Deep-sequencing-guided library screening and profiling of AAV capsids in the primate retina Authors: Youssef Fouani, Dragica Blazevic, Martin Oti, Robert Becker, Benjamin Strobel, Tanja Schönberger, Marius Stephan, Sven Wichert, Felix F. Reichel, Immanuel P. Seitz, Birgit Korbmacher, Kevin Achberger, Julia Roosz, Peter Loskill, Udo Maier, Stefan Michelfelder, Christian Schön Published: 2026 ABSTRACT Intravitreal (IVT) administration of adeno-associated virus (AAV) vectors is considered a promising strategy for retinal gene therapy. Random peptide-display-based directed evolution has yielded novel AAV capsids with potent retinal bioactivity. However, recent reports on safety risks following IVT AAV administration indicate that significant opportunities remain to further refine AAV vector technology. Here, we performed iterative in vivo NGS-guided screening of AAV2 peptide-display libraries in both mice and non-human primates (NHPs), integrating barcoded functional analytics, and single-nuclei RNA sequencing to identify retina-tropic AAV variants. The top-performing capsid in mice, AAV2-GAYPKSP robustly drove eGFP expression in a human retina-on-chip model but failed to drive gene expression in the NHP retina following IVT delivery. In NHPs, a barcoded evaluation of 70 retina-tropic variants revealed a complex correlation between viral genome enrichment and transgene transcriptional output, providing insights that enhance the mechanistic interpretation and methodological design of AAV library selection strategies. Nonetheless, several NHP variants, including AAV2-I.1 and AAV2-I.12, demonstrated broad retinal transduction and potent expression, outperforming benchmarks AAV2 and AAV2-7m8. Single-nucleus RNA sequencing confirmed broad cell-type tropism of the top performing variants. Collectively, these findings expand the AAV toolkit for retinal gene therapy and underscore the importance of optimized screening methodologies in vector discovery. Introduction Vision is one of the most critical senses for humans and any impairment ranging from reduced visual acuity to complete blindness can severely diminish quality of life. A wide array of inherited and acquired ocular diseases has been identified, and advances in genetics, molecular biology, and imaging technologies have significantly enhanced our understanding of the underlying pathomechanisms. These insights have paved the way for the development of novel therapeutic strategies. Among them, anti-VEGF agents such as aflibercept have shown efficacy in managing neovascular eye diseases like age-related macular degeneration.1,2 Still, many retinal disorders remain without effective therapies, highlighting the need for innovative approaches such as adeno-associated viruses (AAVs)-based gene therapy. AAVs have emerged as the leading modality for retinal gene therapy due to their ability to efficiently deliver therapeutic genes and sustain long-term expression, which enables the treatment of inherited and acquired retinal diseases. A landmark achievement in this field is the approval of Luxturna (voretigene neparvovec) in 2017, the first AAV-based gene therapy for Leber’s congenital amaurosis type 2, caused by biallelic mutations in the RPE65 gene. Luxturna’s success not only demonstrated the clinical potential of AAV-mediated gene delivery but also catalyzed a wave of clinical trials targeting a broad spectrum of retinal disorders.3,4,5 Various AAV serotypes have demonstrated robust transduction of the rodent retina, including photoreceptors and retinal pigment epithelium (RPE), when delivered via subretinal injection, a method still widely employed in clinical trials.6,7,8,9,10 However, alternative delivery routes such as intravitreal (IVT) and suprachoroidal injections have gained increasing attention. Both routes of administration are less invasive and enable broader retinal transduction compared to the localized bleb formed during subretinal delivery.11,12,13,14,15 In early studies, AAV2 was identified as the most potent capsid for retinal transduction after IVT injection in rodents.16 Yet, its efficacy following IVT injection in higher order species, particularly non-human primates (NHPs), is significantly diminished. This reduction is primarily