Science - 31 January 2020

to evaluate the role of D2-SPNs in flexible
learning that should not overtly depend on a
negative reward prediction error. We induced
subtle changes in the identity predictions of
preexisting action-outcome relationships by
reversing the outcome congruence between
pairs of action-outcome associations. We trained
mice with bilateral D2-SPN ablations in the
DMS (fig. S8, D and E) and their sham controls
on two action-outcome associations (A1-O1 and
A2-O2), which generated identical performance
in both groups (Fig. 5A and table S5). We then
verified whether both A-O contingencies had
been correctly encoded by giving the mice an
outcome-specific devaluation test, which eval-
uated the effect of sensory-specific satiety on
one or the other outcome on choice between
the two trained actions ( 13 )(Fig.5B).Both
groups correctly encoded the initial contin-
gencies (i.e., A1-O1 and A2-O2); satiety on one
of the outcomes (e.g., O1) reduced performance
of the action associated with that outcome in
training (A1; devalued) relative to the other,
still valued, action (A2; valued) (Fig. 5B and
table S5). We then explored whether these mice
could incorporate new information by train-
ing them with the outcome identities reversed
for 5 days (i.e., A1-O2 and A2-O1) (Fig. 5C and
table S5) prior to a second outcome-specific
devaluation test (Fig. 5D). Whereas Sham mice
were able to show flexible encoding and could
adjust their choice according to the new A-O
associations, mice with D2-SPN ablation failed
to do so (Fig. 5D and table S5).

Discussion

One of the most intriguing characteristics of
the striatum is the random spatial distribution

and high degree of intermingling between its

D1 (direct) and D2 (indirect) projection systems,

a feature that is actively promoted developmen-

tally ( 14 ) and that has been retained throughout

evolution ( 15 ). The result is a highly entropic

binary mosaic that extends through an expan-

sive and homogeneous space and that is mostly

devoid of histological boundaries ( 16 ). Such

organization is unusual in the brain and can

be seen as an adaptation to provide an optimal

postsynaptic scaffold for the integration of

regionally meaningful neuromodulatory sig-

nals ( 17 ). In such a plain, borderless environ-

ment, the rules established locally by D1 and

D2-SPNs are likely to be critical in defining

functional territories throughout the striatum,

and this, we propose, is the key process shaping

striatal-dependent learning.

Our study suggests that the striatum takes

full advantage of this“one-to-one”binary

mosaic structure, in which activated D2-SPNs

access and modify developing behavioral pro-

grams encoded by regionally defined ensembles

of transcriptionally active D1-SPNs (what we

call D2-to-D1 transmodulation). We propose

that this process is slow, as it depends on the

molecular integration of additive neuromo-

dulatory signals ( 5 ), but could, with time, create

the functional boundaries necessary to identify

and shape specific learning in the striatum. A

goodexampleofthissortofdynamic,persist-

ent neuromodulation is the recently described

“wave-like”motion of DA signals throughout

the mediolateral axis of the striatum ( 17 ). Beyond

offering a broad solution to the credit assign-

ment problem, recurrent waves of neuromod-

ulatory activity in defined striatal areas could

provide the kind of unbiased signal that, in

the context of the molecular dichotomies es-

tablished by D1 and D2 receptors ( 8 ), shape

the striatal mosaic into meaningful transcrip-

tional motifs. In the case of extinction learning,

as observed here, noisy alternations between

DA-rich and DA-lean states within the DMS

appear to generate a mixed population of

activated SPNs comprising both D1 and D2

systems. This regional overlap lays the ground-

work for the local one-to-one modulation that

shapes and integrates new learning, limiting

outdated D1-SPN function in the case of ex-

tinction learning, and segregating new and

existing territories of plasticity in the case of

action-outcome identity reversal.

