Molecular codes for cell type specification inBrn3 retinal ganglion cellsSzilard Sajgoa,1, Miruna Georgiana Ghiniaa,2, Matthew Brooksb, Friedrich Kretschmera,3, Katherine Chuanga,4,Suja Hiriyannac, Zhijian Wuc, Octavian Popescud,e, and Tudor Constantin Badeaa,5

aRetinal Circuits Development and Genetics Unit, Neurobiology–Neurodegeneration and Repair Laboratory, National Eye Institute, Bethesda, MD 20892;bGenomics Core, Neurobiology–Neurodegeneration and Repair Laboratory, National Eye Institute, Bethesda, MD 20892; cOcular Gene Therapy Core,National Eye Institute, Bethesda, MD 20892; dInstitute of Biology, Romanian Academy, Bucharest 060031, Romania; and eMolecular Biology Center,Interdisciplinary Research Institute on Bio-Nano-Science, Babes-Bolyai University, Cluj-Napoca 400084, Romania

Edited by Jeremy Nathans, Johns Hopkins University, Baltimore, MD, and approved April 12, 2017 (received for review November 8, 2016)

Visual information is conveyed from the eye to the brain bydistinct types of retinal ganglion cells (RGCs). It is largely unknownhow RGCs acquire their defining morphological and physiologicalfeatures and connect to upstream and downstream synapticpartners. The three Brn3/Pou4f transcription factors (TFs) participatein a combinatorial code for RGC type specification, but their exactmolecular roles are still unclear. We use deep sequencing to define(i) transcriptomes of Brn3a- and/or Brn3b-positive RGCs, (ii) Brn3a-and/or Brn3b-dependent RGC transcripts, and (iii) transcriptomes ofretinorecipient areas of the brain at developmental stages relevantfor axon guidance, dendrite formation, and synaptogenesis. We re-veal a combinatorial code of TFs, cell surface molecules, and deter-minants of neuronal morphology that is differentially expressed inspecific RGC populations and selectively regulated by Brn3a and/orBrn3b. This comprehensive molecular code provides a basis for un-derstanding neuronal cell type specification in RGCs.

retinal ganglion cells | transcription factors | neuronal cell types | Pou4f1 |Pou4f2

The molecular analysis of neuronal circuits benefits signifi-cantly from modern approaches to gene expression profiling

and genetic manipulation. The mechanisms of cell type specificationare still poorly understood, but experiments in model organismssuggest a combination of transcriptional regulation, extracellularsignals, and cell–cell interactions (1–4). Retinal ganglion cells(RGCs) are a particularly powerful system for illustrating themolecular and activity-dependent mechanisms of cell type speci-fication. Based on molecular markers, dendritic arbor morphol-ogies, axonal projections to retinorecipient areas of the brain,synaptic partners, physiological properties, and roles within thevisual circuit, mouse RGCs can be cataloged in 20–30 differenttypes (5–10). Some of the developmental mechanisms by whichRGC features are combined to determine RGC types are begin-ning to be uncovered. Mouse RGCs become postmitotic and startexhibiting specific molecular markers and morphological featuresaround embryonic day 11 (E11). As soon as E12, RGC axons crossthe midline at the optic chiasm, and by E15, the first axons havereached the superior colliculus (SC), the most remote retinor-ecipient area of the brain (11, 12). RGC axons invade their targetnuclei only around birth, and the first 10 postnatal days are themost active period for synapse formation. RGC dendritic arborsdevelop mostly postnatally, with lamination within the innerplexyform layer clearly visible at postnatal days 3–4 (P3–P4) andreaching a nearly mature distribution by P7 (13–16). Combinato-rial transcriptional regulation may play a major role in RGC typespecification. Previous work suggests the following transcriptionalcascade: the basic helix–loop–helix (bHLH) transcription factor(TF) Atoh7 is expressed in RGC precursors and controls the ex-pression of the POU4 family TF Brn3b and the Lim domain TFIsl1, which are both required for the initiation of the RGC tran-scriptional program (17–25). Further downstream TFs includeBrn3a, Brn3c, Eomesodesmin (Tbr2), Ebf1, Ebf3, Onecut1, and

Onecut2 (6, 13, 26–33). Together with Isl1 and Brn3b, thesedownstream factors are expressed in partially overlapping patternsin RGC types, and some were shown to be required for survivaland/or dendrite and axon formation in various RGC types.However, many other TFs may be involved in generating thediversity of RGC types (34–38). We have previously used re-porter knock-in alleles expressing alkaline phosphatase (AP; aglycosylphosphatidylinositol (GPI)-linked, extracellular molecule)at the loci of Brn3a, Brn3b, and Brn3c (Brn3CKOAP) to describetheir cell type distribution among RGCs and other sensory pro-jection neurons. We also identified axonal and dendrite arbordefects in RGCs missing Brn3a, Brn3b, or Brn3c either alone or incombination (6, 13, 31, 39, 40). We now describe an immu-noaffinity purification strategy using anti-AP antibodies to isolateRGCs from Brn3AP RGCs that are either WT or KO for Brn3a orBrn3b. Using our knowledge of partially overlapping RGC pop-ulations expressing Brn3s, we can identify molecules selectivelyenriched in RGCs, selectively expressed in distinct Brn3 RGCsubpopulations, and/or regulated by Brn3a or Brn3b in these RGC

Significance

We report here transcriptome analysis by RNA sequencing(RNASeq) of genetically labeled and affinity-purified mouseretinal ganglion cell (RGC) populations. Using a previouslyestablished conditional knock-in reporter strategy, we labelRGCs from which specific transcription factors have been re-moved and determine the consequences on transcriptionalprograms at different stages critical to RGC development. Wefind that Brn3b and Brn3a control only small subsets of Brn3–RGC–specific transcripts. We identify extensive combinatorialsets of RGC transcription factors and cell surface molecules andshow that several RGC-specific genes can induce neurite-likeprocesses cell autonomously in a heterologous system.

Author contributions: S.S., M.G.G., F.K., K.C., O.P., and T.C.B. designed research; S.S., M.G.G.,M.B., F.K., K.C., and T.C.B. performed research; S.H. and Z.W. contributed new reagents/analytic tools; S.S., M.G.G., M.B., and T.C.B. analyzed data; and S.S., M.G.G., O.P., and T.C.B.wrote the paper.

The authors declare no conflict of interest.

This article is a PNAS Direct Submission.

Data deposition: The next generation sequencing data reported in this paper have beendeposited in the Gene Expression Omnibus (GEO) database, https://www.ncbi.nlm.nih.gov/geo (accession no. GSE87647).1Present address: Yonehara Laboratory, Danish Research Institute of Translational Neu-roscience, Aarhus University, 8000 Aarhus, Denmark.

2Present address: Emerson Laboratory, Biology Department, The City College of NewYork, New York, NY 10031.

3Present address: Scientific Computing Core, Max Planck Institute for Brain Research,Frankfurt am Main 60438, Germany.

4Present address: School of Medicine, Yale University, New Haven, CT 06510.5To whom correspondence should be addressed. Email: badeatc@mail.nih.gov.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1618551114/-/DCSupplemental.

E3974–E3983 | PNAS | Published online May 2, 2017 www.pnas.org/cgi/doi/10.1073/pnas.1618551114

Dow

nloa

ded

by g

uest

on

Mar

ch 2

2, 2

021

populations. We focused this study on E15 RGCs to identify axonguidance molecules (Fig. 1A) and P3 RGCs to detect potentialdeterminants for dendrite formation and synaptogenesis (Fig.1B). Furthermore, we isolated RNA from retinorecipient areasat P3 to identify differential markers for the cellular targets thatmight interact with incoming RGC axons (Fig. 1E). We report acombinatorial code of TFs and cell surface molecules (CSMs)expressed in different RGC subpopulations and retinorecipientareas of the brain. A subset of these genes can intrinsically in-duce arbor-like processes in epithelial cells, suggesting cell-autonomous neuronal arbor formation mechanisms.

ResultsScreen Rationale and Sample Collection. We have induced retina-specific Cre recombination of our conditional knock-in reporteralleles at the Pou4f1 (Brn3aCKOAP) and Pou4f2 (Brn3bCKOAP) lociusing the Pax6α:Cre driver (13). When the sister chromosomecarries a conventional KO allele, recombination of the conditionalallele results in cells lacking both copies of the endogenous genelabeled by the AP marker (Brn3aAP/KO or Brn3bAP/KO RGCs;

KO). When paired with a WT allele, recombination results in AP-tagged heterozygote cells that are phenotypically WT (Brn3aAP/WT

or Brn3bAP/WT RGCs; WT). Although recombination happensthroughout the retina, other retinal cell types do not expresseither Brn3a or Brn3b and therefore, appear AP-negative (6, 13,18, 21, 25, 31, 41). Using this genetic labeling strategy, we cancompare several cell populations (Fig. 1C). (i) Comparing theexpression profiles of Brn3AP/WT RGCs and retinal supernatants,we can identify genes specific for or enriched in RGCs. (ii) RGCgenes regulated by a Brn3 TF should be differentially expressedin Brn3AP/WT vs. Brn3AP/KO RGCs. (iii) Genes specific forBrn3a+Brn3b−, Brn3a−Brn3b+, or Brn3a+Brn3b+ RGC pop-ulations can be identified by comparing expression profiles ofBrn3aAP/WT with Brn3bAP/WT RGCs. We dissociated retinasfrom Pax6α:Cre;Brn3aCKOAP/WT, Pax6α:Cre;Brn3aCKOAP/KO,Pax6α:Cre;Brn3bCKOAP/WT, and Pax6α:Cre;Brn3bCKOAP/KO miceand isolated the AP-expressing Brn3AP RGCs using magneticbeads coupled to anti-AP mouse mAbs (Materials and Methodsand Fig. 1 D and F). We also have labeled the lateral geniculatenucleus (LGN), SC, and pretectal area (PTA) of P3 WT mice byanterograde tracing and dissected and processed them for deepsequencing (Materials and Methods and Fig. 1E).In the following, we will present gene expression data that are

restricted to the RefSeq (https://www.ncbi.nlm.nih.gov/refseq/)subset of mouse transcripts given its highest level of quality andannotation confidence (Material and Methods). We use two al-ternative strategies to identify differentially expressed transcriptsin our datasets. The first, differential expression analysis algo-rithm (DESeq), is the differential expression analysis as de-scribed by Anders and Huber (42), with the false discovery rateset at 0.1 and the fold change at two. The second, referred to inthe text as the “Twofold” criterion, identifies a transcript asdifferentially expressed in a given RGC population if it hasexpression levels in RGCs more than two fragments per kilo-base per million reads (FPKM) and is enriched in RGCscompared with the corresponding retina samples (FPKM levelsin Brn3AP/WT RGCs ≥ 2 × FPKM levels in Brn3AP/WT retinasupernatant). If the transcript also has a more than twofold changebetween Brn3AP/WT and Brn3AP/KO RGCs, it is also considered“regulated” by the respective Brn3.

