The Contribution of Premotor Cortico-Striatal Projections to the Execution of Serial Order Sequences

Mice can learn to execute serial order sequences of lever presses

To investigate the contribution of cortico-striatal synapses to the structuring and execution of self-paced serial order sequences, we trained mice in a serial order task to execute two subsequences of lever presses (Fig. 1A). The execution of a correct serial order sequence was achieved when animals performed at least four presses on a lever 1 (subsequence 1; S1) followed by four presses on lever 2 (subsequence 2; S2), which lead to the delivery of a reward in the magazine (pellet). To execute self-paced sequences, we first trained animals to execute forced sequences (forced: only one of the two levers was exposed at any point, signaling the animals when to press; Fig. 1B). Later in training, blocks of forced and self-paced serial order sequences were intercalated (self-paced: both levers remained exposed so that animals decided when to press and transition between subsequences; Fig. 1C). The requirement of intercalated blocks was necessary as animals’ performance dropped when they were required to perform only self-paced sequences (tested in a group of eight animals: data not shown). The main difference between a forced and a self-paced serial order sequence was that in the former, completing four presses on lever 1 retracted it and exposed lever 2; if four presses were executed on lever 2, it was also retracted, followed by the delivery of a pellet in the magazine (Movie 1). On the other hand, the execution of a correct self-paced sequence also required at least four presses on each lever, but in this case, both levers remained exposed throughout the self-paced trials (Movie 2).

A session containing both different kinds of sequences is identified in the figures as blocks of “forced–self-paced” sessions. During one of these sessions, the animals were required to achieve correct blocks of five forced and five self-paced sequences consecutively for 30 min. Figure 1D shows an example animal during the first (early) and the eleventh (late) day (rewards are labeled with filled red triangles for forced and in blue for self-paced). After 21 d of training, and 9 d performing blocks forced–self-paced sessions, animals showed a stable performance of correct sequences, demonstrated by the fact that performance no longer changed from sessions 9 to 11 [performance forced, χ²(2) = 2.51, p = 0.30, performance self-paced, χ²(2) = 1.0, p = 0.65, Friedman test, wild-type (WT) animals, n = 8, Fig. 1E]. No differences were found in comparisons of a number of task-relevant parameters between forced versus self-paced sequences: latency to initiate, the transitions intrasubsequences, the intervals between presses intra subsequences (p > 0.05, Mann–Whitney U test, WT animals, n = 8; Extended Data Fig. 1-1A).

While animals executing forced sequences reached an 83 ± 1.5% in correct performance (correct sequences/errors + correct sequences), the execution of self-paced sequences only reached 61 ± 1.7% (WT animals, n = 8; Fig. 1E, red and blue tick curves, respectively). To further investigate this performance, we quantified the types of errors and modes in which the animals executed correct sequences. An error was defined as a sequence of presses in which the animal did not achieve four presses on lever 1 followed by four presses on lever 2. The errors were grouped into four categories: (1) incorrect start (starting on the opposite lever); (2) breaks in S1; (3) premature switches from S1 to S2; and (4) breaks in S2 (Fig. 1B,C, right panels 1–4, respectively; note that 1 and three were only possible during self-paced). These four categories were quantified late in training (Fig. 1G). The decrease in Breaks in S1 and S2 was the primary determinant in performance improvement of both forced (ForcBreaks) and self-paced (S-PBreaks) sequences compared with pretraining (p < 0.05, Mann–Whitney U test, WT animals, n = 8; Extended Data Fig. 1-1B, panels 2, 4). A decrease in incorrect-start errors accounted for the major improvement in the self-paced performance (p = 0.007, Mann–Whitney U test, WT animals, n = 8; Extended Data Fig. 1-1B, panel 1). No difference in performance between forced and self-paced sequences was detected when accounting for only the available errors in both forced and self-paced sequences: BreaksS1+BreaksS2 (p = 0.5, Mann–Whitney U test, WT animals, n = 8; Fig. 1E, blue dashed curve vs red curve).

