The intellectual disability protein RAB39B selectively regulates GluA2 trafficking to determine synaptic AMPAR composition

RAB39B is a member of the RAB family of small GTPases that controls intracellular vesicular trafficking in a compartment-specific manner. Mutations in the RAB39B gene cause intellectual disability comorbid with autism spectrum disorder and epilepsy, but the impact of RAB39B loss of function on synaptic activity is largely unexplained. Here we show that protein interacting with C-kinase 1 (PICK1) is a downstream effector of GTP-bound RAB39B and that RAB39B-PICK1 controls trafficking from the endoplasmic reticulum to the Golgi and, hence, surface expression of GluA2, a subunit of alpha-amino-3-hydroxy-5-methyl-4-isoxazole propionic acid receptors (AMPARs). The role of AMPARs in synaptic transmission varies depending on the combination of subunits (GluA1, GluA2 and GluA3) they incorporate. RAB39B downregulation in mouse hippocampal neurons skews AMPAR composition towards non GluA2-containing Ca2+-permeable forms and thereby alters synaptic activity, specifically in hippocampal neurons. We posit that the resulting alteration in synaptic function underlies cognitive dysfunction in RAB39B-related disorders.

T he human RAB39B gene, which maps to the distal Xq28 locus, encodes RAB39B, a member of the RAB GTPases, small monomeric GTPases belonging to the RAS-like GTPase superfamily, that have a key role in the regulation of intracellular vesicular trafficking 1 . Indeed, RAB GTPases are physically associated with specific organelles, through specific effector protein interaction, and act as a network to regulate both spatially and temporally the transport of specific vesicles switching from the active GTP-bound and the inactive GDPbound state 2,3 . RAB and RAB-associated proteins have been shown to play an important role in a number of rare monogenic as well as multifactorial diseases 4,5 characterized by cognitive deficit. In particular, previous work on X-linked intellectual disability (XLID) has identified mutations in the GDI1 gene that cause loss of function of the RAB-interacting protein aGDI 6,7 . RAB GTPases and interacting proteins are abundantly expressed in the central nervous system across species, and Gdi1 mouse models have been relevant to demonstrate the causal link of aGDI loss of function with deficits in glutamate release and cognitive impairment 8,9 . Our group previously demonstrated that loss-offunction mutations in human RAB39B gene are associated with XLID comorbid with autism spectrum disorder and seizures 10 and also comorbid with early onset of Parkinson's disease 11 . Moreover, it was reported that also a 0.5-Mb tandem copynumber gain at distal Xq28 including RAB39B was linked to XLID 12 . We defined that RAB39B is a neuronal-specific protein localized at the Golgi compartment, but its functional role is unknown 10 . The demonstrated link with a human genetic disease urged us to define the functional role of RAB39B in neurons and the potential link to cognitive function.
In the present study, we define that RAB39B interacts with protein interacting with C-kinase 1 (PICK1), which became an attractive target for further studies. Previous studies defined the functional role of PICK1 in synaptic alpha-amino-3-hydroxy-5methyl-4-isoxazole propionic acid receptor (AMPAR) surface expression, trafficking and post-synaptic targeting [13][14][15][16][17][18] , but the mechanism of AMPAR biogenesis, assembly and trafficking is poorly understood. In hippocampal CA1 pyramidal neurons, AMPARs are hetero-tetramers formed in the endoplasmic reticulum (ER) [19][20][21] mostly composed of GluA1/GluA2 or GluA2/GluA3 hetero-dimers 22 . Export of AMPARs from the ER requires the interaction of the C-terminal domain of AMPAR subunits with other proteins. In fact, synapse-associated protein 97 (SAP97) interacts with the immature GluA1 C-terminal domain 23 early in the secretory pathway 24 , while GluA2 C-terminal domain interacts with PICK1, which is necessary for GluA2 exit from the ER. Thus, PICK1 selectively binds GluA2/GluA3 heterodimers for exit from the ER to Golgi compartment 25 .
Given the RAB39B-PICK1 interaction and the Golgi localization of RAB39B, here we hypothesize that RAB39B might be a key molecule fundamental to regulate a specific secretory trafficking step of GluA2 AMPAR subunit. Supporting this idea, Rab39b-silenced mouse hippocampal neurons show increased GluA2 immature form suggesting ER retention. This result provokes, as a final step, a reduction in the amount of GluA2 AMPAR subunits at the post-synaptic membrane, leading to the formation of GluA2-lacking, Ca 2 þ -permeable AMPARs, often associated with immature synapses and cognitive impairment 26 .