attributed to the limited ability of AAV2 to penetrate the inner limiting membrane (ILM), a critical barrier to vector diffusion and retinal access.17,18 To overcome the poor transduction profile of wild-type capsids after IVT injection, extensive efforts have focused on developing second-generation AAV capsids.19,20,21,22,23,24,25,26,27 Directed evolution and peptide-display strategies have yielded second-generation AAV2-based variants, including AAV2-7m8, NHP#9,28 AAV2.NN, and AAV2.GL,28 some of which have demonstrated improved retinal transduction in preclinical models, including NHP and canines,29,30 and progressed to clinical evaluation (NCT04418427, NCT07002398).22,28 More recently, novel strategies emerged toward directly evaluating capsids in NHPs, human retinal explants, and organoids to better predict clinical performance.31,32,33 These advances have yielded a growing repertoire of second-generation capsids optimized for IVT delivery, showing improved transduction of the retinal periphery and fovea in NHPs. While AAV-based IVT gene delivery has shown substantial promise for treating retinal diseases, concerns regarding dose-limiting toxicity have come into focus. In this context, recent high dose administration (6.0 E + 11 vg/eye) in diabetic macular edema (DME) patients using the clinically validated AAV2-7m8 vectors led to rapid and clinically meaningful reductions in intraocular pressure. These adverse effects prompted early termination of the corresponding clinical trial due to dose-limiting toxicity (NCT04418427). Therefore, there remains a clear need for novel AAV vectors capable of driving robust and efficient transgene expression in both preclinical and clinical settings while minimizing adverse events. In this study, we present a comprehensive screening approach using AAV2-based 7-mer peptide insertion libraries in both mice and NHPs. NGS-based rigorous monitoring of peptide library screening, RNA-expression performance in different retina areas as well as single extracted nuclei allowed the identification of novel, potent bioactive AAV2 variants that outperformed AAV2 and the clinical benchmark for IVT injection, AAV2-7m8. Results In vivo selection and characterization of retina-tropic AAV2 variants To establish our peptide-library screening pipeline and identify AAV2 variants with enhanced retinal tropism, we constructed a diverse AAV2 peptide-display library by inserting randomized 7-mer peptides (1.28 E + 9 theoretical variants) at loop VIII residue 588, as previously described (Figure 1A).34,35,36 NGS of the initial viral pool (R0) confirmed their high diversity (Figure 1B).Figure 1NGS-driven evolution of AAV2 peptide display library in mouse retina and barcoded-based evaluation of the top retina-enriched candidates(A) Workflow of AAV2 peptide library screening through three rounds of selection in mouse retina following IVT administration. (B) Peptide library diversity across rounds of selection. Top 10 retina-enriched AAV2 variants are indicated. R0 = input viral pool. R1-3 = library pool after selection round in vivo. (C) Relative contribution of top 10 retina-enriched variants to R3-captured variants in the retina. D.T[K/R] motif enrichment is indicated. (D) Progressive enrichment of selected variants in the retina throughout the three rounds of screening. (E) NGS-based quantification of variant barcodes on DNA and RNA levels in mouse retinae. The relative enrichment of individual variants was calculated as the normalized count of each variant divided by the total normalized barcode reads across all variants. Both variant expression and genome abundance were normalized to variant abundance in the input viral library. (F) Representative cSLO images of eGFP expression in mouse retinae. (G) Immunohistochemistry (anti-GFP, chromogen: DAB) based detection of AAV-mediated expression in FFPE sections of the mouse retina. GCL, ganglion cell layer; INL, inner nuclear layer; ONL, outer nuclear layer. Scale bars, 100 μm. (H) RT-qPCR-based quantification of relative eGFP mRNA expression in mouse retinae. Values indicate mean ± SEM (n = 4–6). Statistical significance compared to BSS buffer was determined using one-way ANOVA with Tukey’s test for multiple comparisons. ∗p < 