relative to V1. Thus, the onset of OS maturation in PMLS lags behind V1, which in turn highlights that the earlier development of PMLS DS is not due to earlier development of the area in general but that it reflects processes specific to processing of motion direction.

The relative time courses of DS and OS maturation reveal another very intriguing difference between V1 and PMLS development. Comparing these time courses within each area rather than across areas (Fig. 2E), first replicated previous findings of OS development before DS development in V1 (Li et al., 2006). In PMLS, however, our data showed no (or only minimal) delay between DS and OS development (e.g., both normalized time courses show immature levels at P31–32 and significant improvement by P33–34). The normalized L_dir and L_ori values can again be used to compare the relative development of OS and DS without contamination by overall differences in tuning levels. For V1, a two-way ANOVA with factors age and tuning property on these normalized data resulted in a significant main effect for tuning property (p < 0.001; Table 6) and a trend for age (p = 0.01), but no significant interaction between the factors (p = 0.4), consistent with differences between the two tuning properties that are maintained throughout development (again, the endpoints used for normalization were excluded from the ANOVA). In contrast, in PMLS only the effect of age reached significance (p < 0.001), reflecting that DS and OS developed in parallel. The changepoint analysis supports this conclusion as well. As stated above, changepoints for V1 L_ori and L_dir were P31 and 35, while they fell at P33 for both tuning properties in PMLS. These differences in the sequence of DS and OS development between V1 and PMLS suggest that there might be fundamental differences in the organization of circuits computing direction information in V1 and PMLS (see Discussion).

Developmental changes in response latency

In the marmoset, MT has been shown to develop as early as V1, at least anatomically (Bourne and Rosa, 2006). This early development involves a transient, direct pathway from the pulvinar to MT (Warner et al., 2012; Mundinano et al., 2018). This raises the interesting possibility that early DS development in PMLS is supported by input sources different from those found in adults. Resolving this question will ultimately require anatomical and causal studies. However, response latencies should at least provide some initial insights into network configurations at different ages. As a next step, we therefore decided to analyze the evolution of response latencies in V1 and PMLS with age. Latencies can be quite variable from neuron to neuron. To provide the most robust estimates even at the youngest ages, we increased our sample size by analyzing MU instead of SU data. Here, MU activity is computed as the thresholded but not spike-sorted activity for each recording channel, ensuring that each spike was only counted on a single channel (see Materials and Methods). To be included in the latency dataset, MU sites had to meet the same visual responsiveness criteria used for SUs above (see Table 1 for the resulting dataset size). Additionally, we limited the dataset to recording sessions in which data were collected simultaneously for V1 and PMLS so that within-animal comparisons were feasible in all cases. This further minimizes confounds in the inter-areal comparisons that could arise from changing response dynamics due to fluctuations in anesthetic state.

For each included MU site, we then determined the onset latency for the site's best stimulus (as the condition evoking the most robust responses). Latencies were computed by binning spikes into 10 ms bins and computing the first bin in which responses exceeded the mean of prestimulus baseline activity by two standard deviations, followed by at least two successive bins exceeding baseline activity (Raiguel et al., 1999). Figure 3A shows raster plots for three example V1 MU sites at different ages, as well as three example PMLS sites recorded in the same animals. These examples capture two important trends present in the overall dataset shown in Figure 3B: First, latencies in both areas became significantly shorter with age, until very short latencies were reached in adults (median V1 latency across all MU sites, 60 ms; median PMLS latency, 70 ms). Second, there was a large delay in PMLS responses relative to V1 in young animals. As latencies in both areas gradually improved over the next days, the gap in latencies between the areas ultimately narrowed and became quite small in adults.

Figure 3, C and D, plots how these findings play out at the per-neuron rather than per-animal level (see Tables 7, 8 for full statistics). First, Figure 3C plots the distribution of latencies for the standard three age groups, demonstrating the gradual improvement in latencies with age in both areas (Kruskal–Wallis test across groups per area and all post hoc rank-sum tests: p < 0.001). Second, Figure 3D shows how average latencies, computed in two-day windows (this time without normalization), evolve with age, highlighting the gradual improvement in latencies between P30 and 35 in both areas, as well as the larger delay in PMLS responses in this earlier phase. Indeed, significant differences between V1 and PMLS emerged only for the three age windows falling between P29 and 34 (two-way ANOVA with factors age and area, all main effects and interaction p < 0.001; post hoc rank-sum tests between areas for P29–30, P31–32, P33–34, p < 0.001) but note that there is still a trend toward larger PMLS latencies for older animals (p ≤ 0.05 for P35–36 and animals aged P42–50). Finally, we made use of the fact that the dataset was limited to cases with simultaneous recordings in V1 and PMLS to test for each animal individually whether V1 and PMLS latencies differed systematically. For this analysis, the PMLS and V1 latency distributions recorded in the same animal were compared against each other using a rank-sum test, and the resulting p values were Bonferroni-corrected by the total number of animals to account for multiple comparisons. Figure 3E plots the corrected p values as a function of each animal's age, with the red line indicating p = 0.01. With the exception of one older animal, significant differences between V1 and PMLS latencies using this criterion occurred only before P35.

