Flexible behavior requires the ability to modulate sensory processing based on task context, yet the circuit-level mechanisms supporting this capacity remain poorly understood. Here, we combine recurrent neural network modeling and neural recordings from mouse visual cortex to investigate how task context shapes sensory coding. Networks trained on an instruction-based discrimination task develop a disinhibitory interneuron-to-interneuron motif that dynamically gates task-relevant sensory information. Perturbation and lesion analyses show that this motif is necessary for task performance and for maintaining distinct sensory representations across contexts. We validate key predictions in mouse visual cortex, where interneuron activity patterns exhibit comparable task-dependent modulation. These results identify a biologically plausible circuit motif that supports flexible sensory processing and link recurrent connectivity structure to adaptive context integration in both artificial and biological systems.

RNN model and dynamic decision-making tasks

To understand the cortical computations required for flexibly and optimally integrating bottom-up and top-down information, we employed a biologically plausible RNN model consisting of continuous-variable firing units (model schematic shown in Fig 1a right; see Methods) and trained the model to perform novel dynamic decision-making tasks (Fig 1a, 1b). We trained the RNNs using a supervised gradient-descent procedure, in which the network learned to adjust its recurrent synaptic weights, output readout weights, and neuron-specific synaptic decay time constants to perform the task accurately (Fig 1; see Methods). To preserve biological interpretability, we enforced Dale’s principle by constraining each unit to be either purely excitatory or purely inhibitory, using a sign-constrained weight parametrization [48]. These constraints ensure that learned network dynamics can be meaningfully related to cortical circuit motifs, enabling mechanistic interpretation of how specific inhibitory circuits support flexible encoding of top-down information.

These RNNs were employed to execute flexible decision-making tasks designed to achieve several distinct objectives. The first aim was to investigate how identical sensory inputs might be processed differently depending on distinct top-down inputs. These tasks also allowed for the independent manipulation of both bottom-up and top-down inputs. Lastly, the task incorporated separate stimulus modalities, each providing independent sources of top-down information. To meet these criteria in task design, we modified the classic delayed match-to-sample (DMS) task, commonly used in experimental studies [62–66], to develop one-modality DMS and two-modality DMS tasks.

For the one-modality DMS task (Fig 1a), the RNN model received one stream of input stimulus signal (bottom-up) containing two sequential stimuli separated by a delay period. Each stimulus was set to either +1 or −1. The model additionally received a task cue signal (top-down), indicating whether the network should employ the pro- or anti-DMS stimulus-response mapping. For the pro-DMS task (a task cue of +1 during the instruction period), the network was required to generate an output of +1 if the signs of the two stimuli matched, and −1 otherwise. For the anti-DMS (a task cue of −1 during the instruction period), the network was required to generate a negative output if the signs of the two stimuli were the same, and a positive output otherwise. The task cue was given either before (early instruction; 150–400 ms) or after the first stimulus (late instruction; 650–900 ms; Fig 1a).

Critically, in nature, not all aspects of sensory input are relevant to the task at hand. To effectively navigate this, organisms rely on additional top-down signals, such as selective attention, alongside direct task signals to optimize decisions. To study this more complex scenario, we developed the two-modality DMS task by incorporating two streams of input stimulus signals (bottom-up; modality 1 and modality 2) and an attention cue (top-down) which instructs the model which stimulus modality is behaviorally relevant on a given trial and thus should be attended to (Fig 1b; see Methods). An attention signal of −1 during the instruction period informs the model to focus on modality 1 while ignoring modality 2, and vice versa for a +1 attention cue signal. The task cue manipulation is identical to that used in the one-modality DMS task. Similar to the one-modality task, the timing of the instruction period was manipulated such that task and attention cues were delivered either before (early instruction) or after the presentation of the first stimulus (late instruction).

For each decision task (one-modality DMS and two-modality DMS), we trained 20 RNNs to perform the task with high accuracy (> 95%; see Methods). For the one-modality DMS, the models performed the task accurately regardless of the timing of the instruction window (Fig 1c; see S1 Fig for example single neuron activity and output signals). The model performance on the two-modality DMS was lower when the top-down cues were provided early (P < 0.001, two-sided Wilcoxon rank-sum test; Fig 1c). This can be attributed to the model’s requirement to retain both the task and attention signals for a longer duration in the early instruction condition.

Dynamic encoding of top-down information by RNNs

After training RNNs to perform the dynamic decision tasks, we next investigated the effects of the top-down factors (task and attention cues) on the network dynamics. To visualize RNN dynamics in response to distinct external inputs of both bottom-up and top-down signals under various conditions, we plotted the kinetic energy landscape (see Methods) of a representative RNN. This was achieved by using two principal components (Fig 2a, 2b; see Methods). The kinetic energy landscape provides a useful guideline for understanding RNNs as dynamical systems, since it offers a visual summary that illustrates how external inputs shape the structure of fixed points and “slow” points (local minima in the energy landscape) as well as neural trajectories within the state space [67–69]. Therefore, seeking to identify these fixed and “slow” points in the two-dimensional energy landscape enabled us to gain insight into how the task and attention signals influenced the RNN network dynamics.

We hypothesized that the presence of top-down cues would condition network dynamics even before the arrival of the bottom-up sensory inputs. In addition, we postulated that opposing cues would correspondingly lead to separable shifts in fixed point structure, which we first visualized through the energy landscape of the network during the first instruction period (150–400 ms, see Fig 1a). For the one-modality DMS task, the energy landscape of the example RNN model displayed an area with an inverted U-shape where the kinetic energy was notably low (darker blue area in Fig 2a). The local minimum of the landscape stayed around the center of this U-shaped profile during the period before the first stimulus presentation when the task cue was not yet presented (no instruction). However, when the task cue was presented early (early instruction), the local minimum shifted to the left and to the right for the pro- and anti-DMS condition, respectively (Fig 2a).

For the two-modality DMS task, the energy landscape again revealed a U-shaped region characterized by a low-energy profile (dark blue in Fig 2b). Consistent with what we observed in the one-modality task, the local minimum was situated at the center of the low energy structure during the first instruction period for late instruction trials, where no task or attention cues were yet presented (no instruction). Analyzing different combinations of the task and attention cue signals displayed distinct local minima points along the low-energy profile. For instance, the local minima corresponding to the pro-DMS signal were positioned to the right of the no-instruction minima. Along this pro-DMS “arm” of the energy trajectory, two local minima were observed: one corresponding to the modality-one attention signal and the other to the modality-two attention signal. These results suggest that network dynamics shift as a function of the combination of top-down attention and task signals, in a dissociable manner, even before the sensory inputs arrive.

Subsequently, we performed a quantitative fixed point search analysis [67] for different values of external input to the network during the first instruction period (150–400 ms, Fig 1a). During this time, the network could receive either a pro-DMS, an anti-DMS, or no task cue (i.e., late instruction trials). We plotted network trajectories as well as fixed points discovered for each of these conditions in the principal component space, for the representative RNN performing the one-modality (Fig 2c) and the two-modality DMS tasks (Fig 2d). We observed that the presence and content of top-down signals given to the network not only altered its trajectories in the state space but also affected its fixed point structure. The network performing the two-modality DMS task exhibited a branching pattern that reflected symmetrical top-down coding across attended modalities for each task cue condition (anti- versus pro-DMS).