REFERENCES AND NOTES

1. M. E. Bouton,Psychopharmacology 236 ,7–19 (2019).


2. M.E.Bouton,B.W.Balleine,Behav. Anal. 19 ,202–212 (2019).


3. B. W. Balleine,Neuron 104 ,47–62 (2019).


4. A. Mohebiet al.,Nature 570 ,65–70 (2019).


5. P. Greengard,Science 294 , 1024–1030 (2001).


6. A. Stipanovichet al.,Nature 453 , 879–884 (2008).


7. C.R.Gerfen,D.J.Surmeier,Annu.Rev.Neurosci. 34 , 441–466 (2011).


8. J. Bertran-Gonzalezet al.,J. Neurosci. 28 , 5671–5685 (2008).


9. M. J. Wanat, I. Willuhn, J. J. Clark, P. E. M. Phillips,Curr. Drug
   Abuse Rev. 2 , 195–213 (2009).


10. H. H. Yinet al.,Nat. Neurosci. 12 , 333–341 (2009).


11. A. E. McGovernet al.,J. Neurosci. Methods 209 , 158–167 (2012).


12. A. E. McGovernet al.,J. Neurosci. 35 , 7041–7055 (2015).


13. B. W. Balleine, A. Dickinson,Neuropharmacology 37 ,407– 419
    (1998).


14. A. Tinterriet al.,Nat. Commun. 9 , 4725 (2018).


15. S. Grillner, B. Robertson,Curr. Biol. 26 , R1088–R1100 (2016).
    16.G. Gangarossaet al.,Front. Neural Circuits^7 , 124 (2013).


16. A. A. Hamid, M. J. Frank, C. I. Moore,bioRxiv729640 [Preprint]
    (2019).
    ACKNOWLEDGMENTS
    We thank Z. Skrbis for technical assistance.Funding:This work was
    supported by the Australian Research Council (Grants DE160101275 to
    J.B.-G., DP19010251 to J.B.-G. and M.M. and DP150104878 to B.W.B.)
    and by a Fellowship from the NHMRC of Australia to B.W.B.
    (GNT1079561).Author contributions:M.M., B.W.B. and J.B.-G.

Matamaleset al.,Science 367 , 549–555 (2020) 31 January 2020 6of7

Fig. 5. D2-SPNs control the updating of learning.
Bilateral genetic lesions of D2-SPNs were performed
in the DMS ofadora2a-Cre::drd2-eGFP hybrid mice
(fig. S8, D and E) (eight mice per group). Initial
learning: (A) Sham and Lesioned mice were trained
to two action–outcome (A-O) contingencies, resulting
in increased performance (press/min) across days.
(B) Initial devaluation test: a choice (A1 versus A2)
was presented after having sated the mice on one
or the other outcome (O1/O2) over consecutive days.
Graph shows performance on the valued (blue:
provides nonsated O) and devalued (gray: provides
sated O) levers. Additional learning: (C) Mice were
then trained to the reversed A-O contingencies,
which rapidly increased press/min performance.
(D) A new round of devaluation and choice tests
were presented [as in (C)]. *, significant overall
effect (black) and interaction (red). n.s.,
not significant (table S5).

Initial training

Initial devaluation

Reversal training

A1 O1

A2 O2

New devaluation

O1: A1 vs. A2

A1 O1

A2 O2

Devalued Valued

Valued Devalued

Valued Devalued

Devalued Valued

??

??

O2: A1 vs. A2

A1 O1

A2 O2

O1: A1 vs. A2

O2: A1 vs. A2

Days 1-14

Days 15-16

Days 17-21

Days 22-23

0

2

4

6

8

10

12

14

0

3

6

9

A

B

C

D

Press/min

Sham Lesioned

Press/min

Training days

n.s. Press/min

Valued

Devalued

Valued

Devalued

Sham Lesioned

Sham

Lesioned

n.s.

Press/min

Training days

n.s.

Sham

Lesioned

02040 60 80

3

9

11

7

1

5

13

17

18

19

20

21

020 60 8040

RESEARCH | RESEARCH ARTICLE