Taken together, our data show that in addition to age-dependent decreases in response latencies in both areas, extra delays occur in PMLS in young animals, possibly reflecting changing input sources and computational demands with age (see Discussion). Interestingly, for PMLS the maturation of latency coincides with the maturation of DS and OS, while for V1, OS develops before latency matures and DS thereafter.

Visual experience accelerates DS development in both areas

An important aspect of the development of any visual function is the degree to which it reflects endogenous mechanisms versus the impact of visual experience. In this regard, previous experiments in the ferret have established a highly useful “training” paradigm that allows controlled tests of the impact of different types of visual experience (Li et al., 2008). In this paradigm, visually naive animals are exposed to a stimulus of interest for a prolonged period of time (while continuously under anesthesia). Baseline tuning levels like DS are determined before and after stimulus exposure to assess whether the chosen training stimulus exerted a measurable influence over neural response properties. For ferret V1, previous studies (Li et al., 2008) reported DS increases after exposure to a drifting grating only—exposure to a gray screen was found to be ineffective, as was exposure to a stationary, flashing grating (a grating with temporal modulation of its contrast). In identifying motion signals as an important cue for V1 DS development, these results demonstrate that the paradigm is a sensitive probe for which types of visual information might influence an area's development. For this reason, we used it here to gain further insights into PMLS DS development.

Experiments using the training paradigm followed the procedures just described. More precisely, we first determined pretraining tuning levels in V1 and PMLS in visually naive animals. As before, we used drifting gratings to quantify DS and OS levels. We then exposed the animals to a specific stimulus for 8 h. To minimize adaptation, the paradigm was divided into 20-min-long training (or stimulation) blocks in which the stimulus was presented, alternating with 10 min of a blank screen (Fig. 4A). Furthermore, each training block itself consisted of alternating stimulus presentations (5 s) and blank screen periods (10 s). At the conclusion of this training phase, DS and OS levels were again measured in V1 and PMLS and compared with the pretraining levels. Because of the duration of the experiments, we did not attempt to track SU across pre- and post-training measurements. Instead, we assessed DS at the MU level in each case as the more stable measure, reasoning that as electrodes were not moved throughout the experiment, MU activity reflects activity in the same population of neurons before and after exposure (see Materials and Methods for inclusion criteria for the training datasets).

As a first test, we investigated the effect of motion cues on tuning development in both areas by using drifting gratings as training stimuli. In each presentation of the training stimulus, the grating could drift in one of two opposing directions, so that the axis of motion and orientation was constant throughout. Using this stimulus set, our data replicated the previous findings of increased V1 DS levels after exposure. Figure 4B plots the cumulative L_dir distribution across MU sites before and after training, demonstrating a significant increase in V1 DS levels after training (rank-sum test, p < 0.001). While this increase is consistent with previous studies, the effect size appears to be significantly smaller in our animals than in those earlier studies (Li et al., 2008; Van Hooser et al., 2012; Ritter et al., 2017; Roy et al., 2020). A possible explanation for this discrepancy is a difference in the amount of natural visual experience per animal across studies. None of the animals studied here had naturally opened their eyes at the start of the training experiments. In contrast, previous studies included animals of the same age range, but with 1–3 d of visual experience (Van Hooser et al., 2012; Van Hooser, personal communication). Indeed, the mean pretraining V1 L_dir level was higher in at least one of these previous studies than in our sample, as would be expected in animals with some visual experience [here, mean pretraining V1 L_dir across all training cohorts, 0.07 ± 0.001 (SEM); Ritter et al., 2017: 0.105 ± 0.017]. It is plausible that even brief amounts of natural visual experience can predispose V1 circuits to be able to change more rapidly in response to the training stimuli, therefore resulting in larger effects. Also note that previous studies using the same training paradigm have shown that changes in DS are strongest at the trained direction in V1, both because neurons initially preferring this direction strengthen in their responses but also because stimulus preferences become aligned to the training stimulus (Li et al., 2008; Van Hooser et al., 2012). Here, we chose the training directions to drive responses at both recording sites, but did not aim to match them precisely to stimulus preferences at either location. It is possible that some variability in the training effect size is caused by a different degree of alignment between the training direction and the stimulus preferences at our recording locations.

