Cortical direction selectivity increases from the input to the output layers of visual cortex

Based on the neuronal responses to drifting gratings with varied directions, we obtained direction tuning of each recording site, from which we calculated the direction selectivity index (DSI), which represents the strength of direction selectivity, and determined the preferred direction, which is the direction that evokes maximum responses (Fig 1B).

We then characterized the laminar distribution of the DSI (Fig 2A). We observed that the output layers exhibited a significantly larger DSI (Fig 2B; output layers: 0.4605±0.0161, N = 134 recording sites; input layers: 0.2566±0.0194, N = 109 recording sites; interlaminar difference: p = 1.7392*10⁻⁸; Wilcoxon rank-sum test) and had a larger proportion of recording sites with strong direction selectivity (Fig 2B; proportion of recording sites with DSI > 1/3, output layers: 73.13%, N = 134 recording sites; input layers: 43.12%, N = 109 recording sites; interlaminar difference: p = 2.0992*10⁻⁶; chi-square goodness-of-fit test), indicating an enhancement of direction selectivity in the output layers. This enhancement was found in almost all probe placements (Fig 2C, N = 25 probe placements, p = 2.5432*10⁻⁵; Wilcoxon signed-rank test). Across all probe placements, we found that the DSIs in the output layers were approximately 78.59% greater than those in the input layers (Fig 2D, DSI enhancement: 0.7859±0.1813, N = 25 probe placements). Additionally, our analysis revealed a strong positive correlation between the DSIs in the output layers and those in the input layers (Fig 2C, N = 25 probe placements, r = 0.7131, p = 9.7213*10⁻⁶; Spearman’s ρ), suggesting that the direction selectivity observed in the output layers may be partially inherited from the input layers. Similar results were obtained when considering only L2/3 and L4 (S3A–S3C Fig).

Fig 2. Direction selectivity is enhanced in the output layers.

(A) Laminar pattern of the DSI. Each dot represents the data from one recording site. Recording sites belonging to the output layers, including L2/3 and L5, are shown in blue. Recording sites in the input layers, including L4 and L6, are shown in red. The vertical dashed line represents a DSI of 1/3, indicating that the response strength in the preferred direction is twice that in the null direction. The running medians of neighboring recording sites are shown as black lines, and shaded error bars indicate the SEM. (B) Cumulative distributions of DSIs in the output layers (blue line) and the input layers (red line). The horizontal dashed line indicates a cumulative proportion of 0.5, and the vertical dashed line represents a DSI of 1/3. The median value and SEM of the DSI for both layers are shown at the top. The proportions of recording sites exhibiting strong direction selectivity (DSI > 1/3) within the input and output layers are denoted by colored arrows on the right. Statistical comparisons for median DSI values (at the top) and proportions of recording sites with strong direction selectivity (on the right) between the output layers and the input layers are also shown. (C) Comparison of DSIs between the output layers and the input layers within the same probe placement. For each probe placement, we obtained the median value and SEM of the DSI across all recording sites in the output layers and in the input layers. Thus, each dot represents data from one probe placement. The Spearman correlation and the significance of paired comparisons between the output layers and the input layers are shown. Data are median ± SEMs. (D) The distribution of DSI enhancements between the output layers and the input layers. The DSI enhancement was calculated as the ratio of the DSI difference between the output and input layers to the DSI in the input layers. Data and code that support these findings are available at: https://doi.org/10.5281/zenodo.14177191. DSI, direction selectivity index.