We report here that RAB39B is the key molecule regulating the translocation of GluA2/GluA3 heterotetramers into the Golgi. Indeed, in the absence of RAB39B, AMPARs arrangement at the post-synaptic site is misregulated, which provides an explanation for the involvement of RAB39B mutations in the aetiology of ID and autism spectrum disorder.

Results
GTP-bound RAB39B interacts with PICK1. To understand the causal link between XLID and gene defects associated with RAB39B, we set out to map the role of RAB39B in vesicular trafficking. As a first step, we searched for downstream effectors of RAB39B by use of a yeast two-hybrid screening ( Supplementary Fig. 1a-c). With the human full-length RAB39B protein as bait, by screening a human fetal brain cDNA library we identified a protein originally described as protein interacting with C-kinase 1 (PICK1) as the strongest interacting molecule. In agreement with our results from the two-hybrid screen, recombinant GST-RAB39B locked in the active state with the nonhydrolysable GTP analogue GTPgS and the dominant active GST-RAB39B Q68L efficiently pulled down PICK1, as compared with GDP-bound or dominant-negative GST-RAB39B S22N, using whole-brain lysate of murine origin (Fig. 1a,b). Instead, recombinant GST-RAB39A, sharing 76% identity at protein level with RAB39B, did not pull down PICK1 (Supplementary Fig. 1d).
GST-RAB39B GTPgS likewise pulled down PICK1, obtained by in vitro transcription and translation, which indicated that PICK1 interacts directly with RAB39B, and mostly in its GTPbound state (Fig. 1c). Furthermore, myc-tagged PICK1 coimmunoprecipitates with flag-tagged RAB39B from lysates of HEK293T cells in which the two constructs were co-expressed (Fig. 1d). Finally, we performed the reverse pull down using GST-PICK1 and found that it pulled down RAB39B from the mouse brain lysate, as well as previously identified interacting proteins PKCa 27,28 and the GluA2 AMPAR subunit 29 (Fig. 1e).
RAB39B effector-binding region interacts with PICK1 PDZ domain. To define how the two proteins interact at the molecular level, COS7 cells were transfected with RAB39B or PICK1 mutant constructs and interactions were tested again by pull down assays. Since GST-PICK1 did not pull down RAB39B construct lacking a stretch (flag-DER, aa [35][36][37][38][39][40][41][42][43][44][45][46][47][48][49] corresponding to the canonical effector-binding region as present in all RAB proteins 30 , we concluded that the association between RAB39B and PICK1 is a conventional RAB-effector protein interaction (Fig. 2a). To better define the residues involved in the PICK1-RAB39B effectorbinding region, RAB39B constructs carrying different point mutations were used (flag-D38A, flag-V41G42V43AAA or flag-D44F45AA). Such RAB39B constructs were generated based on the predicted structural complex between RAB39B and PICK1 PDZ (see below). Consistently with structure predictions, GST-PICK1 did not pull down RAB39B construct carrying D44 and F45 mutated to alanines (aa D44F45AA), indicating that such amino-acid pair in the N-terminal half of the b2-strand of RAB39B play a central role in the association between RAB39B and PICK1 (Fig. 2a). PICK1 is a PDZ-and BAR-domain-containing protein. GST-RAB39B GTPgS pulled down recombinant myc-tagged PDZ domain (myc-PDZ; aa 23-99), but not flagtagged recombinant PICK1 lacking the PDZ domain (flag-D 121, aa 1-121; Fig. 2b). Thus, the interface between the two interacting proteins involves amino acids D44 and F45 of the ER domain of RAB39B and, respectively, the PDZ domain of PICK1.
The PDZ domain of PICK1 has been shown before to interact with a variety of proteins 31,32 , and point mutations within the PDZ domain of lysine 83 (K83) or of both lysine 27 (K27) and aspartate 28 (D28) with alanines abolished protein binding 29,31,33,34 . Since GST-RAB39B GTPgS pulled down fulllength myc-PICK with the K83A substitution (myc-K83A), but not with the K27D28AA substitutions (myc-KDAA), we concluded that K27 and D28 in the PDZ domain is key for the interaction of PICK1 with RAB39B (Fig. 2b).