The energy landscape plots and the low-dimensional trajectories are qualitative ways to visualize which portions of the neural network’s activity landscape are more likely to be stable or have slower trajectories. The intuition behind the energy landscapes is that their local minima correspond to lower trajectory velocity. It is useful to visualize the transformation of these landscapes as a function of external inputs to the network, since that determines how specific instructions reshape the structure of the energy landscapes, and, therefore, the likelihood of each type of trajectory given a stimulus-instruction input pair.

To further examine network dynamics as a function of distinct top-down and bottom-up information, we performed a manifold discovery analysis, utilizing the CEBRA method [70]. This approach enables latent structure discovery and has the advantage of conditioning latent structure discovery over variables of interest. The contrastive learning objective and nonlinear mapping give CEBRA the ability to discover structured, consistent representations that preserve data relationships beyond simple variance maximization along orthogonal axes.

Here, we performed an unsupervised analysis with regard to all task covariates except for time. This ensured that adjacent time points were mapped to closer points in latent space and that smooth latent dynamics were obtained. This method allowed us to determine the latent trajectory bifurcations in the state space following presentation of stimuli and top-down cues in the network performing the one-modality DMS (i.e., one-modality RNN; Fig 2e) and the two-modality DMS tasks (i.e., two-modality RNN; Fig 2f–2i). In the one-modality RNN, the trajectories during the first instruction period were separated into three pathways corresponding to the pro-DMS cue signal (light green and light purple in Fig 2e), anti-DMS cue signal (dark green and dark purple in Fig 2e), and the absence of the task cue instruction (i.e., late instruction condition; gray in Fig 2e). The pro-DMS and anti-DMS pathways were then further bifurcated based on the identity of the first stimulus during the first stimulus period. For the two-modality RNN, we observed similar latent trajectories for the attended modality (Fig 2f). However, the bifurcations driven by the identity of the first stimulus were not as well separated in the unattended modality (Fig 2G). To quantify this effect, we measured the Euclidean distance between latent trajectories corresponding to opposite first-stimulus identities during the stimulus-1 period using repeated random subsampling. The average trajectory separation was significantly greater for the attended modality than for the unattended modality, confirming that stimulus-driven bifurcations are selectively amplified by attention (S2 Fig). This pattern held true across pro- and anti-DMS trials, which exhibited qualitatively similar bifurcations (Fig 2h, 2i). These findings were further confirmed by the cross-temporal discriminability analysis, which revealed that the identity of the first stimulus of the attended modality was more robustly encoded throughout the trial duration compared to the unattended modality (Figs 2j–2m; S3; see Methods). Linear support vector machine (SVM) decoding analysis also revealed that the top-down cue signals were robustly encoded by the RNN models (S4 Fig; see Methods for more details). These results suggest that lower-dimensionality representations of network dynamics sufficiently capture the network’s flexibility in encoding stimuli in a context-dependent manner, depending on stimulus relevance, as flexibly determined by top-down cues.

Inhibitory units play a critical role in both encoding and maintaining top-down information

After establishing the model’s capacity for flexible encoding of top-down information, we next investigated the relative contributions of excitatory and inhibitory units in facilitating this process. Through dissection of the RNNs performing the one-modality and two-modality DMS tasks, we found that the optimized synaptic decay time constants for the excitatory units were significantly larger compared to the time constants of the inhibitory units (P < 0.001, two-sided Wilcoxon rank-sum test, Fig 3a, 3b). Importantly, these optimized decay time constants should be interpreted as effective functional timescales that emerge through training to support the task’s computational demands.

Building on these findings, which indicate distinct roles for excitatory and inhibitory units, we conducted a series of linear SVM decoding analyses. These analyses aimed to decode the task cue identity (anti- versus pro-DMS) from the activity of excitatory and inhibitory units within RNNs engaged in the one-modality DMS task (see Methods). While the task cue identity could be readily decoded from the activities of both excitatory and inhibitory units following the cue presentation on early instruction trials, the SVM decoding accuracy was significantly higher for the inhibitory units throughout the trial duration (P < 0.05, two-sided two-sample KS test, Bonferroni corrected across time; Fig 3c). We observed similar trends for RNNs performing the two-modality DMS task (Fig 3d). In these models, the inhibitory units also exhibited stronger maintenance of the attention signal compared to the excitatory units (S5 Fig). Interestingly, decoding performance for stimulus identity was higher in inhibitory units for both the attended and unattended modalities, when compared to excitatory units (S5 Fig).

Given these findings that highlight the distinct role of inhibitory units on maintaining task-related information, we next investigated their network-level implications through the functional connectivity patterns of inhibitory units. More specifically, we measured the functional closeness centrality of all units in each network across time (see Methods). Assuming that correlations between activation rates over time establish a metric of distance between units, it follows that units with greater centrality emerge as potential candidates for serving as functional hubs in their respective networks. While both excitatory and inhibitory units increased in centrality following the top-down cue presentation (early instruction) for the RNNs performing the one-modality and two-modality DMS tasks, the average centrality was higher for the inhibitory units after the onset of task and attention cue presentation (P < 0.05, two-sided two-sample KS-test, Fig 3e, 3f). Comparing the distribution of the centrality measure from individual units in the one-modality RNNs before (at 145 ms) and after (at 395 ms) the first instruction period on early instruction trials revealed that the centrality values of inhibitory units were distributed more widely post-instructions. This finding indicates the formation of inhibitory functional connectivity hubs following the top-down cue presentation (Fig 3g, 3h). To understand the network structure underlying this increase in inhibitory centrality, we examined the connectivity patterns between rule-selective populations in the one-modality RNNs. Inhibitory neurons preferentially inhibited other inhibitory neurons with the same rule selectivity, with significantly stronger within-selectivity compared to across-selectivity I → I connections (S6 Fig). This structured motif was also observed across connection types, suggesting the emergence of rule-specific inhibitory subnetworks. Taken together, these results suggest a central role for inhibitory units in encoding task-specific top-down information, as well as stimulus identity.

Inhibitory-to-inhibitory connections carry top-down information

Given the crucial role of inhibitory units in encoding and sustaining top-down signals, we next investigated the neural circuitry surrounding these inhibitory units to further elucidate their contribution to the maintenance of task-specific top-down information.

Lesioning inhibitory connections to inhibitory units selective for the anti-DMS task cue (by reducing synaptic weights by 50%; see Methods) in a representative RNN trained to perform the one-modality DMS task resulted in the disruption of anti-DMS task signaling (Fig 4a). Consequently, the network exhibited a tendency to perform the pro-DMS task irrespective of the task cue presented. By visualizing the neural dynamics of this lesioned RNN using CEBRA [70], we confirmed that the neural trajectories were no longer separated by the task cue condition during the first instruction period in early instruction trials (Fig 4b).