https://doi.org/10.1371/journal.pbio.3002947.g002

In addition to comparing DSIs, we quantified the shifts in preferred directions of neuronal pairs located at different depths within the output layers, within the input layers or across the input and output layers. We observed that for most neuronal pairs within a cortical column, the preferred direction remained largely unchanged, irrespective of whether the neurons were within the output layers, within the input layers or across the input and output layers (Fig 3A, proportion of neuronal pairs with preferred direction difference less than 30 degrees, within output layers: 83.33%, N = 330 pairs; within input layers: 69.62%, N = 237 pairs; across input and output layers: 78.16%, N = 577 pairs), reflecting the columnar organization of direction preference previously demonstrated in the early visual cortices of carnivores (cat: [52–54]; ferret: [3,55]). We also found a small subset of neuronal pairs exhibiting a shift in preferred directions of approximately 180 degrees (Fig 3A, proportion of neuronal pairs with preferred direction difference greater than 150 degrees, within output layers: 10.30%, N = 330 pairs; within input layers: 20.25%, N = 237 pairs; across input and output layers: 14.38%, N = 577 pairs), consistent with findings from previous studies [56]. However, upon comparing neuronal pairs with strong direction selectivity (those for which the DSIs of both recording sites were greater than 1/3), we found that nearly all pairs exhibited no change in their preferred directions (Fig 3B, proportion of neuronal pairs with preferred direction difference smaller than 30 degrees, within output layers: 86.74%, N = 181 pairs; within input layers: 88.41%, N = 69 pairs; across input and output layers: 88.99%, N = 218 pairs). Enhancement of direction selectivity in the output layers and maintenance of preferred directions across input and output layers were also observed with sorted single units and when considering only putative excitatory neurons based on their spike waveforms (S4A–S4D Fig).

Fig 3. Preferred directions are maintained within a cortical column.

(A) Differences in the preferred directions between all pairs of recording sites within the output layers (left panel), within the input layers (middle panel), and across the input and output layers (right panel). (B) Similar to A but restricted to pairs in which the DSIs of both recording sites were larger than 1/3. Data and code that support these findings are available at: https://doi.org/10.5281/zenodo.14177191. DSI, direction selectivity index.

https://doi.org/10.1371/journal.pbio.3002947.g003

To further elucidate the processing and relay of direction selectivity from the input layers to the output layers, we compared the population-averaged direction tuning between the input and output layers. Prior to averaging, we aligned the preferred direction of each recording site to 90 degrees. As depicted in Fig 4A, as opposed to the neuronal responses in the preferred direction, the response strength in the other directions was weakened in the output layers compared to that in the input layers. As a control, we also compared the response strength under blank stimuli between the output layers and the input layers. We found that, under blank stimuli, the response strength was also reduced in the output layers (Fig 4A), which is in line with the reduced spontaneous activity in the output layers of awake macaques [57–60] and suggests that neuronal activity in the output layers is inherently weaker than that in the input layers.

Fig 4. The response in the preferred direction is selectively preserved from the input layers to the output layers.

(A) Population-averaged direction tuning of response strength within the output layers (blue lines) and the input layers (red lines). The blue and red horizontal lines represent the average responses to blank stimuli within the output layers and within the input layers, respectively. Data are mean ± SEMs. (B) Population-averaged direction tuning of the response ratio between the output layers and the input layers. Data are mean ± SEMs. (C) Distribution of the response ratio under the indicated stimulus conditions. The vertical dashed line represents a response ratio value of 1. From left to right, the panels show the response ratio under blank stimulus, in the null direction, and in the preferred direction. (D) Comparison of the response ratios between the indicated stimulus condition pairs. From left to right, the panels depict the comparison of response ratios between the null direction and blank stimulus, between the preferred direction and blank stimulus, and between the preferred direction and the null direction. Each dot indicates the data from one probe placement. Data and code that support these findings are available at: https://doi.org/10.5281/zenodo.14177191.

https://doi.org/10.1371/journal.pbio.3002947.g004

We then calculated the response ratio, which is the ratio of the response strength between the output layers and the input layers. The response ratio was greatest in the preferred direction (Fig 4B). Fig 4C shows the distribution of response ratios under 3 stimulus conditions: blank stimulus, null direction, and preferred direction. For both the blank stimuli and null direction, the response ratio was significantly less than 1 (Fig 4C; blank stimuli: N = 24 probe placements, p = 2.98*10⁻⁶; null direction: N = 24 probe placements, p = 0.00154; sign test), indicating a less active neuronal response in the output layers under nonpreferred conditions. In contrast, in the preferred direction, the response ratio approached 1 (Fig 4C, N = 24 probe placements, p = 0.541; sign test), suggesting the preservation of the preferred response in the output layers. A comparison of the response ratios among these conditions revealed no significant difference between the blank stimuli and null direction (Fig 4D, N = 24 probe placements, p = 0.317; Wilcoxon signed-rank test), whereas the response ratio for the preferred direction was substantially greater than that for both the blank stimuli (Fig 4D, N = 24 probe placements, p = 0.00184; Wilcoxon signed-rank test) and null direction (Fig 4D, N = 24 probe placements, p = 2.35*10⁻⁵; Wilcoxon signed-rank test). Additionally, there are diverse relationships between response ratios across different stimulus conditions (Fig 4D; null direction versus blank stimuli: r = −0.29, p = 0.166; preferred direction versus blank stimuli: r = −0.48, p = 0.0185; preferred direction versus null direction: r = 0.85, p = 2.13*10⁻⁶; Spearman’s ρ). The interlaminar comparisons of response strength at these chosen stimulus conditions for each probe placement are shown in S5 Fig. Similar results were obtained when considering only the signal transmission from L4 to L2/3 (S3D–S3G Fig).