Deep Sequencing and Sample Quality Control. Because RGCs are arare retinal cell population (around 0.5%) (43), bead-coupledRGCs from several mice (typically six to eight retinas) of identi-cal genotype were pooled to constitute one RGC sample. For eachgenotype described, two RGC samples and one retina supernatantsample were submitted to RNA extraction, reverse transcription,amplification, and sequencing using an Illumina Sequencingplatform (Materials and Methods). For WT brain regions, threereplicates for LGN and SC (derived each from one mouse) andtwo replicates for whole-brain homogenates (pooled whole-brainnonvisual regions; each from one mouse) were submitted to se-quencing, and one PTA sample was generated by pooling tissuesfrom three individual mice. Correlation coefficients between bi-ological replicates were very high (> 0.95) and between samples(e.g., RGC vs. retina) were lower, ranging from 0.55 to 0.95 (Fig.2A, Dataset S2, and Fig. S1A). Hierarchical clustering across alltranscripts expressed in our samples reveals a good correlation ofreplicates by RGC genotype, retina, age, or brain region (Fig. 2Band Dataset S3), confirmed also by principal component analysis(Fig. S1 B–D). We determined the success of our cell purificationand deep sequencing strategy in several ways. Visualization of thereads mapping to the endogenous Brn3a and Brn3b loci (Fig. 2C–E)shows dramatically higher levels in Brn3aAP/WT over Brn3aAP/KO

[adjusted means = 85.68 WT, 2.81 KO; DESeq P = 3.59 e-27, ad-justed p value (padj) = 5.02e-23; t test P = 8.67e-04] and Brn3bAP/WT

over Brn3bAP/KO RGCs at both P3 (adjusted means = 78.67 WT,6.43 KO; DESeq P = 3.13 e-23, padj = 4.4e-19; t test P = 0.045) and

P3 Dendrite FormationAxon - target interaction Synapse formation

E15 Axon Guidance

Brn3aAP/WT RGCs Brn3bAP/WT RGCsvs.

iii) Genes Specific for distinct RGC populations

Supernat

RGC

P 0.5P 3.5

A

+

Other cells Brn3AP

RGCsαAPBeads

B

C

D

E

F i ii

iii iv

Brn3AP/WT RGCs Brn3AP/KO RGCsvs.

ii) Brn3-dependent RGC Genes:

Brn3AP/WT RGCs Other (supernatant)vs.

i) Genes enriched in RGCs:

ChTB

Fig. 1. Experimental goal and design. (A) E15 retina containing heteroge-neous undifferentiated cells (gray) and RGCs (purple), which are mostly post-mitotic and extend axons. (B) P3 retina with RGCs extending dendrites. RGCaxons are involved in synapse formation. (C) Comparison strategies. (i) RGC-enriched genes were identified by comparing E15 or P3 RGCs (Brn3aAP orBrn3bAP) with AP-negative cells. (ii) Genes regulated by Brn3s were inferred bycomparing Brn3bAP/WT or Brn3aAP/WT (heterozygote) RGCs with Brn3bAP/KO orBrn3aAP/KO (KO) RGCs, respectively. (iii) To identify genes expressed selectivelyin RGC subpopulations, we compare Brn3bAP/WT with Brn3aAP/WT RGCs.(D) Immunomagnetic purification of RGCs. E15 or P3 retinas are dissociated, andBrn3aAP or Brn3bAP RGCs are separated using anti-AP–coated magnetic beads.RNA was extracted from either supernatant or Brn3AP RGCs-coupled beadsand processed for RNASeq. (E) Retinorecipient brain tissue isolation fromP3.5 pups. Cholera toxin B (ChTB-AF488) was injected in the eyes of P0.5 mousepups. At P3.5, brains were isolated and vibratome-sectioned, and green fluo-rescent regions [LGN, shown in Right, olivary pretectal nucleus (OPN), medialterminal nucleus (MTN), and SC] were dissected. All residual brain regions werepooled and used as controls. (F) The 10-μL samples from each RGC immuno-purification were spread on slides and stained for AP and DAPI: (i–iii) threeexamples of Brn3AP RGCs coupled to magnetic beads (top), DAPI nuclearcounterstain (middle), and merged images (bottom). Note occasional DAPI-positive AP-negative cells (bead cluster on the right in i). (iv) Retina superna-tant after immunopurification showing a Brn3AP RGC (arrowhead) and aretinal pigment epithelium cell (arrow). Estimation of yield and purity is inMaterials and Methods and Dataset S1. (Scale bars: E, 250 μm; F, 40 μm.)

Sajgo et al. PNAS | Published online May 2, 2017 | E3975

DEV

ELOPM

ENTA

LBIOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

Mar

ch 2

2, 2

021

E15 (adjusted means = 86.93 WT, 20.41 KO; DESeq P = 6.45e-13,padj = 8.67e-10; t test P = 0.0405). The residual reads in Brn3aAP/KO

and Brn3bAP/KO RGCs are mapping to the 5′ and 3′ UTRs,consistent with the replacement of the endogenous coding exonswith the AP ORF and preservation of the 5′ and 3′ Brn3a andBrn3b UTRs (Fig. 2 C, Upper and D, Upper). Compared with theretina or the brain samples, RGC samples have higher expressionlevels for Brn3a (adjusted means = 86.63 for WT RGCs, 5.34 forP3 retinas; DESeq P = 9.21 e-52, padj = 1.61e-49; t test P = 6.57e-05)and Brn3b transcripts at both P3 (adjusted means = 78.85 for WTRGCs, 3.19 for P3 retinas; DESeq P = 4.37 e-51, padj = 2.18 e-48;t test P = 0.0015) and E15 (adjusted means = 86.95 for WT RGCs,12.21 for E15 retinas; DESeq P = 4.53 e-15, padj = 2.79 e-12; t testP = 0.036) (Fig. 2E). Reads aligned to the AP cDNA [using a“minigenome” generated with Bowtie (44)] (Materials and Methods)display the expected distribution, with essentially no reads in retinasamples and high levels of the transcripts in the RGC samples (Fig.2F). Of note, in E15 but not P3 RGCs, the AP reporter is expressedat much higher levels in the Brn3bAP/KO than in the Brn3bAP/WT

RGCs. Thus, the higher levels of Brn3b reads in the Brn3b KO atE15 could be explained by the AP transcript carrying partial Brn3b5′ and 3′ UTRs along. Finally, previously described RGC markersare visibly enriched in Brn3AP RGCs (Fig. 2G). Thus, we are con-fident that our cell sorting, RNA isolation, and deep sequencingapproach is successful.

Outcome of RGC Purification. We compared expression data fromBrn3aAP/WT and Brn3bAP/WT RGCs to identify genes that are

selectively expressed in these partially overlapping cell populations.We find that, at P3, a large number of transcripts are enriched inRGCs, with most being selective for Brn3aAP/WT RGCs (1,423byDESeq and 1,667 by Twofold), some being common to Brn3aAP/WT

and Brn3bAP/WT RGCs (994 by DESeq and 1,285 by Twofold),and only a small number being selective for Brn3bAP/WT RGCs (71 byDESeq and 407 by Twofold) (Fig. 3A, Dataset S4, and Fig. S1E).Only a small fraction of these RGC-enriched transcripts seem to beregulated by Brn3b or Brn3a. Strikingly, most of these are selec-tively affected by Brn3b ablation, and only very few are Brn3a-dependent (Fig. 3 C and D, Dataset S4, and Fig. S1F) in keepingwith the more dramatic effect of Brn3b ablation on RGC devel-opment. Interestingly, the repertoire of transcripts enriched inBrn3bAP/WT RGCs changes significantly between E15 and P3 (Fig.3B, Dataset S4, and Fig. S1G). This fact may reflect the distinctRGC-specific programs required at the two ages. The set of Brn3b-dependent genes expressed by Brn3bAP RGCs is also dramaticallydistinct at the two ages (Fig. 3 C andD, Dataset S4, and Fig. S1H).The differences between Brn3a- and Brn3b-positive RGC tran-scriptomes may reflect differences in cell type distribution,whereas the temporal differences may define the distinct func-tionality required in the early (E15; axon guidance) vs. late (P3;dendrite formation, synaptogenesis, and myelination) stages ofRGC maturation. An analysis of Gene Ontology terms among ourcandidate genes shows significant enrichment for known neuronal-associated processes and pathways, such as neurotransmitter re-ceptors and release, voltage-gated channels, signaling cascades,synaptic transmission, neuronal projection, etc. Among these pathways,

RG

CR

etin

aP

3E

15E

15P

3

0 20 40 60 80 100 120

AP

-2 -1 3210

P3 Brn3aAP/WT RGC

-2-1

3210

RGC-Brn3aAP/WT-P3

RGC-Brn3bAP/WT-P3

AP

AP

887

36

611

74RGC-Brn3bAP/KO-P3

RGC-Brn3aAP/KO-P3

S2

S1

P3 Brn3aAP RGC

WT

KO

P3 Brn3aAP/WT

RGC

retina-2 -1 3210 -2 -1 3210

A C

FD

Pou

4f1

Pou

4f2

Pou

4f3

Isl1

Isl2

Eom

esT

br1

Thy

1S

ncg

Snc

aR

bfox

3N

efl

Nef

mR

bpm

sO

pn4

Fst

l4Ja

m2

Cdh

6C

artp

tT

rhr

10

20

30

40

50

60G

E

0 20 40 60 80

Brn3a(Pou4f1)

0 20 40 60 80 100

Brn3b(Pou4f2)