Next, we then demonstrated receptor trafficking to the plasma membrane by TIRF. Co-expression of flag-RAB39B did not drive the plasma membrane translocation of either GFP-GluA1 or GFP-GluA2 (Fig. 5a,b). Remarkably, in triple transfected cells, flag-RAB39B mediated the GFP-GluA2 translocation to the plasma membrane in the presence of myc-PICK1 ( Fig. 5d) but not GFP-GluA1 (Fig. 5c).
All together our results clearly indicate that the effectorbinding region of RAB39B interacts with the PDZ domain of PICK1, also required for interaction with the GluA2 AMPAR subunit 36 . We also demonstrated that the RAB39B À PICK1 interaction is necessary for the RAB39B-driven trafficking of GluA2 cargo. PICK1-mediated functional linkage between RAB39B and GluA2-containing AMPAR relies on PICK1 dimer to act as a scaffold that hosts the other two proteins. In fact, GluA2 co-immunoprecipitates with PICK1 and RAB39B in a wild -type mouse brain lysate, but not in the Pick1-knockout brain, demonstrating the existence of the endogenous complex and that PICK1 acts as the bridge between RAB39B and GluA2 (Fig. 5e).
A model of interactions among the RAB39B-PICK1-GluA2. To gain insight, at the atomic level, into the supramolecular organization of RAB39B, PICK1 and GluA2 AMPAR subunit, we modelled the structure of RAB39B:2PICK1:GluA2 complex ( Fig. 6a and Supplementary Methods). This required comparative modelling and rigid-body docking to predict structure and interaction modes of PICK1 and RAB39B (see Methods). In contrast, the complex between the GluA2 C-terminus (GluA2Ct; aa 858-ESVKI-862) and PICK1 was extracted from a conformational ensemble solved by NMR 37 . In our structural model, the N-terminal half of the b2-strand of RAB39B (that is, amino-acid stretch 43-47 of the effector-binding region) interacts with the PDZ domain of PICK1 in a similar and partially overlapping manner as GluA2Ct, which is a b-strand as well (Fig. 6a-c). The major portions of RAB39B participating in the interface with PICK1 PDZ include (a) the 43-47 segment of the b2-strand, which makes inter-backbone and inter-side chain interactions with PICK1 bB and with aB; (b) switch 1 (swI) that interacts with the PICK1 bA/bB loop and aB; and (c) the switch 2 (swII) that interacts with PICK1 bC as well as the bA/bB and bB/aA loops (b and c). The predicted complex highlights D44 and F45 in the N-terminal half of the b2-strand as playing a central role in the interface, D44 making a salt bridge with K83 of PICK1 PDZ. Such a centrality of D44 and F45 has been validated by in vitro experiments (Fig. 3). Remarkably, the PICK1 bA/bB loop of PDZ holds the K27-D28 pair that we found essential for PICK1 À RAB39B interaction. Structure predictions indicate D28 as more important than K27 in RAB39B recognition. Indeed, K27 is buried at the bB/aB interface, its positively charged nitrogen atom being about 9 Å from the carboxylate oxygen atoms of D38 on the swI of RAB39B, and we postulate that the effect of the K27A mutation on RAB39B recognition, if any, is indirect. In contrast, D28 recognizes the G protein via a salt bridge with R70 in swII. Structure prediction therefore suggests that the interactions of RAB39B and GluA2 with the PDZ domain of PICK1 are mutually exclusive. In line with evidence that PICK1 can dimerize 33 , the structural model suggests that dimerization of PICK1 is a prerequisite for simultaneous recognition of both RAB39B and GluA2 each by one of the PICK1 molecules in the PICK1 dimer (Fig. 6a-c). The existence of such complex is supported by our co-immunoprecipitation experiments shown above.
To assess whether the RAB39B-PICK1 tandem directs trafficking of GluA2/GluA3 hetero-tetramers in their endogenous neuronal context, we first analysed the intracellular localization of RAB39B and PICK1 in flag-RAB39B-transfected primary mouse  hippocampal neurons ( Supplementary Fig. 3). Flag-RAB39B appears to partially co-localize with PICK1 in the cell body, instead of along dendrites ( Supplementary Fig. 3a). In a similar manner, flag-RAB39B and PICK1 co-localize in part with the ER or Golgi markers, calreticulin and GM130, respectively ( Supplementary Fig. 3b,c). No co-localization was observed between flag-RAB39B and adaptor protein 2 (AP2), which was used as a negative control 10 . Although RAB39B and PICK1 are certainly involved in many other intracellular pathways, our data support that RAB39B-PICK1 interaction has a role in cargo trafficking between the ER and the Golgi compartments. We then analyse the effect of Rab39b downregulation in transduced murine primary hippocampal neurons with previously described shRab39b and shScramble lentiviral particles 10 . Silencing with shRab39b led to a 40% downregulation of RAB39B expression levels (mean ± s.e.m., shScramble: 1.1 ± 0.13, shRab39b: 0.68±0.08; n ¼ 3; t-test P ¼ 0.035), without affecting protein expression levels of PICK1, GluA1, GluA2 and GluA3 ( Supplementary Fig. 4).
