Communication between the medial frontal cortex and mediodorsal thalamus is required for timing performance in rats

Abstract

Predicting when future events will occur is critical for adaptive behavior, yet the brain networks underlying this ability remain unclear. The prefrontal cortex (PFC) and its rodent analogue, the prelimbic cortex (PL), are known to support the ability to track time and execute decisions in anticipation of upcoming events. Given this, we can expand our understanding of the circuits involved in time-based behavior by looking for regions that interact with the PFC during timing tasks. The mediodorsal nucleus of the thalamus (MD) is a strong candidate, as it shares dense reciprocal connections with the PFC and is widely implicated in higher-order cognition. However, data implicating the.

MD itself in timing is still minimal, and to our knowledge, causal data directly implicating MD-PFC communication in timing is lacking entirely. To address this, we trained nineteen male Long Evans rats on a dual interval timing task requiring decision-making and tested the effects of reversibly inactivating the MD/PL individually or blocking communication between them using GABA-A receptor agonist, muscimol. All manipulations produced a similar deficit—flattening the characteristic Gaussian-shaped reponse curves while leaving overall response rates intact. Single-trial analyses revealed that this was not simply a broadening of temporally-controlled response bursts, but rather a fundatment breakedown in the canonical burst structure of timing behavior, with responses becoming more uniflormly distrubted across the trial. These findings further implicate the MD and PL in timing and, to our knowledge, provide the first causal evidence that communication between the MD and PFC is specifically required. More broadly, these findings add to growing evidence that MD-PFC interactions support higher-order cognition. Timing impairments are common in a variety of neuropsychiatric diseases including schizophrenia, Parkinson’s disease, autism, and bipolar disorder so this work provides key implications for the involvement of a novel MD-PFC pathway.

1. Introduction

Time is fundamental to cognition. When making a decision, we must not only select the right action but also execute it at the right moment (Fan et al., 2012). When forming a memory, we encode not only what happened and where, but also when the event occurred (Tulving, 1986). Even our understanding of causality depends on the ability to recognize that causes precede their respective effects (Hume, 1739). Given its broad relevance, time perception is a topic of interest across philosophy (Mctaggart, 1908), psychology (Fraisse, 1964), neuroscience (Paton & Buonomano, 2018), and clinical efforts to treat cognitive impairments in neurological conditions (Parker, 2015). Our interest is in how the brain represents time, where—despite its importance—we still know surprisingly little.

One major challenge is that timing appears to depend on a complex network of brain regions (for review see Mondok & Wiener, 2023). For example, timing has been studied extensively in association cortices, the basal ganglia, and midbrain dopamine centers (Jazayeri & Shadlen, 2015; Mello et al., 2015; Soares et al., 2016). However, many other areas have been implicated, such as the hippocampus (Gupta et al., 2019; Shikano et al., 2021; De Corte et al., 2022), cerebellum (Narain et al., 2018), and even the primary sensory cortices (Shuler, 2006). Explaining how these areas work together to mediate timing is a difficult endeavor.

This challenge is compounded by the diversity of time-related signals seen among neurons in most of these areas while individuals track time (for reviews see Balcı & Simen, 2024; De Corte, Akdoğan, et al., 2022). For example, ‘ramping’ neurons that monotonically change in firing rate as a learned time-interval elapses have been well documented (Chen et al., 2017; Jazayeri & Shadlen, 2015; Leon & Shadlen, 2003; Simen et al., 2011). However, ‘time-cells’ that show consistent Gaussian-like firing-profiles around specific time intervals have also been observed (Mau et al., 2018; Mello et al., 2015; Sawatani et al., 2023; Zhou et al., 2020). Furthermore, there is even indication that some neural populations encode time through idiosyncratic, non-linear firing profiles (Wang et al., 2018; Buonomano & Mauk, 1994; Buonomano & Maass, 2009).

Converging evidence suggests that the prefrontal cortex (PFC), including its rodent analog the prelimbic cortex (PL) is necessary for timing intervals in the seconds-to-minutes range—referred to as ‘interval timing’ (Matell & Meck, 2000). Disruption of the human PFC or PL, consistently produces impairments in a variety of timing tasks (Buhusi et al., 2018; Koch et al., 2003; Wang et al., 2018), and all forms of time-related signaling noted above have been observed in this area (Matell & Meck, 2004; Schonhaut et al., 2023; Wang et al., 2018). Given the established role of the PFC, a principled strategy for uncovering the networks that support timing is to search for areas that interact with the PFC during time-based behavior. Past work focused on studying the role of frontostriatal circuits during timing tasks (Buhusi & Meck, 2005; Matell & Meck, 2004) with data suggesting that the PFC encodes the passage of time while the striatum uses this information to appropriately time decisions (Matell et al., 2003; Mello et al., 2015). More recent data indicates that the frontal cortex and thalamus functionally interact to facilitate timing performance (Parker et al., 2017; Wang et al., 2018) yet, the precise thalamic nuclei contribution and interaction with the PFC remain to be established.