0.05, ∗∗p < 0.005, p > 0.05, (ns, not significant). NGS-driven evolution of AAV2 peptide display library in mouse retina and barcoded-based evaluation of the top retina-enriched candidates (A) Workflow of AAV2 peptide library screening through three rounds of selection in mouse retina following IVT administration. (B) Peptide library diversity across rounds of selection. Top 10 retina-enriched AAV2 variants are indicated. R0 = input viral pool. R1-3 = library pool after selection round in vivo. (C) Relative contribution of top 10 retina-enriched variants to R3-captured variants in the retina. D.T[K/R] motif enrichment is indicated. (D) Progressive enrichment of selected variants in the retina throughout the three rounds of screening. (E) NGS-based quantification of variant barcodes on DNA and RNA levels in mouse retinae. The relative enrichment of individual variants was calculated as the normalized count of each variant divided by the total normalized barcode reads across all variants. Both variant expression and genome abundance were normalized to variant abundance in the input viral library. (F) Representative cSLO images of eGFP expression in mouse retinae. (G) Immunohistochemistry (anti-GFP, chromogen: DAB) based detection of AAV-mediated expression in FFPE sections of the mouse retina. GCL, ganglion cell layer; INL, inner nuclear layer; ONL, outer nuclear layer. Scale bars, 100 μm. (H) RT-qPCR-based quantification of relative eGFP mRNA expression in mouse retinae. Values indicate mean ± SEM (n = 4–6). Statistical significance compared to BSS buffer was determined using one-way ANOVA with Tukey’s test for multiple comparisons. ∗p < 0.05, ∗∗p < 0.005, p > 0.05, (ns, not significant). For each in vivo screening round, both eyes of three C57BL/6NCrl mice were intravitreally injected with the AAV2 library at a dose of 4.2 E + 09 vg/eye (Figure 1A). Eight days post-injection, retinal, and RPE tissues were harvested, and NGS was used to quantify the abundance of AAV genomes, enabling assessment of variant enrichment. The recovered viral genomes were then repackaged for subsequent screening rounds. As selection progressed, the intermediate viral libraries became increasingly enriched for a subset of retina-tropic variants, evidenced by a marked reduction in peptide diversity: from 99.9% unique sequences in round 1 (R1), to 66.1% in R2, and 40% in R3 (Figure 1B). By the third round, approximately 28,300 unique peptides were recovered from retinal tissue, including 1,077 candidates with abundance >100 counts per million. The top 10 peptides accounted for 61.1% of all retina-enriched variants, with AAV2-GDKTKYP accounting for 42.9% of the total reads (Figures 1B and 1C). AAV2-GDKTKYP showed the most pronounced progressive enrichment throughout the screening procedure, exhibiting ×17.7-fold enrichment in R2 compared to R1 and ×46.2-fold increase in R3 compared to R1 (Figure 1D). A detailed analysis of retina-captured peptides from R3 revealed several distinguishing biochemical features, including enrichment of aromatic and basic amino acids, and increased hydrophobicity. Cluster-based motif analysis of retina-enriched AAV2 variants identified a D.T[K/R] motif, defined as an aspartate (D), followed by any amino acid, then threonine (T), and either lysine or arginine (K/R) (Figures S1A–S1C), with 7 out of the top 10 retina-enriched variants harboring such motifs (Figure 1C). Subsequently, we nominated retina-enriched variants to test their retinal tropism and transcriptional activity in a barcoded screening campaign.37 Based on their abundance and progressive enrichment in the mouse retina, four AAV2 variants (AAV2-GDKTKYP, AAV2-PDATKHG, AAV2-DDATKQK, and AAV2-GAYPKSP) were nominated and vectorized, alongside four benchmark capsids (AAV2, AAV8, AAV9, and AAV2-7m8). Additionally, two RPE selective variants (AAV2-DNLHRKE and AAV2-PTDKKMP) were also included (Figure S1D). Each vector encapsulated a unique DNA barcoded reporter (eGFP) driven by a cytomegalovirus (CMV) promoter and was individually produced. All vectors were then pooled with the aim of uniformly representing individual AAV variants in the library.37 The barcoded library was then IVT administered into female C57BL/6NCrl mice at a total dose