RG

CR

etin

aP

3E

15E

15P

3B

rain

P3

log FPKM

log

FP

KM

FPKM

FPKM FPKM

Rea

dsR

eads

Rea

dsR

eads

0.5 Kb

0.5 Kb

SC

-2S

C-3

SC

-1P

TA

Bra

in-2

B

rain

-1LG

N-1

LGN

-2LG

N-3

1

KO

-1K

O-2

WT

-1W

T-2

Brn3b P3-RGC

Brn

3a-K

OB

rn3a

-WT

Brn

3b-K

OB

rn3b

-WT

P3-ret

WT

-1W

T-2

KO

-1K

O-2

KO

-1K

O-2

WT

-1W

T-2

KO

WT

Brn3a P3-RGC

RGC ret

Brn3b-E15

65432 7

−3 −2 −1 0 1 2 3

B

Fig. 2. Sample characterization and validation. (A) Log-scale scatter plots comparing FPKM levels. (Left) Comparison of two samples (S1 vs. S2) derived from P3Brn3aAP/WT RGC; R = 0.9915. (Center) Means of two P3 Brn3aAP/KO RGC samples (KO) vs. means of two P3 Brn3aAP/WT RGC samples (WT); R = 0.9922. (Right)Comparison of a P3 Brn3aAP/WT retinal supernatant (retina) with the mean of two P3 Brn3aAP/WT RGC samples (RGC); R = 0.7761. Red diagonals separate thetwofold comparison lines, and the red corners enclose genes with less than two FPKM for both samples in the plot. (B) Clustergram across 18,185 transcripts thatwere expressed at greater than or equal to one FPKM in at least one of the samples. Clustering was performed on standardized sample values, first along thesample dimension (columns) and then along the transcript dimension (rows) (Materials andMethods). Branches are color-coded and labeled 1–7 as follows: branch 1,Brn3aAP P3 RGCs; branch 2, E15 retina and Brn3bAP RGCs; branch 3, P3 retinas; branch 4, Brn3bAP P3 RGCs; branch 5, SC and PTA; branch 6, whole-brain controls;branch 7, LGN. For each sample (along the bottom), numbers indicate biological replicates. Color scale represents units in SDs of the distribution across all ob-servations for each given row (gene). [Dataset S2 shows cross-correlation matrix of all samples, and Fig. S1 shows additional scatter plots and principal componentanalysis (PCA).] (C and D) Visualization of mapped reads. (C) Reads from (Upper) Brn3aAP/KO and (Lower) Brn3aAP/WT P3 RGCsmapping to the Brn3a locus. (D) Readsfrom either (Upper) Brn3bAP/KO or (Lower) Brn3bAP/WT P3 RGCs mapping to the Brn3b locus. The x axis is in kilobases (notches every 0.5 kb). The y axis is scaled tothe highest read stack (indicated in the bottom right corner). The AP cDNA inserted in the recombined alleles is indicated. Gray bars flanked by black notchesrepresent reads. Thin blue lines represent spliced reads reaching across two exons. Exons (rectangles) and introns (lines) are shown for Brn3a (Pou4f1, three exons)and Brn3b (Pou4f2, two exons) in C, Lower and D, Lower. Coding regions within exons are blue. (E) Expression levels (FPKM) for Brn3a and Brn3b genes. MouseWT P3 brain samples are (from top to bottom) whole brain (white; median of two samples), PTA (black; one sample derived from three mice), and SC and LGN(dotted and gray bars, respectively; each medians of three samples). For retina and RGCs, samples are (from top to bottom) P3 Brn3bAP/WT (dark red), Brn3bAP/KO

(light red), Brn3aAP/WT (dark green), Brn3aAP/KO (light green), E15 Brn3bAP/WT (dark blue), and Brn3bAP/KO (light blue). Retina values represent individual retinalsamples, and RGC samples represent medians of two samples. (F) Expression (FPKM) of the knocked in AP cDNA color-coded as in E. (G) Heat map for knowngeneral and subtype-specific RGC markers. Expression levels are normalized to the maximum level for each gene and displayed on a 64-level scale (red is high).

E3976 | www.pnas.org/cgi/doi/10.1073/pnas.1618551114 Sajgo et al.

Dow

nloa

ded

by g

uest

on

Mar

ch 2

2, 2

021

adhesion molecules and TFs represented each about 5–10% of theidentified genes, regardless of the selection criteria (Dataset S4).

In Situ Validation. Because the starting RGC population fromwhich RNA was extracted is relatively small and the degree ofamplification is quite significant, we sought to validate a subsetof about 10% of the identified RGC-enriched molecules with anindependent technique. We designed in situ hybridization (ISH)probes for the 3′ UTR of 233 genes from our candidate pool andtested them against WT P3 retinas (Fig. 3 E, F, and I, Dataset S4,and Fig. S1 I, K, and L). Genes were selected based on literaturesearches for potentially interesting yet less well-studied pathwaysrelated to neuronal development and exhibit a broad range ofexpression levels. In addition, we searched the Allen Brain In-stitute mouse brain development ISH atlas (developingmouse.brain-map.org/) with our Brn3bAP/WT RGC E15 candidate genelist and identified 265 genes for which E15 eye sections wereavailable (Fig. 3 G, H, and J, Dataset S4, and Fig. S1 J, M, andN). There did not seem to be a strong correlation between RNAsequencing (RNASeq) FPKM expression levels and intensity ofin situ signal in the ganglion cell layer (GCL) at P3. As an ex-ample, in P3 Brn3AP/WT RGCs, Ig superfamily 6 (Igsf6) isexpressed at only 4–7 FPKM, whereas Rab6b has ranges between150 and 300 FPKM (Fig. 3E). Whereas both genes are enrichedin the GCL, the signal to noise ratio for Igsf6 is much higher thanthat in Rab6b, although probes are comparable in length andmelting temperature. In a similar fashion, Rgs4 and Myt1l havecomparable ISH GCL-specific signals at E15 (Fig. 3H), butRgs4 has almost six times higher expression in E15 Brn3bAP/WT

RGCs compared with Myt1l (75 vs. 13 FPKM). Nevertheless,about 60–75% of candidate genes predicted by RNASeq to beenriched in RGCs over the retina were confirmed by our in situscreens at either P3 or E15. The enrichment with positive hitswas not dramatically affected by using either the DESeq orTwofold selection criteria (compare Fig. 3 I and J with Fig. S1 Iand J; Dataset S4). Confirmation by ISH was consistently betterfor Brn3-regulated genes identified by DESeq (Fig. 3 I, Right andJ, Right) compared with the Twofold criterion (Fig. S1 I and J),but the total number of hits was smaller (Dataset S4). Thus, it

seems that DESeq is a more stringent selection method; how-ever, it may miss many useful candidates compared with ourTwofold criterion. Overall, more than one-half of the targetsidentified by deep sequencing show RGC specificity in ISH re-sults, but expression levels detected with the two methodsdiffer significantly.

Retinorecipient Nuclei Transcriptomes and ISH Validation. To iden-tify molecules that may be conferring specificity for the distinctretinorecipient brain areas, we have mined our RNASeq data fortranscripts enriched in SC, LGN, or PTA compared with thewhole-brain homogenate. Fig. 4 A and B show that essentially allpossible combinations can be found, with transcripts common toall three nuclei, common to only two of them, or selective foreach nucleus individually (complete lists are in Dataset S5). Ofthese candidate genes, 122 (LGN), 116 (PTA), and 134 (SC)were tested by ISH by the Allen Brain Institute (examples are inFig. 4 C–H). For the LGN and SC, between 55 and 75% ofRNASeq-predicted genes were specific or regionally expressedin the expected nucleus, with others being broadly expressed ornegative (Fig. 4 I and J). The number of specific transcripts wasmuch lower for the PTA. Interestingly, several transcriptsshowed lamination within the SC (Gpc3, Barhl1, and Foxb1),consistent with the possibility that these markers are selective forSC functional laminae (Fig. 4 E and F). Of note, applying DESeqmodestly increased the ratio of positive hits by ISH comparedwith our Twofold criterion, while reducing the total number ofcorrect hits, similar to what we observed for the RGC data. We,therefore, present throughout the text results based on bothselection criteria.

TF Program of Brn3AP RGCs and Retinorecipient Areas. TFs play asignificant role in neuronal cell type diversification. We, there-fore, compared our data with a merged list of 2,437 TFs andtranscriptional regulation-associated genes compiled by com-bining recently published surveys (45, 46). Of these genes, almostone-third (1,647) had expression levels of more than one FPKMin our RGC samples, but a more restricted subset was enrichedin RGCs compared with the retina (DESeq = 153, Twofold =322) (Fig. 5 A, C, and E and Dataset S6). An even smaller set of

P3Brn3bAP

Brn3a and/or Brn3b dependent RGC transcripts

RGC specific RGC enriched Whole retina Negative

Brn3b-RGC-enriched5 10 15

Lrfn3

2 4 6 8

Itga6

2 4 6 8

Igsf6

10 30 50 70

Cpne4

5 10 15 20

Tsc22d3

5 10 15

Myt1

20 60 100

Slc18a2

50 150 250

Rgs4

20 40 60 80

Pax6

10 20 30 40

Myt1l

10 30 50 70

Nrxn1

50 150 250

Rab6b

Brn3-RGC-regulated

Brn3aAP

P3

Transcripts enriched in Brn3AP RGCs vs. retina

P3Brn3bAP

Tsc22d3Myt1Slc18a2Rgs4 Pax6Myt1l

Lrfn3Itga6Igsf6Cpne4 Nrxn1Rab6b

E15 In Situ Hybridization (Allen Brain Data)P3 In Situ Hybridization Screen

Brn3-RGC-enriched Brn3b-RGC-regulated

A B

E

F

G

H

I J

C D

70

202 8631423 994

71

44

186 36158

32

123

20

22

58 30

4

5

7 37

5

16 191

497

33 47

2

Brn3bAP

E15

Upregulated Downregulated

Brn3a - WTvsKO - P3Brn3b - WTvsKO - P3Brn3b - WTvsKO - E15

RG

CR

etin

a

P3E15

E15P3

Bra

in

P3

RG

CR

etin

a

P3E15

E15P3

Bra

in

P3

Fig. 3. Only a small fraction of genes enriched in Brn3+ RGCs depend on Brn3a or Brn3b. (A–D) Venn diagrams representing transcripts enriched in Brn3AP

RGCs. Transcripts reported here are derived from DESeq (42), with a 0.1 false discovery rate and a twofold change between compared conditions. (A and B)Transcripts enriched in Brn3AP RGCs over retina supernatants. (A) Partially overlapping gene expression profiles in P3 Brn3aAP/WT RGCs and Brn3bAP/WT RGCs.(B) Partially overlapping gene expression programs in E15 and P3 Brn3bAP/WT RGCs. (C and D) Transcripts (C) down- or (D) up-regulated in Brn3AP RGCs as aresult of either Brn3a or Brn3b loss. (E–H) RGC-enriched transcripts validation by ISH. (E and G) FPKM values (indicated on the X axis) for RGC-enrichedcandidate genes. Plots are scaled to individual maximum levels. Sample color coding as in Fig. 2E. (F) P3 ISHs with probes directed against the 3′ UTR. Note thespectrum of outcomes from RGC-specific (Rab6b and Cpne4) to RGC-enriched (Igsf6 and Nrxn1) and from full retina expression (Itga6) to lack of expression(Lrfn3). (H) In situ analysis of RGC-specific genes at E15.5. Expression images reproduced with permission from the Allen Brain Institute atlas. (I) Pie chartsummaries of 223 genes analyzed by ISH at P3 broken down by genes predicted by RNASeq to be RGC-enriched (Brn3–RGC-enriched from A) or Brn3-regulated (Brn3–RGC-regulated from C and D). (J) Pie chart summaries of 58 genes found in E15 ISHs. Fig. S1 K–N shows control experiments for the P3 andE15 in situ screens. Dataset S4 lists all of the transcripts in the Venn diagrams and pie charts. (Scale bar: F, 100 μm; H, 400 μm.)