The mediodorsal nucleus of the thalamus (MD) is a strong candidate for thalamic involvement in timing. The MD and PFC share dense, reciprocal connections across species, including the rodent PL (Alcaraz et al., 2016; Georgescu et al., 2020; Ray & Price, 1993). Furthermore, MD-PFC interactions are heavily implicated in a variety of cognitive functions, including executive function and working memory (Bolkan et al., 2017; Golden et al., 2016; Markowitsch, 1982; Parnaudeau et al., 2018; Peräkylä et al., 2017). Importantly, the MD itself has been shown to be necessary for timing (Lusk et al., 2020; Yu et al., 2010).

However, critical questions remain unresolved. Although the MD and PFC have each been implicated in timing, most prior evidence regarding their functional interactions has been correlative, typically monitoring changes in task-related activity in on area after manipulating another (Alexander & Fuster, 1973; Parnaudeau et al., 2013). While recent causal studies have implicated MD-PFC communication in cognitive processes such as context integration (Marton et al., 2018) and task switching (Mukherjee et al., 2021), whether MD-PL communication is specifically required for timing behavior has not been established.

We addressed these gaps in the current study using the ‘peak-interval procedure’, an established time-based decision making task (Roberts, 1981). During this task, the onset of a cue instructs subjects to make a response after a specific duration elapses (e.g., tone-8 s). A trained rat’s responses will cluster around the cue’s respective interval, providing a simple assay of how decision making is guided by time. We trained rats on this task and then used reversible inactivations to test two specific hypotheses: first that inactivating the MD or PL individually would impair timing performance, and second that blocking MD-PL communication specifically would also produce a deficit. For the latter, we employed a functional disconnection design. This approach allowed us to not only further test the necessity of the MD and PL in timing individually, but also to provide the first causal test of whether MD-PL communication is specifically required for timing behavior.

2. Results

2.1. The MD and PL are necessary for timing performance

As outlined in Fig. 1, we first assessed whether the mediodorsal nucleus of the thalamus (MD) and prelimbic cortex (PL) are individually necessary for timing performance and, using a within-subjects design, asked whether any inactivation deficits would differ across areas. Specifically, we trained nine rats on a timing task known as the ‘peak-interval’ procedure (Fig. 1A; Roberts, 1981; Swanton & Matell, 2011; De Corte et al., 2018). During this task, trials start with the presentation of a cue that signals reward can be earned for responding after a specific interval of time has elapsed (e.g., tone-8 s). Once reward is earned, the cue terminates, and the next trial begins after an inter-trial interval. In contrast, during probe trials, no reward is delivered, and the cue remains on for much longer than normal (e.g., 96–128 s), terminating regardless of whether rats respond. During these trials, rats’ responses typically cluster around the trained interval, leading mean response rates over time to resemble a Gaussian curve centered over the cue’s interval. In this experiment, we trained rats to associate two cues with distinct time intervals (e.g., tone-8 s/light-16 s), presenting each cue individually during separate trials. Once trained, we implanted cannulae bilaterally into the MD and PL (Fig. 1C). We then tested the effects of reversibly inactivating either area during separate sessions using muscimol (Fig. 1B).

Fig. 1. Task summary and histological analysis for within-subject bilateral inactivations of the MD and PL.

(A) Overview of the peak-interval procedure, including operant chamber layout (houselight, tone generator, nosepoke, and reward port) and trial structure. During rewarded-trials, the cue terminates and reward is delivered for the first response at or after the current cue’s interval (e.g., tone = 8 s, light = 16 s). During probe trials, reward is omitted and the cue is left on for an extended duration (96–128 s), terminating independent if responding. Bottom: conceptual illustration of typical behavior during probe trials, with average response curves at the top and single-trial burst patterns below. (B) Infusion conditions for within-subject design: saline (left), PL inactivation (middle), and MD inactivation (right). (C) Cannula tip locations in MD and PL shown on atlas diagrams (left) and representative histological sections (right).

After saline infusions, mean response rates across time during probe trials showed typical Gaussian-like curves, peaking near each cue’s respective interval (Fig. 2A). However, both MD and PL inactivations flattened these curves, with responses being less common around each cue’s interval and more common at irrelevant times throughout the trial (Fig. 2A).

Fig. 2. Inactivating the prelimbic cortex (PL) or mediodorsal thalamus (MD) bilaterally flattens mean response curves (n = 9).

(A) Mean response rate across time during probe trials (normalized by sum) for the 8 s and 16 s cues (top and bottom, respectively), split by infusion type. Data were grouped into 1 s bins and have been smoothed over a 3-bin window (for visualization only). (B) Percent change in the fit-quality of a gaussian distribution to response rates across time and overall mean response rates (i.e., ignoring time) during probe trials for each area, relative to saline sessions (top and bottom, respectively). In all cases, bars represent mean +/− SEM of the data, and asterisks indicate significance at p < 0.05, relative to saline sessions.