The correct execution in forced sequences was S1 = 4→S2 = 4; in self-paced the execution of correct sequences grouped in several options (as anything with S1 = >4→S2 >= 4 was correct). The majority of correct sequences were Long-S1 sequences (S1 > 4→S2 >= 4; WT animals, n = 8; Fig. 1F).

Besides the execution structure of the sequences, we quantified the length parameters. The latency to start, the transition between subsequences, the interpress intervals, the duration, all became faster as training progressed (Friedman test, p < 0.05; WT animals, n = 8; Extended Data Fig. 1-1A). Later in training, the latency to initiate the sequences was not different between forced and self-paced (forced: 1.1 ± 0.1 s, self-paced: 1.5 ± 0.3 s, p = 0.535, Mann–Whitney U test, WT animals, n = 8; Fig. 1H). Neither the transition time between subsequences (S1_end→S2_start) or the interpress intervals intersubsequences between the two modes (p > 0.05, Mann–Whitney U test; Fig. 1J,K). However, the duration and the number of presses were longer in the S1 of self-paced than in forced [forced_duration: 3.2 ± 0.2 s vs self-paced_duration: 4.8 ± 0.4 s, forced_S1presses = 4 ± 0 vs 6.1 ± 0.6 in self-paced_S1presses, p = 0.0002, Mann–Whitney U test; Extended Data Fig. 1-1C,D).

This task parameterization showed that mice could learn and execute forced and self-paced serial order sequences in blocks. The main difference in the execution of these two modes of sequences (besides that self-paced have more possibilities for errors) is that animals execute longer chunks of presses in the subsequence 1 of self-paced sequences.

Activity modulation in premotor and motor cortical neurons during the execution of serial order sequences

Once we established that mice are capable of executing serial order sequences, we questioned if, as predicted, the activity of M2 and M1 cortical neurons was modulated during the execution of these trained sequences (Tanji and Evarts, 1976; Mushiake et al., 1990; Churchland et al., 2006; Shima et al., 2007). For this purpose, we trained eight animals as in Figure 1. After 3 d in the blocks of forced–self-paced sessions, we performed brain surgeries to implant a mobile electrode array either in M2 (n = 6; Fig. 2A, middle panel) or M1 (n = 2; here, we verified the region to evoke forepaw movements by microstimulation; see Materials and Methods; Fig. 2A, right panel). After 4 d of recovery, the training restarted until animals reached a stable performance in the execution of forced and self-paced sequences (10–15 blocks sessions postsurgery). From these animals, 34 well-isolated units in M2 and 26 units in M1 were analyzed. Representative examples of these neurons’ activity during the executions of sequences are presented in Figure 2B,C. The z score heatmaps of Figure 2D,E show all recorded units during the execution of self-paced sequences (the same units during the execution of forced sequences are presented in Extended Data Fig. 2-1).

To answer whether the same units were active during the different phases of the sequences, we measured the mean activity of neurons above two z scores during the execution of the sequences. We classified neuronal activity into categories based on different parameters of sequence execution (Fig. 2F). We found no statistical difference in the comparison of M2 versus M1 activity, only a tendency for more modulated/engaged M2 neurons related to the initiation of a sequence (Fig. 2G). To explore this difference during sequence initiation, we compared the mean z score between M2 and M1 activity from the units that showed a significant modulation 1 s before starting the sequences (Fig. 2H,I). This comparison showed a bigger z score positive modulation in M2 than in M1 during the execution of the sequences (M2 = 14 of 34 units, M1 = 8 of 26 units; mean z score in self-paced_sequences: M2_S1 = 1.32, M2_S2 = 0.85 vs M1_S1 = 0.55, M1_S2 = 0.22; p < 0.05, z score difference test; Fig. 2J, bars labeled “S1 and S2”). Remarkably M2 units showed a stronger modulation in self-paced sequences, even before the initiation of the sequence (M2_S-Pinit = 1.4 vs M1_S-Pinit = 0.8, p = 0.04 z score difference test; Fig. 2J, bars labeled “Init” in blue) and a similar tendency during the initiation of forced sequences (Extended Data Fig. 2-1H, bars labeled “Init” in red). No difference in the number of recruited units throughout the execution of the sequences was detected (p > 0.05, χ² test; Extended Data Fig. 2-1I).