We first determined the effect of RAB39B silencing on the intracellular localization of PICK1 GluA1, GluA2 and GluA3 in the murine hippocampal neurons by immunofluorescence at 14 DIV. We found an increase in PICK1, GluA2 and GluA3 levels in the cell body of shRab39b-treated neurons as compared with the neurons treated with the control shRNA (mean±s.e.m., PICK1 shScramble n ¼ 21  Fig. 7a and Supplementary Fig. 5a), which suggests that there is a defect in protein trafficking when levels of RAB39B are downregulated. Next, we analysed the effect of RAB39B silencing on the intracellular trafficking in murine hippocampal neurons of the AMPAR subunits GluA1, GluA2 and GuA3 by examining their glycosylation status. Both AMPAR subunits are N-glycosylated and their N-glycans are subsequently modified in the ER before they are targeted to the plasma membrane. We therefore deglycosylated lysates from RAB39B silenced or control neurons with either Endo-b-N-acetylglucosaminidase H (EndoH), which removes only unmodified N-glycans 25 , or, as a control, with peptide-N-Glycosidase F (PNGasef), which removes all N-glycans, before analysis of the AMPAR subunits by immunoblotting. In the control lysates two isoforms, differing in their mobility in gel, were detected for GluA1, GluA2 and GluA3 on treatment with EndoH: an upper band, corresponding to the mature isoform (as it acquired partial EndoH resistance), and a lower band, corresponding to the immature isoform (as it acquired no EndoH resistance). The ratio between the two bands thus is a measure for AMPAR subunit maturation. As such, we determined that GluA2 and GluA3 maturation was impaired in shRab39b-treated neurons compared with the control neurons (mean±s.e.m., GluA2 shScramble: 0.96±0.05, shRab39b: 0.52 ± 0.05; n ¼ 3; t-test P ¼ 0.002; GluA3 shScramble: 1.0±0.07, shRab39b: 0.0±0.0; n ¼ 3; t-test P ¼ 0.0001; Fig. 7b), while GluA1 maturation was not affected by the shRab39b silencing (mean±s.e.m., shScramble: 1.29±0.21, shRab39b: 1.28 ± 0.15; n ¼ 3; t-test P ¼ 0.9; Fig. 7b).
These results indicate that in neurons the RAB39B-PICK1 tandem directs trafficking specifically of GluA2/GluA3 heterotetramers to the Golgi and, consequently, regulates GluA2 cell surface expression and spine morphology. RAB39B determines AMPAR composition and synaptic transmission. Since our results so far demonstrated that RAB39B is key to ensure efficient trafficking of GluA2/GluA3 to the Golgi compartment, and, hence, its cell surface expression, we reasoned that ultimately RAB39B controls the availability of GluA2 and thus AMPAR composition at the post-synapses. Because the subunit composition, in turn, determines Ca 2 þ permeability of the AMPARs and therefore impacts on synaptic transmission, we asked whether RAB39B-PICK-directed trafficking of GluA2/ GluA3 is key for glutamatergic synapses. To this end, we performed electrophysiological recordings on primary murine hippocampal neurons that were infected with the shRab39b or control shScramble lentiviruses at low doses to ensure that most pre-synaptic inputs would come from non-infected (unaltered)  Fig. 6a,b). The amplitude of miniature EPSCs was slightly but significantly larger in RAB39Bsilenced neurons (mean ± s.e.m. in pA, shScramble n ¼ 14: À 17.1 ± 1.1, shRab39b n ¼ 13: À 22.9 ± 1.9; t-test Po0.01) (Fig. 8a-c). The miniature frequency, however, was unaffected, suggesting that indeed RAB39B silencing occurred mostly in post-synaptic neurons due to the low-density lentiviral infection (Fig. 8c). Interestingly, kinetic analysis of the miniature currents indicated that RAB39B silencing accelerated decay (mean±s.e.m. in ms, shScramble n ¼ 14: 2.60 ± 0.32, shRab39b n ¼ 13: 1.71±0.05; t-test Po0.01) but not rise times of synaptic currents (Fig. 8b,c). The specificity of shRab39b effect was determined using a rescue experiment. Electrophysiological recordings were carried out on primary murine hippocampal neurons that were infected with the shRab39b or control shScramble lentiviruses at multiplicity of infection (MOI) 1 to ensure that pre-synaptic inputs and post-synaptic outputs come from infected (silenced) neurons and after 6 days in culture transfected with a flag-RAB39B-rescue, as