Consistent with this, the fit-quality of a Gaussian distribution to these response curves dropped following MD and PL inactivations [Fig. 2B; Infusion: F(2,14) = 15.51, p < 0.001; MD: t(6) = −4.50, p < 0.005; PL: t(6) = −4.79, p < 0.005; R² M +/− SEM: Sal = 0.90+/− 0.03, MD = 0.48 +/− 0.08, PL = 0.74 +/− 0.03]. Importantly, inactivations did not reliably change overall mean response rates during probe trials [Fig. 2B; Infusion: F(2,14) = 0.24, p = 0.794]. Together, these data suggest that rats were unable to time their responses appropriately following MD and PL inactivations, with no detectable impact on their motivation or motor function. Given the drop in fit-quality, the validity of measures derived from the fits is questionable (e.g., peak-time, spread). Nonetheless, for exploratory purposes, we provide standard analyses in the supplemental results for all experiments covered here (Fig. S1).

While these data show a clear performance-deficit, there are multiple explanations for why the response curves would appear flat. Specifically, unlike the averaged response curves, rats do not produce a Gaussian-shaped response pattern during individual probe trials. Rather, rats emit a discrete ‘burst’ of responses near the cue’s interval, with a sharp onset/offset (Church et al., 1994). The timing of the burst varies slightly from trial-to-trial. Consequently, when averaged across trials, mean response rates show a smooth, Gaussian shape (see Fig. 1A for visual illustration). Flatter mean response curves could result from either 1) disruption of the burst-pattern itself or 2) increased variability in the timing of the burst.

To distinguish between these possibilities, we analyzed single-trial responding next. Fig. 3A shows heat maps of responses across time for all probe trials collected in the experiment. We quantified the burst using the standard approach, in which a step function is fit to single-trial response data (see methods; Church et al., 1994). Importantly, even when response times are random, burst-like patterns will still emerge by chance (i.e., burst-finding algorithms will, of course, always find bursts, even in random data). To illustrate, we repeated the analysis on time-shuffled data from saline sessions and present the resulting heat map for comparison. The key question is whether any bursts found in the MD/PL inactivation data reflect meaningful structure in timed behavior or arise coincidentally from random variation.

Fig. 3. Inactivating the MD or PL degrades single-trial patterns of timed responding (n = 9).

(A) Heat maps showing response patterns during individual probe trials, sorted by stop-time and binned in 1-s intervals. Data are shown separately by infusion type. Shuffled saline data (second column) served as a control to illustrate expected patterns if no burst-pattern occurred during individual trials. (B) Percent change in fit-quality (R²) of a step function fit to individual trials for each inactivation type, relative to saline. (C) Accuracy of a support vector machine trained to discriminate shuffled data from unshuffled data, split by infusion type. In all cases, bars represent mean +/− SEM of the data, and asterisks denote statistical significance (p < 0.05).

During saline sessions, responses show a sharp distinction between the pre/post burst periods, clustering around the expected times (Fig. 3A). The contrast is less clear in the shuffled data, as bursts were, by definition, detected randomly across time. Notably, a similar disruption is also apparent in the MD and PL inactivation data: the primary burst was less pronounced, and rats showed more sporadic responses throughout the trial and/or secondary bursts outside of the expected window.

Consistent with this, the fit-quality of a step-function to individual trial responding dropped following both MD and PL inactivations [Fig. 3B; Infusion: F(2,14) = 13.34, p < 0.005; MD: t(6) = −4.38, p < 0.005; PL: t(6) = −2.51, p < 0.05; R² M +/− SEM: Sal = 0.72 +/− 0.03, MD = 0.46 +/− 0.03, PL = 0.61 +/− 0.03]. As a complementary test, we trained a machine learning algorithm (support vector machine) to discriminate shuffled and unshuffled single-trial data for each infusion type. This provides a ‘data-driven’ approach to evaluating the organization of behavior, casting few prior assumptions regarding how the data should be organized. As expected, accuracy was high for saline data and dropped following MD and PL inactivations [Fig. 3C; Infusion: F(2,14) = 10.16, p < 0.005; MD: t(6) = −3.77, p < 0.01; PL: t(6) = 3.22, p < 0.05; M +/− SEM: Sal = 0.73 +/− 0.03, MD = 0.54 +/− 0.02, PL = 0.61 +/− 0.03]. While accuracy under MD and PL inactivations began to approach chance levels, all infusion types were either significantly above chance or marginally so in the case of the MD [Sal: t(8) = 7.09, p < 0.001, MD: t(8) = 2.12, p = 0.067, PL: t(8) = 3.65, p < 0.01]. However, as a whole, these data indicate that the MD and PL are necessary for using time as a framework to efficiently organize goal-directed behavior.

2.2. Communication between the MD and PL is necessary for timing performance

The MD and PL share dense, reciprocal connections and often co-project to areas known to be involved in timing, such as the striatum. Our prior results clearly show that both areas contribute to timing behavior. However, whether these structures facilitate timing through independent or overlapping circuits is unclear. To address this, we next asked whether communication between the MD and PL is specifically required for using time to efficiently organize goal-directed behavior.

