To test whether visual experience can evoke sequence representations in the visual cortex, we assigned mice to yoked experimental and control groups. On each of four training days, mice in the experimental group were shown 200 presentations of a single sequence of oriented sinusoidal gratings (termed ABCD, where each letter represents a unique orientation; Fig. 1a,b) and control animals were shown 200 random permutations of the same sequence elements (CBDA, DACB, etc.). On the fifth day, both groups were shown the trained sequence and a novel sequence constructed by reordering the same elements (DCBA). We measured visual evoked potentials (VEPs) in binocular layer 4 (see Online Methods) and found that ABCD elicited a markedly larger response after training than DCBA in the experimental group (Fig. 1c and Supplementary Fig. 1), but not in the control animals, which, as a result of the randomized nature of their training, had no reason to expect the sequence elements to appear in any particular order. Thus, repeated exposure to a visual sequence is sufficient to encode a neural representation of that sequence.

Figure 1: Learned spatiotemporal sequence representations in V1. (a) Schematic representation of head-fixed stimulus presentation. (b) On each of four training days, the experimental group (n = 6 mice) was shown 200 presentations of the sequence ABCD (where each letter indicates a uniquely oriented sinusoidal grating) and the control group (n = 4 mice) was shown 200 random permutations of the same sequence elements. Each element was held onscreen for 150 ms and sequences were separated by 1.5 s of gray screen. All mice were tested on the fifth day with the sequences ABCD, DCBA and ABCD 300 (subscript indicates a 300-ms element hold time). (c) Sequence-evoked local field potentials recorded on the fifth day showed that ABCD drives larger responses (blue) than DCBA (red) in the experimental mice, whereas there were no differential responses in control mice. Voltage traces represent the average response of all mice in each group and triangles mark the onset of each sequence element. (d) ABCD 300 drove relatively small responses in both groups. (e) Training regime had a significant effect on sequence response magnitude (quantified as the average peak-to-peak response to each of the four elements; Supplementary Fig. 1) potentiation (two-way RM-ANOVA, F 1,8 = 22.560, P = 0.001). There was a significant interaction (two-way RM-ANOVA, F 1,8 = 6.638, P = 0.008) between sequence and experimental group on day 5 and post hoc analysis revealed that the response to ABCD was significantly larger (indicated by asterisk) than either DCBA (t 5 = 5.738, P = 0.002) or ABCD 300 (t 5 = 4.923, P = 0.005). Sequence effects were not significant in the control group (P > 0.5 for all post hoc comparisons). Error bars show s.e.m. (f) Potentiation time course. (g) Sequence effects were evident in spiking neural activity. In this representative example, ABCD drove higher peak firing rates than DCBA (multi-unit spike rasters above peristimulus time histograms, dashed lines indicate element onset times). Full size image

The same mice were also tested with the familiar sequence presented with novel timing (ABCD 300 , where the subscript indicates that each stimulus element was held on the screen for twice the 150-ms duration used during training). The initial response to the first sequence element was very similar to that seen with the trained timing, but responses to subsequent sequence elements were clearly smaller (Fig. 1d). Comparing the average sequence evoked response magnitudes (Fig. 1e) confirmed what was qualitatively obvious from the VEP waveforms; in the experimental group, serial order and timing both strongly influenced evokedresponse magnitudes. The effects of reordering were not specific to sequence reversal; other tested sequence permutations also caused decreased response magnitudes similar to those shown in Figure 1 for DCBA. These data suggest that any manipulation of sequence content after training disrupts the response magnitude. In contrast, there was no effect of sequence order or timing in the control group, although there was a magnitude increase relative to day 1 (Fig. 1f). Sequence-specific effects were also visible in cortical spiking activity, as demonstrated by the trained sequence driving higher multi-unit spike rates than a novel sequence (Fig. 1g and Supplementary Fig. 2).

To further investigate the temporal specificity with which sequences can be learned and to rule out the possibility that there is something inherently special about the 150-ms timing used in the previous experiments, we trained a cohort of mice using a protocol in which the four sequence elements were held on-screen with alternating short and long durations (Fig. 2a). After training, the mice were tested with the trained sequence presented with both familiar (short-long-short-long) and novel (long-short-long-short) timing. Although the difference between familiar and novel timing was subtle, the cortical response to the trained sequence presented with familiar timing was larger than the response to either a reordered or re-timed sequence (Fig. 2b,c). That this specificity was a consequence of training was clear from the minimal effect of timing evident in responses driven by a novel sequence (Fig. 2c,d and Supplementary Fig. 3).