of 2.0 + E09 vg/eye (2.0 E+ 08 vg/variant/eye). DNA and RNA were extracted from the RPE and the retina 28 dpi (days post-injection) and AAV genome abundance and eGFP transcript levels were quantified using NGS (Figure S1D). Among the novel variants, AAV2-GDKTKYP and AAV2-GAYPKSP were the top transducers, reaching levels comparable to the benchmark vectors AAV2 and AAV2-7m8 (Figure 1E, DNA). At the transcriptional level, AAV2-GAYPKSP achieved the highest eGFP mRNA expression across all variants, outperforming both AAV2 and AAV2-7m8 (Figure 1E, RNA). In contrast, the RPE-tropic variants AAV2-DNLHRKE and AAV2-PTDKKMP showed poor retinal transduction and eGFP expression (Figure 1E, DNA/RNA). In RPE, AAV2-7m8 was the most efficient transducer, followed by AAV2-GDKTKYP and AAV2-GAYPKSP (Figure S1E, DNA), while AAV2-PDTKKMP, AAV2-GAYPKSP, and AAV2 drove the highest eGFP expression (Figure S1E, RNA). To validate these NGS-based findings, we evaluated the top five retina-specific AAV2 variants in individual eGFP reporter experiments in vivo. Each AAV was administered IVT into female C57BL/6NCrl mice at 1.0 E + 9 vg/eye. Expression of eGFP was visualized and quantified at 28 dpi using light fundus autofluorescence (FAF), immunohistochemistry, and quantitative reverse-transcription PCR (RT-qPCR). Consistent with the barcoded screen, AAV2-GAYPKSP drove robust and evenly distributed eGFP expression throughout the eye, outperforming AAV2-7m8 (Figures 1F–1H). In contrast, retinae IVT injected with the other top-ranked capsids from our screen, AAV2-GDKTKYP, AAV2-GQDTTRF, and AAV2-PDATKHG, displayed only sparse, punctate eGFP expression (Figure 1F). Immunohistochemistry revealed that both AAV2-GAYPKSP and AAV2-7m8 drove eGFP expression throughout the ganglion cell layer (GCL), partially in the inner nuclear layer (INL), and to lesser extent in the outer nuclear layer (ONL) (Figure 1G). Moreover, a transduction gradient starting from the injection site could be observed with AAV2-7m8, whereas GAYPKSP transduced the whole circumference evenly, also indicating higher potency. Based on its performance in barcoded screening and single variants, we next assessed AAV2-GAYPKSP, expression kinetics by FAF imaging up to 107 dpi and compared its transcriptional profile to that of AAV2-7m8 in C57BL/6NCrl mice (Figure S1F). AAV2-GAYPKSP consistently induced higher eGFP expression levels than AAV2-7m8, with differences already apparent at 14 dpi. Confocal scanning laser ophthalmoscope (cSLO) imaging revealed high expression levels of eGFP were reached at 28 days after injection, with no further significant increase at later time points of 70 and 107 days post-injection (Figure S1F). To further evaluate the translational potential of AAV2-GAYPKSP, we assessed its performance in a human RoC model. The RoC platform integrates human induced pluripotent stem cell (iPSC)-derived retinal organoids (ROs), comprising rods, cones, horizontal, amacrine, Müller glial, and bipolar cells, with microfluidic perfusion to model physiological AAV delivery.38,39 Transduction with AAV2-7m8 or AAV2-GAYPKSP resulted in comparable eGFP expression levels. eGFP signals in both conditions were detectable as early as day 3 post-transduction in the RPE and became more pronounced in both the RO and RPE by day 7 (Figure S2A). Assessing the cellular transduction profile of AAV2-GAYPKSP in the RoC revealed eGFP expression in cones, rods as well as Müller glia cells (Figures S2B–S2D). Given its robust bioactivity in C57Bl/6CNrl mice and the RoC model, we advanced AAV2-GAYPKSP to evaluation in NHP. Sera from 40 animals were prescreened for neutralizing antibodies against AAV2, AAV2-7m8, and AAV2-GAYPKSP.40,41 Only seronegative animals against AAV2-7m8 or AAV2-GAYPKSP were selected. Both capsids were IVT administered at doses of 1 E + 11 and 1 E + 12 vg/eye. Fundus imaging showed dose-dependent eGFP expression with AAV2-7m8, detectable from 14 dpi and plateauing by 42 dpi (Figures 2A and 2B). Conversely, AAV2-GAYPKSP yielded no detectable eGFP as analyzed by Fundus (Figure 2C). Consistently, immunohistochemistry analysis showed that AAV2-7m8 mediated eGFP expression across the fovea, mid-periphery, and far-periphery. In contrast, AAV2-GAYPKSP drove only