Motivated by these findings, we employed a systematic approach where we lesioned synaptic connections based on task cue selectivity (pro- versus anti-DMS) and unit type (excitatory versus inhibitory) for the one-modality DMS task. We found that disrupting inhibitory-to-inhibitory () connections led to profound impairment in task performance (Fig 4c top; **P < 0.01, ***P < 0.001 by two-sided Wilcoxon rank-sum test). Disrupting similarly tuned inhibitory connections impaired performance in both pro- and anti-DMS trials (Ia → Ia and Ip → Ip in Fig 4c top; **P < 0.01, ***P < 0.001 by two-sided Wilcoxon rank-sum test). However, lesioning oppositely tuned connections resulted in greater impairment in one type of task trials (Ia → Ip and Ip → Ia in Fig 4c top; **P < 0.01, ***P < 0.001 by two-sided Wilcoxon rank-sum test). We also observed significant changes in task performance when the connections were lesioned (Fig 4c bottom; **P < 0.01, ***P < 0.001 by two-sided Wilcoxon rank-sum test). Importantly, the disproportionately large behavioral effects observed following lesioning of task cue-specific connections cannot be explained by a greater number of connections being perturbed, as synapses constitute the smallest fraction of outgoing connections compared to other synaptic classes (S7 Fig). The effects of lesioning task cue-specific connections were similar in the RNN models trained to perform the two-modality task (Figs 4d; S8). In addition, connections also carried attention cue-specific information in RNNs performing the two-modality DMS task (Figs 4e; S8). To further assess the relative contributions of excitatory versus inhibitory outputs, we extended our lesion analyses to include selectivity-matched perturbations of excitatory-to-excitatory () and excitatory-to-inhibitory () connections. Across both one- and two-modality RNNs, perturbation of connections produced minimal performance changes, while lesioning connections led to modest impairments that were substantially smaller than those observed following disruption of connectivity (S9–S13 Figs). These results indicate that, although excitatory projections contribute to shaping inhibitory competition, disinhibitory circuitry remains the dominant mechanism supporting the maintenance and flexible deployment of task- and attention-related top-down information.

These findings suggest that inhibitory-to-inhibitory connections are essential for integrating context-defining top-down information.

Two-module RNN model with hierarchical organization

Recent empirical studies have shown that inhibitory-to-inhibitory connections within early sensory areas receive strong task-specific and context-based modulation by higher cortical areas [40,61,71].

Based on our results thus far, which underscore the significance of connections in contextual coding, we next investigated whether inhibitory-to-inhibitory connections in low-level sensory areas would emerge as a critical target for task-oriented top-down modulation in our model. To accomplish this, we developed a two-module RNN model inspired by the cortical hierarchy. In this model, the first module (sensory module) is modeled after the early sensory cortex and receives the stimulus input signal only (Fig 5a). The first module subsequently conveys the processed sensory signal to the second, nonsensory module through exclusively feedforward excitatory connections. Given previous findings from animal studies showing that long-range feedforward projections are overwhelmingly excitatory, whereas top-down modulation can additionally engage sparse long-range inhibitory influences [31,33], we trained our two-module RNNs to perform the one-modality DMS task without intermodule feedforward inhibitory projections. The second module of the model, designed after higher cortical areas, receives the task cue signal (Fig 5a). Using this architecture, we trained 20 two-module RNNs to perform the one-modality DMS task (see S14 Fig for example single neuron activity and output signals). On average, the two-module RNNs took longer to learn the one-modality task compared to the one-module networks (mean ± stdev, 38,571 ± 6,306 trials for the one-module RNNs versus 55,221 ± 10,187 trials for the two-module RNNs; S15 Fig).

To investigate if the two-module RNNs processed the task cue signal in a manner similar to that observed in the one-module RNNs, we characterized the kinetic energy landscape and performed the fixed-point analysis on the two-module networks during the first instruction period (150–400 ms; see Methods). Similar to the landscape and energy minima seen in the one-module RNNs (Fig 2a), we observed an inverted U-shaped energy landscape and displacement of energy minima as a function of the task cue identity (Fig 5b). While the fixed-point analysis suggests a line-like structure as an attractor for the visualized example network (Fig 5c), this is not a consistent pattern across random initializations. The most common observation was for fixed points to cluster along unordered point clouds as opposed to line attractors, similar to the one-module network. Still, these results suggest that network trajectories are conditioned by a systematic shift in fixed point structure, which is brought about by the presence of distinct top-down signals in the form of task instructions.

Characterizing the synaptic decay dynamics of the two-module RNNs revealed slower excitatory dynamics in the nonsensory module compared to the sensory module (Fig 5d; P < 0.001 by two-sided Wilcoxon rank-sum test). The inhibitory units exhibited the opposite trend (Fig 5e; P < 0.001, two-sided Wilcoxon rank-sum test). While further work is required to understand the heterogeneous timescales across excitatory and inhibitory neurons, there are a few possible interpretations for this finding. Longer excitatory time constants in the sensory module may support integration over longer times, beneficial to context integration via top-down feedback to sensory areas. Faster excitation relative to inhibition, on the other hand, could support rapid sensory stimulus encoding. In the nonsensory module, more closely related to the final decision process, it is possible that an inverted dynamic is at play. Decision making often requires integrating evidence over time, which functionally necessitates longer effective integration windows. Slower excitatory synaptic dynamics are often incorporated to ensure that network dynamics smooth out noisy inputs and accumulate information gradually.

Investigating the average synaptic connection strength revealed strong excitatory connections within the sensory module and from nonsensory to sensory modules (Fig 5f; P < 0.001, Kruskal–Wallis test followed by Dunn’s multiple comparison test). Interestingly, the average inhibitory signaling was the strongest within the nonsensory module (Fig 5g; P < 0.001, Kruskal–Wallis test followed by Dunn’s multiple comparison test). Both and connections within the nonsensory module were significantly stronger than their counterparts in the sensory module (S16 Fig). These findings are consistent with previous empirical evidence suggesting increased inhibitory strength and diversity along the cortical hierarchy [60,72]. A possible interpretation for these findings is that stronger inhibition in the decision areas allows a tighter control of excitation, which supports precise modulation of internal computations that depend on context and goals rather than direct sensory drive. In higher-order modules where computations mainly rely on recurrent internal dynamics (e.g., evidence accumulation, rule maintenance), stronger inhibition may improve stability and prevent excessive activity. Recent literature suggests such a stabilizing role for inhibition in the context of strongly connected excitatory populations [53], expanding the parameter space where decision-making circuits can operate without instability.

Linear SVM decoding analyses in the two-module RNNs showed that the task signals were more readily decodable from the inhibitory units than from the excitatory units in both sensory and nonsensory modules (Fig 5h, 5i; see S17 Fig for the nonsensory module results, P < 0.05, two-sided two-sample KS test with Bonferroni corrected across time). Most notably, the SVM decodability of the task signal decayed more slowly for the inhibitory units compared to the excitatory units in both modules.

Inhibitory-to-inhibitory connections in both sensory and nonsensory modules encode top-down information

Analyzing the neural trajectories of an example two-module RNN during the first stimulus period for early instruction trials revealed trajectory grouping based on the task cue identity (Fig 6a). Lesioning connections in both sensory and nonsensory modules by reducing the weights by 50% abolished this top-down modulation (Fig 6b), mirroring the results observed in the one-module RNN (Fig 4b). Lesioning other synaptic connections (i.e., , , ) did not lead to disruption of the top-down information encoding (S18–S21 Figs). Thus, these results suggest that the hierarchical, two-module RNNs also rely on inhibitory-to-inhibitory connections to encode contextual information.