We used the same procedure as in the prior experiment, training a new group of rats (n = 10) and implanting bilateral cannulae to target the MD and PL. However, instead of targeting each region individually, we used a ‘disconnection’ procedure (Fig. 4A; Barker et al., 2007). In this design, inactivations are always unilateral. However, we administered either ipsilateral or ‘crossed’ muscimol infusions (e.g., Ipsilateral: right-PL + right-MD; Crossed: left-PL + right-MD).

Fig. 4. Task summary and histological analysis for crossed and ipsilateral inactivations of the MD and PL.

(A) Schematic of the disconnection protocol showing three infusion conditions: saline, crossed, and ipsilateral. Check marks and Xs indicate preserved or disrupted MD–PL communication in each hemisphere, respectively. Only one spatial permutation is illustrated for each disconnection condition; all counterbalanced permutations were used across sessions (see Methods). (B) Diagram of infusion cannula tip locations for all rats in the experiment for the MD (top) and PL (bottom).

As illustrated in Fig. 4A, ipsilateral inactivations should have weak effects, as the two areas can still communicate in the contralateral hemisphere. By contrast, crossed inactivations will disrupt any MD-PL communication bilaterally. Therefore, if MD-PL communication facilitates timing, crossed infusions should yield a larger deficit than ipsilateral ones. However, if they act through fully independent circuits, unilateral communication will always be preserved, producing equivalent ipsilateral/crossed inactivation effects. As such, this design provides a selective means of disrupting mono- or poly- synaptic MD-PL interactions.

Fig. 5 illustrates the mean response-curve data from the experiment. As expected, saline infusions had no apparent impact on Gaussian-like shape for either cue. Similarly, ipsilateral inactivations produced no noticeable disruption to this pattern. Critically, crossed infusions clearly appeared to flatten the curves, closely mirroring the effect observed following bilateral inactivations from the prior experiment.

Fig. 5. Disconnecting the MD from the PL flattens mean response curves (n = 10).

(A) Mean response density across time during probe trials for the 8 s and 16 s cues (top and bottom, respectively), separated by infusion condition (Saline, Ipsilateral, Crossed). Data were grouped into 1 s bins and have been smoothed over a 3bin window (for visualization only). (B) Top: Percent change relative to saline in the fit-quality (R²) of a Gaussian distribution to response curves following crossed and ipsilateral infusions. Bottom: Same as top, but data come from overall response rates across probe trials (collapsed across time). Asterisks denote significant differences (p < 0.05).

Consistent with these impressions, only crossed infusions produced a drop in Gaussian fit-quality [Fig. 5B; Infusion: F(2,16) = 15.53, p < 0.005; Sal vs. Cross: t(7) = −3.57, p < 0.05; Cross vs. Ipsi: t(7) = −4.42, p < 0.005; Sal vs. Ipsi: t(7) = 2.22, p = 0.057; R² M +/− SEM: Sal = 0.94 +/− 0.01, Crossed = 0.61 +/− 0.07, Ipsi = 0.95 +/− 0.01]. Furthermore, none of the inactivations reliably affected overall trial response rates [Infusion: F(2,16) = 0.613, p = 0.487].

As noted above, analyzing single-trial data is essential to fully characterize the deficits underlying the flattened response curves caused by crossed infusions. Fig. 6 illustrates these data. As with the bilateral inactivations, crossed infusions appeared to selectively disrupt the burst pattern of responding, producing less structured and more sporadic response timing (Fig. 6A).

Fig. 6. Disconnecting the mediodorsal thalamus (MD) from the prelimbic cortex (PL) disrupts the temporal organization of single-trial behavior (n = 10).

(A) Heat maps showing responses during individual probe trials, sorted by the best identified stop-time and binned in 1-s intervals. Data are shown for each infusion condition. Shuffled saline data are included to illustrate expected data if no burst-pattern were present. (B) Percent change in fit-quality of a step function fit to individual trials for each inactivation type, relative to saline. (C) Accuracy of a support vector machine trained to discriminate shuffled data from unshuffled data, split by infusion type.

Consistent with this, only crossed infusions attenuated the fit-quality of a step function to individual trials [Fig. 6B; Infusion: F(2,16) = 24.15, p < 0.001; Sal vs. Cross: t(7) = −4.75, p < 0.005; Ipsi vs. Cross: t(7) = −6.65, p < 0.001; Sal vs. Ipsi: t(7) = −1.11, p = 0.300; R² M +/− SEM: Sal = 0.75 +/− 0.02, Cross = 0.53 +/− 0.03, Ipsi = 0.73 +/− 0.01]. Regardless of infusion-type, overall fit quality was higher for the 8 s cue [Cue: F(1,8) = 69.48, p < 0.001]. However, crossed infusions attenuated this difference, possibly indicating a floor-effect [Cue X Infusion: F(2,16) = 6.38, p < 0.01; 8 s vs. 16 s cue: Sal: t(8) = 8.82, p < 0.001; Cross: t(8) = 2.30, p = 0.05; Ipsi: t(8) = 8.15, p < 0.001].