At the same time, since the effect size in our experiments is smaller, there might be a concern that the population data mostly reflects changes in a single animal. In addition to analyzing data across animals, we therefore also used a LME model to directly control for grouping of data by animal. More precisely, we fit the before and after training L_dir data with a LME model with a fixed effect for training phase (before/after training), allowing random intercepts and slopes per animal (see Materials and Methods and Fig. S3). On the full dataset, the effect of training phase indeed was not significant (p = 0.2). However, four out of the five animals showed an increase in L_dir values after training (Table 9). If the fifth animal (which showed a decrease) is excluded, the effect of training phase becomes significant (p = 0.009).

Importantly, our data show for the first time that exposure to drifting gratings also has a measurable impact on PMLS DS levels. Combined across all animals, L_dir levels significantly increased in PMLS with training (rank-sum test, p < 0.001). The LME model similarly returned a significant effect of training phase (p < 0.001) and increases in L_dir values for all five animals (Table 9). Thus, it appears that drifting gratings contain the necessary information to drive DS development not just in V1 but also in PMLS.

To support this conclusion, it is important to compare the results to the effects of exposing animals to a blank screen for 8 h, to determine whether the observed DS changes indeed were due to information contained in the stimulus versus a more general effect of unspecific visual stimulation (since the eyes are opened at the start of the experiment) or simply the passage of time. Figure 4C and Figure S3 plot the L_dir changes observed before and after exposure to a homogenously gray screen for 8 h for both V1 and PMLS. Based on analyzing the population data as a whole, V1 L_dir levels actually decreased during this time period (rank-sum test, p = 0.003). In PMLS, a trend toward higher L_dir levels did not reach significance (rank-sum test, p = 0.01). The LME models, which account for per-animal effects, showed no significant change in the L_dir values for either area (V1, p = 0.2; PMLS, p = 0.7), and positive and negative L_dir changes occurred with the same frequencies across animals for both areas (Table 9). Thus, for PMLS as for V1, the DS changes induced by the drifting grating indeed reflected the impact of the particular visual information presented during the training, not unspecific effects.

In addition to changes in DS, we also tested whether OS levels were affected by the training paradigm (Fig. 4E,F; Fig. S3). In V1, increases in OS were previously observed after exposure to a drifting grating but also after exposure to a gray screen (Li et al., 2008; Ritter et al., 2017), suggesting that these improvements occur independent of the particular visual information provided. Here, we observed no changes in V1 L_ori levels after exposure to a drifting grating across the full dataset (rank-sum test, p = 0.3; LME model, p = 0.8; Table 10). However, the same animal that was an outlier in terms of L_dir changes in this cohort also was the only animal that showed decreases in L_ori after training (see LME slopes in Table 10). Excluding this animal resulted in significant changes in V1 L_ori levels before and after training (population level rank-sum test, p < 0.001; LME model, p = 0.002; Table 10). In contrast, exposure to a gray screen actually resulted in a small but significant decrease in V1 L_ori at least at the per-site level in our experiments (rank-sum test across all MU sites, p = 0.006; LME model, p = 0.08; Table 10). Thus, in our dataset, the V1 OS changes are specific to the stimulus used, not unspecific as shown previously. Again, the differences between our and previous results may lie in the difference in starting point. In this context, it is worth noting that, just like baseline DS levels, baseline OS levels were lower in our sample than in previous studies (here, mean pretraining V1 L_ori across all cohorts, 0.18 ± 0.002; Ritter et al., 0.383 ± 0.036).

In PMLS, L_ori levels improved significantly after exposure to drifting gratings (rank-sum test, p < 0.001; LME model, p < 0.001; 5/5 animals with positive slope; Table 10). No change was observed in the control cohort exposed to a gray screen (rank-sum test, p = 0.8; LME model, p = 0.4; Table 10), indicating that in PMLS, changes in OS, like changes in DS, were largely driven by the stimulus information provided during training.

