Main

The spike activity patterns of midbrain dopamine-containing neurons signal unexpected and salient cues and outcomes1,2,3,4,6,7, and the dynamics of these phasic neural signals have been found to follow closely the principles of reinforcement learning theory3,4,5,6. In accordance with this view, selective genetic manipulation of the phasic firing of dopamine neurons alters some forms of learning and cue-guided movements8,9. Episodes of transient dopamine release in the ventral striatum have been detected with FSCV, and these also occur in response to primary rewards and, after learning, to cues predicting upcoming rewards10,11,12,13. Thus, dopamine transients in the striatum share many features of the phasic spike activity of midbrain dopamine neurons.

Classic studies of such dopamine transients have focused on Pavlovian and instrumental lever-press tasks, in which rewards were within arm’s reach1,2,3,10,11,12,13. However, in many real-life situations, animals must move over large distances to reach their goals. These behaviours require ongoing motivational levels to be adjusted flexibly according to changing environmental conditions. The importance of such control of ongoing motivation is reflected in the severe impairments suffered in dopamine deficiency disorders, including Parkinson’s disease. In addition, in pioneering experimental studies, dopamine signalling has been implicated in controlling levels of effort, vigour and motivation during the pursuit of goals in maze tasks14,15,16,17. It has been unclear how phasic dopamine signalling alone could account for persistent motivational states18. We adapted chronic FSCV to enable prolonged measurement of real-time striatal dopamine release as animals learned to navigate towards spatially distant rewards.

We measured dopamine levels in the dorsolateral striatum (DLS) and ventromedial striatum (VMS) (Extended Data Figs 1 and 2, and Methods) as rats navigated mazes of different sizes and shapes to retrieve rewards (Figs 1, 2, 3, 4, and Methods). The rats were trained first on an associative T-maze task to run and to turn right or left as instructed by tones to receive a chocolate milk reward at the indicated end-arms19 (n = 9, Figs 1, 2 and 4). Unexpectedly, instead of mainly finding isolated dopamine transients at the initial cue or at goal-reaching, we primarily found gradual increases in the dopamine signals that began at the onset of the trial and ended after goal-reaching (Fig. 1a, b). These ‘ramping’ dopamine responses, identified in session averages by linear regression (Pearson’s R > 0.5, P < 0.01), were most common in the VMS (75% of sessions) but were also present at DLS recording sites (42% of sessions). They were evident both in single trials (Fig. 1a–c) and in population averages (Fig. 1e, f, and Extended Data Figs 2g, h and 3), bore no clear relationship to run speed within or across trials (Fig. 1d), and matched, in electrochemical profile, dopamine release evoked by tonic electrical stimulation in vivo (Extended Data Fig. 2i, j). Before goal-reaching, the ramps had similar amplitudes in correct (65% overall) and incorrect trials (Fig. 1e, f). After goal-reaching, the signals were significantly larger in correct trials, particularly in the VMS (paired t-test, P = 0.01, Fig. 1e, f). Notably, a subset of the session-averaged signals in the DLS (22%, 58 out of 262; 7 probes in 5 rats) showed sustained inhibition up to goal-reaching (Extended Data Fig. 3). Such negative signals were rare in the VMS (5%, 15 out of 300 recordings), suggesting that ramping dopamine signals in the DLS, but not VMS, exhibit heterogeneity in polarity.

Figure 1: Ramping striatal dopamine signals occur during maze runs.
figure 1

a, b, Baseline subtracted current (a) and dopamine concentration ([DA], b) measured by FSCV in VMS during a single T-maze trial. c, d, Trial-by-trial changes in dopamine concentration (c) and velocity (d) relative to goal-reaching. e, f, Dopamine concentration (mean ± s.e.m.) for VMS (e, n = 300 session-averaged recordings from 18 probes across 214 sessions) and for DLS (f, n = 262, 13 probes) for correct (blue) and incorrect (red) trials, averaged over all 40 trial sessions.