previously described 10 ( Supplementary  Fig. 6c,d). The reintroduction of RAB39B recovered the amplitude and kinetic of the miniature currents ( Fig. 8d-g). Since RAB39B directs GluA2 traffic and thus cell surface expression, we attributed the electrophysiological changes on RAB39B silencing to an altered subunit composition of synaptic AMPARs. To test this hypothesis, we monitored inward rectification of AMPAR-mediated synaptic responses in organotypic hippocampal slices from mice that were injected in the CA1 with the relevant lentiviruses (shScramble or shRab39b) by evoked EPSCs, recorded at À 60 mV and þ 40 mV. The ratio of synaptic responses at þ 40 mV over responses at À 60 mV was calculated to serve as a rectification index (R.I.). We found a significant increase in the R.I. of RAB39B silenced neurons (mean±s.e.m., uninfected n ¼ 6: 1.801±0.175; shScramble n ¼ 6: 1.403 ± 0.166; shRab39b n ¼ 6: 2.998 ± 0.297; ANOVA: Po0.0001) (Fig. 8h,i), indicative of enrichment at synapses of inward rectifying AMPARs. Since GluA2-lacking AMPARs in fact are inward rectifying 38 we concluded that the alterations in synaptic transmission are due to a lowered GluA2 availability. This interpretation also fits with our observation that RAB39B down regulation leads to increased miniature EPSC amplitude ( Fig. 8a-g), in agreement with the higher conductance of GluA2lacking AMPARs 39 . Finally, we tested whether long-term depression (LTD) was altered in RAB39B down regulated cells, given the involvement of PICK1 in activity-dependent AMPAR endocytosis and LTD 15,40 . To this end, we assessed NMDARdependent LTD in shRab39b-, shScramble-and non-infected CA1 neurons (Fig. 8j). After LTD induction (500 stimuli at 1 Hz, with postsynaptic depolarization at À 40 mV), AMPAR-mediated synaptic responses were equally depressed in the three conditions. These results suggest that, in contrast with PICK1, RAB39B is not required for LTD. Therefore, RAB39B does not appear to be involved in activity-dependent AMPAR endocytosis or recycling.

Discussion
Our results define RAB39B in tandem with its downstream effector PICK1 as a key molecule dedicated to GluA2 AMPAR subunit trafficking. We demonstrate a molecular complex where the dimerization of PICK1 is a prerequisite for simultaneous recognition of both RAB39B and GluA2 by each one of the PDZ domains of PICK1, in a PICK1 dimer conformation.
GluA1/A2 heteromers are mostly confined to post-ER compartments. Finally, to become competent for ER export, AMPARs need to co-assemble with auxiliary proteins. In the brain, auxiliary subunits of the stargazing/TARP (transmembrane AMPAR regulatory protein) family and the cornichon family are physically associated with the channel and regulate their trafficking and gating [45][46][47][48][49] . Recently, Cornichon-2/-3 was described as cargo receptors to selectively bind GluA1, and not GluA2, in hippocampal neurons, allowing the forward trafficking of GluA1-containing AMPARs from the ER to Golgi and finally their synaptic expression 50 . Instead, our study identifies RAB39B as the first RAB GTPase regulating a vesicular trafficking step for the transport of GluA2/GluA3 heteromers into the secretory pathway.
Finally, our analysis of the steady-state levels of AMPAR subunits at the post-synaptic membrane of Rab39b silencing hippocampal neurons demonstrates that the alteration in the RAB39B-mediated secretory pathway leads to a decrease in surface GluA2 density and an increase in GluA1 AMPAR subunits. This observation was confirmed functionally by the ARTICLE analysis of the rectification properties of AMPAR-mediated synaptic transmission resulting in an increased rectification index suggestive of an increased fraction of GluA2-lacking AMPARs at synapses related to the increased amplitude of miniature EPSCs. Because it was previously described that, in the hippocampus, heterotetramers of GluA1/GluA2 and GluA2/GluA3 subunits, together with a smaller contribution from GluA1 homomers, represent the most common combinations in excitatory synapses 22 , we could speculate that altering RAB39B-mediated early trafficking step of GluA2/GluA3, the result will be an increase in GluA1 Ca 2 þ -permeable homomers.