Sajgo et al. PNAS | Published online May 2, 2017 | E3977

DEV

ELOPM

ENTA

LBIOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

Mar

ch 2

2, 2

021

genes was Brn3a- and/or Brn3b-dependent (DESeq = 43,Twofold = 95) (Fig. 5 B, D, and E and Dataset S6). Fig. 5Hshows an unsupervised clustering of 38 TFs previously implicatedin RGC development. Interestingly, samples are first separated byage (P3 vs. E15) regardless of retina or RGC origin (Fig. 5H,branches a + b vs. c). In addition, at both P3 and E15, Brn3bAP/KO

(Brn3b KO) RGCs are grouped closer to the respective retinasupernatants (e.g., Fig. 5H, branch b) than to the other RGCsamples (e.g., Fig. 5H, branch a) (containing Brn3aAP/WT,Brn3aAP/KO, and Brn3bAP/WT RGCs at P3). This similarity sug-gests that loss of Brn3b turns Brn3bAP/KO RGCs into a moreundifferentiated “whole-retina” state. When contrasting the hor-izontal branches of the hierarchical tree, we can distinguish a“general RGC” group (branch 2, Pou4f1, Pou4f2, Irx2, Irx3,Irx4 and Irx6, Ebf1, and Ebf3), a “P3 RGC” group (branch 1, Isl2,Pou4f3, Klf2, Klf8, Myt1l, and Tbr1), and a “Brn3b-only RGC”group (branch 4, Nhlh1, Eomes, Irx1, and Tbx20). A subset of TFsin branch 3 shows up-regulation in Brn3bAP/KO vs. Brn3bAP/WT

RGCs at E15, consistent with their suppression by Brn3b in RGCs(Atoh7, Dlx1, Dlx2, Onecut2, Barhl2, Eya2, and Zic2). Many ofthese TFs are known to be required upstream or in parallel withBrn3b in the RGC class specification cascade. Similar patterns areseen over the full datasets containing RGC-enriched or Brn3-dependent TFs (Dataset S6 and Fig. S2). A similar-sized set ofTFs is enriched in the retinorecipient areas, with partially over-lapping expression between all three nuclei (Fig. 5 F and G,Dataset S6, and Fig. S3). The identified sets of TFs and thepowerful combinatorics provided by our approach thus yield alarge set of potential transcriptional regulators of RGC and/orretinorecipient nuclei cell types.

CSMs Involved in Cell–Cell Interactions, Axon Guidance, and NeuriteFormation. A variety of transmembrane proteins involved in cell–cell, cell–matrix, and receptor–ligand interactions (e.g., Integrins,Cadherins, Igs, Leucine Rich Repeats, Ephrins, Semaphorins,Plexins, Robo, and Tenneurins) are required for guiding axon anddendrite formation, and/or establishing specific synaptic interac-tions between neuronal cell types. We, therefore, queried the Genedatabase (National Center for Biotechnology Information) andestablished a comprehensive list containing 822 genes that includeone or more of these protein domains in their structure. Of these

genes, about one-quarter (DESeq = 156, Twofold = 237) weredifferentially expressed in RGCs (Fig. 6 A and C and Dataset S7),and an even smaller number (DESeq = 13, Twofold = 93) wereregulated by Brn3 TFs in RGCs (Fig. 6 B and D and Dataset S7).Only very few (DESeq = 20, Twofold = 69) were differentiallyexpressed in retinorecipient areas of the brain (Fig. 6E and DatasetS7). Partially overlapping sets of CSMs were selective for Brn3aand Brn3b (Fig. 6A) and dynamically expressed during embryonicand early postnatal development in RGCs (Fig. 6C). The numberof differentially regulated transcripts in Brn3bAP RGCs increasesbetween E15 and P3 (Fig. 6D). A clustergram example of 37 ad-hesion molecules and/or guidance receptors implicated in RGCdevelopment (Fig. 6F) reveals several interesting patterns. P3Brn3bAP RGCs are more closely grouped with the retina samples(Fig. 6F, vertical branch d and horizontal branch 1) and share in-creased expression of several molecules implicated in marking and/ordetermining lamination in the retina (Jam2, Plxna2, Robo3,Epha8, Sema5a, and Sema5b) (47–50). A second group is visiblyenriched in E15 Brn3bAP RGCs (Fig. 6F, branch 2). Amongthem, many receptors and/or ligands implicated in axon guid-ance are enriched in RGCs on Brn3b loss (Sema3a, Plxna3,Plxna1, Epha6, Efnb2, Efna4, Efna5, Cntn2, Nfasc, Slit1,and Tenm3) (Fig. 6F, column 1: compare Brn3bAP/WT withBrn3bAP/KO RGCs) (51–53). Finally, clusters 3 and 4 seem to behighly expressed in most brain regions as well as RGCs at P3.Fig. S4 (Dataset S7) shows the clustergram for all Brn3a- andBrn3b-dependent RGC-enriched transcripts identified in thisstudy. It highlights largely three groups. Branch 1 is P3 RGC-specific and mostly Brn3b-dependent. Branch 2 is specific to E15Brn3b RGCs and either negatively or positively regulated byBrn3b. Finally, branch 3, by far the largest group, is commonto brain regions and P3 RGCs. Most of the molecules in thissubgroup are down-regulated in Brn3bAP/KO compared withBrn3bAP/WT RGCs. The extended set of 237 RGC-enriched ad-hesion molecules (Dataset S7 and Fig. S4) follows similar trends.Thus, our screen defines combinatorial expression of many sur-face ligands and receptors implicated in cell–cell adhesion,neurite formation, and synapse specificity. Brn3b seems to play arole in regulating many of these neuronal identity determinants,thus explaining its important role in axon guidance and dendriteformation, whereas only a few are regulated by Brn3a in keeping

386

107

221

85

324 74

117

SCLGN

PTA

0 5 15 25

Gpc30 4 8 12

Barhl10 100 200 300

Cck0 5 10 15

Esrrb0 5 10 15

Foxb1

Gpc3

Barhl1

Foxb1

Cck

Esrrb

A C

D

E

F

G

Specific/Regional Broad Negative

17

3

1158

16

60

63 11

48 5

3

827

16

73 18

2

5

SCPTA

DE

Seq

LGNNucleus Selective Genes

SCPTALGNNucleus Enriched Genes

I

H143

45

183

40

99 2317

SCLGN

PTA

B 31

5

14 12

9

18 30

6

10 25

4

5 4

5

7 19

3

5

P4 ISH (Allen)

Twof

old

J

LGN

PTAPTA

SC

SC

SC

RG

CR

etin

a

P3E15

E15P3

Bra

in

P3

Fig. 4. Combinatorial gene expression in P3 retinorecipient brain nuclei. (A and B) Venn diagram comparisons of enriched unique or shared genes in LGN-,SC-, and PTA-derived samples. (A) Significantly differentially expressed transcripts identified by DESeq. (B) Transcripts passing the Twofold criterion. (C–G) ISHpatterns from Allen Brain Institute atlas at P4 for genes predicted to be nucleus-specific. (C) Cck expression in LGN. (D) Esrrb expression in PTA. (E–G) Gpc3,Barhl1, and Foxb1 expression in three distinct layers of SC. Foxb1 is also expressed in the PTA. Insets show complete sagital brain sections for the genes,documenting expression in additional brain regions. (Scale bars: C–G, 200 μm.) (H) FPKM values across brain regions, retina supernatants, and RGCs for genespresented in C–G. Sample color coding as in Fig. 2E. (I and J) Allen Brain Institute atlas validation outcomes for genes identified in our screen by criteria used inA or B broken down by genes predicted to be expressed only in one retinorecipient nucleus (selective) or expressed at higher levels in two or more reti-norecipient nuclei (intersection sets, enriched). From our candidate lists, 122 (LGN), 116 (PTA), and 134 (SC) were present in the Allen Brain Institute atlas, andof those, more than one-half were nucleus-specific for the LGN and SC, but only about one-quarter were nucleus-specific for the PTA. Transcripts in all Venndiagrams and pie charts are listed in Dataset S5.

E3978 | www.pnas.org/cgi/doi/10.1073/pnas.1618551114 Sajgo et al.

Dow

nloa

ded

by g

uest

on

Mar

ch 2

2, 2

021

with its more specific role in RGC type specification. The clus-tergram of transcripts enriched in retinorecipient areas (Fig. S5)reveals clusters specific for three investigated nuclei (LGN, PTA,and SC) that seem to have little expression in the retina or RGCs.However, about 30 transcripts are expressed in RGCs at levelsequal to or higher than those in retinorecipient areas. Collectively,the identified CSMs and guidance receptors could represent homo-or heterotypic interacting cell–cell adhesion or ligand–receptorpairs involved in targeting the axons of specific RGC types to thesenuclei or establishing cognate synaptic connections.