Performing a systematic lesioning analysis, similar to the one utilized for the one-module RNNs in Fig 4, demonstrated the importance of inhibitory units with opposite tuning to the task cue (i.e., anti- versus pro-DMS) in the sensory module for reliably encoding the task signal (Fig 6c). In addition, disrupting connections within the nonsensory module severely impaired the task performance of all the RNNs, confirming that top-down information is primarily encoded by connections in the nonsensory module (Fig 6d). Lesioning feedback inhibitory connections had minimal influence on the task performance (Fig 6e, 6f). Lesioning other synaptic connections did not lead to disruption of top-down information encoding (S18–S21 Figs), although perturbations of connections led to modest performance impairments, supporting a role for long-range excitatory drive in recruiting inhibitory populations in higher-order areas.

These results suggest that the task cue information is encoded by synapses in the non-sensory module. What is even more unexpected is that the top-down information is transmitted to the sensory module via feedback connections and encoded by connections within the sensory module (Fig 6c).

Disinhibitory circuits in mouse primary visual cortex and RNN sensory module encode top-down information

Building on the findings from the previous section, which indicate the role of sensory areas in encoding both stimulus-driven (bottom-up) and task-related (top-down) information, we proceeded to test our hypothesis. Specifically, we proposed that cortical connections within early sensory processing regions represent top-down information via feedback from higher cortical areas. To investigate this, we analyzed a publicly available experimental data which demonstrated the capability of the mouse primary visual cortex (V1) in representing contextual information [61]. This electrophysiology dataset was collected with a linear array recording electrode covering all layers of V1 and preprocessed with standard filtering and spike sorting pipelines.

In this study, during V1 activity recording, mice were trained to identify the location of a salient grating visual pattern (“figure”) that stood out from the visual background (“ground”) [61]. These patterns were distinguished based on either orientation, phase, or texture (Fig 7a). Crucially, in each of these conditions, the sensory inputs chosen to fall within the exact receptive fields of the recorded neurons remained strictly identical in both the figure and background conditions (see Methods).

This allows for probing the role of recurrent connections from higher visual areas into V1 on top-down representations, given previous results which causally link recurrent feedback from higher visual areas and top-down feedback into V1, beyond lateral connections within this area [61,73,74]. This mirrors our approach in configuring our RNNs to process identical bottom-up sensory information under varying contexts informed by top-down cues. Therefore, disparities in neural activity observed across conditions in the mouse dataset could be attributed to the context indicated by stimulus regions outside of the measured neural receptive field.

First, we tested whether the activity of V1 neurons could differentiate between figure and ground conditions, even when the sensory information within their receptive fields remained identical. A successful figure versus ground decodability would be evidence that feedback from regions outside receptive fields projects into the measured V1 sites. Performing a linear SVM decoding analysis on the V1 activity to decode the condition label (figure versus ground) revealed that the V1 responses were modulated by the task context (Fig 7b; P < 0.05, permutation test, Bonferroni corrected). The high decoding accuracy indicates that feedback signals from regions outside the receptive fields robustly modulate V1 activity in response to information about task context.

In sensory cortical areas, including V1, vasoactive intestinal peptide-expressing (VIP) interneurons send inhibitory projections to somatostatin-positive (SST) interneurons. This forms microcircuitry characterized by inhibitory-to-inhibitory connections, similar to the connections observed in the sensory module of our two-module RNN model (Fig 5a). The canonical VIP-SST disinhibitory circuit in cortex is a natural candidate for the biological implementation of the top-down regulatory mechanism elicited by inhibitory-to-inhibitory connections that arose in our artificial models because VIP neurons are known to receive long-range regulatory feedback from frontal areas [75,76]. However, it is important to acknowledge that other types of inhibitory-to-inhibitory connections exist in sensory circuitry, such as SST to parvalbumin (PV) neuron connections [77].

Leveraging the experiments within the dataset where VIP interneurons of mouse V1 were optogenetically silenced, we analyzed the role of VIP neurons in figure-background discrimination, in trials where top-down cues were presented through the orientation of visual grating patterns. Through the SVM decoding analysis, we observed a decrease in decoding accuracy following optogenetic silencing of V1 VIP neurons (P < 0.05, one-sided t test; Fig 7c). These results demonstrate the importance of intact inhibitory-to-inhibitory projections for successfully encoding top-down information in V1 circuits.

We next utilized CEBRA [70] to analyze the neural dynamics of the sensory module in an example two-module RNN model during the first stimulus presentation in early instruction trials of the one-modality DMS task. The recovered latent space revealed distinct neural trajectories delineated by the task cue signal (Fig 7d). The trajectories were initially divided into two arms based on the identity of the task cue (light green and light purple versus dark green and dark purple trajectories in Fig 7d). Within each arm, the trajectories were further separable by the first stimulus identity (+1 or −1). Removing all the inhibitory-to-inhibitory connections within the sensory module abolished the effects of the task signal in the sensory module (Fig 7e). These results highlight a convergence between theoretical models and experimental evidence for the importance of feedback into lower sensory areas in coding top-down information.

Performing a time-resolved linear SVM decoding analysis on the sensory module across all 20 trained two-module RNNs revealed that the task cue identity could be reliably decoded from sensory population activity, providing additional confirmation of top-down modulation imposed by the nonsensory module (Fig 7f; see Methods). When inhibitory-to-inhibitory () connections within the sensory module were removed, task cue decoding accuracy was markedly reduced compared to the intact network (Fig 7f), indicating that these disinhibitory connections play a critical role in enabling reliable top-down information encoding within the sensory circuitry.

Introducing retro-cue condition elicits shift in strategy

To test whether the disinhibitory circuit mechanisms identified above generalize to settings in which attentional relevance is specified after stimulus encoding, we introduced a retro-cue variant of the two-modality task as illustrated in Fig 8a. The task still involves two streams of stimuli (modality 1 and modality 2) along with the task cue (pro- and anti-DMS). In this task version, the task cue was fixed to be given before the first stimulus (see Methods). Notably, in this modified version, the attention cue may be presented before the first stimulus (early instruction), after the first stimulus (late instruction), or after the second stimulus (retro-cue), as depicted in Fig 8a. This modification provides a more nuanced exploration of attentional processes by allowing flexibility in the timing of attention cue presentation.