PMLS DS development is sensitive to more visual cues than V1

The training experiments so far demonstrate that patterned visual experience can impact not just V1 but also PMLS DS (and OS) development. The next question then is whether certain types of visual information are more important in this process than others. As mentioned above, exposure to a training stimulus lacking motion signals previously proved ineffective in changing V1 DS levels. To probe whether motion signals are similarly crucial for PMLS DS development, we performed a third set of training experiments, in which animals were exposed to flashing gratings. These gratings were generated by temporally modulating grating contrast between 0 and 100%. Importantly, no contrast reversals were included. This ensured that while the stimulus contrast varied dynamically, the spatial position of the stimulus remained fixed, and no motion cues were provided.

The results of these experiments are plotted in Figure 4D and Figure S3. In agreement with the earlier study by Li et al. (2008), V1 L_dir levels remained unchanged after 8 h of exposure to the flashing grating (rank-sum test, p = 0.9; LME model, p = 0.9; 3/6 animals with L_dir increases; Table 9). In contrast, the same stimulus was highly effective in PMLS, with significant improvements in L_dir at the end of the training experiment (rank-sum test, p < 0.001; LME model, p < 0.001, 7/7 animals with L_dir increases; Table 9). In fact, at least in the cohorts studied here, flashing gratings resulted in larger changes in PMLS L_dir levels than drifting gratings. We used the LME models to estimate training effect size per animal (see Materials and Methods). Based on this metric, training effects were significantly larger for flashing than drifting gratings in PMLS (rank-sum test, p = 0.003). Our data therefore strongly suggest that contrary to V1, explicit motion cues were not required for promoting DS development in PMLS. While the flashing gratings are stationary spatially, they nonetheless dynamically changed over time. Based on our data, PMLS is sensitive to these temporal cues, while V1 is not (or to a much lesser degree). It remains to be determined whether temporal cues alone are strong enough that a spatially uniform stimulus that changes luminance in time, but lacks spatial information, would nonetheless be sufficient in causing PMLS DS increases.

In addition to highlighting the role of different types of visual information, the differential impact of flashing gratings on V1 and PMLS represents a striking example of dissociation in DS development between the two areas reminiscent of the observations during normal development. The training data generally provide an opportunity to further test this idea, because they can be used to ask whether L_dir changes induced by the training protocol in V1 are correlated with L_dir changes in PMLS in the same animal. Figure 4H plots the L_dir changes in V1 against the changes in PMLS for each animal (with training effect sizes determined from the per-animal LME models as above). Note that for this plot, we only used animals in which data from both areas were available. The lack of correlation of effect sizes is evident in this figure (r = −0.09; p = 0.8) and supports the notion that DS development in PMLS does not track V1 DS development.

Lastly, the difference in response to the flashing gratings between V1 and PMLS also extended to OS (Fig. 4G; Fig. S3). In PMLS, L_ori levels increased after training (rank-sum test, p < 0.001; LME model, p = 0.005; 6/7 animals with L_ori increases; Table 10), while no change was detected in V1 (rank-sum test, p = 0.2; LME model, p = 0.3; Table 10). As for L_dir, there was no correlation between the changes in V1 and PMLS L_ori levels across animals (Fig. S4A; r = 0.01; p = 0.97).

As intriguing as the diverging effects of the flashing gratings on V1 and PMLS are, it needs to be ruled out that the flashing gratings simply fail to activate V1, therefore rendering the stimulus ineffective in driving L_dir change. To address this question, we turned to MU responses recorded during the stimulus presentation periods in the training blocks and compared responses to flashing gratings against those to drifting gratings as an effective training stimulus. Figure 4I plots the average response time courses evoked by both types of stimuli for the two areas (averaging across all blocks and all MU sites in all animals for the two training cohorts). These time courses clearly show that flashing gratings did not fail to drive V1 but rather that both types of stimuli evoked reliable activity in both areas. In fact, while peak rates differed between the two areas (Fig. S5), they did not differ between stimulus types (two-way ANOVA with factors area and stimulus type; area, p < 0.001; type, p = 0.1; interaction, p = 0.3; Table S1). The major difference between the stimulus types appeared to be the duration of the evoked response (quantified as the time with firing rates above baseline; see Materials and Methods) and as a consequence the average rates (Fig. S5, Table S1). More precisely, response durations were significantly shorter for flashing than drifting gratings in both areas (ANOVA with factors area and stimulus type, both main effects and interaction p < 0.001; individual rank-sum tests per area, p < 0.001; Table S1). Interestingly, the difference in response duration was more pronounced in V1 than PMLS. However, if lower V1 drive due to shorter response durations indeed was the reason for a lack of DS change after training with flashing gratings, then response duration should be correlated with the training-induced change in L_dir. As is obvious from Figure 4J, in V1 this is not the case for both stimuli combined or individually (both conditions combined, r = −0.07; p = 0.8; drifting, r = −0.7; p = 0.2; flashing, r = −0.3; p = 0.5). The same is the case for PMLS (both conditions, r = −0.6; p = 0.04; drifting, r = 0.001; p = 1; flashing, r = −0.5; p = 0.3). Thus, the different outcomes of exposure to flashing gratings for V1 and PMLS tuning levels cannot simply be attributed to a failure to drive V1 response but indeed appear to reflect sensitivity to different stimulus attributes in the two areas.