Both M2 and M1 contain units encoding the length of the serial order sequences of actions

After the evaluation of the mean z score related to the execution of the serial order sequences, we tested the hypothesis that M2/M1 encode the length parameters of the execution of these sequences (Riehle and Requin, 1989; Riehle et al., 1994; Lucchetti and Bon, 2001; Sakurai et al., 2004; Churchland et al., 2006; Jin et al., 2009; Hernández et al., 2010; Merchant et al., 2013; Crowe et al., 2014; Murakami et al., 2014; Mendoza et al., 2018). To evaluate this possibility, we performed linear regressions between the duration of the segments of the sequences (total sequence, S1, transition, or S2) and the neural activity recorded from individual neurons in either M2 or M1. We reasoned that if these cortices contained activity encoding the sequence length, their cortico-striatal projections could convey this information to the striatum.

Figure 3A shows two examples of units that presented significant regressions (R² > 0.6 and p < 0.05) in specific time bins during the execution of serial order self-paced sequences (Fig. 3A, middle row panels, gray shadows). Figure 3B shows the proportion of units that presented with significant regressions based on five questions: column i, does the activity before the first press (1 s) encode the sequences’ length? ii, does the mean activity during the sequence encode the length of the sequence? iii, iv, and v, Is there a relationship between the duration of S1, the transition or the duration of S2, and the mean firing rate during these epochs? The answer to these five questions is presented per recorded unit in either M2 or M1 (Fig. 3B, upper panels) and as the % of units with significant regressions per region (Fig. 3B, bottom panels). Notably, the biggest proportion of units with significative regressions was between the sequence duration and the firing rate before the initiation of the sequence in both M2 and M1 (Fig. 3B, columns i) or the sequences duration and the mean firing rate of M2 during the sequences (Fig. 3B, columns ii).

The percentage of recruited units over time (units presenting significant regressions, 200-ms bin, sliding 10 ms; R² > 0.6 and p < 0.05) is presented in Extended Data Figure 3-1B. Despite small tendencies, no difference was detected in the proportion of units recruited overtime either when comparing M2 versus M1 nor when comparing within each region between forced and self-paced sequences (1 s before or one after the start of the sequences). Similarly, no differences were found between neuronal activity and the duration of the sequence, duration of S1, duration of the transition, or the number of presses (comparison of bars p > 0.05, χ² test; Extended Data Fig. 3-1B).

To address whether specific cortico-striatal M2 or M1 neurons encode for sequence parameters, we recorded from antidromic photo-identified cortico-striatal neurons in vivo (Lima et al., 2009; Díaz-Hernández et al., 2018). To achieve this, we first confirmed the lateral region of the striatum (LS) that is innervated by both M2 and M1 projections (Extended Data Fig. 3-2A–D; Oh et al., 2014; Sohur et al., 2014; Hintiryan et al., 2016; Hunnicutt et al., 2016). Then we trained mice in which the expression of Channelrhodopsin-2 (ChR2) was targeted to the excitatory cortical neurons in M2 (n = 4) or M1 (n = 2) and their projections to the striatum (using Emx1-Cre mice; Guo et al., 2000; Gorski et al., 2002). We then implanted an electrode array above the ChR2 expression and an optical fiber into LS to antidromically activate the cortico-striatal units (Extended Data Fig. 3-2E–G; Friedman et al., 2015). Following this procedure and after a stable performance in the execution of blocks of forced–self-paced sequences, the activity of M2 or M1 units was recorded. During these recordings, we recorded cortical and cortico-striatal photo-identified units (PID units). A PID unit was that, at the end of the session, responded to antidromic light stimulation with short latency (<10 ms; Díaz-Hernández et al., 2018). The mean latency of M2+M1→LS was 5.1 ± 0.4 ms; n = 21; Extended Data Fig. 3-2F, bottom right panel). Following these criteria, we identified 10 M2 and 11 M1 cortico-striatal PID units (Extended Data Fig. 3-2G). As in the non-PID units (Fig. 3B), the major proportion of PID units presented significant regressions between the sequences duration and the firing rate before the initiation, the duration, and the mean firing rate during the execution of the sequences (Extended Data Fig. 3-2G, columns i, ii).