In fact, the AMPA-type glutamate receptors play a critical role in synaptic plasticity underlying learning and memory by mediating the majority of fast excitatory synaptic transmission and by trafficking into and out of the synapse 51 . The presence of GluA2 in heteromeric AMPAR renders the channel impermeable to Ca 2 þ and Zn 2 þ , thus influencing channel kinetic, channel conductance and synaptic transmission. In contrast, GluA2lacking AMPAR is Ca 2 þ -permeable and is mostly observed in young animals or after an epoch of plasticity-inducing neuronal activity [52][53][54] . In fact, our results describe a slight alteration in spine morphology, depicting a scenario of immature spines in shRab39b-treated neurons, as they are shorter and wider than control neurons. Previous studies described that Ca 2 þ -permeable AMPARs are important in experience-dependent and synaptic plasticity 55 . These studies suggest that Ca 2 þ -permeable AMPAR play a prominent role in maintaining circuits in a labile state where further plasticity can occur, thus promoting metaplasticity 56 . Moreover, it was shown that the abnormal expression of Ca 2 þ -permeable AMPARs is implicated in drug addiction and memory disorders, with significant implication in the development of therapeutic approaches to these disorders 56 . In addition, alteration of the relative proportion of AMPAR subunits after pathological insults 54,57 or in response to physiological stimuli 58 will shift the balance towards homomers and in turn alter the signalling landscape of the neuron.
Altogether our findings shed light on the molecular mechanisms responsible for the maturation and the trafficking of GluA2/ GluA3 AMPAR in the secretory pathway, and pave the way to explain the functional role of RAB39B in cognitive disorder.
Unsolved questions remain still open. At first, what other molecules convey the transport competence to RAB39B? One possibility could be the involvement of additional downstream effectors proteins that give the directionality and the timing for RAB39B-PICK1-GluA2 complex movement. GM130 a cis-Golgi protein involved in accepting vesicles from the ER 59,60 , as well as dynein/dynactin motor complex, driving microtubule-dependent trafficking from the ER to Golgi 61 and along dendrites to the post-synaptic sites 62 or myosin Va 63 involved in dendritic and post-synaptic actin-dependent trafficking 64 could be the most reliable factors. At present, RAB39B is required for the transport of GluA2-containing AMPAR from the ER to the Golgi, but we cannot assess its involvement from the Golgi to the neuronal surface. Future studies will undoubtedly be necessary to answer these questions as well as the availability of Rab39b-null mouse model to better define the link between RAB39B-mediated intracellular trafficking and cognitive disorders.

Methods
Most of the experiments were on mouse hippocampal neurons prepared from mouse embryos at E18 days. Astroglial feeder layer cultures were prepared from mouse cortices at post-natal day 2. Mice were C57Bl/6N genetic background and were obtained from Charles River, Italy. Mice were euthanized in accordance with 'Institutional Animal Care and Use Committee San Raffaele (IACUC)' at San Raffaele Scientific Institute, Milan, Italy and approved by the Italian National Ministry of Health, IACUC ID 470 and in accordance with the guidelines established by the European Community Council Directive of 24 November 1986 on the use of animals in research (86/609/EEC).
The site-directed mutagenesis using the Quick Change Lightning kit (Stratagene) was used to generate nucleotide deletion or mutations on flag-RAB39B or pGEX-RAB39B to obtain the following plasmids . pDBleu-RAB39B was created amplifying the human RAB39B cDNA, obtained from retro-transcription of linfoblastoid cells RNA from control patients, using specific primers (pPC86 For: 5 0 -GCGGTCGACCATGGAGGCC ATCTGGCTGTACC-3 0 ; pPC86 Rev: 5 0 -GCGGCGGCCGCTCTAGCACAAACA TCTCCTCTCTGA-3 0 ). The PCR product was than cloned in frame into SalI and NotI sites of pDB-Leu (ProQuest Two-hybrid system, Gibco BRL). FLAG-RAB39B-rescue, cherryRAB39B-rescue and GFP-RAB39A was created as previously described 10 .