What mechanisms can lead to a direction-tuned response ratio between the output layers and the input layers and ultimately result in enhanced direction selectivity in the output layers? Similar to the neural circuits responsible for generating direction selectivity [10,61], the direction-tuned response ratio between the output layers and the input layers we observed could result from several factors: (1) Nonlinearity [50,62], such as spiking nonlinearity [63] or gating of sensory information across layers [64]. Under this assumption, a direction-independent gain of signal transmission from the input to output layers, combined with a nonlinearity like a threshold-linear function, is sufficient to explain the observed direction-tuned response ratio. (2) A direction-dependent change in the relative strength of excitation and inhibition, which could involve increased excitation in the preferred direction [13], increased inhibition in the null direction [14], or both, as seen with the influence of brain state on direction selectivity in rabbit LGN direction-selective neurons [38]. This mechanism suggests that, in contrast to a direction-independent gain, the gain of interlaminar signal transmission varies with the direction of the stimulus.

To differentiate between these 2 potential mechanisms, we constructed 2 models simulating signal transmission from the input layers to the output layers (Fig 5A). In both models, the neuronal response in the input layers was first scaled by a linear gain and then subjected to static nonlinearity, specifically a threshold-linear transfer function [65–67]. The parameters of both models were optimized to explain neuronal responses in the output layers. The distinction between these 2 models lies in the linear gain: in the first model, the gain was constant across all stimulus conditions (Model I: untuned gain), whereas in the second model, the gain varied with stimulus condition (Model II: tuned gain).

Fig 5. A model with an interlaminar direction-tuned gain mechanism can explain neuronal responses in the output layers.

(A) Two models aim to explain neuronal responses in the output layers. In both models, responses from the input layers are first scaled by a linear gain and then pass through a static nonlinearity before producing the output responses. In the first model (Model I, Untuned Gain), the linear gain is fixed across all directions. In the second model (Model II, Tuned Gain), the linear gain varies in different directions. The exemplar neuronal responses in the input and output layers are the same as those in Fig 4A, and are also presented as mean ± SEMs. (B) Comparison of DSIs between the output of Model I and the input layers. (C) Comparison of DSIs between the output of Model I and the output layers. (D) The distribution of the contribution of nonlinearity to the enhancement of direction selectivity across layers. This contribution is quantified by calculating the ratio of the DSI difference between Model I (untuned gain) and the input layers to the DSI difference between the output layers and the input layers. (E) Comparison of adjusted goodness of fit (adjR²) between Model I and Model II. Since Model I and Model II differ in the number of free parameters, we used adjR², which accounts for the number of free parameters in each model, to enable fair comparisons. The horizontal and vertical dashed lines represent an adjR² value of 0.8. Only results from probe placements where Model II has an adjR² greater than 0.8 were shown in other figures. Similar results were obtained using the AIC and BIC. (F) Comparison of DSIs between the output of Model II and the output layers. (G) The distribution of the contribution of the direction-tuned gain to the enhancement of direction selectivity across layers. This contribution is quantified by calculating the ratio of the DSI difference between Model II (tuned gain) and Model I (untuned gain) to the DSI difference between the output layers and the input layers. For B, C, E, and F, each dot represents data from one probe placement. Data and code that support these findings are available at: https://doi.org/10.5281/zenodo.14177191. AIC, Akaike information criterion; BIC, Bayesian information criterion; DSI, direction selectivity index.