Molecular Determinants of Neuronal Morphology. Dendritic arborbranching patterns of neurons as diverse as RGCs and somato-sensory neurons of the peripheral nervous system [dorsal rootganglia (DRG)] bear striking resemblances (6, 40, 54, 55), despitedeveloping in distinct cellular and extracellular matrix environ-

ments. This observation prompted us to ask whether some of ourRGC-specific genes could cell-autonomously support process(neurite) extension and branching as proposed in other contexts(56–58). To test this possibility, we overexpressed in HEK293 cellsa set of 10 candidates identified in our RNASeq data and analyzedthe effects on cellular morphology (Fig. 7). The 10 genes had beenconfirmed by in situ and belong to molecular families known to af-fect cellular morphology, but they were relatively unstudied in thiscontext. They range from TFs (Irx4) to cytoskeletal adaptors(Ablim3 and Stmn3), molecules mediating cytoskeleton–membraneinteractions [S100a10, copines (Cpne), and Epb4.1l3], membranefolding or trafficking [Reep5 and reticulon receptor 4rl1 (Rtn4rl1)],or putative cell–cell adhesion molecules (Rtn4rl1, PcdhA1, andIgsf) (Fig. 7 and Dataset S8). Target genes were overexpressedusing transfection into HEK293-Cre cells using a Cre-dependentFLEX approach (Fig. 7A) (59). The intracellular localization ofthe expressed gene can be tracked using its HA tag, whereas amembrane-attached EGFP (meGFP) reveals the effects on cell

50(11)

18(10)

41(23)

16(5)

45(26)

19(3)

16(0)

SCLGN

PTA

Brn3aAP

123(92)

84(55)

42(8)

Brn3bAP

P3E15

167(9)

54(18)

72(44)

17(1)

7(1)

45(26)

Brn3a WTvsKO

Brn3bWTvsKO

39(16)

14(6)

38(21)

P3 RGCs Brn3aAP vs Brn3bAP

Brn3bAP RGCsE15 vs. P3

Overall Transcription Factor Expression

All TF

Expressed in RGCs

Enriched in RGCs

Regulated by Brn3a or Brn3b

Tbx18Zfhx4

Fbxl19Sox14

Pax3Bhlhe23

Otx2Tfap2bPou4f1Pou4f2

Sp5Smyd1

Sp9Zic1Zic4

Brain

PTA

SC

LGN

WT P3

P3WTvsKO

E15WTvsKO

A C

E

FG

H

-2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5

Tfap2bTfap2aIrx5Tbx20Irx1EomesNhlh1Nhlh2Sox4Tle1Sox2Onecut3Onecut1Sox11Eya2Atoh7Onecut2Dlx1Zic2Dlx2Barhl2Klf7Isl1Pou6f2Ebf3Irx3Irx2Ebf1Pou4f2Irx6Irx4Pou4f1Tbr1Myt1lKlf2Klf8Pou4f3Isl2

RGC RetinaP3 E15P3

Brn3b

Brn3b

Brn3a

Brn3a

Brn3b

Brn3b

WT

KO

KO

WT

KO

WT

KO

WT

KO

WT

KO

WT

RGC

a b c

1

2

3

4

5

B D

1647

2437

322(153)

95(43)

10

20

30

40

50

60

Fig. 5. TF repertoire of Brn3aAP and Brn3bAP RGCs and retinorecipient brainareas. (A–D) Venn diagrams representing TF genes enriched in Brn3AP RGCs.Numbers represent transcripts identified using the Twofold and DESeq(parentheses) criteria. Note that the DESeq criteria find a much smallernumber of significantly differentially expressed TFs. (E) Venn diagram rep-resenting global TF expression in Brn3AP RGCs. Of a list of 2,437 TFs andtranscriptional activity molecules (45, 46), 1,647 are expressed at more thanone FPKM in Brn3AP RGCs. TF genes enriched in Brn3aAP and/or Brn3bAP RGCsat E15 and/or P3 number 322 by the Twofold criteria (more than two FPKMin RGCs and more than twofold in RGCs compared with retina) and153 by the DESeq protocol. TF genes differentially expressed in RGCs in aBrn3a- and/or Brn3b-dependent manner are either 95 (Twofold criteria) or43 (DESeq). (F) Venn diagram with TFs selectively expressed in LGN, SC, andPTA. Numbers represent transcripts identified using the Twofold criteria andDESeq (parentheses). (G) Examples for TFs expressed either selectively orjointly in three retinorecipient nuclei. Expression levels normalized to themaximum level of each gene and displayed on a 64-level heat map (red ishigh). (H) Clustergram of TFs believed to be important for RGC development.Sample names are labeled at the top, and hierarchical cluster major branchesare color-coded and labeled a–c. TFs (highest expressed transcript) are an-notated to the right, and hierarchical tree branches of interest are color-coded and labeled 1–5. Color scale bar is at the bottom, and units are in SDs(clustergram details are in Materials and Methods and Fig. 2B). Clustergramsof the complete sets of TFs regulated by Brn3s, enriched in RGCs, or selectivefor particular retinorecipient areas are provided in Figs. S2 and S3. DatasetS6 lists all transcripts in the Venn diagrams.

PTA

8(0)

4(0)

63(10)

Brn3aWTvsKO

Brn3bWTvsKO

P3

E15

59(4)

64(13)

62(73)

P3 RGCs Brn3aAP vs Brn3bAP

Brn3bAP RGCsE15 vs. P3

19(3)

11(0)

56(10)

P3WTvsKO

E15WTvsKO

Brn3aAP

72(69)

105(80)

21(3)

Brn3bAP

SCLGN

PTA

16(7)

3(0)

21(8)

1 (2)

4(1)

12(2)

12(0)

Retinorecipient areas

Cdh5Cdh8Chl1AlcamCdh13Cdh4NrcamCntn1Efna3Cntn5Cdh6Cdh12DscamRobo2Sema6aSema3eF11rFstl4Sdk1Nrp1Plxna1Tenm3Slit1Sema3aNfascEpha6Plxna3Cntn2Efna5Efna4Efnb2Sema5bSema5aEpha8Robo3Plxna2Jam2

−3 −2 −1 0 1 2 3

LGN

RGC RetinaP3E15 E15P3

Brn3b

Brn3b

Brn3a

Brn3b

Brn3b

Brn3a

WT

KO

WT

KO WT

KO

KO

WT

WT

KO

KO

WT

BrainP3

WT

Brain

SC

RGCP3

A

C

B

D

E

F

ab c

d

1

2

3

4

Fig. 6. CSM repertoire of Brn3aAP and Brn3bAP RGCs and retinorecipientbrain areas. (A–D) Venn diagrams representing adhesion molecule genesfrom several molecular families believed to be important in neurite forma-tion that were enriched in Brn3AP RGCs. Expression criteria and comparisonsets are identical to those in Fig. 5 A–D. The survey includes 822 genes, andthe extended analysis is provided in Dataset S7. (E) Venn diagram for ad-hesion molecules enriched or selective for retinorecipient areas. Expressioncriteria and comparison are the same as in Fig. 5F (Dataset S7). For A–E,numbers represent transcripts identified using the Twofold and DESeq (pa-rentheses) criteria. (F) Clustergram of a subset of adhesion molecules be-lieved to be important for RGC development. Clustering algorithm andannotations are same as in Fig. 4. Sample names are labeled along the top,and hierarchical cluster major branches are color-coded and labeled a–d.Genes (highest expressed transcript) are annotated on the right, and hier-archical tree branches of interest are color-coded and labeled 1–4. The colorscale bar is at the bottom, and units are in SDs. Clustergrams covering thecomplete sets of differentially regulated or expressed adhesion moleculesand guidance receptors are provided in Dataset S7 and Figs. S4 and S5.

Sajgo et al. PNAS | Published online May 2, 2017 | E3979

DEV

ELOPM

ENTA

LBIOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

Mar

ch 2

2, 2

021

membrane shape (Fig. 7B). As evident from Fig. 7C, large cel-lular processes reminiscent of lamellipodia are evident whenStmn3 and Igsf6 are overexpressed, whereas more branched,dendritic-like processes are evident in cells expressing S100a10,Epb4.1l3, Rtn4rl1, or Cpne4, often resulting in a nominal in-crease of cell area as defined by the bounding polygon (Fig. 7Dand Dataset S8). More modest but still significant changes arevisible in the case of Irx4, Ablim3, Reep5, or PcdhA1. Thus,isolated overexpression of several molecules identified in ourscreen may induce morphological changes consistent with neu-rite formation in cultured epithelial cells.

Subcellular Localization of Candidate Genes in RGCs in Vivo. Con-sistent with the extensive genetic diversification of molecularfamilies in vertebrates, many of the genes identified in our screenare members of large molecular families. Thus, loss of functionscreens are not expected to yield dramatic results. However, westudied the subcellular distribution and gain of function effectfor some of our candidate molecules by overexpressing them invivo in RGCs. Cre-dependent adeno-associated virus 1 (AAV1) viralvectors for S100a10, Rtn4rl1, Cpne4, and Igsf6 were injected in theretinas of Brn3bCre/WT and Brn3bCre/Cre mice (60) at P0, and sub-cellular distribution of the expressed genes together with dendriticarbor morphologies of infected neurons were revealed by immu-nostaining in adult mice (Fig. 8, Dataset S8, and Fig. S6). Althoughmany RGC bodies were meGFP-positive, accurate imaging waspossible mostly for the large, sharply laminated dendritic arbor types(characteristic of ON and OFF alpha RGCs). None of the overex-pressed molecules induced dramatic changes in dendritic arbormorphologies in either Brn3bCre/WT (WT) or Brn3bCre/Cre (KO)neurons. As an example, dendritic arbor areas of labeled RGCs didnot show any changes on either Brn3bCre/Cre (KO) or Brn3bWT/Cre (WT) RGCs (Fig. 8, Dataset S8, and Fig. S6). The subcellularlocalization of the four molecules varied widely. S100a10, a non-calcium-binding member of the S100 family involved in membraneprocesses, including signaling and membrane fusion (61, 62), wasrestricted to RGC bodies (including the nucleus) and axon (Fig.8A and Fig. S6A). Rtn4rl1 (Nogo receptor 1), implicated in axonnavigation across myelinated substrates and endoplasmic re-ticulum function (63, 64), is restricted to the cell body and someproximal dendrites (Fig. 8B and Fig. S6B). Cpne4, a Ca2+ binding

PTPY Irx4

S100a10

Cpne4

Stmn3

Ablim3

Epb4.1l3

Reep5

Rtn4rl1

PcdhA1

Igsf6

HA

ITR

ITR

CMV Promoter SV40 polyA

GeneXmeGFP P2AA

C

D

Cre

pAAV_FLEX_GeneHA_meGFP

Cre Cre Cre

Bor or

PT

PY

Irx4

S10

0a10

Cpn

e4

Stm

n3

Abl

im3

Epb

4.1l

3

Ree

p5

Rtn

4rl1

Pcd

hA1

Igsf

6

104

103

102

Are

a ( μm

2 )

a b c d e

f g

h i j k

b’

c’

d’

e’

f’g’

h’

j’

k’

i’

Fig. 7. Heterologous overexpression of several RGC-derived genes can affectHEK293 morphology. (A and B) Adeno-associated vector and the overexpressionstrategy used. A minimized CMV promoter drives an expression cassette flankedby tandem inverted lox sites (black, loxP; white, lox2272) and followed by aminimized SV40 polyadenylation signal. The expression cassette is in reverseorientation and will be activated by inversion/excision induced by the tandemlox sites (FLEX strategy). It consists sequentially of the cDNA for the gene to beexpressed (GeneX), a triple HA tag (HA), a self-cleaving T2A peptide, andEGFP coupled to a GAP43 membrane localization signal (meGFP). ITRs are viralinverted terminal repeats. After Cre-mediated inversion–excision, the twopeptides are transcribed and separated on translation; the meGFP reveals theplasma membrane, whereas GeneX distributes in its expected subcellularcompartment (e.g., nucleus, intracellular compartment, or plasma membrane)and can be traced by immunostaining with αHA. (C) Examples of HEK293-Crecells transfected with the pAAV_FLEX_GeneX_meGFP vectors. PTPY is a controlvector containing teal fluorescent protein tagged with a PSD95 domain andmembrane attached yellow fluorescent protein (meYFP). The overexpressedgenes are indicated (b–k′). Note extensive process formation or cell area en-largements in HEK293-Cre cells expressing S100a10, Stmn3, Igsf6, Epb4.1l3,Rtn4rl1, and Cpne4. For b–k, highlighted marquees are enlarged, and greenand red channels are shown separately in b′–k′. (D) All overexpressed vectorsexhibited highly significant cell area enlargements compared with the PTPYcontrol (P < 0.005) (Dataset S8). The y axis is in log10 scale. Box and whiskersplots: the tops and bottoms of each box are the 25th and 75th percentiles ofthe samples, respectively (interquartile ranges). The lines in the middle ofboxes are the sample median. Whiskers are drawn from the ends of theinterquartile ranges to the farthest observations within the whisker length(the adjacent values). Short red lines represent outliers. (Scale bar: C, 20 μm.)