Yeast two-hybrid assay. Yeast two hybrid screening was performed following the manufacturer's instruction (ProQuest Two-hybrid system, Gibco BRL). The fulllength human RAB39B cDNA was cloned into the pDBLeu vector, in frame with the GAL4-binding domain, and used as bait to screen a human foetal brain cDNA library (ProQuest Pre-made cDNA Libraries) cloned into the pPC86 vector in frame with the GAL4-activating domain. Positive colonies grew on plates containing 10 mM 3-amino-1,2,4-triazole(3-AT) without tryptophan and leucine and expressed different reporter genes: HIS3 and URA3 allowed the growth, respectively, in medium lacking histidine and uracil, LacZ induced a colorimetric reaction in the presence of 5-bromo-4-chloro-3-indolyl-b-D-galactopyranoside (X-gal). cDNA plasmids from positive clones were recovered via TOP10F' E. coli, plated on ampicillin and sequenced. Yeast two-hybrid screening was repeated two times and only positive colonies expressing three reporter genes (HIS3, LacZ and URA3) were taken in consideration. The test was performed following the manufacturer's instructions (ProQuest Two-hybrid system, Gibco BRL).
GST pull-down. GST fusion proteins were prepared in E. coli strain BL21DE3, isolated and immobilized on glutathione-Sepharose 4B beads (GE Healthcare; as decribed 65 ). GST pull-down experiments were carried out with mouse brain lysates or COS7 cells transfected with different constructs or in vitro transcribed and translated PICK1.
PICK1 was in vitro transcribed and translated modifying the standard protocol (GeHealthcare) as follows: the in vitro reaction was reduced to 30 min instead of 90 min.
All the lysates were pre-cleared with 120 mg GST immobilized on glutathione-Sepharose 4B beads (GE Healthcare) for 1 h at 4°C.
Forty micrograms of GST-RAB39B in presence of GDP or GTPgS (Sigma-Aldrich) or 40 mg of GST-RAB39B S22N and GST-RAB39B Q68L were incubated on rotation 3 h at 4°C with 8 mg of mouse brain lysed in HKT or RIPA buffer, or 2 mg of COS7 cells expressing myc-PDZ, flag-D121, myc-KDAA, myc-K83A or 10 ml of the in vitro translational reaction of PICK1. Eighty micrograms of GST-PICK1 immobilized on glutathione-Sepharose 4B beads were incubated on rotation 3 h at 4°C with 16 mg of mouse brain lysed in HKT buffer or 2 mg of COS7 cells expressing flag-RAB39B or flag-RAB39B DER, flag-D38A, flag-V41G42V43AAA or flag-D44F45AA.
In all the experiments, GST alone and/or GST-RAB5 GDP and GTPgS immobilized on glutathione-Sepharose 4B beads were used as control.
Immunoprecipitation. Immunoprecipitation studies were carried out with lysates prepared from HEK293T cells co-expressing flag-RAB39B and myc-PICK1 or from wild-type mouse brain or Pick1-knockout mouse brain. Transfected cells were washed in phosphate-buffered saline (PBS) Ca 2 þ /Mg 2 þ (0.1 mM CaCl 2 , 1 mM MgCl 2 ) and scraped in CHAPS buffer (1% CHAPS; Sigma-Aldrich; 0.1 mM EDTA pH 8, Protease Inhibitor Cocktail in PBS solution) and rotated at 4°C for 1 h before centrifuging at 13,200 g for 20 min to pellet debris. Mouse brains were lysed in HBST buffer (10 mM HEPES pH 7.4, 150 mM NaCl, 0.5% Triton X-100, Protease Inhibitor Cocktail in H 2 O solution), homogenized 10 times by hands, rotated at 4°C for 10 min and centrifuged at 13,500 g for 20 min at þ 4°C to pellet debris.
All the lysates were pre-cleared with washed protein G-Sepharose-4 fast flow beads (GE Healthcare/Amersham Biosciences) for 2 h at 4°C.
Cell cultures, transfection and lentiviral transduction. COS7 and HEK293T cells were grown in Dulbecco's modified Eagle's medium (Gibco) supplemented with 10% fetal bovine serum (Gibco), 1% glutamine, 1% penicillin and streptomycin and were regularly passaged to maintain exponential growth. COS7 cells were transfected using Lipofectamine 2000 (Invitrogen) following the manufacturer's instructions and fixed 24 h after transfection.
Mouse hippocampal neurons were transduced (MOI1) at DIV1 with lentiviral particles expressing shScramble or shRab39b as previously described 10 . See Supplementary Experimental Procedures for details.