https://doi.org/10.1371/journal.pbio.3002947.g005

We first evaluated whether nonlinearity alone could explain the neuronal responses observed in the output layers. Although the DSI of the output responses in Model I exceeded that of the input layers (Fig 5B, N = 24 probe placements, p = 0.000228; Wilcoxon signed-rank test), it remained lower than that of the output layers (Fig 5C, N = 24 probe placements, p = 0.000162; Wilcoxon signed-rank test). The overall contribution of nonlinearity to the enhancement of direction selectivity is about 29.09%±5.17% (Fig 5D). This finding indicates that while nonlinearity contributes to enhancing direction selectivity beyond the input layers, it does not entirely account for the enhanced direction selectivity observed in the output layers. In contrast, Model II offers a more comprehensive explanation of neuronal responses in the output layers, as indicated by the adjusted goodness of fit (adjR²), which accounts for the number of free parameters in each model (Fig 5E, adjR² of Model II versus adjR² of Model I: N = 25 probe placements, p = 1.23*10⁻⁵; Wilcoxon signed-rank test). Based on the Akaike information criterion (AIC) and Bayesian information criterion (BIC), which both account for the number of free parameters in each model, we obtained similar results (S6 Fig). The DSI of the output response in Model II shows no significant difference compared to that in the output layers (Fig 5F, DSI of the output responses in Model II versus DSI of the output layers: N = 24 probe placements, p = 0.407; Wilcoxon signed-rank test). The diverse relationships of response ratios between different stimulus conditions, as demonstrated in Fig 4D, can also be accurately replicated in Model II (S7 Fig). Additionally, the contribution of direction-tuned gain to the enhancement of direction selectivity is 70.15%±6.30% (Fig 5G), suggesting that an inter-laminar direction-tuned gain mechanism plays a critical role in this process.

Subsequently, we analyzed the population-averaged direction tuning of the linear gain extracted from Model II. We observed that the linear gain was direction-tuned, peaking in the preferred direction (Fig 6A). A strong positive correlation was found between the DSI of the linear gain and the DSI of the output layers (Fig 6B, N = 17 probe placements, r = 0.57, p = 0.0179; Spearman’s ρ), indicating that this direction-tuned gain mechanism contributes to the enhancement of direction selectivity in the output layers. Fig 6C shows the comparison of linear gains among the blank stimulus, null direction, and preferred direction. The gain in the null direction exceeded that of the blank stimulus (Fig 6C, N = 17 probe placements, p = 0.00309; Wilcoxon signed-rank test), suggesting interlaminar amplification of the response in the null direction. Furthermore, the gain in the preferred direction was greater than those for both the blank stimulus (Fig 6C, N = 17 probe placements, p = 0.000503; Wilcoxon signed-rank test) and the null direction (Fig 6C, N = 17 probe placements, p = 0.000293; Wilcoxon signed-rank test), demonstrating selective amplification in the preferred direction. Similar results were obtained when considering only the signal transmission from L4 to L2/3 (S3H and S3I Fig) and when using a different type of nonlinearity, the power-law nonlinearity (S8 Fig).

Fig 6. A direction-tuned gain contributes to the enhancement of direction selectivity in the output layers.

(A) Population-averaged direction tuning of linear gains in Model II. The horizontal line is the linear gain under the blank stimulus. Data are mean ± SEMs. (B) Relationship of DSIs between the linear gain and the output layers. (C) Comparison of linear gains between the indicated stimulus condition pairs. From left to right, the panels depict the comparison of linear gains between the null direction and the blank stimulus, between the preferred direction and the blank stimulus, and between the preferred direction and the null direction. For B and C, each dot represents data from one probe placement. Data and code that support these findings are available at: https://doi.org/10.5281/zenodo.14177191. DSI, direction selectivity index.

https://doi.org/10.1371/journal.pbio.3002947.g006

In the nonlinearity we used, whether a threshold-linear transfer function or a power-law nonlinearity, a threshold-like component is included, which could act like a suppression. To explicitly clarify the roles of linear gain and the suppression, we developed a third model. In this model (Model III), input layer responses are first multiplied by a linear gain, then subtracted by a constant, and finally processed through a rectified-linear nonlinearity (ReLU) to simulate output layer responses (S9A Fig). We found that Model II outperformed Model III by a very small margin (S9B Fig). In Model III, the gain was also direction-tuned (S9D Fig), and there was a strong relationship between the DSIs of the tuned gain and the output layers (S9E Fig).