sparse expression, limited to a few cells mostly limited to the far-peripheral retina (Figure S3A) which cannot be visualized by Fundus in vivo.Figure 2Differential eGFP expression mediated by AAV2-GAYPKSP in human retina-on-chip model and NHP retina(A) Semi-quantitative assessment of eGFP expression using cSLO-based detection in NHP retina. AF-SLO images were evaluated using an ophthalmologist-assigned categorical grading system: trace, weak, moderate, strong. (B) Schematic representation of cSLO scanning workflow of the NHP retina. Central, mid- and far-periphery regions are indicated. The retinal center includes fovea and macular. (C) cSLO-based detection of eGFP in NHP retina at the baseline and 10 weeks following the IVT administration of the vehicle, AAV2-7m8 and AAV2-GAYPKSP. Two doses were injected; 1 E + 11 and 1 E + 12 vg/eye. Values indicate mean ± SD. Differential eGFP expression mediated by AAV2-GAYPKSP in human retina-on-chip model and NHP retina (A) Semi-quantitative assessment of eGFP expression using cSLO-based detection in NHP retina. AF-SLO images were evaluated using an ophthalmologist-assigned categorical grading system: trace, weak, moderate, strong. (B) Schematic representation of cSLO scanning workflow of the NHP retina. Central, mid- and far-periphery regions are indicated. The retinal center includes fovea and macular. (C) cSLO-based detection of eGFP in NHP retina at the baseline and 10 weeks following the IVT administration of the vehicle, AAV2-7m8 and AAV2-GAYPKSP. Two doses were injected; 1 E + 11 and 1 E + 12 vg/eye. Values indicate mean ± SD. AAV2-7mer library screening in cynomolgus monkey The unexpected finding that the strong performance of AAV2-GAYPKSP in the mouse retina did not translate to the NHP retina prompted us to perform a direct in vivo screening of the AAV2-7mer peptide-display library in NHP. The same screening workflow employed in the mouse was followed in the NHP campaign (Figures 1A and 3A).AAV2-seronegative animals were selected for both peptide-display and the subsequent barcoded library studies. The peptide-display library was IVT administered at a dose of 2.0 E + 11 vg/eye (n = 3, Figure 3A). Central and peripheral retina were harvested 8 dpi, and viral genomes were PCR-amplified and analyzed by NGS. The central retina comprises both the fovea and macular (Figure 2B).Figure 3NGS-driven screening of AAV2 peptide display library in NHP retina(A) Workflow of AAV2 library screening through successive rounds of selection in NHP retina following IVT administration. (B) Peptide library diversity across rounds of selection. Top 12 retina-enriched AAV2 variants are indicated. R0 = input viral pool. R1-3 = Library pool after selection round in vivo. (C) Retina-enriched AAV variants after R3. Table shows the grouping criteria for hit validation of nominated candidate groups for barcode-based evaluation. (D–E) NGS-based quantification of relative transduction (DNA; %AAV2) in NHP retina (D) periphery and (E) center. (F–G) NGS-based quantification of relative eGFP expression (RNA; %AAV2) in NHP retina (F) periphery and (G) center. n = 3–14. Data are represented as mean ± SEM. Statistical significance compared to AAV2 was determined using one-way ANOVA with Dunn’s test for multiple comparisons. ∗p < 0.05, ∗∗p < 0.005, ∗∗∗p < 0.0005, ∗∗∗∗p < 0.0001. Dashed line indicates relative AAV2-mediated expression. NGS-driven screening of AAV2 peptide display library in NHP retina (A) Workflow of AAV2 library screening through successive rounds of selection in NHP retina following IVT administration. (B) Peptide library diversity across rounds of selection. Top 12 retina-enriched AAV2 variants are indicated. R0 = input viral pool. R1-3 = Library pool after selection round in vivo. (C) Retina-enriched AAV variants after R3. Table shows the grouping criteria for hit validation of nominated candidate groups for barcode-based evaluation. (D–E) NGS-based quantification of relative transduction (DNA; %AAV2) in NHP retina (D) periphery and (E) center. (F–G) NGS-based quantification of relative eGFP expression (RNA; %AAV2) in NHP retina (F) periphery and (G) center. n = 3–14. Data are represented as mean ± SEM. Statistical significance compared to AAV2 was determined using