PowerPoint slide

Full size image
Figure 2: Ramping dopamine signals proximity to distant rewards.
figure 2

a, Distribution of trial times (from warning click to goal-reaching, n = 3,933 trials). b, c, Dopamine release modelled as a function of time elapsed since maze-running onset (b) and as a function of spatial proximity to visited goal (c) for short (purple) and long (orange) trials (see Methods). Vertical lines indicate trial start (red) and end (purple and orange) times. d, Peak dopamine concentration versus trial time for all ramping trials (n = 2,273, Pearson’s R = 0.0004, P = 0.98). e, Experimentally recorded dopamine release (mean ± s.e.m.) in short (n = 327, purple) and long (n = 423, orange) trials. Dopamine peaks at equivalent levels, as in the proximity model in c. f, Normalized peak dopamine levels (mean ± s.e.m.) predicted by time-elapsed (red) and proximity (light blue) models, and measured experimental data (dark blue).

PowerPoint slide

Full size image
Figure 3: Dopamine ramping is sensitive to reward magnitude.
figure 3

a, b, Average dopamine signals from a VMS probe, for consecutive T-maze (a) and M-maze (b) sessions with asymmetric rewards. Asterisks indicate the goal with the larger reward; red arrows (and Switch) indicate reversal of reward amounts. c, Dopamine signals from a different rat running in the S-maze. White arrows indicate run direction. d, Average (± s.e.m.) peak dopamine across all value experiments (n = 4 rats). e, Average (± s.e.m.) VMS dopamine during T-maze (n = 44 sessions in 3 rats, black) and M-maze (n = 17, blue) sessions in the same rats. f, g, Average (± s.e.m.) peak dopamine signals for the sessions plotted in a (f) and b (g) for trials to left (blue) and right (red) goals. Shading indicates arm with larger reward. h, i, Average normalized dopamine (h) and running speed (i) for runs to high (light green) and low (dark green) reward goals in the M-maze. Vertical lines indicate turns. j, k, Average normalized dopamine (j) and running speed (k) in the S-maze (n = 9 sessions in 2 rats), plotted as in h and i.

PowerPoint slide

Full size image
Figure 4: Ramping dopamine selectivity can emerge with training without experimentally imposed reward discrepancies.
figure 4

a, Average normalized dopamine at a VMS site as a function of maze location (n = 19 sessions). b, Dopamine selectivity indices (Methods) for all individual sessions averaged in a. c, Average running speed for sessions in a. d, Selectivity indices for all VMS (left) and DLS (right) recordings (red) compared to shuffled data (blue) for all rats (n = 9). e, f, Average per cent correct performance (e) and average Z-score normalized dopamine selectivity (f) across training blocks. Error bars, s.e.m.

PowerPoint slide

Full size image

We identified isolated phasic transients at warning click indicating trial start and after goal-reaching. These were clearly distinct from the slower ramping responses in approximately 10% of single trials (Extended Data Fig. 4) but were often superimposed on the ramping signals, indicating that the signals recorded could include combinations of transient increases after warning click, slower ramps to goal-reaching, and transient increases after goal-reaching (Extended Data Fig. 4d). The peak magnitudes of the dopamine ramps were comparable to, or slightly smaller than, those of isolated phasic dopamine signals recorded here (Extended Data Fig. 4) and in other studies11,12, and they were correlated with the peak magnitudes of free-reward evoked dopamine measured on the same probes (Pearson’s R = 0.45, P < 0.001, Extended Data Fig. 5), indicating that the ramping signals could be subject to similar regulatory mechanisms and display similar anatomic heterogeneity as classical phasic reward-evoked dopamine signals.

We took advantage of the trial-to-trial variability in the rats’ run times (Fig. 2a) to determine whether ramping dopamine release reflected elapsed time or reward proximity, or whether the ramps reflected sums of multiple, accumulated transients to fixed maze cues20. If the dopamine ramps tracked elapsed time, peak dopamine values should have scaled directly with trial time (same slope, different peak height; Fig. 2b, f). If the ramping reflected distance or spatial location relative to goal-reaching (proximity), peak dopamine levels should have been equivalent for shorter and longer trials (different slope, same peak height; Fig. 2c, f). If the ramps were generated by summation of multiple transients, then for characteristic transient dynamics, the signals should have tended to peak at lower values for longer runs than for shorter runs (different slope, different peak height; Extended Data Fig. 6a, b, and Supplementary Discussion). The measured peak dopamine values at goal-reaching were nearly equivalent for short and long trials (Fig. 2e), and were not correlated with trial length (Fig. 2d–f) or with run velocity or acceleration (Extended Data Fig. 6e, f). Moreover, on trials in which rats paused mid-run, the signals remained sustained (or dipped slightly) and resembled the actual proximity to reward (Extended Data Fig. 7). These observations indicated that the ramping signals could represent a novel form of dopamine signalling that provides a continuous estimate of the animal’s spatial proximity to distant rewards (Fig. 2, Extended Data Fig. 6, and Supplementary Discussion).