Finally, we conducted GC analysis to identify the potential neural circuit origin of the direction-tuned gain. For a specific pair of connections, we calculated the pairwise-conditional GC value. In this calculation, all joint dependencies among the other known variables are conditioned out, effectively mitigating the impact of potential confounding factors, such as common inputs. Compared to blank stimuli, stimuli in the null direction elicited stronger feedforward connections from L4 to L2/3 and from L4 to L5 (Fig 7A). Moreover, stimuli in the preferred direction not only reinforced these connections from L4 to L2/3 and from L4 to L5 but also intensified the recurrent connections from L2/3 to L5 (Fig 7B). Notably, compared to those in the null direction, stimuli in the preferred direction further strengthened the connections from L4 to L2/3 and from L2/3 to L5 (Fig 7C). S10 Fig shows the paired comparisons of normalized GC values between different stimulus conditions. These findings suggest that the direction-tuned gain likely stems from the combined effect of feedforward connections from the input layers to the output layers, as well as the recurrent connections within the output layers.

Fig 7. The preferred direction evokes stronger feedforward connections from the input layers to the output layers and recurrent connections within the output layers.

(A) Differences in the normalized pairwise-conditional GC values between the null direction and blank stimulus. The pairwise-conditional GC value for a specific pair of connections is calculated by conditioning out all joint dependencies among the other known variables. This approach helps mitigate the impact of potential confounding factors, such as common inputs. The grayscale inset depicts the p value of the differences in GC values within each connection pair. The source layer (the sender of the projection, from) is positioned along the abscissa, while the target layer (the receiver of the projection, to) is represented along the ordinate. (B) Differences in the normalized GC values between the preferred direction and blank stimulus. (C) Differences in the normalized GC values between the preferred direction and the null direction. Data and code that support these findings are available at: https://doi.org/10.5281/zenodo.14177191. GC, Granger causality.

https://doi.org/10.1371/journal.pbio.3002947.g007

In both carnivores [68,69] and primates [47], direction selectivity is considered an emergent property of the visual cortex and was initially observed in the V1 input layers. Unlike in cat early visual cortices, where direction-selective neurons are prevalent in all layers, in monkey V1 these neurons are only prevalent in the input layers (L4 and L6, [47]). This is because motion information is not primarily transmitted to the V1 output layers but is instead directed to the MT via L4B and L6 of V1 in monkeys [70]. Additionally, within the ventral pathway of monkey, motion information is transmitted to V2 thick stripes from L3 of V1 [71]. In contrast to monkey V1 [72,73], neurons in cat early visual cortices are organized into 2D direction maps [52–54,74–76]. These differences in the laminar and tangential organization of direction selectivity make the cat early visual cortices an ideal model for studying the laminar processing of direction selectivity.

Leveraging laminar recordings from multielectrode linear probes, our study characterized the laminar distribution of DSI and demonstrates, for the first time, that direction selectivity is enhanced in the output layers compared to the input layers within the same cortical column (Fig 2). Based on a database of 899 cells, Kim and colleagues characterized the distribution of DSI across different layers of cat area 17 (A17) [50]. They also found a higher proportion of direction-selective neurons in L2/3 and L5 compared to L4. In L6, however, they observed that the DSI distribution peaked at high DSI values, a pattern not observed in our data. Since our sample size of L6 neurons is relatively small compared to theirs, our sample may underrepresent some highly direction-selective neurons in L6. In a recent study, Hawken and colleagues identified 6 main functional clusters in L6 of monkey V1 [77], suggesting multiple parallel circuits within L6. One of these clusters, which is highly direction-selective, was located near the upper half of L6. Additionally, a two-photon study found that the direction map in cat A17 is not homogeneous but patchy [74]. These findings suggest that sampling at different cortical depths or positions may result in varying DSI distributions in L6.