We found similar results when using a support vector machine trained to discriminate between shuffled and unshuffled data. For example, classification accuracy dropped for both cues following crossed infusions, relative to both saline and ipsilateral infusions [Fig. 6C; Infusion: F(2,16) = 47.70, p < 0.001; Sal vs. Cross: t(7) = −8.29, p < 0.001; Ipsi vs. Cross: t(7) = 7.88, p < 0.001; Sal vs. Ipsi: t(7) = 2.30, p = 0.05; M +/− SEM: Sal = 0.79 +/− 0.03, Cross = 0.58 +/− 0.02, Ipsi = 0.74 +/− 0.02]. In all cases, overall accuracy was above chance [Fig. 6C; Sal: t(9) = 12.33, p < 0.001, Cross: t(9) = 5.06, p < 0.005, Ipsi: t(9) = 14.07, p < 0.001]. Furthermore, we observed better overall accuracy for the 8 s cue. However, crossed inactivations showed a weaker difference, again consistent with a floor effect [Fig. 4C; Cue X Infusion: F(2,16) = 8.90, p < 0.005; 8 s vs 16 s: Sal: t(8) = 7.53, p < 0.001; Cross: t(8) = 3.08, p < 0.05; Ipsi: t(8) = 5.29, p < 0.005].

Together, these results provide causal evidence consistent with the hypothesis that communication between the MD and PL is necessary for temporally organized behavior.

3. Discussion

Our data demonstrate that the mediodorsal thalamus (MD), prelimbic cortex (PL), and their interaction are necessary for timing performance. Inactivating either area individually—or blocking communication between them—disrupted timing, without affecting overall motivation or motor function needed to respond. Beyond extending prior work on each region’s role in timing, these results provide, to our knowledge, the first causal evidence that MD–PL communication is necessary for timing behavior. There are several potential limitations to our work including. First, while our sample size (n = 19) was sufficient to detect robust effects of all manipulations, future studies with larger sample sizes may reveal more subtle differences between inactivation conditions. Second, although we used moderate muscimol doses to minimize spread, some diffusion to adjacent structures cannot be entirely ruled out. Third, disconnection designs cannot distinguish between direct MD-PL projections and indirect pathways (e.g., via striatum). Points two and three are elaborated below. Despite these limitations, the consistent pattern of deficits across all three manipulation types provides strong converging evidence for the necessity of MD-PL communication in timing behavior.

These data build on prior work suggesting that the MD and PFC interact to support adaptive behavior. For example, MD-related cognitive deficits often resemble those seen following PFC dysfunction, such as impaired executive function, working memory, and related processes (Golden et al., 2016; Parnaudeau et al., 2018). Moreover, MD inactivation disrupts PFC activity during cognitive tasks and vice versa (Alexander & Fuster, 1973; Parnaudeau et al., 2013). However, most prior evidence has been primarily correlative, typically monitoring changes in purported task-related activity in one area after inactivating the other. Our results contribute to the recent, but growing, body of causal data showing that MD-PFC interactions mediate cognitive function (Bolkan et al., 2017; de Kloet et al., 2021).

These recent studies have causally implicated MD-PFC communication in cognitive processes such as context integration (Marton et al., 2018) and task switching (Mukherjee et al., 2021). Broadly, these findings have led to the view that MD–PFC circuits maintain and flexibly update internal representations that are critical for adaptive behavior. Given that maintaining internal representation of time and updating information based on temporal information is widely recognized as essential for goal-directed behavior (Balsam & Gallistel, 2009; De Corte & Matell, 2016), our findings may align well with this broader framework.

Our findings might also offer a valuable tool for studying the MD in rodents, where progress has been limited by a lack of reliable cognitive assays (for discussion see Mitchell & Chakraborty, 2013). Prior work in this area has primarily focused on spatial workingmemory tasks. Several studies have found effects of MD manipulations on performance (Bailey & Mair, 2005; Young et al., 1996). However, the literature is admittedly ‘mixed’, as several others have reported null findings (Burk & Mair, 1998; Hunt & Aggleton, 1998a, 1998b; Neave et al., 1993). The peak-interval procedure may be a promising alternative assay, yielding large effects for both MD or PL inactivations and MD-PL disconnection. The task also recruits similar cognitive sub-processes to spatial tasks, such as working memory (Gümüş & Balcı, 2023), attention (Buhusi & Matthews, 2014), and decision-making (Simen et al., 2011). Furthermore, MD-effects with the task have been reliable across studies thus far (see below).

Further, our results have important implications for public-health research. Timing deficits are common in neuropsychiatric diseases including schizophrenia, Parkinson’s disease, autism, and bipolar disorder (Densen, 1977; Lee et al., 2009; Singh et al., 2019; Foerster et al., 2024; Ewald et al., 2023; Parker et al., 2013; Falter et al., 2012), and disrupted MD–PFC connectivity has been implicated in broader cognitive dysfunction seen in patients (Ayub et al., 2021; Chai et al., 2011; Giraldo-Chica et al., 2018; Parnaudeau et al., 2018; Welton et al., 2023). This raises the possibility that MD-PFC communication is a key circuit underlying timing deficits in neuropsychiatric disease. The peak-interval procedure provides an added advantage here, as the task translates easily between rodents and humans with minimal modifications (Malapani et al., 2002; Rakitin et al., 1998). Future work could explore this topic in humans using fMRI and/or EEG, and more invasive techniques can be employed with rodents.