Brn3bCre/WT + P0 : AAV1.CMV.FLEX.GeneX-HA.T2A.mGFP.SV40polyA

Den

dS

oma

Axo

n

S100a10 Rtn4rl1 Cpne4 Igsf6

2.5

2

1.5

1

0.5

0

S100a10 Cpne4Rtn4rl1 Igsf6

Are

a (x

105

mm

2 )

WT KO

A C DB

E

WT KO WT KO Het KOBrn3bCre

Fig. 8. Subcellular localization of candidate genes in RGCs in vivo.(A–D) Examples of adult RGCs from retinas of Brn3bCre/WT mice infected atP0 with AAV1 viral vectors for S100a10, Rtn4rl1, Cpne4, and Igsf6 (constructsare in Fig. 7). Retinas were whole mount-fixed and immunostained with αHA(red) and αGFP (green). (Upper) Merged image of RGC in flat mount per-spective; white marquee squares outline soma and segments of dendritic arborand axon. Lower shows red, green, and merge channels for the highlightedareas. White arrowheads point to meGFP marking dendrite and axon, andwhite arrows point to HA-tagged gene localization in dendrite and/or axon.Note that, in D, axons and dendritic arbors of two RGCs are visible, and Igsf6 islocalized to the axon arbor of only one of them. (E) Dendritic arbor areameasurements for Brn3bCre/WT (WT) or Brn3bCre/Cre (KO) RGCs infected withexpression vectors for four genes (Datasets S1 and S2). Box and whiskers plotsconventions are the same as in Fig. 7D. Examples for the Brn3bCre/Cre (KO) RGCinfections are shown in Fig. S6. (Scale bars: A–D, 100 μm; Insets, 35 μm.)

E3980 | www.pnas.org/cgi/doi/10.1073/pnas.1618551114 Sajgo et al.

Dow

nloa

ded

by g

uest

on

Mar

ch 2

2, 2

021

protein with potential synaptic functions (65, 66) was distributedlargely in cell bodies (including the nucleus) and dendritic arbors(punctate pattern) and to a lesser extent, axons (Fig. 8C and Fig.S6C). Igsf6, an Ig gene of undefined function, was detectable indendrite and cell body; however, its axonal distribution seemed tovary from cell to cell (Fig. 8D and Fig. S6D). These distinct sub-cellular distributions were confirmed for S100a10, Cpne4, andIgsf6 by staining WT uninfected retinal sections with antibodiesagainst the endogenous proteins (Fig. S6 E–G). Thus, tissue cul-ture overexpression, subcellular distribution, molecular structure,and previously described functions in other systems collectivelypoint to a role for these molecules in the establishment of neu-ronal arbor morphology. Because S100a, Cpne, Rtn4r, and Igsfare large molecular families, it is possible that combinatorial ex-pression of one or a few members of these families contributes tothe diversification of neuronal arbors. Transcriptional regulationby Brn3 TFs could, therefore, contribute to RGC type specifica-tion by driving the selective expression of these downstream tar-gets and hence, inducing distinct features in RGCs.

DiscussionWe use RNASeq of purified RGC populations and retina con-trols to identify combinatorial expression codes for TFs andCSMs proposed to be important for the development andspecification of RGC types.Previously, RGCs have been purified and profiled using

immunopanning onto substrates coated with anti–Thy-1 mAb,laser capture microdissection of the GCL, or FACS sorting ofgenetically labeled GFP-positive neurons (67–71). Althoughimmunopanning yields a nearly pure population, it relies onmaintaining RGCs in culture for prolonged periods of time,potentially altering expression profile. GCL microdissection (orlaser capture) yields fresh cells; however, the sample will containmany amacrine neurons. None of the GFP-labeled lines arelimited to RGCs at this point, and relatively small amounts ofRGCs are typically collected. Compared with these methodolo-gies, our immunomagnetic purification approach is relatively fast(about 90 min from mouse to freezer or lysis buffer) and has arelatively high recovery rate (about 25% of potential targetcells), but it results in less pure samples. RNA profiling involvingmicroarray expression analysis only covers a subset of genes andprovides relative gene expression level based on probe hybridization.RNASeq represents the entire collection of transcripts obtainedthrough reverse transcription in a relatively unbiased manner, andeach transcript is typically covered by multiple hits. For both ap-proaches, a limiting amount of starting RNA could be an issue,because no matter the depth of sequencing, saturation in number ofrecovered genes is achieved as recently shown by single-cell RNASeqfrom retinal populations (72, 73). In addition, amplification biasescould occur early in the amplification process, skewing transcriptrepresentation. We, therefore, advocate for validation of RNASeqexperiments with ISH or protein detection techniques.Our dataset confirms between 25 and 50% of the RGC-

specific genes previously identified by FACS sorting, immuno-panning, or GCL laser capture (67, 68, 71), with the intersectionconsisting mostly of well-established RGC markers (Dataset S9and Fig. S7). We, however, expand the potential number oftargets by an order of magnitude (Dataset S9 and Fig. S7). Thisincrease could be caused by the particular comparisons per-formed in the different screens (e.g., developmental time pointsof immunopanned RGCs), the purity of the samples (e.g., RGCsvs. Amacrines in the GCL laser capture experiment), the depthof profiling in microarrays vs. RNASeq, or other experimentaldifferences. Expression profiling experiments using microarrayswere previously performed on Brn3bKO/KO retinas (74, 75). Ofthe combined 234 identified target genes, only 49 are also pre-sent in our combined Brn3b regulated dataset comprising1,008 genes (Dataset S9 and Fig. S7D). These 49 common genes

contain many well-characterized RGC markers and Brn3b targetgenes. The relative lack of overlap could be explained by(i) dilution of the Brn3b-dependent, RGC-specific genes by thewhole retinal tissue or (ii) secondary effects of RGC loss onother retinal precursors.In our hands, only a small fraction of RGC-enriched transcripts

is regulated by Brn3b or Brn3a in RGCs; hence, other TFs mayparticipate in achieving the full RGC phenotype. Of 226 retinalgenes affected by Atoh7 loss (76), only 52 where enriched in E15Brn3bAP/WT RGCs (Dataset S9 and Fig. S7E). However, from165 genes enriched in Atoh7+ cells (77), 120 were also enriched inBrn3bAP/WT RGCs (Dataset S9 and Fig. S7F). The RGC programmay not, therefore, be completely defined by either Atoh7(Math5) or Pou4f2 (Brn3b), and Atoh7 may be required but maynot be sufficient for Brn3b–RGC specification in cell-autonomousand -nonautonomous fashions. Several TFs enriched in Atoh7+

Brn3b− precursors (Dlx1, Dlx2, Onecut2, and Onecut3) (DatasetS9) are negatively regulated by Brn3b at E15 (Fig. 5H) togetherwith Atoh7 itself. Loss of function phenotypes of these genes in-clude RGC defects linked to defects in Horizontal cells (Onecut),and Amacrines (Dlx) (27, 30, 78, 79). Among the TFs enriched inAtoh7+Brn3b+ early RGCs and regulated by Atoh7 and Brn3b,some may control specific RGC subtypes (Eomes), whereas othershave broader roles in the retina (Onecut1 and Isl1) combined withfunctions in RGCs (22, 30, 31). Because negative feedback loopsseem to exist between Brn3b, and Atoh7, Dlx, and Onecut (Fig.5H), it is possible that Brn3b cooperates with more broadlyexpressed TFs to define RGCs, while suppressing others (22, 28,79). Our previous work suggests that RGC types might be de-termined by a combination of TF profiles, each encoding distinctfeatures of the RGC type definition. Similar combinatorial codeshave been described in the spinal cord of vertebrates, the visualand olfactory systems of flies, and the nervous system of Caeno-rhabditis elegans (1–3). Our approach enables us to look at the cell-autonomous effects of Brn3a and Brn3b loss from RGCs. At least43 TF genes were differentially expressed in Brn3 KO RGCscompared with the WT controls, a majority in Brn3b. Becausethese transcripts could indirectly depend on Brn3, say by themodulation of other TFs, ChipSeq analysis would help establishthe direct transcriptional relationship. By contrasting Brn3a+,Brn3b+, and Brn3a+Brn3b+ populations, one can gain insights intomore narrow RGC subgroups. As an example, Brn3b+Brn3a−

RGCs make up Opn4+ intrinsically photosensitive RGCs(ipRGCs)-positive cells (6, 13, 80). Here, we show that this sub-set of RGCs expresses Eomes, Tbox20, and Irx1, consistent with apossible role of these TFs in ipRGC specification (29, 32). Fur-thermore, the few genes regulated by Brn3a may be directly rel-evant to cell type specification of the few cell types missing fromBrn3aKO RGCs (5, 13, 31). The degree of similarity between RGCtranscriptional profiles and retina depends on both the de-velopmental age and Brn3 expression status. In several of ourclustergrams, Brn3bAP/KO RGCs are segregating together with theretina samples and separated from other RGC populations.We report a large set of CSMs overexpressed in RGCs com-

pared with the retina. CSMs affected by Brn3 ablation in Brn3+

RGCs are limited in numbers, especially for Brn3aAP RGCs.Thus, very specific candidate genes were identified for the subtledendritic arbor distinctions in Brn3aAP/KO RGCs. In contrast tothe TFs clusters, CSMs seem to more clearly distinguish RGCsamples from retina controls (Figs. S4 and S5). About 30% of theidentified adhesion molecules have already been implicated invarious aspects of RGC development, but the separation byBrn3 assignment and the differential expression in Brn3b WT vs.KO RGCs will no doubt help to better target the assignment ofindividual CSMs to RGC subpopulations. Of particular interestis the negative regulation by Brn3b of a subset of adhesionmolecules known to mediate axon guidance decisions (Efna4,Efna5, Efnb2, Epha6, Cntn2, Nfasc, Tenm3, Plxna1, Plxna3,