Figure 2: Sequence learning is temporally specific. (a) Mice (n = 13) were trained using ABCD presented with a short-long-short-long temporal profile. On the fifth day, the mice were tested with ABCD and DCBA presented with both familiar (black) and novel (long-short-long-short, gray) timing. (b) The largest responses occurred when the trained sequence was presented with the trained timing (top). Timing made little apparent difference when a novel sequence was shown (bottom). (c) There was a significant interaction between sequence order and timing (two-way RM-ANOVA, F 1,12 = 22.925, P < 0.001). Post hoc analysis revealed the response to ABCD with trained timing was significantly larger than ABCD with novel timing (t 12 = 8.760, P < 0.001). There was also a small effect of timing in DCBA (t 12 = 2.722, P = 0.012). Error bars show s.e.m. *P < 0.05. (d) The relative effect of timing as a function of sequence demonstrated by paired-response plots (dashed lines connect responses for single animals, black indicates the mean). Full size image

One notable aspect of this plasticity is the small amount of sensory experience necessary to potentiate the cortical response. The largest increase in sequence magnitude occurs after the first training day (Fig. 1f), at which point each mouse had seen the sequence only 200 times (corresponding to 2 min of active visual stimulation). This rapid change is similar to a form of cortical plasticity called SRP (stimulus-selective response potentiation), which is characterized by a daily increase in VEP magnitude following repeated exposure to a sinusoidal grating8. This increase is stimulus specific and involves local plasticity in V1 (refs. 8,9). Consistent with forms of learning that occur early in the visual processing hierarchy10, SRP does not transfer between the eyes. To determine whether sequence learning shares this property, we trained mice with sequence presentation restricted to one eye and tested them with monocular presentation to both eyes (Fig. 3a). Although there was a clear effect of sequence on cortical responses driven by the trained eye, learning did not transfer to the untrained eye (Fig. 3b,c). These findings indicate that the modifications elicited by training occur at a site where information from the two eyes can be separated.

Figure 3: Learning does not transfer between eyes and requires muscarinic acetylcholine receptors in V1, but not NMDA receptors. (a) Mice (n = 8) were trained with an occluder restricting visual stimulation to the left eye (LE). Responses were recorded in the hemispheres contralateral and ipsilateral to the viewing eye. On the fifth day, sequences were presented to both eyes. (b) ABCD drives larger responses than DCBA in both hemispheres only when viewed through the trained eye. (c) There was a significant interaction between viewing eye and sequence in both hemispheres (two-way RM-ANOVA, contra: F 1,7 = 25.041, P < 0.001; ipsi: F 1,7 = 10.426, P = 0.002). The response to ABCD was significantly larger than DCBA in both hemispheres only when viewed through the trained eye (contra: t 7 = 8.246, P < 0.001; ipsi: t 7 = 5.091, P < 0.001). (d) Systemic CPP treatment (left, n = 9, 30–60 min before stimulus presentation during training) had no significant effect (two-way RM-ANOVA, F 1 = 1.660, P = 0.220) on sequence potentiation compared with vehicle (n = 6) and response potentiation was significant within both treatment groups (main effect: F 1,13 = 35.525, P < 0.001; post hoc analysis: t 8 = 3.186, P = 0.007 and t 5 = 5.093, P < 0.001). The same CPP blocked subsequent SRP induction in the same mice (right, regrouped after washout, CPP n = 6, vehicle n = 9). There was a significant interaction between treatment and SRP recording session (two-way RM-ANOVA, F 1,13 = 42.210, P < 0.001) and potentiation was significant only in the vehicle control group (t 8 = 9.692, P < 0.001). (e) Muscarinic receptor antagonism during training blocked sequence potentiation. There was a significant day 5 interaction between treatment and sequence in scopolamine-treated (n = 5) and vehicle-treated (n = 5) mice (two-way RM-ANOVA, F 1,8 = 5.827, P = 0.013) and ABCD was significantly larger than DCBA (t 4 = 3.661, P = 0.004) or ABCD 300 (t 4 = 3.813, P = 0.005) only in vehicle-treated mice. (f) Local unilateral infusion of scopolamine in V1 (n = 7 mice) blocked potentiation relative to the opposite vehicle-treated hemisphere (two-way RM-ANOVA, F 1,6 = 30.189, P = 0.002). Error bars show s.e.m. *P < 0.05. Full size image

SRP is mechanistically similar to classical long-term synaptic potentiation, including the requirement for NMDA receptor activation8. To test whether sequence learning shares similar mechanisms and might represent a higher order expression of SRP, we systemically treated mice with either the NMDA receptor antagonist 3-(2-carboxypiperazin–4yl)propyl-1-phosphonic acid (CPP, 10 mg per kg of body weight, intraperitoneal) or saline before sequence presentation on each training day. Notably, expression of sequence learning was comparable between CPP-treated and control mice (Fig. 3d). To confirm the effectiveness of the CPP in blocking NMDA receptors under our experimental conditions, we subsequently reassigned the same mice after a 3-d washout period into new CPP and vehicle control groups and exposed them to the SRP induction protocol. We found that the same CPP prevented induction of SRP (Fig. 3d and Supplementary Fig. 4). Thus, sequence learning is a phenomenon distinct from SRP and does not require NMDA receptor activation.