Within a cortical column, we observed that a minority of the neuronal pairs demonstrated a shift in preferred directions of approximately 180 degrees (Fig 3A), corroborating earlier research [56]. The abrupt shift in preferred directions can be partially explained by the presence of direction fractures within the direction map [52–54,75] and the observation that some neurons exhibit opposite preferred directions for the 2 eyes [78] in cat early visual cortices. However, neuronal pairs with pronounced direction selectivity consistently maintain their preferred direction (Fig 3B). This suggests that strong direction selectivity is coupled with a more stable representation of motion information. Since the DSIs of the output layers and the interlaminar gain are positively correlated (Fig 6B), this stability may stem from the stronger direction-tuned gain of interlaminar signal transmission. These insights offer a vital point of reference for subsequent research on the plasticity of direction selectivity, as well as its potential modulation through learning or adaptive processes.

In addition to direction selectivity, other response properties also exhibited distinct laminar distributions, including the response modulation ratio (F1/F0) and orientation selectivity. Based on F1/F0, visual neurons are classified as simple cells (F1/F0 ≥ 1) and complex cells (F1/F0 16]; monkey V1: [18]), as is shown in S11 Fig. Within both the simple and complex cell populations, direction selectivity is enhanced in the output layers (S11C and S11D Fig). Furthermore, we found no correlation between F1/F0 and DSI in either the output layers or the input layers (S11E and S11F Fig), in line with previous studies in cat early visual cortices [76] and monkey V1 [30,33]. These results suggest that the enhancement of direction selectivity is primarily due to interlaminar processing.

Similar to direction selectivity, previous studies have reported stronger orientation selectivity in the output layers compared to the input layers (cat early visual cortices: [79]; monkey V1: [17,18]), as demonstrated in S12 Fig. A strong correlation was found between orientation selectivity and direction selectivity, suggesting that the generation of orientation selectivity and direction selectivity may share partially overlapping mechanisms, such as intracortical inhibition [62,80,81] or spike threshold [19]. These diverse relationships between different response properties and their laminar dependence highlight the need for comprehensive large-scale models [82–84], which could provide insights into whether these properties share distinct mechanisms and the roles of diverse laminar circuits.

In this study, we combined data from cat area 17 (A17) and area 18 (A18). Previous studies have found a higher proportion of direction-selective neurons in A18 compared to A17 [85,86], with A18 displaying a homogeneous direction map [52,54] and A17 showing a patchy direction map [74]. Consistent with these findings, we also observed a higher proportion of direction-selective neurons in A18, but only in the output layers, not in the input layers (S13 Fig). This difference in DSI distributions within the output layers of A17 and A18 may be explained by the finding that, while the response ratios between the output and input layers are largest at the preferred direction in both areas, the response ratios in A18 tend to be higher than those in A17. It would be interesting to explore whether interlaminar processing varies between different cortical areas, such as V2 thick stripes [72,73] and MT [87] in monkey, where direction maps are also observed.

We identified 2 factors contributing to the enhanced direction selectivity in the output layers. The first is nonlinearity (Fig 5), potentially arising from gating of sensory information across layers [64] or nonlinear transformations from membrane potential to spike activity [63,88]. The second factor is a direction-tuned gain for interlaminar signal transmission (Figs 5 and 6). The functional relevance of this direction-tuned gain may lie in enhancing the signal-to-noise ratio in visual processing and allowing for more accurate and efficient encoding of motion information.

Through GC analysis, we demonstrated increased feedforward connections from the input layers to the output layers, as well as increased recurrent connections within the output layers at the preferred direction (Fig 7). The feedforward connections are inherently excitatory, while the increased recurrent connections could result from increased excitation, increased inhibition, or a combination of both.

The increase in excitation at the preferred direction may be supported by the presence of direction maps. Given the presence of direction maps, neurons within a direction column are likely to receive stronger excitatory inputs upon exposure to their preferred direction, which could result from enhanced feedforward convergence or recurrent connections, or both [89–91]. Future investigations are needed to determine whether the degree of enhancement in direction selectivity varies across different locations within the direction map and to explore whether the direction tuning of interlaminar signal transmission depends on these locations.