These data also give further insight into the networks that support timing, which are currently unclear. With respect to prior work, Meck (2006) observed a behavioral effect that closely parallels ours after lesioning midbrain dopamine centers during the peak-interval task with rats. Specifically, much like our data, rats maintained their response rates during the cue-period, yet were unable to control response timing, flattening the curves. Meck (2006) hypothesized that dopamine signaling functions as a ‘start-gun’, activating downstream timing circuitry when cues are presented. When this signal is disrupted, the timing circuitry fails to activate at cue-onset. Therefore, responses are purely guided by motivational systems, which are generally independent of timing (Kurti & Matell, 2011; Roberts, 1981).

Our results may extend Meck’s (2006) findings. Specifically, both the MD and PL are densely innervated by midbrain dopamine centers (Descarries et al., 1987; García-Cabezas et al., 2007). If Meck’s (2006) proposed dopaminergic ‘start gun’ signal is responsible for initiating MD-PL timing circuitry, we would expect a comparable behavioral effect, as our inactivations would constitute a more direct manipulation of its downstream targets. Future studies could examine this with more temporally precise methods (e.g., optogenetic inhibition of midbrain dopamine projections to either area at cue-onset).

These data also build on prior work investigating the PL and MD in timing independently. For instance, Buhusi et al. (2018) used a dual-cue peak procedure to examine how graded PL inactivation affects timing. They administered low to moderate doses of muscimol (0.11–1 mM) and found that increasing the dose progressively broadened the spread of the response curves. In contrast, our study aimed to assess necessity by using an asymptotic inactivation dose (2 mM), which remains moderate in the context of the broader literature (with high doses typically defined as 5–8 mM; Katz et al., 2016; Randall et al., 2021). Pilot data from our lab suggest that higher doses do not yield larger behavioral effects, and we also avoided them due to anatomical concerns. Specifically, in the MD, muscimol can potentially diffuse into the adjacent ventrolateral thalamus, which produces profound dystonia in rats (Di Chiara et al., 1979).

It is tempting to interpret our data as a pronounced version of the spread effect reported by Buhusi et al. (2018), with the response burst simply extending across a broader time- window. However, this interpretation would predict that the burst-pattern was preserved during single trials—just widened. While visual inspection of the heat maps certainly suggests burst-like responding occasionally occurred, our single-trial analyses indicated a more fundamental disruption—a breakdown in the canonical structure of timing behavior. To reconcile these findings, our assumption is that, when the timing system is sufficiently impaired, behavioral control defaults to motivational systems, which are generally less temporally structured, as discussed above.

Another relevant study comes from Lusk et al. (2020), who performed MD muscimol inactivations in mice during a peak-interval procedure using a single cue/interval. Due to either a species difference or a smaller dose (.08 mM), they found adequate preservation of Gaussian-like response rates. Similar to Buhusi et al. (2018), they found an increase in spread. However, they also found an increase in peak times. This could suggest that the MD and PL play partially independent roles in timing. Indeed, in our data, behavioral effect sizes appeared larger following MD inactivations or MD-PL disconnections, relative to the PL inactivations alone. However, we urge caution with such comparisons, as the MD is physically smaller than the PL. Therefore, a plausible alternative is that we were able to achieve greater coverage of the MD, leading to larger effects. Nonetheless, if the MD’s role extends beyond its interaction with the PL, these findings raise the possibility that it also engages other cortical regions—such as cingulate or premotor cortex—that have received less attention in the rodent timing literature.

The present study opens several avenues for future research. Chief among them is clarifying the precise anatomical pathways through which MD-PFC communication supports timing behavior. It may be tempting to attribute the interaction to their well-known reciprocal projections (i.e., MD → PFC and PFC → MD). However, both structures also project to the dorsal striatum, a region with a well-established role in interval timing (De Corte et al., 2019; Gouvêa et al., 2015). Thus, convergent input from the MD and PFC onto striatal neurons could serve as a critical node of interaction (i.e., MD → Striatum ← PFC). Our data cannot resolve this distinction, as disconnection designs cannot differentiate between mono- and poly-synaptic mechanisms. Nonetheless, the current results clearly demonstrate that MD-PFC communication is required for timing, providing fundamental groundwork for future studies using more targeted approaches, such as optogenetics or projection-specific chemogenetics.