In summary, in Figures 2, 3 and Extended Data Figure 3-2, we show that the activity of M2 and M1 is related to the execution of the trained sequences. The mean z score analysis showed that M2 had a bigger modulation than M1 during the execution and even before the initiation of the serial order sequences. Furthermore, the regression analysis showed that both M2 and M1 contain units that encode the sequences’ execution length.

Inhibition of lateral striatal neurons before initiation of a serial order sequences impairs its execution

This study’s main goal was to establish whether cortico-striatal projections of the premotor and motor cortex contribute to the execution of sequences. To address this point, we first verified that neuronal activity in the lateral striatal (LS) region that receives inputs from M2 and M1 (Extended Data Fig. 3-2A–D) contributes to the execution of serial order sequences. For this purpose, we trained a group of animals in which the inhibitory opsin archaerhodopsin (Arch3.0) was injected bilaterally into LS followed by fiber-optic implantation above the injection sites (Extended Data Fig. 4-1A). Then, once the animals reached around 80% success (correct sequences) in the performance of forced and 60% in self-paced sequences (Extended Data Fig. 4-1A, panel 3), we performed sessions of optogenetic inhibition, delivering the light to inhibit Arch3.0-expressing neurons in a state-dependent manner during the execution of the sequences. We randomized three protocols (each protocol on a different day): (1) before the initiation; (2) during the execution; or (3) during the transition between subsequences (Extended Data Fig. 4-1B–G, H–M, N–S, respectively). We observed that optogenetic inhibition before starting the sequence delayed the initiation (Extended Data Fig. 4-1C). Inhibition during the execution of the sequence increased premature switches (Extended Data Fig. 4-1J, panel 3) and decreased the correct Long-S1 self-paced sequences (Extended Data Fig. 4-1K). Inhibition during the intersubsequence transitions increased the transition time between forced subsequences and showed a tendency to increase the time between self-paced subsequences (Extended Data Fig. 4-1R). Note that the data for the forced sequences are only presented in the figures when significant effects were detected; otherwise, they are presented only in Table 1.

Presequence initiation inhibition of either motor or premotor cortico-striatal projections impairs initiation, but only the latter disrupts the execution

Once we verified that direct inhibition of LS neurons impaired the execution of sequences, we asked whether the premotor (M2→LS) or the motor (M1→LS) cortico-striatal projections contributed to the execution of these sequences. We first verified the inhibition of striatal neurons by the optogenetic inhibition of cortico-striatal projections using Arch3.0 with 2-s pulses of light (Fig. 4A). This pulse length is in a proper range far from the biophysical constraints of using Arch for optogenetic inhibition (Mahn et al., 2016; Díaz-Hernández et al., 2018). Furthermore, we evaluated the proportion of labeled cells in M2/M1 that contain corticofugal axons crossing by the area of the LS manipulations (Extended Data Fig. 4-2).

Next, we tested whether the contribution of cortico-striatal projections is time dependent (Bartolo et al., 2014; Bartolo and Merchant, 2015; Li et al., 2016; Nakayama et al., 2018). We performed optogenetic inhibitions of the M2→LS (n = 9) or M1→LS (n = 8) projections in a state-dependent manner, using three protocols: (1) before initiation (Fig. 4); (2) during execution (Fig. 5); and (3) during the transition between subsequences (Fig. 6).