Visualization of intracellular endogenous AMPAR subunits in neurons was done by hiding the portion of receptors exposed on the membrane surface with primary monoclonal antibodies recognizing the N-terminal portion (anti-GluA1: Millipore, #MAB2263, 1: 1:200 IF; anti-GluA2: Millipore, #MAB397, 1:200 IF; anti-GluA3: made into the laboratory of Prof. R. Huganir Johns Hopkins University, Baltimore, MD, USA, 1:100 IF) diluted in GSDB without Triton X-100 as well as the HRP-conjugated secondary antibody (Biorad, 1:200), and then cells were incubated with the primary monoclonal antibodies recognizing the N-terminal portion diluted in GSDB with 0.3% Triton X-100 as well as the appropriate secondary antibody (Molecular Probes, Invitrogen).
To visualize GluA1 and GluA2 exposed on the cell surface, the primary monoclonal antibodies recognizing the N-terminal portion were diluted in GSDB without Triton X-100, as well as the secondary antibody (Molecular Probes, Invitrogen).
Image acquisition and analysis. For Western blots, ImageJ Analysis Software ('Analyze gels' plugin) was used; full scan of western blots is shown in Supplementary Fig. 6. For GST pull-down experiments, PICK1 binding to GST-RAB39B GDP versus GST-RAB39B GTPgS is quantified as follow: signal from GST-RAB39B GTPgS binding protein is normalized on the 50 kDa GST signal and set at 100%; consequently, the level of GST-RAB39B GDP binding protein, normalized on the 50 KDa GST band, is expressed as percentage of 100.
Confocal images of COS7 cells were obtained using Leica SP8 SMD laser scanning confocal. For spine morphology fluorescent images were acquired with a Biorad MRC1024 confocal microscope, using a Nikon 60 Â objective with sequential acquisition setting at 1,280 Â 1,024 pixel resolution and image data were a z series projection of about 5-10 images, each averaged 4 times and taken at 0.7 mm depth intervals. Pictures of immunostained hippocampal neurons were captured with the same exposure conditions using DeltaVision microscope (Applied Precision) equipped with a Â 60 or Â 100 objectives. At least 10 Z-space slices of 0.20 mm were deconvolved and flattened by maximum projection.
TIRF images, setting 110 nm as the distance from the coverslip, were acquired using Leica SR GSD 3D TIRF microscope (Leica).
Pearson's correlation coefficient was calculated using Volocity software to measure the correlation between two proteins.
Morphometric spine measurements were made with NeuronStudio software. ImageJ Analysis Software ('Gran filter' plugin setting the size from 1 to infinity) was used to measure AMPAR subunits and PICK1 density relatively to the area of infected cell body and dendrites.
Electrophysiology in cultured hippocampal neurons. Seventeen to twenty DIV hippocampal neurons transduced at low or high MOI with shScramble or shRab39b, and in some case transfected with Rab39b as rescue, were patched and whole-cell recordings were made on infected neurons held at À 70 mV in the presence of 3 mM TTX. mEPSCs were acquired until the cell was stable and analysed using pClamp10 (Molecular Devices). Threshold mEPSC amplitude was set at 5 pA, and 300-500 events were collected and averaged to calculate the mean mEPSC amplitude, frequency and kinetics for each culture preparation examined. See Supplementary Experimental Procedures for details.
Electrophysiology in organotypic hippocampal slices. Hippocampal slice cultures were prepared from postnatal day 5-7 mice C57BL/6N. After 1 day in culture every slice was transduced with lentiviral particles (shScramble or shRab39b). After 10-12 days in culture, simultaneous voltage-clamp whole-cell recordings were obtained from nearby infected and uninfected CA1 pyramidal neurons. Whole-cell recordings were made with a Multiclamp 700B amplifier (Axon Instruments) and electrophysiological data were collected with pCLAMP software (Molecular Devices). See Supplementary Experimental Procedures for details.
Statistical analysis. Data are expressed as mean ± s.e.m. Statistical significance was assessed using paired Student's t-test as appropriate (for two groups comparison). We verified the difference between groups by means of t-type test statistic and exact P values computed with permutation methods to avoid any distributional assumption or asymptotical approximation. P values were, then, adjusted for multiplicity applying the Benjamini-Hochberg procedure (false discovery rate) 66 .
Molecular modelling. The structural models of PICK1 and RAB39B were achieved by comparative modelling (that is, by MODELLER 67 ), whereas the interaction between them was predicted by protein docking (that is, by ZDOCK 68 ). The CTD of GluA2 was built by comparative modelling as well. Details of Computational modelling are provided as Supplementary Information.