Finally, while not a primary focus of the present study, our task incorporated crossmodal cues (i.e., tone and light). We found no evidence that our inactivations selectively disrupted timing for auditory versus visual stimuli (or vice versa). Although there were occasional baseline interactions between cue-modality and duration, these effects were constant across different inactivation conditions (see Supplemental Results). These results imply that the MD and PL form part of an amodal timing network. This speaks to a longstanding debate over whether the timing system is amodal or sensory specific (Paton & Buonomano, 2018; Roberts, 1982)—favoring the former. Notably, time-related signals have been documented even in primary sensory areas (e.g., V1; Shuler, 2006), raising the broader question of how signals from distinct sensory modalities ultimately converge onto the timing network characterized here.

4. Methods

4.1. Behavioral training

All training protocols follow closely from De Corte et al. (2018) and were consistent with the University of Iowa’s Animal Care and Use Committee guidelines and the Declaration of Helsinki. Animals were housed with cagemates until surgery on a 12 hr light/dark cycle with access to 15–20 g food daily. We trained male Long-Evans rats (total n = 19) in standard operant chambers (MedAssociates) under modest water deprivation. Body weights were monitored and maintained to 90% of non-deprived levels. Training progressed through three phases. During the initial ‘shaping’ phase (3 sessions), rats learned to make responses on a nose poke to earn water reward (0.057 ml). Sessions lasted 60 min (mins). No cues were presented in this phase, and any response on the nose poke yielded reward. Next, a ‘fixed-interval’ training phase began (6 sessions; 120 mins each). During this phase, trials started with the presentation of either a tone (4.5 kHz; 80 dB) or houselight. Each cue predicted reward availability after either an 8 or 16 s (s) duration elapsed (Fig. 1A). We counterbalanced the modality-duration relationship across rats (e.g., tone-8 s/light-16 s vs. light-8 s/tone-16 s). Once the rat made a response after the respective cue’s interval passed, the cue terminated and reward was delivered. Early responses were not penalized. Trials were followed by a dark inter-trial interval, lasting 90 s on average (30 s minimum + a variable, 60 s geometrically distributed interval). After fixed-interval training, a ‘probe training’ phase began. Probe training was identical to fixed-interval training. However, during a subset of ‘probe trials’, the cue remained on for 96–128 s, regardless of when rats responded, and no reward was delivered (Fig. 1A). Probe trials were randomly intermixed with rewarded trials. We gradually increased the proportion of probe trials during this phase (25% for the first 3 sessions and 50% for the remaining 6 sessions). Furthermore, the inter-trial interval was gradually decreased to 45 s (30 s-fixed + 15 s-geometric). After initial training, rats underwent surgery.

4.2. Surgery/Histological analysis

Rats underwent stereotaxic surgery under isoflurane anesthesia (5% induction, 2.5% maintenance) and then individually housed. We implanted four 23-gauge guide cannulae bilaterally over either the prelimbic cortex (PL) or mediodorsal nucleus (MD) of the thalamus (PL: AP: +3.2, ML: +/− 0.7, DV: −2.0; MD: AP: −2.6 ML: +/− 0.7 DV: −4.4; relative to bregma/skull surface). We inserted 30-gauge dummy cannulae that extended 1 mm beyond the base of the guide cannulae. To protect the junction between the guide tops and dummy cannulae, we placed 19-gauge tubing over the guides, extending 3 mm above their surface. Infusion cannulae used for drug delivery (see below) extended 1.5 mm beyond the base of the guide. All cannulae were made of stainless steel 304 and custom-fabricated (Component Supply Inc.). We secured the implants with dental cement, and rats recovered in their home cage for one week without water deprivation before resuming experimentation. After the experiment, rats were perfused transcardially with chilled saline followed by 120 ml of 4% paraformaldehyde. Brains were extracted, stained with thionin, and imaged using light microscopy to verify cannula placements.

4.3. Retraining and restraint habituation

Following recovery, we retrained rats on the task for seven sessions. During this period, we gradually acclimated them to the restraint procedures used during drug infusions (see also De Corte et al., 2019). On day one, we placed rats in a 0.5′ × 1′ plexiglass enclosure, where all infusions took place, and allowed them to move freely for one minute. On day two, we lightly restrained the rats by hand in the enclosure for one minute. On days three and four, restraint durations increased to three and five minutes, respectively. On days five and six, we conducted mock infusions using procedures that closely mirrored those used during actual infusion sessions (see below). Specifically, we removed the dummy cannulae and cleaned them with 70% ethanol. Then, we inserted two 30-gauge infuser cannulae into the guides (one PL and one MD; 1.5 mm extension from the guide-base). We ran an infusion pump (Stoelting), connected to the infusers via polyethylene tubing, at a rate of 0.1 μL/min for two minutes, although no substance was infused. We ran an infusion pump (Stoelting), connected to the infusers via polyethylene tubing, at a rate of 0.1 μL/min for two minutes, although no substance was infused. No restraint habituation was conducted on the seventh day, as all actual infusion sessions were preceded by a behavioral recovery day.