Given that phasic responses of dopamine-containing neurons can reflect the relative value of stimuli21, we asked, in a subset of rats, whether the ramping dopamine signals could also be modulated by the size of the delivered rewards (Methods). We used mazes with T, M or S configurations and different total lengths (Fig. 3, and Extended Data Fig. 8). We required the animals to run towards one or the other end of the maze and varied the rewards available at the alternate goal regions. With all three mazes, dopamine ramping became strongly biased towards the goal with the larger reward (Fig. 3, and Extended Data Fig. 8). Run speed was slightly higher for the high-reward maze arms (Fig. 3i, k), but these small differences were unlikely to account fully for the large differences in the dopamine signals recorded. When we then reversed the locations of the small and large rewards, the ramping signals also shifted, across sessions or just a few trials, to favour the new high-value maze arm (Fig. 3, and Extended Data Fig. 8). These bias effects were statistically significant for each experimental paradigm (Extended Data Fig. 8h–j, Mann–Whitney U-test, P < 0.05) and across all rats (Fig. 3d, n = 4, Mann–Whitney U-test, P = 0.02).

In the M-maze, the ramps became extended to cover the longer end-arm distances to goal-reaching, and critically, peaked at nearly the same level before goal-reaching as did the ramping signals recorded in the T-maze, despite the longer distance travelled (Fig. 3e). This result suggested that the ramping dopamine signals do not signal reward proximity in absolute terms but, instead, scale with the path distance to a fixed level that depends on the relative reward value.

To determine whether such value-related differences in the ramping dopamine signals would occur when the actions to reach the distant goal sites were equivalent, we used the S-shaped maze. The ramping signals were larger for the run trajectories leading to the larger rewards (Fig. 3c, j, and Extended Data Fig. 9), despite the fact that the sequence of turns and the lengths of the runs needed to reach the larger and smaller rewards were equivalent for both trajectories (n = 2 rats, 4 and 5 sessions per rat, Fig. 3c, j, k, and Extended Data Figs 8 and 9).

In rats performing the free-choice associative version of the T-maze task, robust dopamine signal biases existed in approximately 20% of sessions (Mann–Whitney U-test, P < 0.05) and significantly more often than chance overall (Z-test, P < 0.00001 versus bootstrapped variances; Methods and Fig. 4a, b, d). These biases were maintained across consecutive training sessions for individual animals (Fig. 4b), did not relate to run-speed biases (Fig. 4c, and Extended Data Fig. 10c) or recording hemisphere (Fig. 4d, and Extended Data Fig. 10a, b, f), and, notably, emerged gradually over days as performance improved and training progressed (Fig. 4e, f). Although not obviously related to imbalances in maze cues or differences in left–right performance, they displayed a weak association with right end-arm choice biases that developed late in training (Extended Data Fig. 10d, e, g). Thus, end-arm biases in the ramping dopamine signals could develop even in the absence of experimentally imposed discrepancies in value, possibly reflecting developing internal value estimates (Supplementary Discussion).

Ramping spike-firing has been recorded for putative midbrain dopamine neurons in head-fixed primates under conditions of reward uncertainty22 and for nigral non-dopamine-containing neurons1. We asked whether the magnitudes of the ramping dopamine signals that we recorded in the striatum changed as performance improved on the free-choice associative T-maze task (Fig. 4e). They did not (Pearson’s R = −0.08, P = 0.19; Extended Data Fig. 10h-j), suggesting that uncertainty about reward probability was unlikely to have controlled the magnitude of the ramping signals22 (Supplementary Discussion).