Most of the effects of optogenetic inhibition were on self-paced sequences. Therefore, for Figures 4-6, forced sequences are only presented when effects were detected; otherwise, only self-paced sequences are presented. The full data for forced and self-paced inhibitions is in the Table 1.

To carry out these experiments, a group of Emx1-Cre mice was subject to bilateral Arch3.0-eYFP expression into M2 or M1 (or eYFP for control animals) and bilateral fiber optic implantation into the LS (Fig. 4B; see Figs. 4-6). After surgery, animals were allowed to recover for 3 d before training started and continued until the performance was stable (Fig. 4D). During the optogenetic inhibition sessions, and depending on the protocol, a continuous pulse of green light was applied randomly in 50% of the trials (2 s; triggered by the behavior), allowing us to compare the effects of light inhibition on each animal within the same session.

For the inhibition “before” initiation protocol, we aimed to assess whether the cortico-striatal projections contribute to sequence execution by interfering with the initiation/preparation of the sequences for which we took advantage of the stereotyped behavior of the trained animals. An infrared beam was placed between the magazine and the levers (red arrow-dashed line coming out of the magazine; Fig. 4C), it was possible to trigger a pulse of light when the animal crossed the infrared beam before starting the sequences (Tecuapetla et al., 2016).

The inhibition of the M2→LS projections before initiation increased premature switches (M2→LS_Self-pacedon = 20 ± 5 vs M2→LS_Self-pacedoff = 8 ± 1; the paired mean difference was 12.2 [95.0%CI 6.8, 27] and p = 0.0001, q = 0.01 two-sided permutation t test; Fig. 4E, panel 3, light blue data; Movie 3). This effect was not observed in control animals (Δ on-off comparison to the control group: the unpaired mean difference between control and M2→LS was 10.4 [95.0%CI 3.7, 2.4]. The p value of the two-sided permutation t test was 0.0224, q = 0.06; Fig. 4F, panel 3, bottom part). The increased premature switches were accompanied by decreased presses within the self-paced sequence (M2→LS_Self-pacedon = 9 ± 0.6 vs M2→LS_Self-pacedoff = 10 ± 0.4; the paired mean difference was −1.16 [95.0%CI −2.42, −0.495], p = 0.016, q = 0.05 two-sided permutation t test; Fig. 4F, light blue data).

This inhibition protocol also increased the latency to start the sequences (M2→LS_{Self-paced_Latency}on = 2.9 ± 0.4 vs M2→LS_{Self-paced_Latency}off = 1.8 ± 0.3; the paired mean difference was 1.1 [95.0%CI 0.556, 1.96], p = 0.0026, q = 0.01 two-sided permutation t test; Fig. 4G, light blue left panel; forced sequences were also delayed, see Table 1, data related to Fig. 4), an effect not observed in control animals (Fig. 4G, Δ on-off panel, gray data). This increased latency to start was accompanied by an increase in the animals’ return to the magazine in the trials that received light inhibition, suggesting that inhibition interrupted the proper serial order sequence initiation (see Movie 4; M2→LS_{SP-returns_to_start}on = 9 ± 2% vs M2→LS_{SP-returns_to_start}off = 1 ± 0.7%, the paired mean difference between control and M2→LS was 8.29 [95.0%CI 5.14, 11.7], p = 0.006, q = 0.03 two-sided permutation t test. Unpaired mean difference between control and M2→LS was 9.86 [95.0%CI 6.0, 13.8], p = 0.0002, q = 0.002 two-sided permutation t test; Fig. 4G, Δ on-off panel).