We trained 20 one-module RNNs to perform this modified two-modality DMS task. We then compared the resulting network dynamics for the case where top-down cues (both task and attention cue signals) were given before the first stimulus against the dynamics observed from the RNNs trained for the original two-modality DMS task (Fig 2). Applying CEBRA to an example RNN trained for the modified two-modality DMS task revealed that the trajectories for the attended modality closely resembled those obtained from the example RNN trained for the original two-modality DMS (i.e., compare Fig 8b with Fig 2f). The trajectories were initially distinguished by the task cue (pro- versus anti-DMS). Subsequently, each task “arm” bifurcated based on the identity of the attended first stimulus (Fig 8b). Interestingly, in the case of pro-DMS, where the unattended first stimulus had an identity of +1, the trajectories diverged based on the identity of the attended first stimulus (illustrated by the green trajectories in Fig 8c). This stands in stark contrast to the trajectories observed in the RNN model performing the original two-modality DMS task, where the trajectories of the unattended modality were unaffected by the attended modality (Fig 2g). In other words, the introduction of the retro-cue condition to the two-modality task forced the RNNs to maintain not only the attended but also the unattended stimuli throughout the trial period. This is further underscored by the robust temporal discriminability of the first stimulus of the unattended modality across the three attention cue conditions (i.e., early, late, and retro cue conditions) in the 20 RNNs trained for the modified two-modality DMS task (Fig 8d–8f; compare these to Fig 2l, 2m). Carrying out systematic lesioning analyses revealed that reducing inhibitory-to-inhibitory () connections led to a pronounced decrease in the temporal discriminability of the first stimulus (Fig 8g–8i), whereas lesioning the other synaptic classes produced substantially weaker effects (S22 Fig). These results demonstrate that, despite the altered task demands imposed by the retro-cue condition, stimulus representations continue to rely critically on disinhibitory circuitry for their maintenance.

Overall, introducing the retro-cue condition resulted in the RNN model employing a different strategy to flexibly prioritize and maintain relevant information throughout the task execution. Notably, this shift in strategy preserves the same underlying circuit mechanism, with disinhibitory inhibitory-to-inhibitory interactions remaining essential for stabilizing stimulus representations under dynamically changing attentional demands.

Continuous-rate RNN model

In this study, we constructed the following continuous-variable RNN model:

(1)(2)(3)

where refers to the synaptic time constants, denotes the synaptic current variable from N units across T time points. By applying a sigmoid nonlinearity, we estimated the firing rates based on the synaptic current values (). The connection weights from the time-varying inputs (; , the number of input channels) to the network were represented by the weight matrix . Additionally, contains connection weights between the N units.

To compute the network output (), we linearly combined all the firing rates specified by the output connection weight matrix, , along with the constant bias term, b. The network size (N) was set to 1,000 for all the networks. Output weights could be positive or negative.

For the one-modality DMS task, the input signals () contained three channels: two channels for the two sequential stimuli and one channel for the task cue signal (Fig 1a). For the two-modality DMS task, contained 6 channels, corresponding to two stimulus modalities (two channels for each modality), one channel for the task cue signal, and another one for the attention cue (Fig 1b). For the one-module RNNs, was projected to all the units in the network. However, for the two-module RNNs, the stimulus signals were injected into the first 200 units, while the task and attention cues were provided to the second module comprising 800 units (Fig 5a).

For the two-module RNN model, the recurrent connectivity weight matrix () was constrained to remove the feedforward inhibitory connections from the first module (first 200 units; sensory module) to the second module (nonsensory module) by applying a binary mask.

Training and task details

The dynamics in Equation 1 were discretized using the first-order Euler approximation method and using the step size () of 5 ms:

(4)

where . Given that the units in this network model communicate through differentiable and continuous signals, a gradient-descent supervised method, known as backpropagation through time (BPTT; [94]), was employed to train our RNNs to perform cognitive tasks. Specifically, the trainable parameters of the model included , , , and b. We used Adam (adaptive moment estimation) optimization algorithm to update these parameters. The learning rate was set to 0.01, and the TensorFlow default values were used for the rest of the parameters, including the first and second moment decay rates. To further impose biological constraints, we enforced Dale’s law (uniform neurotransmitter release characteristics within separate excitatory and inhibitory neurons) using methods similar to those implemented in previous studies [48,54,95]. To adhere to empirical findings regarding the ratio of excitatory and inhibitory neurons observed in the brain, each RNN consists of 80% excitatory and 20% inhibitory units (i.e., E–I ratio of 80/20; [30,96,97]).

Importantly, instead of fixing the synaptic decay constant () to a fixed value for all the units, we optimized the parameter for each unit. The parameter was trained to range from 20 ms to 125 ms to model heterogeneous synaptic dynamics of different receptors in the cortex [98,99]. We initialized the synaptic decay time constant parameter () using

(5)

where is the sigmoid function and refers to the standard normal Gaussian distribution. ms and ms were used to constrain the parameter to range from 20 ms to 125 ms.

Our model training was deemed successful if the following two criteria were satisfied within the first 70,000 epochs: 1) loss value (defined as the root mean squared error between the network output and target signals) < 7; and 2) task performance > 95% for the one-modality DMS task and > 85% for the two-modality DMS task. If the network did not meet the criteria within the first 70,000 epochs, the training was terminated. For each successfully trained RNN, we simulated 400 trials for the one-modality task and 1,000 trials for the two-modality task, randomizing each trial to contain one possible combination of instruction order, top-down cues, and stimulus identities. Using these criteria, we trained 20 RNNs for each task.

SVM decoding and state space analyses

We performed decoding on RNN activity with a 5-fold cross-validated linear support vector machine analysis (SVM; sklearn function SVC). Given a trained RNN model, we first generated firing rate timecourses ( in Equation 3) of 1,000 units across 1,000 simulated trials. Next, the data were randomly sampled for instruction order (early versus late), stimulus identity (+1 versus −1), and top-down task signals (pro-DMS versus anti-DMS; as well as attended versus unattended for the RNNs performing two-modality DMS). The SVM classifier was trained on the firing activities from either excitatory or inhibitory units and then tested on the left-out trials to distinguish each of the signals of interest.

The decoding plots represent single trial decoding accuracy for the signal of interest. All model units of a given kind (inhibitory or excitatory) are included simultaneously in the SVM decoder as features. The measures of variability (standard error of the mean) are included as shaded areas around the decoding accuracy trajectories, which are computed in cross-validated fashion across 5 folds.

To characterize the network dynamics in response to distinct external inputs, we performed a fixed-point analysis [67] using the FixedPointFinder toolbox [68]. This method involves numerically minimizing a proxy energy function derived from the RNN update equations to identify fixed points and subsequently obtaining linearized dynamics around these points to describe their behavior. In our case, this is equivalent to minimizing q(x), defined from the RNN dynamics of Equation 1:

(6)

We repeated this search for every set of external inputs u(t), which included null inputs for late instruction trials, task cue (pro-DMS versus anti-DMS), and attention cues (for the two-modality DMS).

For each optimization run, we randomly selected 200 out of the 1,000 simulated trials to create initial conditions for the optimizer. From each trial, one initial condition was generated by taking the 1,000-dimensional system state at the end of the first instruction period. Each optimization run had the following parameters: number of iterations = 30,000; initial learning rate = 0.1; outlier distance scale = 10; unique tolerance = 10. The outlier distance scale served as a cutoff distance for discarding outlier putative fixed points, measured from the centroid of the vector of initial conditions. The unique tolerance is the numerical precision used to determine if two fixed points are unique. Fixed points that were too close to previously identified ones were discarded. Two fixed points were considered unique if their distance exceeded this threshold.