Classic studies of dopamine neuron firing and striatal dopamine release have largely focused on transient responses associated with unpredicted rewards and reward-predictive cues. Here we demonstrate that, in addition to such transient dopamine responses, prolonged dopamine release in the striatum can occur, changing slowly as animals approach distant rewards during spatial navigation. These dopamine signals seem to represent the relative spatial proximity of valued goals, perhaps reflecting reward expectation23. It remains unclear whether these signals represent goal proximity on the basis of environmental cues, effort, or internally scaled estimates of distance. However, the brain possesses mechanisms for representing both allocentric spatial context and relative distance from landmarks24, which could, in principle, be integrated with dopaminergic signalling to produce such extended dopamine signals.

Transient dopaminergic responses to learned reward-predictive cues have been proposed to initiate motivated behaviours25,26, but with this mode of signalling alone, it is difficult to account for how dopamine acts to maintain and direct motivational resources during prolonged behaviours (Supplementary Discussion). The ramping dopamine signals that we describe here, providing continuous estimates of how close rewards are to being reached, and weighted by the relative values of the rewards when options are available, seem ideally suited to maintain and direct such extended energy and motivation.

Methods Summary

Male Long Evans rats (n = 9) were deeply anaesthetized and were surgically implanted with headstages carrying voltammetry microsensors placed in the VMS (anteroposterior +1.5 mm, mediolateral ±2.1 mm, dorsoventral 6–7 mm) and DLS (anteroposterior +0.5 mm, mediolateral ±3.5 mm, dorsoventral 3.5–4.0 mm) bilaterally (n = 1 rat) or in the left (n = 5) or right (n = 3) hemisphere. Triangular voltage sweeps (−0.4 V to 1.3 V) were applied at 10 Hz to the carbon fibre microsensor probes27 relative to an Ag/AgCl reference electrode implanted in the neocortex. Current changes due to dopamine oxidation–reduction reactions were verified by principal component regression28 using a training set of current profiles from evoked dopamine release and pH changes recorded in 5 rats in response to electrical stimulation of the medial forebrain bundle (MFB). Training on the T-maze task19 (Figs 1, 2 and 4) and chronic voltammetry recording began 4 weeks after surgery, and continued for 15 to 35 days for each rat. The position of each rat was monitored continually by tracking software as the rat navigated the maze (Neuralynx), and the positions were used to trigger auditory cues and reward pumps when the animal entered specific maze locations. After T-maze training, a subset of rats (n = 3) was trained in an extended M-maze. Rats were required either to continue performing the associative tone-cued task (M31) or to visit one of the two end-arms on each trial by blockade of the other end-arm (M36 and M47). Another subset of rats (n = 2) underwent training on an S-maze version of the task in which they only had to run back and forth from end to end to retrieve rewards. Voltammetry data were analysed with in-house MATLAB codes (Mathworks), and probe positions were verified by standard histology19.

Online Content Any additional Methods, Extended Data display items and Source Data are available in the online version of the paper; references unique to these sections appear only in the online paper.

Online Methods

All experimental procedures were approved by the Committee on Animal Care at the Massachusetts Institute of Technology and were in accordance with the US National Research Council Guide for the Care and Use of Laboratory Animals. Sample sizes were chosen based on signal variability estimates from other published studies using FSCV.

Implant procedures

Implantations were performed on deeply anaesthetized fully mature male Long Evans rats (n = 9; 380 to 450 g), in sterile conditions according to approved surgical procedures19 with headstages carrying 1 to 3 independently movable voltammetry probes targeting the DLS (anteroposterior +0.5 mm, mediolateral ± 3.5 mm, dorsoventral 3.5–4.0 mm), of the right (n = 3) or left (n = 5) hemisphere, or the DLS bilaterally (n = 1), 1 to 3 probes targeting the VMS of the same hemispheres (anteroposterior +1.5 mm, mediolateral ±2.1 mm, dorsoventral 6–7 mm), and a unilateral Ag/AgCl reference electrode in the posterior cortex (anteroposterior −2.3 mm, mediolateral ±3.5 mm, dorsoventral ∼0.5 mm). Five rats that underwent maze training and three additional rats for acute stimulation experiments were implanted with tungsten bipolar stimulation electrodes (FHC) straddling the ipsilateral MFB (anteroposterior −4.6 mm, mediolateral ±1.3 mm, dorsoventral 7–8 mm) to verify striatal dopamine release (see below).