Conversely to the inhibition of M2→LS projections, the increase in premature switches and the decrease in the number of presses within the sequence were not observed when we inhibited M1→LS projections (Fig. 4E,F, dark blue data). Instead the inhibition of the M1→LS projections increased the latency to start the self-paced sequences (M1→LS_{Self-paced_Latency}on = 3.2 ± 0.4 vs M1→LS_{Self-paced_Latency}off = 2.6 ± 0.3; the paired mean difference was 0.614 [95.0%CI 0.40, 1.1] and p = 0.009, q = 0.04 of the two-sided permutation t test; Fig. 4G, comparison to control: panel Δ on-off same figure, dark blue vs gray data, unpaired mean difference between control and M1→LS was 0.88 [95.0%CI 0.4, 1.7] and p = 0.003, q = 0.01 of the two-sided permutation t test), and the returns to start (comparison to control: panel Δ on-off in the same figure, dark blue vs gray data; unpaired mean difference between control and M1→LS was 11.8 [95.0%CI 3.3, 26.1] and p = 0.02, q = 0.06 of the two-sided permutation t test; Fig. 4H).

Together these results suggest that the M2→LS but not the M1→LS projections contribute to the execution/structuring of self-paced sequences while both M2→LS and M1→LS contribute to the initiation.

Inhibition of premotor cortico-striatal projections at the start of the execution decreased Long-S1 sequences

One prediction for the cortico-striatal projections’ contribution to serial order sequences execution is that their requirement would be time dependent (see Figs. 2, 3; Li et al., 2016; Nakayama et al., 2018). To further prove this idea, we performed a second protocol: inhibition during “execution,” which consisted of performing light inhibition of the cortico-striatal projections once the execution of the sequences started (triggered by the first press in S1; Fig. 5A). Using this protocol, we observed that the inhibition of the M2→LS projections did not modify the proportion of the different categories of errors (Fig. 5B) nor the number of presses in the sequences (Fig. 5C), but it did decrease the proportion of Long-S1 sequences (M2→LS_{Long_S1_Seq}on = 35 ± 5 vs M2→LS_{Long_S1_Seq}off = 50 ± 6, the paired mean difference was −14.6 [95.0%CI −18.9, −9.8] and p = 0.002, q = 0.04 of the two-sided permutation t test; Fig. 5D, upper panels; comparison to control eYFP-group: Δ on-off panel, unpaired mean difference between control and M2→LS was −18.4 [95.0%CI −28.3, −8.8] p = 0.003, q = 0.04 of the two-sided permutation t test Fig. 5D, bottom panel), with no overall effect on the duration or the number of presses in the sequences (see Table 1; data related to Fig. 5). The inhibition of the M1→LS projections did not yield any significant effects with the inhibition during execution protocol (Fig. 5, dark blue data).

The results from the inhibition of M2→LS or M1→LS projections during the beginning of the sequence execution highlight that the activity of the M2→LS projections is important for the length of the sequences (particularly for S1). Furthermore, these results also suggest that, at least later in training, the M1→LS projections are not required for sequence execution.

Inhibition of premotor cortico-striatal projections during the transition between serial order subsequences increases the transition time

So far, we have shown that the cortico-striatal projections from M2→LS contribute to the appropriate initiation and execution of self-paced serial order sequences (Figs. 4, 5). However, in the previous experiments, the light inhibition never occurred during the transition between subsequences. Given that a small proportion of units from either M2 or M1 showed significant modulation during the transition-moment (Fig. 2F,G), we asked whether inhibiting directly during the transition could reveal whether the M2→LS or M1→LS contribute to the transition. For this purpose, we set up an inhibition during “transition” protocol, in which the press before the penultimate press of S1 triggered light inhibition for 2 s (Fig. 6A, for self-paced; in forced sequences, it was the third press; see Table 1, data related to Fig. 6).