To visualize the fixed points and RNN neural trajectories, we plotted them in principal component space spanned by the top three principal components extracted from the average activities during the first instruction period. In addition, we visualized the entire landscape of the energy function q(x) in the top two principal components during the first instruction period along with their respective energy minima for each external input u(t) condition (Figs 2a, 2b, 5b). For qualitative visualization purposes only, we created a grid of possible (x,y) coordinates in two-dimensional principal component space and performed an inverse PCA transform with sklearn function , which returns the unique orthogonal projection of the PC point onto native space. This function works by multiplying the PCA-space representation by the transpose of the PCA rotation matrix and adding the dataset mean. This implies that any variance that is perpendicular to the principal components is treated as zero and discarded. This is still lossy but is the best approximation possible for our visualization purposes.

CEBRA analysis

To visualize network dynamics and analyze the impact of top-down signaling on RNN dynamics, we applied unsupervised manifold discovery CEBRA [70] on the 1,000 simulated trials of representative RNN models during the first instruction and first stimulus periods. This method is capable of uncovering latent structures in multidimensional dynamical datasets, either unsupervised or conditioned on specific variables of interest. Throughout this study, we conducted unsupervised latent structure discovery based solely on the temporal sequence of data points across the entire dataset, without considering trial conditions, tasks, or stimulus identities. More specifically, we utilized the default model parameters for the ‘offset10-model’ architecture, using a batch size of 256, time offsets of 10, learning rates of , and cosine as the distance measure.

Cross-temporal discriminability analysis

In addition to the SVM decoding analysis (see above), we also quantified the amount of information encoded by each unit within RNNs using cross-temporal discriminability analysis [100–102]. For Figs 2j–2m, the firing rate estimate time courses of each unit for each first stimulus identity were initially divided into two splits (odd versus even trials) and averaged across trials within each split. Subsequently, for each unit in each split, the difference in the average firing rates between the two conditions based on the first stimulus identity (i.e., + 1 and −1) was computed. Next, Pearson’s correlation coefficient between the two splits was calculated. The discriminability score was then obtained by applying Fisher’s z-transformation. This process was carried out separately for the attended modality (Fig 2j, 2k) and the unattended modality (Fig 2l, 2m). For Fig 7f, the first stimulus identity was fixed to −1, and the discriminability of the task cue signal was computed using the two task cue identities (pro- versus anti-DMS). The diagonal values of the discriminability matrices are shown. For Fig 8d–8f, similar procedures to those employed for Fig 2j–2m were implemented.

Centrality measure

We utilized a functional closeness centrality measure to determine the centrality of each unit within RNNs. For this, we first defined a functional distance metric between unit i and j a network, based on the absolute of the Pearson correlation () between their synaptic current activity over time, denoted as and :

(7)

This function ensures that unit pairs that are more predictive of each other’s activity, either via positive or negative correlation, have a distance function closer to 0, while unit pairs with weak correlations have a distance function closer to 1.

Next, we used the complete distance matrix as the edge distances between nodes in a graph and employed the Python package networkx [103] to compute the network closeness centrality measure for each unit i from all n units:

(8)

Here, S(i,j) represents the shortest path distance between two units i and j, computed utilizing the distance matrix via Dijkstra’s algorithm with the distance attributes as edge weights. Concretely, a node has high centrality when it is, on average, only a few steps away along shortest paths. Intuitively, a neuron is a good vantage point or information broadcast hub if communication follows shortest paths. We computed centrality for all neurons within the entire network and plotted it separately for excitatory versus inhibitory neurons.

Identifying task and attention cue selective units

To identify units selective for pro-DMS versus anti-DMS task signals, we first generated 400 trials for the one-modality DMS and 1,000 trials for the two-modality DMS. We next extracted firing rate estimates of all units in each of the trained RNNs. Then, the trials and firing rate estimates were sorted based on the early task cue signal, separating them into pro-DMS and anti-DMS groups. For each unit, we averaged the firing rate estimates during the first instruction window for each trial. Subsequently, we performed the one-sided Wilcoxon rank-sum statistical test on the average firing rates from the two task groups for each unit. Units with significantly higher average firing rates in the pro-DMS group compared to the anti-DMS group were classified as pro-DMS selective units. Similar procedures were performed to identify attention-selective units for the two-modality RNNs.

Lesioning analyses

For Figs 4 and 6, trained synaptic connections were lesioned systematically to identify specific connections important for encoding and maintenance of top-down information. For each trained network, connections were initially classified into four subgroups: (inhibitory units inhibiting other inhibitory units), (inhibitory units inhibiting excitatory units), (excitatory units exciting inhibitory units), and (excitatory units exciting other excitatory units). Within each subgroup, connections were further categorized based on their selectivity for the task and attention cue signal (see above). The selective connections were then lesioned by reducing their strength by 50%.

Experimental data

To further characterize the impact of top-down signaling over information processing, we utilized publicly available experimental data on the role of recurrent processing in the activity of mouse primary visual cortex [61]. In this task, mice had to determine the location of a salient grating visual pattern which contrasted from a visual background, distinguishing them based either on the orientation, the phase, or the texture of the grating pattern. For the purposes of our analyses, we averaged results across orientation, phase, and texture conditions. Target patterns could appear either in the right or left visual field, prompting mice to indicate their locations by licking a spout on the corresponding side to receive water or milk droplets as rewards. Crucially, neural recordings and stimuli were set up in such a way that image contents falling within the receptive fields of the recorded neurons were strictly identical in the figure and background conditions. Thus, any difference in neural activity measured across conditions was caused by the context indicated by image regions outside of the measured neural receptive field.

We first sought to determine whether V1 activity contained information distinguishing target grated stimuli from nontarget grated backgrounds even when the sensory input within receptive fields was identical. This would indicate that feedback from regions outside receptive fields projects into the measured V1 sites. For this analysis, neural activity consisted of multiunit activity from 198 recording sites in 13 different V1 electrode penetrations in 6 awake mice. To determine which information was contained in V1 multiunit activity, we performed a 4-fold cross-validated linear SVM decoding analysis (MATLAB function fitcsvm), for figure versus background labels from 0.5 s before to 1 s after stimulus presentation, utilizing activity in separate channels as features. To obtain a null distribution for decoding accuracy, we generated 1,000 random label permutations and performed the same decoding analysis.

Furthermore, we leveraged an additional analysis using the aforementioned dataset. Specifically, we determined whether activity of inhibitory neurons in V1 is necessary for figure versus background discrimination, as predicted by the central role of inhibitory neurons in our modeling approach. The analysis contained multiunit activities of different subgroups of inhibitory interneurons in V1: vasoactive intestinal peptide–expressing (VIP) and somatostatin-positive interneurons (SST). In addition, the dataset also contains recordings when VIP or SST neurons were optogenetically inhibited (see [61] for more details).