Behavioural training

All behavioural training was conducted on a custom built ‘grid maze’ with fully reconfigurable tracks and walls. Training on an associative T-maze task with auditory instruction cues (Figs 1, 2, and 4)19 began 4 weeks after implantation. Voltammetric recordings began when rats learned to run smoothly down the track to retrieve reward. Early sessions with sporadic maze behaviour, such as wall rearing and sluggish initiation of maze running, were discarded. Daily behavioural sessions consisted of 40 trials. Trials began with a warning click, followed 0.5 s later by the lowering of a swinging gate, allowing the rat to run down the maze. Half-way down the long arm, a tone was triggered (1 or 8 kHz), indicating which end-arm to visit in order to retrieve chocolate milk reward (0.3 ml) delivered through automated syringe pumps (Pump Systems) upon the rat’s arrival. The spatial position of each rat was monitored continually by video tracking (Neuralynx). Tone delivery and syringe pumps were controlled by in-house behavioural software written in MATLAB (Mathworks). After 15 to 35 T-maze sessions per rat, a subset of rats (n = 3) received 17 training sessions (4 to 6 sessions each) on the M-maze task in which the end-arms of the T-maze were extended (Fig. 3). These rats received a larger amount of reward (0.4 ml) at one goal site than at the other (0.1 ml for 2 rats and 0.2 ml for 1 rat). After 2 to 3 sessions with a given set of spatial reward contingencies, the reward amounts at the two goals were reversed. One rat (M31) was required to make turn choices in response to tones as in the previous T-maze task, whereas the other two rats (M36 and M47) were directed pseudo-randomly to one end-arm of the maze on each trial by removing the track to the opposite arm (20 trials to each arm) without tone presentation. Two rats were trained on the S-maze task (Fig. 3). These rats were required simply to run back and forth to retrieve a large volume of chocolate milk (0.4 ml) at one goal and a small volume (0.1 ml) at the other goal. Consecutive visits to the same reward site did not trigger the reward pumps.

Voltammetry data acquisition and analysis

Waveform generation and data acquisition for voltammetry recordings were done with two PCI data acquisition cards and software written in LabVIEW (National Instruments). Triangular voltage waveforms were applied to chronically implanted carbon fibre electrodes, relative to the reference electrode, at 10 Hz. Electrodes were held at −0.4 V between scans, and were ramped to 1.3 V and back to −0.4 V during each scan27. Current produced by redox reactions was recorded during voltage scans.

We compiled a library of current versus applied voltage templates for dopamine and pH changes of varying magnitudes by stimulating the MFB (60 Hz, 24 pulses, 300 µA) to induce dopamine release in the striatum in 5 rats maintained under isoflurane anaesthesia. We used these templates from all 5 rats as a training set to perform chemometric analysis28 on voltammetry measurements obtained during behaviour with in-house MATLAB software. This procedure allowed us to distinguish changes in current due to dopamine release from changes due to pH or to other electroactive substances28. In a separate set of rats, we stimulated the MFB (10 Hz, 60 pulses, 100–120 µA) under urethane anaesthesia to mimic the slower, low amplitude ramping signals that we observed in behaving animals (Extended Data Fig. 2). Current changes were converted to estimated dopamine concentration by using calibration factors obtained from in vitro measurements of fixed dopamine concentrations. Behavioural video tracking was synchronized with voltammetry recordings by marker transistor–transistor logic signals sent to the voltammetry data acquisition system.

For each trial, voltammetry data were normalized by subtracting average background current at each potential measured during the 1-s baseline period before warning click. Session averaged traces (Figs 1 and 3, and Extended Data Figs 3 and 10) were computed by averaging the dopamine signals recorded in a single session across 40 trials, and then averaging these traces to obtain global averages across all rats and electrodes. Each session-averaged trace (one from each probe from each session) was considered as an independent measurement for computing s.e.m. Concatenation of dopamine and proximity signals (see below) was performed by scaling the peri-event windows using the median inter-event intervals between consecutive events across all trials (Figs 1, 2, 3, and Extended Data Figs 3, 4 and 10). Traces between two consecutive events were plotted by taking data from each event to half of the median inter-event interval. Maze arm selectivity (Fig. 4 and Extended Data Fig. 10) was computed by the following equation:

where [DA]left and [DA]right represent dopamine concentration during trials to the left and right arms of the maze, respectively.