As a final check, we compared the baseline levels of selectivity pre-training across all training cohorts for both areas. Animals in the three cohorts were matched in terms of age range (drifting gratings, P29–33, mean 30.8; gray screen, P28–31, mean 29.8; flashing gratings, P29–32, mean 30.6) and their lack of visual experience. Nonetheless, it is possible that a different starting point (in terms of tuning development) might predispose some animals for larger training effects than others. We indeed observed systematic differences in the baseline DS and OS levels across the three cohorts. At the same time, these differences did not appear to be related to the differences in training effects. At the per-site level, V1 L_dir levels were significantly higher in the cohort exposed to a gray screen than both other groups (rank-sum test, p < 0.001 for both tests; Table S2), and there was a trend toward higher L_dir levels in the cohort trained with drifting gratings than that trained with flashing gratings (rank-sum test, p = 0.03; Table S2). The cohort exposed to drifting gratings therefore showed baseline tuning levels that fell between both other groups. Nonetheless, this was the only group that showed improvements in DS with training, arguing against a simple relationship between baseline tuning levels and training effects. When analyzed at the per-animal level using the pretraining mean and training effect slope determined from the LME (Fig. S4B), a negative correlation between baseline tuning levels and training effect size arose across all three groups (r = −0.7; p = 0.003). This correlation is driven by the results in the cohort exposed to the gray screen: excluding the gray screen cohort from the analysis results in a nonsignificant correlation (r = 0.2; p = 0.5), as is the correlation for each group on its own (drifting, r = 0.5; p = 0.4; flashing, r = 0.1; p = 0.8; gray, r = −0.9; p = 0.02). The per-animal data therefore also argue that the differences in effectiveness of different training stimuli do not merely reflect different starting conditions across the cohorts.

In PMLS, we similarly observed a systematic difference in baseline L_dir levels across the three cohorts (Kruskal–Wallis test, p < 0.001; Table S2). Here, tuning levels were lowest in the cohort exposed to flashing gratings when analyzed at the population level, followed by the drifting grating cohort and the gray screen cohort (rank-sum tests, all p < 0.001; Table S2). When analyzed at the per-animal level, we consequently observed a trend for a negative correlation between pretraining means and the training effect slope across all three cohorts (r = −0.4; p = 0.08; Fig. S4D). However, as for V1, this correlation was driven by the gray screen cohort, and correlations were not significant once that cohort was excluded from the analysis (r = −0.2; p = 0.6) and for each cohort individually (drifting, r = 0.8; p = 0.1; flashing, r = 0.1; p = 0.8; gray, r = −0.2; p = 0.7). Thus, again, it seems highly unlikely that the effectiveness of different stimuli on PMLS tuning purely reflects different baseline tuning levels. L_ori results, finally, generally paralleled the results for L_dir for both areas and similarly point to stimulus content rather than baseline tuning levels, as the major source for different training outcomes (Fig. S4C,E; Table S2).

The following reasons further support the notion that the consistent negative correlations we observed between baseline levels and training effects are spurious rather than a real effect: First, a negative correlation between baseline level and training effect size might be expected if baseline levels were close enough to mature levels to limit possible training effect sizes (i.e., causing a ceiling effect). In all cases here, baseline levels were immature enough that even after training, tuning levels were below the mature values, making it unlikely that a ceiling effect caused the negative correlation. Second, it is possible that higher baseline levels indicate a more mature developmental state that is less susceptible to the training paradigm. As discussed above, a comparison of our results with those of previous studies shows that to the contrary, larger effects are observed in animals with more visual experience and higher baseline tuning levels than found in our data, arguing against this explanation (see also Discussion).