Several forms of experience-dependent plasticity in V1 have been shown to require the cholinergic input arising from the basal forebrain11,12. To test whether sequence potentiation requires acetylcholine, we systemically treated mice with either the muscarinic receptor antagonist scopolamine or vehicle. Mice in the scopolamine-treated cohort showed no evidence of sequence potentiation over the training period or recognition of the trained sequence on day 5 (Fig. 3e and Supplementary Fig. 5). Likewise, local microinfusion of scopolamine into V1 of one hemisphere blocked potentiation in that hemisphere even as the vehicle-treated hemispheres of the same mice potentiated normally (Fig. 3f and Supplementary Fig. 6). These results demonstrate the involvement of the cortical cholinergic system in the mechanisms underlying sequence learning.

It is clear from these data that the mice learned neural representations of the familiar visual sequence, but it is not clear whether this representation was sufficient to reproduce the sequence absent external stimulation. To test this possibility, we trained a cohort of mice with the sequence ABCD and tested with two sequences where the second element was omitted and replaced by a gray screen (Fig. 4a). In the first test sequence (A_CD) the omitted element was preceded by A, established during training to predict element B, whereas the second test sequence (E_CD) was initiated by a novel element E that had not been established to predict anything. The cortical response to a gray screen preceded by E was small and consisted solely of a late positive-going bump (Fig. 4b). In contrast, the response following A shared a similar morphology and timing with the response actually evoked by the element B: the average latency to peak negativity during the second element was almost identical when the response was driven by element B (60.8 ± 2.6 ms) or anticipatory based on the presence of element A (60.3 ± 3.5 ms). There was no statistical difference in the average sequence magnitude between ABCD and A_CD (t 6 ± 0.964, P ± 0.354), but both were larger than E_CD (Fig. 4c). Restricting statistical analysis to the second element revealed that the anticipatory response following the predictive element A, while smaller than the response to the actual element B, was larger than the response following the nonpredictive element E. The data therefore suggest that a memory of stimulus element B is recalled in V1 when it is cued by stimulus element A.

Figure 4: Learned sequence representations are predictive and involve multiple cortical layers. (a) Mice (n = 7) were trained with ABCD and tested with two sequences, A_CD and E_CD, where the second element was replaced with a gray screen. (b) When the omitted element was preceded by A (red), a response occurred in position 2 (marked by the dashed gray box) that was similar in form and latency to the response evoked when B was actually presented (blue). This predictive response was absent when the omitted element was preceded by the novel element E (green). (c) There was a significant effect of sequence on both the average magnitude across the four elements (left, one-way RM-ANOVA, F 2,6 = 12.186, P = 0.001) and the response of the second element alone (right, one-way RM-ANOVA, F 2,6 = 31.597, P < 0.001). Significant differences determined by post hoc analysis are indicated by brackets (full sequence: t 6 = 4.675, P = 0.002 and t 6 = 3.711, P = 0.006; Elmnt 2: t 6 = 4.175, P = 0.003 and t 6 = 3.771, P = 0.003). Error bars show s.e.m. *P < 0.05. (d,e) Laminar field recordings (d) and CSD analysis (e) showed characteristic activation patterns evoked by trained and novel sequences. The DCBA sink-source pattern was similar to ABCD, but with smaller magnitudes. Activation patterns during omitted elements (marked gray triangles) closely matched those produced by real stimuli when the sequence was initiated with A, but not E. When each sequence element was held onscreen for twice the trained duration, activation patterns resembling those that would have occurred had element B been shown appeared at the expected time (highlighted with dashed gray box). Full size image

To investigate how sequence evoked activity varies as a function of cortical depth, we implanted mice with linear arrays of 16 recording electrodes spanning the cortical layers from the surface to the white matter and trained them as before on the sequence ABCD (Online Methods). Sequence-driven VEPs spanned the cortical depth with positive-going responses in the superficial layers and relatively large negative-going responses in the middle and deeper layers. Both the familiar and novel sequences evoked clear responses, although those driven by the trained sequence were larger in all layers (Fig. 4d). Current source density (CSD) analysis, which estimates current source and sink locations and magnitudes by calculating the second spatial derivative of recorded voltages13, was performed to determine the laminar distribution and temporal order of the transmembrane currents that produced the recorded field potentials. The earliest current sinks driven by the first sequence element occurred in thalamorecipient layers 4 and 6 approximately 50 ms after stimulus onset (Fig. 4e). These sinks then spread to layers 2/3 and were followed by deep layer sources. This characteristic activation pattern, with an additional initial superficial current sink, was repeated for subsequent sequence elements and was approximately the same, albeit with different magnitudes, for both ABCD and DCBA. CSD analysis also revealed the cued activation described above, with a clear differentiation between the omitted element response following the predictive element A and nonpredictive element E. Activation characteristic of the B response was also observed when A was held onscreen for twice the trained duration. The observation that anticipatory current sinks can be resolved at short latencies in the thalamorecipient layers suggests the possibility of anticipatory activation of thalamic relay neurons via corticothalamic feedback.