Sajgo et al. PNAS | Published online May 2, 2017 | E3981

DEV

ELOPM

ENTA

LBIOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

Mar

ch 2

2, 2

021

Sema3a, and Slit1) at E15 but not P3. Intriguingly, Brn3b seemsto also suppress Zic2 expression in E15 RGCs, a TF required forcorrect ipsilateral vs. contralateral segregation at the optic chi-asm (Fig. 5H) (81). Many of the disregulated guidance cuesfunction as repellants, potentially explaining the abnormalbranching into the inner plexyform layer or nonvisual nuclei ofthe thalamus observed in Brn3bAP/KO RGC axons.Whole-tissue profiling of retinorecipient nuclei yielded many

interesting markers, including some that appeared to be lami-nated in the SC. Many of the identified TFs confirm previouswork on brain patterning; however, the relatively reduced num-ber of differentially expressed adhesion and guidance moleculesis somewhat surprising. It could be that RGC axons are triagedduring the axon guidance process in the various intermediatepoints (optic nerve, chiasm, tract, and brachium) and thus, needfewer cues after they have reached the target areas. Alterna-tively, RGCs may make connections in several nuclei and hence,use similar guidance/recognition codes. Undoubtedly, cell puri-fication, sorting, and profiling approaches using specific markerswill distinguish more defined molecular cues. For now, ourdata provide a useful entry point into the molecular diversity ofretinorecipient neurons.Cell-autonomous mechanisms could determine the shape and

size of neuronal arbors. Dissociated cultures of cortical and hip-pocampal neurons, starburst amacrine cells, and cerebellar Pur-kinje neurons can adopt morphologies similar to those found invivo (82). Our overexpression screen identifies several moleculescapable of inducing neurite-like extensions in HEK293 cells. Themolecular nature of these genes includes cytoskeleton interactors,transmembrane proteins with roles in intracellular membranetraffic, and membrane-associated molecules with somewhatcomplex roles in membrane remodeling. Most of the studied genesare members of large molecular families. It is possible that neu-ronal arbor morphologies could initially be shaped by the com-binatorial expression of these genes and then, adjusted andsculpted by negative or positive cues or activity. In some instances,RGCs expressing particularly large amounts of Rtn4rl1, Cpne4,and S100a10 did exhibit changes of the dendritic arbors; however,these defects were rare and inconsistent enough not to be includedin this work. However, our expression system allowed us to studythe subcellular localization of our candidates in RGCs. It shouldbe noted that subcellular localization of the various candidates

could be influenced by RNA processing and trafficking mecha-nisms, tagging, and/or overexpression and should be confirmedusing antibodies against the endogenous protein.The follow-up and individual investigation of the identified

target genes are beyond the scope of this study, but our datasetcould benefit a large group of researchers interested in cell typespecification, cell surface molecular codes, and pathogeneticmechanisms of glaucoma. The strategy and reagents outlinedhere could be used to isolate and profile many subpopulations ofneurons in the DRG, vestibular and auditory ganglia, and tri-geminal ganglia and some sympathetic and parasympathetic af-ferent neurons, all expressing Brn3s.

Materials and MethodsMouse lines and genetic recombination strategy were previously described (13).Retina dissociation was previously described (83), and immunoaffinity purifica-tion was developed by us using magnetic beads coupled to anti-AP antibodies.RGC recovery was tested by AP staining of bead-coupled RGCs and supernatantsand estimated to about 25% of all RGCs in a retina, whereas fold enrichment ofRGC samples over retinal supernatants was about 65× (Fig. 1F and Dataset S1).Retinorecipient nuclei fromWT P3 mice were visualized by anterograde tracing,dissected under a fluorescent microscope, and processed for RNA isolation. Theentire remaining brain tissue was homogenized and used as control (Fig. 1E).RNA extraction was done using the RNEasy Kit (Qiagen), and RNASeq librarypreparation sequencing on Illumina platforms and Read mapping were pre-viously described (44, 84, 85). Alignments were visualized with Igviewer (86).DESeq, hierarchical clustering, and principal components analysis (PCA) wereperformed using dedicated components of the Matlab Bioinformatics toolbox(42, 87, 88). Gene Ontology analysis was done using the Panther suite. ISH val-idation was performed using the protocol described by Schaeren-Wiemers andGerfin-Moser (89) and queries in the Allen Brain Institute brain atlas.HEK293 overexpression studies, transfections, viral expression vectors, and Brn3-Cre mice AAV infections were previously described. Extensive description ofmaterials and methods is provided in SI Materials andMethods. Mouse handlingprocedures were approved by the National Eye Institute Animal User Committeeunder protocol NEI640. Next generation sequencing data reported here areavailable under Gene Expression Omnibus accession number GSE87647.

ACKNOWLEDGMENTS. We thank Norimoto Gotoh for retina dissociationprotocols; Harsha K. Rajasimha for assistance with Bowtie analysis; PeterColosi, Scott Sternson, and Rachel O. Wong for AAV constructs; Beverly Wufor help with gene expression cloning; and Nadia Parmhans for genotyping.Funding was provided by the Intramural Research Program of the NationalEye Institute (Z.W. and T.C.B.).

1. Hobert O (2011) Regulation of terminal differentiation programs in the nervoussystem. Annu Rev Cell Dev Biol 27:681–696.

2. Jessell TM (2000) Neuronal specification in the spinal cord: Inductive signals andtranscriptional codes. Nat Rev Genet 1:20–29.

3. Komiyama T, Luo L (2006) Development of wiring specificity in the olfactory system.Curr Opin Neurobiol 16:67–73.

4. Bassett EA, Wallace VA (2012) Cell fate determination in the vertebrate retina. TrendsNeurosci 35:565–573.

5. Badea TC, Nathans J (2004) Quantitative analysis of neuronal morphologies in the mouseretina visualized by using a genetically directed reporter. J Comp Neurol 480:331–351.

6. Badea TC, Nathans J (2011) Morphologies of mouse retinal ganglion cells expressingtranscription factors Brn3a, Brn3b, and Brn3c: Analysis of wild type and mutant cellsusing genetically-directed sparse labeling. Vision Res 51:269–279.

7. Baden T, et al. (2016) The functional diversity of retinal ganglion cells in the mouse.Nature 529:345–350.

8. Coombs J, van der List D, Wang GY, Chalupa LM (2006) Morphological properties ofmouse retinal ganglion cells. Neuroscience 140:123–136.

9. Helmstaedter M, et al. (2013) Connectomic reconstruction of the inner plexiform layerin the mouse retina. Nature 500:168–174.

10. Sümbül U, et al. (2014) A genetic and computational approach to structurally classifyneuronal types. Nat Commun 5:3512.

11. Dräger UC (1985) Birth dates of retinal ganglion cells giving rise to the crossed anduncrossed optic projections in the mouse. Proc R Soc Lond B Biol Sci 224:57–77.

12. Godement P, Salaün J, Imbert M (1984) Prenatal and postnatal development of reti-nogeniculate and retinocollicular projections in the mouse. J Comp Neurol 230:552–575.

13. Badea TC, Cahill H, Ecker J, Hattar S, Nathans J (2009) Distinct roles of transcriptionfactors brn3a and brn3b in controlling the development, morphology, and functionof retinal ganglion cells. Neuron 61:852–864.

14. Coombs JL, Van Der List D, Chalupa LM (2007) Morphological properties of mouse retinalganglion cells during postnatal development. J Comp Neurol 503:803–814.

15. Edwards MA, Caviness VS, Jr, Schneider GE (1986) Development of cell and fiberlamination in the mouse superior colliculus. J Comp Neurol 248:395–409.

16. Chen C, Regehr WG (2000) Developmental remodeling of the retinogeniculate syn-apse. Neuron 28:955–966.

17. Brown NL, Patel S, Brzezinski J, Glaser T (2001) Math5 is required for retinal ganglioncell and optic nerve formation. Development 128:2497–2508.

18. Gan L, et al. (1996) POU domain factor Brn-3b is required for the development of alarge set of retinal ganglion cells. Proc Natl Acad Sci USA 93:3920–3925.

19. Wang SW, et al. (2001) Requirement for math5 in the development of retinal gan-glion cells. Genes Dev 15:24–29.

20. Xiang M, et al. (1997) Essential role of POU-domain factor Brn-3c in auditory andvestibular hair cell development. Proc Natl Acad Sci USA 94:9445–9450.

21. Xiang M, Gan L, Zhou L, Klein WH, Nathans J (1996) Targeted deletion of the mousePOU domain gene Brn-3a causes selective loss of neurons in the brainstem and tri-geminal ganglion, uncoordinated limb movement, and impaired suckling. Proc NatlAcad Sci USA 93:11950–11955.

22. Mu X, Fu X, Beremand PD, Thomas TL, Klein WH (2008) Gene regulation logic inretinal ganglion cell development: Isl1 defines a critical branch distinct from butoverlapping with Pou4f2. Proc Natl Acad Sci USA 105:6942–6947.

23. Sun Y, et al. (2008) A central role for Islet1 in sensory neuron development linkingsensory and spinal gene regulatory programs. Nat Neurosci 11:1283–1293.

24. Wu F, et al. (2015) Two transcription factors, Pou4f2 and Isl1, are sufficient to specifythe retinal ganglion cell fate. Proc Natl Acad Sci USA 112:E1559–E1568.

25. Erkman L, et al. (1996) Role of transcription factors Brn-3.1 and Brn-3.2 in auditoryand visual system development. Nature 381:603–606.

26. Jin K, Jiang H, Mo Z, XiangM (2010) Early B-cell factors are required for specifying multipleretinal cell types and subtypes from postmitotic precursors. J Neurosci 30:11902–11916.

27. Klimova L, Antosova B, Kuzelova A, Strnad H, Kozmik Z (2015) Onecut1 andOnecut2 transcription factors operate downstream of Pax6 to regulate horizontal celldevelopment. Dev Biol 402:48–60.

E3982 | www.pnas.org/cgi/doi/10.1073/pnas.1618551114 Sajgo et al.

Dow

nloa

ded

by g

uest

on

Mar

ch 2

2, 2

021

28. Mao C-A, et al. (2008) Eomesodermin, a target gene of Pou4f2, is required for retinalganglion cell and optic nerve development in the mouse. Development 135:271–280.