The involvement of inhibition in direction selectivity is evidenced by experiments showing the influence of direction selectivity through manipulations of inhibition or modeling results. Direction selectivity of cortical neurons is reported to decrease with the blockade of inhibition [92–96] and to increase with enhanced inhibition [80,97]. By modeling the cat’s early visual cortices, Freeman revealed that inhomogeneities in the orientation map can cause spatiotemporal offsets between excitation and inhibition [98], and these offsets have been demonstrated to contribute to direction selectivity across various species (cat: [63]; monkey: [99]; ferret: [95]; mouse: [35,100]).

There are different possibilities for the relative direction tuning of inhibition and excitation: inhibition can be untuned [82], co-tuned with excitation [63,101–103], or tuned differently from excitation [95]. To study the role of inhibition in the interlaminar direction-tuned gain, future studies could use either pharmacology or optogenetics [95,104] to selectively activate or inactivate inhibition within certain cortical layers. If the impact of intracortical inhibition is significant, it would be beneficial to examine the roles of diverse types of inhibitory cells [105–107] and how the interplay between excitation and inhibition leads to the interlaminar direction-tuned gain.

In our single-unit data, when considering only narrow-spiking units, we found no significant difference in direction selectivity between the input and output layers, as direction selectivity was already strong in the input layers (S4E Fig). However, it is uncertain whether this finding can be attributed to the direction tuning of inhibitory neurons in cat early visual cortices, due to 2 factors: (1) the sample size of narrow-spiking units in the input layers is relatively small (28 single units); and (2) classification of neuronal types based on spike waveform remains controversial. Recent studies have shown that some narrow-spiking units in cat early visual cortices and monkey V1/premotor/motor cortex may belong to excitatory neurons [108–110], while some GABAergic neurons in mouse V1 exhibit broad-spiking waveforms [111]. Moreover, results on the direction selectivity of inhibitory neurons are diverse (cat early visual cortices [112,113]; ferret V1 [114]), highlighting the need for future studies to characterize the response properties of inhibitory neurons in the early visual cortices and their functional roles within neocortical networks.

In our model results, although responses in the output layers are derived by scaling responses from the input layers in our model (Fig 5), this does not conclude that the underlying neural circuit employs a pure multiplication algorithm. Alternatively, the tuned gain can be equivalent to a fixed gain combined with tuned suppression, as illustrated in Model IV of S9 Fig. It has been shown that normalization—a process where an excitatory drive is divided by the summed activity from neighboring neurons—is a canonical neural computation [115–119] that can explain various neuronal response properties, including contrast gain control [120,121], cross-orientation suppression [122,123] and surround suppression [24,26,124]. Understanding whether normalization contributes to laminar processing of direction selectivity, and how interactions between excitatory and inhibitory neurons shape underlying neural circuits, is crucial for advancing our knowledge of visual information processing.

In the present study, we recorded neuronal responses in anesthetized animals, which is well-suited for studying cortical processing of basic visual features. However, the use of anesthetized animals presents 2 inherent limitations. First, anesthesia may attenuate feedback connections from higher visual areas [125–127]. These feedback connections may influence the direction selectivity of neurons in the early visual cortices [128,129]. Second, the impact of propofol, the anesthetic used in the present study, on neuronal responses throughout the cortical column is a matter of ongoing debate, and it remains unclear whether the differences in response strengths across layers are affected [130–133]. To address these limitations, future studies could record laminar responses in awake animals to determine whether the results are consistent with those of the present study or if additional connections are engaged in the absence of anesthesia.

By using drifting gratings with fixed stimulus properties such as contrast, size, spatial frequency, and temporal frequency, and by varying only the direction, we can directly compare direction selectivity between the input and output layers through simultaneous recordings. Previous studies have demonstrated that direction selectivity is modulated by varied spatiotemporal configurations of visual stimuli [7,134,135]. Additionally, both intralaminar and interlaminar connections are known to exhibit variability in response to changes in stimulus configurations [26,45,60,136,137]. Therefore, it would be worthwhile for future studies to employ a richer array of stimuli to determine whether the strategy for columnar processing of motion information varies across different spatiotemporal configurations.

In total, we obtained 26 probe placements that had a high SNR and were perpendicular to the cortex. The perpendicularity of each probe placement was verified based on the variation of RF center positions and the consistency of preferred orientations across recording sites. To quantify the variation of RF center positions, we calculated the RF Center Dispersion:
(1)
where (x_i, y_i) and σ_i are the center position and radius of the ith recording site, N is the number of recording sites in one probe placement, and is the averaged center position across N probe placements. This parameter quantifies the variation in RF centers relative to the RF size within one probe placement, with the numerator representing the variation of RF center positions and the denominator representing the population-averaged RF diameter. A smaller value indicates less variation in RF center positions across all recording sites.