Session-averaged dopamine traces were identified as having positive or negative ramping characteristics (Extended Data Fig. 3) if they exhibited a significantly positive or negative linear regression coefficient (Pearson’s, R > 0.5 or R < −0.5 and P < 0.01) over the entire trial period. Trials with phasic responses around the trial start (Extended Data Fig. 4) were identified by calculating the relative difference between consecutive time points (100 ms per sample) in a 1-s window with its centre sliding in 0.1-s steps for a 1-s period from 0.5 s to 1.5 s after warning click. For a given window position, if the differences were all positive values across the first half of the window (0.5 s) and negative across the second half, we determined that a significant inflection point was present in that trial. Comparison of dopamine signals on long and short trials (Fig. 2) was done by selecting trials that fell in the bottom third (short trials) and in the top third (long trials) of the trial-time distribution for each rat that displayed a dopamine peak within 0.5 s of goal-reaching. Trials with noisy video tracking data were discarded from this analysis. The simulations for the time-elapsed model (Fig. 2b, f) were made by calculating the average slope of the ramping signals across all trials on each session and by using linear extrapolation to predict the peak dopamine values on each trial within that session. Peak values were normalized to the median peak value for each session individually and averaged for short and long trials to generate the predictions in Fig. 2f. The multi-transient model (Extended Data Fig. 6) was implemented in MATLAB and tested using a range of physiologically realistic estimates for the slope and decay times for previously observed transient signals in vivo11. For the model results shown in Extended Data Fig. 6a, c, simulations were run 100 times, and temporal offset times for the transients were determined by drawing randomly (normrnd function in MATLAB) from a normal distribution with means of 0.8 s (with standard deviation of 0.5 s) and 1.4 s (with standard deviation of 7 s) for short and long trials, respectively (Extended Data Fig. 6a, c). Spatial proximity to goal (Fig. 2c and Extended Data Fig. 7b, e) was calculated by summing the pixel differences in the x and y spatial dimensions for each recorded rat's position. These traces were averaged across all short and long trials separately to generate the traces shown in Fig. 2c. Session-by-session estimations of peak dopamine concentration were made by randomly generating peak trial values using the mean and standard deviation of peak values present in the experimental data. All peak values for short and long trials (Fig. 2f and Extended Data Fig. 6b, d) were normalized to the mean dopamine concentration for each session for both simulated and experimental data.

The presence of population selectivity in the dopamine signals on the associative T-maze task (Fig. 4) was determined by comparing the variance of selectivity indices from the T-maze recordings to the distribution of variances obtained by shuffling the dopamine concentrations on the two end-arms and bootstrapping 10,000 times (Z-test comparing variance of the data to the variances of the bootstrapped distributions, P < 0.00001). To identify changes in selectivity and ramping magnitude across training, Z-scores of selectivity indices (Fig. 4f) were computed for each rat by taking the absolute values of all selectivity indices, then by normalizing across sessions before combining all values across rats. Choice selectivity (Extended Data Fig. 10) was computed similarly to the selectivity score for dopamine:

Run time biases and per cent correct biases (Extended Data Fig. 10) between the two arms were also computed in this way. Correlations between peak dopamine magnitude and per cent correct performance were calculated by normalizing the average peak dopamine values on each trial to the average peak value across all trials within that session.

Histology

Probe positions were verified histologically19. Brains were fixed by transcardial perfusion with 4% paraformaldehyde in 0.1 M NaKPO4 buffer, post-fixed, washed in the buffer solution, and cut transversely at 30 μm on a freezing microtome, and they were stained with cresylecht violet to allow reconstruction of the recording sites (Extended Data Fig. 1). For a subset of the probes, a constant current (20 mA, 20 s) was passed through the probe before fixation to make micro-lesions at probe-tip locations.