Using this dataset, we decoded figure versus background labels from V1 multiunit activity with a 5-fold cross-validated linear SVM, both with and without optogenetic VIP suppression (time window length: 250 ms, time steps: 25 ms). In the original data manuscript, authors tested 3 different figure-ground segregation schemes for visual gratings: orientation, phase, and texture. Behaviorally, authors found that orientation elicited the highest accuracy rates. Optogenetically, orientation was the one condition that elicited the highest differences in spike rates comparing laser on versus laser off conditions. Therefore, for our purposes of demonstrating the impact of optogenetic silencing on SVM decoding, we sub-selected only trials which had orientation as a top-down signaling cue and performed 5-fold cross-validation. Other than that, no further sub-selection was performed. We then compared decoding accuracy between the laser-on and laser-off conditions by performing a two-sample one-sided t test for whether accuracy was larger in trials without laser suppression. We performed this test within the main time window of interest, when the laser was delivered (250 ms to 500 ms).

In parallel to the neural data analysis, we applied an analogous linear SVM decoding framework to activity from the sensory module of the trained two-module RNNs. For each model, task cue identity (pro versus anti) was decoded from population activity using a time-resolved linear SVM. Neural activity was first averaged within sliding temporal windows (window size: 50 ms; step size: 25 ms) to generate feature vectors for each trial and time window. Classifiers were trained using a random 70% subset of trials and evaluated on the remaining 30% held-out trials, and this procedure was repeated across 100 random train-test splits. Decoding accuracy was computed independently for each neuron and time window, then averaged across neurons within the sensory module and across trained models.

Statistical analyses

All RNNs trained in the present study were randomly initialized (with random seeds) before training. Nonparametric statistical methods were employed throughout the study. In figures containing boxplots, two-sided Wilcoxon rank-sum or signed-rank tests were performed to ascertain statistically significant differences between two groups.

Following linear SVM decoding and functional centrality analyses, we aimed to determine whether decoding accuracy or functional centrality values were higher in excitatory or inhibitory neurons. To achieve this, we utilized the scipy.stats function ks_2samp and performed two-sided Kolmogorov–Smirnov (KS) tests across all time bins and Bonferroni corrected for the number of tested time bins. Similarly, we performed a one-sided t test in the time bin of interest to specifically test the hypothesis that decoding accuracies in V1 would decrease following optogenetic silencing.

For cross-temporal discriminability analyses, we performed binomial tests and Bonferroni corrected for the number of tests to determine whether each decoding accuracy value was higher than chance. Assuming a false positive rate of 5%, and a Bernoulli process to generate significant decoding events at this rate, the number of instances S falsely classified as significant within a population of size N is given by a binomial distribution: . Accordingly, we derived a binomial p-value for the probability of obtaining an observed sensitive instance count K larger than expected by chance, using the cumulative binomial distribution: .

S1 Fig. Example single-unit activity and output dynamics from a trained one-modality RNN.

a. Neuron-by-time heatmaps showing activity of all 1,000 units from a representative single-module, one-modality RNN during Pro-DMS (left) and Anti-DMS (right) trials. For visualization purposes only, each neuron’s activity was z-scored across time. White vertical lines demarcate key task epochs, including instruction, stimulus, delay, and response periods. b. Corresponding RNN output signals for Pro-DMS (teal) and Anti-DMS (purple) trials. Gray shaded regions indicate the presentation of stimulus 1 (Stim 1) and stimulus 2 (Stim 2). The output reflects the learned stimulus–response mapping conditioned on the task cue. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s001

(EPS)

S2 Fig. Quantification of latent trajectory separation for attended versus unattended modalities.

Boxplots show the Euclidean distance between CEBRA latent trajectories corresponding to opposite first-stimulus identities, computed during the stimulus 1 epoch. Distances were calculated separately for attended and unattended modalities by randomly subsampling latent trajectories across task conditions and repeating the procedure 100 times. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s002

(EPS)

S4 Fig. Temporal encoding of top-down signals during DMS tasks.

Cross-temporal decodability of task and attention signals is shown as the average generalization accuracy for each pair of training/testing time windows. Highlighted boundaries indicate significant decoding accuracy, , binomial test, Bonferroni corrected. White lines denote the reference event of each trial period (task and attention cues, stimulus onsets, response). a. One-modality DMS, task signal, early instruction. b. Same, for late instruction. c. Two-modality DMS, task signal, early instruction. d. Same, for late instruction. e. Two-modality DMS, attention signal, early instruction f. Same, for late instruction. g. Two-modality DMS, attended modality signal, early instruction, attended modality. h. Same, for unattended modality. i. One-modality DMS, stimulus identity. j. Two-modality DMS, stimulus identity, stimulus modality 1. k. Same, for stimulus modality 2. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s004

(EPS)

S5 Fig. Inhibitory units robustly maintain top-down signals.

a. SVM decodability of the attention signal from inhibitory (blue) and excitatory (red) units from the two-modality RNNs. b. SVM decodability of the attended first stimulus identity from inhibitory (blue) and excitatory (red) units from the two-modality RNNs. c. SVM decodability of the unattended first stimulus identity from inhibitory (blue) and excitatory (red) units from the two-modality RNNs. Black bars indicate significant differences between excitatory and inhibitory units (P < 0.05, two-sided two-sample KS test). Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s005

(EPS)

S6 Fig. Structured connectivity across rule-selective populations.

Average synaptic weights (magnitude) between rule-selective populations across all trained one-module, one-modality RNNs (n = 20), separated into within-selectivity (“Same”; anti → anti and pro → pro) and across-selectivity (“Cross”; anti → pro and pro → anti) connections. Panels show results for inhibitory-to-inhibitory (I → I), excitatory-to-inhibitory (E → I), inhibitory-to-excitatory (I → E), and excitatory-to-excitatory (E → E) connections. Each gray line represents an individual network, and black markers denote mean ± s.e.m. ***P < 0.001 by paired Wilcoxon signed-rank test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s006

(EPS)

S7 Fig. Quantification of outgoing synaptic connections in single-module one-modality RNNs.

Boxplots show the total number of outgoing synaptic connections for task-selective excitatory and inhibitory neuron populations across 20 single-module RNNs trained to perform the one-modality DMS task. a. Outgoing connections from pro-DMS-selective populations; b. Outgoing connections from anti-DMS-selective populations. For each condition, outgoing connections are grouped by presynaptic and postsynaptic cell type (, , , ). Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. Statistical comparisons were performed using a Kruskal–Wallis test followed by Dunn’s post hoc multiple-comparisons test. Significance levels are indicated as **P < 0.01 and ***P < 0.001. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s007

(EPS)

S8 Fig. Inhibitory-to-inhibitory connections primarily encode task and attention cue signals.

a. Performance of all 20 RNNs trained to perform the two-modality DMS when task-specific (top) or (bottom) synaptic weights were reduced by 50% when modality 2 was attended. b. Performance of all 20 RNNs trained to perform the two-modality DMS when modality-specific (top) or (bottom) synaptic weights were reduced by 50% when the task signal was fixed to anti-DMS. , inhibitory units preferring anti-DMS cue; , inhibitory units preferring pro-DMS cue; , excitatory units preferring anti-DMS cue; , excitatory units preferring pro-DMS cue; , inhibitory units preferring modality-1 attention cue; , inhibitory units preferring modality-2 attention cue; , excitatory units preferring modality-1 attention cue; , excitatory units preferring modality-2 attention cue. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. *P < 0.05, **P < 0.01, ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s008

(EPS)

S9 Fig. Perturbational analysis of excitatory and inhibitory synaptic connections in one-modality RNNs.