We quantified the consistency of preferred orientations by calculating Preferred Orientation Consistency:
(2)
where θ_i is the preferred orientation of the ith recording site and |C| is the amplitude of a complex value. The range for Preferred Orientation Consistency is from 0 to 1, with higher values indicating greater consistency in preferred orientations across recording sites; a value of 1 occurs when all recording sites within a probe placement share the same preferred orientation. In S1A Fig, we presented RF Center Dispersion and Preferred Orientation Consistency for all probe placements. Low RF Center Dispersion values and Preferred Orientation Consistency values close to 1 confirm perpendicular penetration of these probes. For long-lasting probe placements, perpendicularity was further verified postmortem through histology, as probe tracks were clearly delineated by CO staining.

For each probe placement, we assigned all recording sites to corresponding layers according to stimulus-evoked MUAs, LFPs, CSDs, and spatial RFs from SPN experiments [21].

Concerning the differences in cortex thickness across probe placements, we assigned a relative depth to each recording site [41,42,138,139]. The range of relative depth was from 0 to 1, with 0 indicating the border between the cortical surface and L2/3, and 1 indicating the border between L6 and white matter. The border between the cortical surface and L2/3 can always be identified from LFPs because the LFP waveform is nearly identical for recording sites out of the cortex, and an abrupt change exists between neighboring recording sites once it enters the cortex. The border between L6 and white matter was determined by MUA and spatial RF. The deepest site with a high SNR of MUA and consecutive spatial RF was deemed 50 μm above this border.

We separated the cortex into 4 layers (L2/3, L4, L5, and L6) using 3 relative depths (0.35, 0.58, and 0.76 from top to bottom) to mark the borders between neighboring layers. These borders were determined primarily based on MUA response latency, defined as the time point at which the response reached half of its maximum (Fig 1B). Notably, L4 and L6 exhibited faster latencies than L2/3 and L5 (Fig 1C–1E).

In S2A Fig, we showed MUA, LFP, and CSD responses evoked by black and white stimuli within the spatial RF for an example probe placement. MUA responses revealed advances in latency and a stronger rebound in both L4 and L6 [21]. LFP responses indicated a polarity reversal in L6, while CSD responses identified a current sink in L4 [140,141]. S2B Fig displays the number of electrodes within each layer across all probe placements, consistent with the relative thickness of each layer.

To enhance the statistical power, we combined the data from L2/3 and L5 as the “output layers” and those from L4 and L6 as the “input layers” due to their roles in signal transmission along the visual hierarchy [41] and their similarities in response properties [21]. In S3A Fig, we further showed that the DSI in L2/3 and L5 was significantly larger than in L4 and L6. Similar results were observed when comparing between L4 and L2/3, as seen between the output layers and input layers, including enhanced direction selectivity in L2/3 (S3B and S3C Fig), a direction-tuned response ratio between L2/3 and L4 (S3D–S3G Fig), and a direction-tuned gain mechanism from L4 to L2/3 (S3H and S3I Fig).

Spike sorting was initially performed automatically using KiloSort [144] and then manually curated using Phy [145]. The quantity of each unit was quantified by the SNR of the spike waveform, which was defined as the ratio of the peak-to-peak amplitude of the mean waveform to twice the standard deviation of the noise [146]. Any unit with a spike waveform SNR greater than 1.5 was classified as a single unit (SU).

We separated all SUs into 5 classes (narrow-spiking units, broad-spiking units, triphasic-spiking units, compound-spiking units, and positive-spiking units) based on recent work studying extracellular spike waveforms in cat early visual cortices [147]. The details of the classification can be found in our previous work [21].

For each SU, we calculated the response modulation ratio (F1/F0) under the preferred stimulus condition, where F1 represents the response at the same temporal frequency as the stimulus drifting, and F0 is the averaged response. We then categorized each SU as either a simple cell (F1/F0 ≥ 1) or a complex cell (F1/F0 S11 Fig.