4.4. Drug infusions

Following retraining and restraint habituation, we began infusion sessions using the same protocols as described for the mock infusions. We infused 0.22 μL of either saline or 2 mM muscimol (a GABA-A receptor agonist; Sigma Aldrich). Infusion targets differed by experiment. For the bilateral inactivation experiment, we infused muscimol bilaterally into either the PL or MD and infused saline into the other site. For baseline sessions, we infused saline into all targets. This yielded three infusion types: muscimol-PL/saline-MD, saline-PL/muscimol-MD, and saline-PL/saline-MD. For baseline sessions in the disconnection experiment, saline sessions still targeted all four sites. However, during muscimol sessions, we always targeted one PL and one MD for inactivation. Specifically, during ‘ipsilateral’ inactivation sessions, the inactivated PL and MD were in the same hemisphere (left-PL/left-MD; right-PL/right-MD). Conversely, during ‘crossed’ inactivation sessions, the inactivated PL and MD were in opposing hemispheres (left-PL/right-MD; right-PL/left-MD). In both cases, we infused saline into the remaining areas. All rats received each ipsilateral/crossed infusion combination in a randomized order. One session of normal training separated all infusion sessions to allow for behavioral recovery and prevent GABA receptor sensitization. Rats began the task 30 min after the final infusion.

4.5. Behavioral measures

4.5.1. Mean response rate analysis

We computed mean response rates by grouping responses from probe trials into 1-second bins and averaging across trials. To quantify timing behavior, we fit each response curve with a Gaussian + kurtosis parameter function described in the following equation (De Corte et al., 2018; Swanton & Matell, 2011):

$$Y=Y_0+S*\text{exp}\left(\frac{abs{\left(\frac{t-PT}{SD}\right)}^k}{2}\right)$$

Y₀ represents the baseline rate of responding, S is a scale parameter, t is time, PT is the mean (i.e., ‘peak-time’), SP is the spread, and K allows the function to fit peaks with varying degrees of kurtosis/skew. As described below, we were primarily interested in the fit-quality, measured by the R². For all statistical analyses on R² data, we applied a Fisher’s r-to-z transformation to the square root of each value, as is standard for improving normality in correlation-based measures. This is standard and required to obtain normality in data based on Pearson’s correlation coefficients. We also report peak-time and coefficient of variation (CV; spread/mean) measures from the fits in the supplemental results section. Finally, to assess potential changes in motivation or motor output during inactivation sessions, we computed the overall mean response rate across the trial period (i.e., ignoring time as a factor).

4.5.2. Single-trial analyses

We analyzed single-trial patterns of responding, again grouping responses into 1-second bins. We identified putative response bursts using a standard approach in which a dual step-function (first step-up, second step-down) is fit to responding (Church et al., 1994). The algorithm finds bursts of responding by iteratively fitting the step function across time in the trial, minimizing absolute residuals. We extracted the fit-quality of the function (R²), as well as the start and stop times, defined as the time where the initial up-step and the terminal down-step were found, respectively.

For the machine learning analysis, we trained a support vector machine (SVM) to classify original single-trial response patterns versus time-shuffled versions of the same trials (i.e., binary classification; Matlab, fitcsvm function). Each trial was represented as a vector of response rates, grouped into 1-second time bins. These binned response rates served as the features for the classifier. We used a radial basis function (RBF) kernel to allow the SVM to derive a non-linear decision boundary between real and shuffled data. We used a leave-one-out cross-validation approach for each rat, training the SVM on all but one trial (both real and shuffled) and testing it on the held-out pair.

Because shuffling will always introduce random variation from one classifier to the next, we repeated the training/testing process 30 times (matching the average number of probe trials per cue at each infusion type) and used the average classification accuracy as the final datum. We verified that increasing or decreasing the number of runs did not materially affect the results. We also explored more sophisticated classifiers that are better suited for sequential data, such as recurrent neural networks (RNNs) and long short-term memory (LSTM) models. However, these approaches produced results highly similar to those obtained with the SVM. Given the relatively limited number of probe trials and our aim to avoid overfitting, we opted to report the SVM-based analysis as a more interpretable and computationally efficient method.

4.6. Statistical analyses

We analyzed each measure with a mixed-model Analysis of Variance (ANOVA) using SPSS. We assessed sphericity using Mauchly’s test. In cases where sphericity was violated, we employed the Greenhouse-Geisser correction. For both experiments, modality-duration pairing (tone-8 s/light-16 s vs. light-8 s/tone-16 s) served as a between-subjects factor. For the bilateral inactivation experiment, Infusion (saline, MD muscimol, PL muscimol) and Cue (8 s vs. 16 s) served as a within-subjects variables. The model was the same for the disconnection experiment. However, the Infusion factor corresponded to the spatial arrangement of the inactivations (saline, ipsilateral, crossed). To simplify reporting, we screened for laterality effects with individual ANOVAs (e.g., Crossed: left-MD/right-PL vs. right-MD/left-PL). As expected, there were no reliable effects [all Fs < 1.75]. Therefore, we pooled the data. We used simple effects analyses to probe any significant main effects or interactions and report the corresponding t values/degrees of freedom. Where appropriate, we used Dunnett’s test to control for multiple comparisons.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.nlm.2026.108142.

Data availability

Data will be made available on request.