Task performance following selective lesioning of synaptic connections based on presynaptic/postsynaptic cell type and task selectivity. Rows correspond to different connection classes (, , , ), and columns show lesions applied to selectivity-matched or crossed populations (anti versus pro). Lesions were implemented by reducing the corresponding synaptic weights by 50%, and performance is shown separately for anti and pro trials. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. **P < 0.01 and ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s009

(EPS)

S10 Fig. Perturbational analysis of excitatory and inhibitory synaptic connections in two-modality RNNs when modality 1 is attended.

Task performance following selective lesioning of synaptic connections in two-modality RNNs, with the attention cue fixed to modality 1. Figure organization follows the same layout and conventions as S9 Fig. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. *P < 0.05, **P < 0.01, ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s010

(EPS)

S11 Fig. Perturbational analysis of excitatory and inhibitory synaptic connections in two-modality RNNs when modality 2 is attended.

Task performance following selective lesioning of synaptic connections in two-modality RNNs, with the attention cue fixed to modality 2. Figure organization follows the same layout and conventions as S9 Fig. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. *P < 0.05, **P < 0.01, ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s011

(EPS)

S12 Fig. Perturbational analysis of excitatory and inhibitory synaptic connections in two-modality RNNs with the task signal fixed to the pro-DMS condition.

Task performance following selective lesioning of synaptic connections in two-modality RNNs, with the task cue fixed to the pro-DMS condition and performance compared between trials in which modality 1 or modality 2 was attended. Figure organization follows the same layout and conventions as S9 Fig. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. *P < 0.05, **P < 0.01, ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s012

(EPS)

S13 Fig. Perturbational analysis of excitatory and inhibitory synaptic connections in two-modality RNNs with the task signal fixed to the anti-DMS condition.

Task performance following selective lesioning of synaptic connections in two-modality RNNs, with the task cue fixed to the anti-DMS condition and performance compared between trials in which modality 1 or modality 2 was attended. Figure organization follows the same layout and conventions as S9 Fig. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. *P < 0.05, **P < 0.01, ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s013

(EPS)

S14 Fig. Example single-unit activity and output dynamics from a trained two-module, one-modality RNN.

a. Neuron-by-time heatmaps showing activity of all 1,000 units from a representative two-module, one-modality RNN during Pro-DMS early (left) and late (right) instruction trials. For visualization purposes only, each neuron’s activity was z-scored across time. White vertical lines demarcate key task epochs, including instruction, stimulus, delay, and response periods. b. Corresponding RNN output signals for Pro-DMS early (red) and late (blue) instruction trials. Gray shaded regions indicate the presentation of stimulus 1 (Stim 1) and stimulus 2 (Stim 2). The output reflects the learned stimulus–response mapping conditioned on the task cue. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s014

(EPS)

S15 Fig. Training two-module RNNs takes longer for the one-modality DMS compared to the one-module RNNs.

a. Comparison of the number of training trials required for training between the one-module and two-module RNNs, with 20 RNNs trained for each case. b. Both one-module and two-module RNNs achieved high accuracy in performing the one-modality DMS task (mean±stdev; 97.7±1.3% for the one-module model; 97.6±1.2% for the two-module model). Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s015

(EPS)

S16 Fig. Comparison between and connection strength in two-module RNNs.

(a) and (b) connections were significantly stronger within the non-sensory module than the sensory module. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s016

(EPS)

S17 Fig. Decoding task signals in nonsensory module of two-module RNNs.

a. Average decoding accuracy in early instruction trials for pro- versus anti-DMS signal in the nonsensory module of two-module networks, including either excitatory (red) or inhibitory (blue) units. Shaded areas indicate standard error of the mean. Black bars indicate significant differences between excitatory and inhibitory units (P < 0.05, two-sided KS test). b. Same, for late instruction trials. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s017

(EPS)

S18 Fig. Perturbational analysis of excitatory and inhibitory synaptic connections within the sensory module in two-module RNNs.

Task performance following selective lesioning of synaptic connections within the sensory module of two-module RNNs trained to perform the one-modality task. Rows correspond to different connection classes (, , , ), and columns show lesions applied to selectivity-matched or crossed populations (anti versus pro). Lesions were implemented by reducing the corresponding synaptic weights by 50%, and performance is shown separately for anti and pro trials. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. *P < 0.05 and **P < 0.01 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s018

(EPS)

S19 Fig. Perturbational analysis of excitatory and inhibitory synaptic connections from non-sensory module to sensory module in two-module RNNs.

Task performance following selective lesioning of synaptic connections from the nonsensory module to the sensory module of two-module RNNs trained to perform the one-modality task. Rows correspond to different connection classes (, , , ), and columns show lesions applied to selectivity-matched or crossed populations (anti versus pro). Lesions were implemented by reducing the corresponding synaptic weights by 50%, and performance is shown separately for anti and pro trials. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. *P < 0.05 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s019

(EPS)

S20 Fig. Perturbational analysis of excitatory and inhibitory synaptic connections from sensory module to non-sensory module in two-module RNNs.

Task performance following selective lesioning of synaptic connections from the sensory module to the nonsensory module of two-module RNNs trained to perform the one-modality task. Rows correspond to different connection classes (, , , ), and columns show lesions applied to selectivity-matched or crossed populations (anti versus pro). Lesions were implemented by reducing the corresponding synaptic weights by 50%, and performance is shown separately for anti and pro trials. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. **P < 0.01 and ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s020

(EPS)

S21 Fig. Perturbational analysis of excitatory and inhibitory synaptic connections within non-sensory module in two-module RNNs.

Task performance following selective lesioning of synaptic connections within the nonsensory module of two-module RNNs trained to perform the one-modality task. Rows correspond to different connection classes (, , , ), and columns show lesions applied to selectivity-matched or crossed populations (anti versus pro). Lesions were implemented by reducing the corresponding synaptic weights by 50%, and performance is shown separately for anti and pro trials. Boxplot: central lines, median; bottom and top edges, lower and upper quartiles; whiskers, 1.5 × interquartile range; outliers are not plotted. ***P < 0.001 by two-sided Wilcoxon rank-sum test. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s021

(EPS)

S22 Fig. Effects of synaptic lesioning on unattended stimulus representations in two-module RNNs.

Average cross-temporal discriminability of the first stimulus identity of the unattended modality in two-module RNNs trained to perform the two-modality DMS task. Discriminability matrices are shown for trials in which the attention cue was presented early (left), late (middle), or as a retro-cue (right). Rows correspond to selective lesioning (50% reduction in synaptic weights) of different connection types: a. Inhibitory-to-inhibitory (). b. Inhibitory-to-excitatory (). c. Excitatory-to-inhibitory (). d. Excitatory-to-excitatory () connections. White lines denote task epoch boundaries. Results are averaged across 20 trained RNNs. Source data can be found in Zenodo at https://doi.org/10.5281/zenodo.20148306.

https://doi.org/10.1371/journal.pbio.3003831.s022

(EPS)