Figure 6B shows the mean press in which the animals of the M2→LS or the M1→LS groups received the inhibitory stimulation (or the corresponding timestamp in the “off” trials); on average, it was the press −2 counting from the final press in the S1 (M2→LS_{transition light}on = −2.0 ± 0.1 vs M2→LS_{transition light}off = −1.9 ± 0.1, p = 0.25, Mann–Whitney U test; Fig. 6B). With this protocol, we observed that the inhibition of M2→LS, but not M1→LS, showed a tendency to increase the duration of the transition (M2→LS_Transitionon = 1.0 ± 0.06 vs M2→LS_Transition off = 0.8 ± 0.06, the paired mean difference was 0.08 [95.0%CI 0.02, 0.16] and p = 0.050, q = 0.4 of the two-sided permutation t test; Fig. 6C, upper panels), that reached significance when compared with the eYFP-control group: unpaired mean difference between control and M2→LS was 0.18 [95.0%CI −0.09, 0.3] and p = 0.003, q = 0.03 of the two-sided permutation t test; Fig. 6C, Δ on-off bottom panel). This effect was time-specific since neither the inhibition before nor the inhibition execution protocol on the M2→LS yielded a similar effect (Fig. 6D). This effect on the transition raised the question of whether the M2→LS projections treat each subsequence as independent chunks or concatenate them into a single chunk once the whole sequence has been acquired (Geddes et al., 2018; Martiros et al., 2018). We hypothesized that if the transition was affected by the light inhibition, the proportion of breaks in the sequences might increase as well. Contrary to this hypothesis, no modifications on the breakings of sequences were detected by this protocol (Fig. 6E). We also observed no differences in the number of presses or sequence length (Table 1), supporting the idea that the M2→LS projections contribute to the transition intersubsequences without affecting the second part of the serial order sequence.

Altogether, the results from the recording and inhibition experiments reveal that the cortico-striatal projections to the LS from the premotor and motor cortices have specific contributions to the execution of sequences. Both projections contribute to the initiation, while M2→LS also contributes to the correct execution and transition between subsequences of self-paced sequences. Importantly the M2→LS contribution was essential during the beginning of the execution, suggesting its contribution is relevant before the initiation, at the beginning of execution, and decreasing in relevance as the sequence execution progressed.

Premotor cortico-striatal projections innervate the direct pathway with a stronger synaptic weight than the indirect pathway

Finally, we investigated whether the M2→LS projections differentially impact the two subcircuits of striatal projection neurons, the direct or indirect striatal pathways (Wall et al., 2013), by looking at synaptic connectivity. For these experiments, we crossbred Emx1-Cre (allowing us to express ChR2 in M2) with BAC D2-GFP mice. By injecting red retrobeads into the substantia nigra, we could identify in the striatum neurons from direct pathway (red retrobeads; dMSN), or the indirect pathway (GFP; iMSN). This allowed us to record from identified striatal neurons in brain slices ex vivo while stimulating the M2→LS ChR2 axons (Extended Data Fig. 4-3). Using this procedure, we recorded pairs of neurons in the same brain slice in the LS (one iMSN and one dMSN; counterbalancing the recording order per slice). We evoked postsynaptic responses by activating the M2→LS ChR2 projections with brief pulses of light (1 ms; Extended Data Fig. 4-3C). From these experiments, we recorded a total of 6 pairs of striatal neurons (in 6 different slices from four animals). Postsynaptic evoked responses had a bigger amplitude in the dMSN than in the iMSN (dMSN: 290 ± 100 vs iMSN = 30 ± 9 pA, Mann–Whitney U test, p = 0.031; Extended Data Fig. 4-3F), with no difference in their latencies (dMSN: 3.3 ± 0.1 ms vs iMSN = 3.8 ± 0.1 ms, Mann–Whitney U test, p > 0.05; Extended Data Fig. 4-3F). These evoked postsynaptic responses were glutamatergic, being blocked by the AMPA receptor antagonist CNQX (Extended Data Fig. 4-3G). This finding suggests that the M2→LS projections may have stronger synaptic connections onto the dMSNs than onto the iMSNs in vivo when the animals are executing the serial order action sequences.