29. Mao CA, et al. (2014) T-box transcription regulator Tbr2 is essential for the formationand maintenance of Opn4/melanopsin-expressing intrinsically photosensitive retinalganglion cells. J Neurosci 34:13083–13095.

30. Sapkota D, et al. (2014) Onecut1 and Onecut2 redundantly regulate early retinal cellfates during development. Proc Natl Acad Sci USA 111:E4086–E4095.

31. Shi M, et al. (2013) Genetic interactions between Brn3 transcription factors in retinalganglion cell type specification. PLoS One 8:e76347.

32. Sweeney NT, Tierney H, Feldheim DA (2014) Tbr2 is required to generate a neuralcircuit mediating the pupillary light reflex. J Neurosci 34:5447–5453.

33. Wang SW, et al. (2002) Brn3b/Brn3c double knockout mice reveal an unsuspected rolefor Brn3c in retinal ganglion cell axon outgrowth. Development 129:467–477.

34. Li X, et al. (2016) Loss of AP-2delta reduces retinal ganglion cell numbers and axonalprojections to the superior colliculus. Mol Brain 9:62.

35. Rousso DL, et al. (2016) Two pairs of ON and OFF retinal ganglion cells are defined byintersectional patterns of transcription factor expression. Cell Reports 15:1930–1944.

36. Triplett JW, et al. (2014) Dendritic and axonal targeting patterns of a genetically-specified class of retinal ganglion cells that participate in image-forming circuits.Neural Dev 9:2.

37. Jiang Y, et al. (2013) Transcription factors SOX4 and SOX11 function redundantly toregulate the development ofmouse retinal ganglion cells. J Biol Chem 288:18429–18438.

38. Moore DL, et al. (2009) KLF family members regulate intrinsic axon regenerationability. Science 326:298–301.

39. Sajgo S, Ali S, Popescu O, Badea TC (2016) Dynamic expression of transcription factorBrn3b during mouse cranial nerve development. J Comp Neurol 524:1033–1061.

40. Badea TC, et al. (2012) Combinatorial expression of Brn3 transcription factors in so-matosensory neurons: Genetic and morphologic analysis. J Neurosci 32:995–1007.

41. Xiang M, et al. (1993) Brn-3b: A POU domain gene expressed in a subset of retinalganglion cells. Neuron 11:689–701.

42. Anders S, Huber W (2010) Differential expression analysis for sequence count data.Genome Biol 11:R106.

43. Jeon CJ, Strettoi E, Masland RH (1998) The major cell populations of the mouse retina.J Neurosci 18:8936–8946.

44. Langmead B, Salzberg SL (2012) Fast gapped-read alignment with Bowtie 2. NatMethods 9:357–359.

45. Ravasi T, et al. (2010) An atlas of combinatorial transcriptional regulation in mouseand man. Cell 140:744–752.

46. Zhang HM, et al. (2015) AnimalTFDB 2.0: A resource for expression, prediction andfunctional study of animal transcription factors. Nucleic Acids Res 43:D76–D81.

47. Kim I-J, Zhang Y, Yamagata M, Meister M, Sanes JR (2008) Molecular identification ofa retinal cell type that responds to upward motion. Nature 452:478–482.

48. Matsuoka RL, et al. (2011) Class 5 transmembrane semaphorins control selectiveMammalian retinal lamination and function. Neuron 71:460–473.

49. Matsuoka RL, et al. (2011) Transmembrane semaphorin signalling controls laminarstratification in the mammalian retina. Nature 470:259–263.

50. Wong GK, Baudet ML, Norden C, Leung L, Harris WA (2012) Slit1b-Robo3 signalingand N-cadherin regulate apical process retraction in developing retinal ganglion cells.J Neurosci 32:223–228.

51. Feldheim DA, O’Leary DD (2010) Visual map development: Bidirectional signaling,bifunctional guidance molecules, and competition. Cold Spring Harb Perspect Biol 2:a001768.

52. Mosca TJ (2015) On the Teneurin track: A new synaptic organization moleculeemerges. Front Cell Neurosci 9:204.

53. Tessier-Lavigne M, Goodman CS (1996) The molecular biology of axon guidance.Science 274:1123–1133.

54. Li L, et al. (2011) The functional organization of cutaneous low-threshold mechano-sensory neurons. Cell 147:1615–1627.

55. Wu H, Williams J, Nathans J (2012) Morphologic diversity of cutaneous sensory af-ferents revealed by genetically directed sparse labeling. eLife 1:e00181.

56. Jan YN, Jan LY (2010) Branching out: Mechanisms of dendritic arborization. Nat RevNeurosci 11:316–328.

57. Zipursky SL, Grueber WB (2013) The molecular basis of self-avoidance. Annu RevNeurosci 36:547–568.

58. Corty MM, Matthews BJ, Grueber WB (2009) Molecules and mechanisms of dendritedevelopment in Drosophila. Development 136:1049–1061.

59. Chuang K, Nguyen E, Sergeev Y, Badea TC (2015) Novel heterotypic Rox sites forcombinatorial Dre recombination strategies. G3 (Bethesda) 6:559–571.

60. Sajgo S, et al. (2014) Dre - Cre sequential recombination provides new tools for retinalganglion cell labeling and manipulation in mice. PLoS One 9:e91435.

61. Rescher U, Gerke V (2008) S100A10/p11: Family, friends and functions. Pflugers Arch455:575–582.

62. Dempsey BR, et al. (2012) Structure of an asymmetric ternary protein complex pro-vides insight for membrane interaction. Structure 20:1737–1745.

63. Wang X, et al. (2015) Intravitreal delivery of human NgR-Fc decoy protein regeneratesaxons after optic nerve crush and protects ganglion cells in glaucoma models. InvestOphthalmol Vis Sci 56:1357–1366.

64. Dickendesher TL, et al. (2012) NgR1 and NgR3 are receptors for chondroitin sulfateproteoglycans. Nat Neurosci 15:703–712.

65. Creutz CE, et al. (1998) The copines, a novel class of C2 domain-containing, calcium-dependent, phospholipid-binding proteins conserved from Paramecium to humans.J Biol Chem 273:1393–1402.

66. Reinhard JR, et al. (2016) The calcium sensor Copine-6 regulates spine structuralplasticity and learning and memory. Nat Commun 7:11613.

67. Kim CY, Kuehn MH, Clark AF, Kwon YH (2006) Gene expression profile of the adulthuman retinal ganglion cell layer. Mol Vis 12:1640–1648.

68. Wang JT, et al. (2007) Disease gene candidates revealed by expression profiling ofretinal ganglion cell development. J Neurosci 27:8593–8603.

69. Winzeler A, Wang JT (2013) Purification and culture of retinal ganglion cells fromrodents. Cold Spring Harb Protoc 2013:643–652.

70. Gao F, et al. (2016) Comparative analysis of three purification protocols for retinalganglion cells from rat. Mol Vis 22:387–400.

71. Siegert S, et al. (2012) Transcriptional code and disease map for adult retinal celltypes. Nat Neurosci 15:487–495.

72. Shekhar K, et al. (2016) Comprehensive classification of retinal bipolar neurons bysingle-cell transcriptomics. Cell 166:1308–1323.e30.

73. Macosko EZ, et al. (2015) Highly parallel genome-wide expression profiling of indi-vidual cells using nanoliter droplets. Cell 161:1202–1214.

74. Mu X, et al. (2004) Discrete gene sets depend on POU domain transcription factorBrn3b/Brn-3.2/POU4f2 for their expression in the mouse embryonic retina. Development131:1197–1210.

75. Qiu F, Jiang H, Xiang M (2008) A comprehensive negative regulatory program con-trolled by Brn3b to ensure ganglion cell specification from multipotential retinalprecursors. J Neurosci 28:3392–3403.

76. Mu X, et al. (2005) A gene network downstream of transcription factorMath5 regulates retinal progenitor cell competence and ganglion cell fate. Dev Biol280:467–481.

77. Gao Z, Mao CA, Pan P, Mu X, Klein WH (2014) Transcriptome of Atoh7 retinal pro-genitor cells identifies new Atoh7-dependent regulatory genes for retinal ganglioncell formation. Dev Neurobiol 74:1123–1140.

78. de Melo J, et al. (2005) Dlx1 and Dlx2 function is necessary for terminal differentiationand survival of late-born retinal ganglion cells in the developing mouse retina.Development 132:311–322.

79. Feng L, et al. (2011) Brn-3b inhibits generation of amacrine cells by binding to andnegatively regulating DLX1/2 in developing retina. Neuroscience 195:9–20.

80. Chen SK, Badea TC, Hattar S (2011) Photoentrainment and pupillary light reflex aremediated by distinct populations of ipRGCs. Nature 476:92–95.

81. Herrera E, et al. (2003) Zic2 patterns binocular vision by specifying the uncrossedretinal projection. Cell 114:545–557.

82. Lefebvre JL, Sanes JR, Kay JN (2015) Development of dendritic form and function.Annu Rev Cell Dev Biol 31:741–777.

83. Tabata T, et al. (2000) A reliable method for culture of dissociated mouse cerebellarcells enriched for Purkinje neurons. J Neurosci Methods 104:45–53.

84. Brooks MJ, Rajasimha HK, Swaroop A (2012) Retinal transcriptome profiling by di-rectional next-generation sequencing using 100 ng of total RNA. Methods Mol Biol884:319–334.

85. Kaewkhaw R, et al. (2015) Transcriptome dynamics of developing photoreceptors inthree-dimensional retina cultures recapitulates temporal sequence of human coneand rod differentiation revealing cell surface markers and gene networks. Stem Cells33:3504–3518.

86. Robinson JT, et al. (2011) Integrative genomics viewer. Nat Biotechnol 29:24–26.87. Bar-Joseph Z, Gifford DK, Jaakkola TS (2001) Fast optimal leaf ordering for hierar-

chical clustering. Bioinformatics 17:S22–S29.88. Eisen MB, Spellman PT, Brown PO, Botstein D (1998) Cluster analysis and display of

genome-wide expression patterns. Proc Natl Acad Sci USA 95:14863–14868.89. Schaeren-Wiemers N, Gerfin-Moser A (1993) A single protocol to detect transcripts of

various types and expression levels in neural tissue and cultured cells: In situ hybrid-ization using digoxigenin-labelled cRNA probes. Histochemistry 100:431–440.

Sajgo et al. PNAS | Published online May 2, 2017 | E3983

DEV

ELOPM

ENTA

LBIOLO

GY

PNASPL

US

Dow

nloa

ded

by g

uest

on

Mar

ch 2

2, 2

021