one-way ANOVA with Dunn’s test for multiple comparisons. ∗p < 0.05, ∗∗p < 0.005, ∗∗∗p < 0.0005, ∗∗∗∗p < 0.0001. Dashed line indicates relative AAV2-mediated expression. Again, NGS monitoring revealed a progressive reduction in library diversity across screening rounds, reflecting the selective enrichment of retina-tropic variants. By R3, the top 12 peptides accounted for approximately 60% of all viral DNA reads (Figure 3B). Remarkably, AAV2-PDATKHG, previously identified as the second-most enriched variant in the mouse retina (Figures 1B and 1C) emerged as the most abundant variant in the NHP retina, representing 31.7% of all captured sequences in round 3 (Figure 3B).Cluster analysis of NHP retina-enriched variants after round 3 revealed enrichment of the D.T[K/R] motif cluster (Figure S3B), similar to variants enriched in the mouse retina following round 3 (Figures S1A–S1C). Barcoded library-based evaluation of the novel AAV2 variants in NHP retina Based on our initial screening in mice, we observed a pronounced enrichment of AAV2 variants harboring the D.T[K/R] motif among the most abundant retina-captured variants (Figures 1B and 1C; Figure S1B). These variants exhibited limited transgene expression in both the barcode-based and single-variant campaigns. Since a similar pattern emerged in the third round of NHP screening among the top 12 most abundant variants (Figure 3B), we broadened our selection criteria beyond the variants. Specifically, our criteria considered progressive variant enrichment in retina and/or RPE from R1 to R3, and the presence or absence of the D.T[K/R] motif (Figure 3C). On this basis, we performed a barcode-screening campaign comprising 70 candidate variants, representing six functional groups: (1) the MLDKTRS variant, enriched in both retina and RPE and containing the D.T[K/R] motif (group I.5); (2) variants enriched in both retina and RPE lacking the D.T[K/R] motif (group I, 17 variants); (3) retina-enriched variants containing the D.T[K/R] motif (group II, 15 variants); (4) retina-enriched variants lacking the D.T[K/R] motif (group III, 10 variants); (5) RPE-enriched variants containing the D.T[K/R] motif (group IV, 13 variants); and (6) RPE-enriched variants lacking the D.T[K/R] motif group V, 14 variants) (Figure 3C; Table 1).Table 1Candidates’ selection and grouping for barcoded-AAV library screening campaignsAAV IDAa sequenceGroupAAV2-I.1EEDGKRRgroup IAAV2-I.2DKSQKPPgroup IAAV2-I.3IEPHGRNgroup IAAV2-I.4GDKVRMPgroup IAAV2-I.5EQGKHREgroup IAAV2-I.6DTKFTISgroup IAAV2-I.7VKDLRQPgroup IAAV2-I.8DRQEDRKgroup IAAV2-I.9KDLSRPSgroup IAAV2-I.10IPKERPAgroup IAAV2-I.11KDMQKPPgroup IAAV2-I.12DMSHDRRgroup IAAV2-I.13HNPDRVGgroup IAAV2-I.14DHTIIFSgroup IAAV2-I.15NVTGDSAgroup IAAV2-I.16GMVEQGMgroup IAAV2-I.17AKYQTDHgroup IAAV2-II.37VQDMTKGgroup IIAAV2-II.38SNDATKMgroup IIAAV2-II.39MSDHTKTgroup IIAAV2-II.40EMDITKDgroup IIAAV2-II.41DDSTKMQgroup IIAAV2-II.42PDATKQKgroup IIAAV2-II.50NHDATKAgroup IIAAV2-II.51LEDQTKSgroup IIAAV2-II.52NDATKWFgroup IIAAV2-II.53LVDDTKMgroup IIAAV2-II.54LHDTTKNgroup IIAAV2-II.55NHDTTKHgroup IIAAV2-II.56SYDTTKNgroup IIAAV2-II.57IQDTTKNgroup IIAAV2-II.58PDATKHGgroup IIAAV2-III.43EESRRREgroup IIIAAV2-III.44DDHKRREgroup IIIAAV2-III.45PDDERRKgroup IIIAAV2-III.46KDKLDIFgroup IIIAAV2-III.47AKDMRQPgroup IIIAAV2-III.48AMDMKSVgroup IIIAAV2-III.49REIIRPPgroup IIIAAV2-III.58PDAAKHGgroup IIIAAV2-III.59PHATKHGgroup IIIAAV2-III.60PDATNHGgroup IIIAAV2-IV.27DDATKQKgroup IVAAV2-IV.28GDKTKYPgroup IVAAV2-IV.29DRTKMHDgroup IVAAV2-IV.30DMTKWVQgroup IVAAV2-IV.31TSHDTTKgroup IVAAV2-IV.32DMTKVTFgroup IVAAV2-IV.33VDPTRKPgroup IVAAV2-IV.34LGDMTRVgroup IVAAV2-IV.35DDTTRIMgroup IVAAV2-IV.36QVDGTRVgroup IVAAV2-IV.37GDKTKYPgroup IVAAV2-IV.38GQDTTRFgroup IVAAV2-IV.61HADQTRVgroup IVAAV2-V.15AEKGKVPgroup VAAV2-V.16DQRMKPPgroup VAAV2-V.17QTDVTMVgroup VAAV2-V.18SQMMIKAgroup VAAV2-V.19NKMVPMPgroup VAAV2-V.20DKTHHRDgroup VAAV2-V.21GAYPKSPgroup VAAV2-V.22MDMKKPSgroup VAAV2-V.23KPIERQPgroup VAAV2-V.24DKSPGRVgroup VAAV2-V.25GVSQRAPgroup VAAV2-V.26FEKSRPAgroup VAAV2-V.62PTDKKMPgroup VAAV2-V.63DNLHRKEgroup VAAV2-I5.14MLDKTRSgroup I.5 Table 1Candidates’ selection and grouping for barcoded-AAV library screening campaignsAAV IDAa sequenceGroupAAV2-I.1EEDGKRRgroup IAAV2-I.2DKSQKPPgroup IAAV2-I.3IEPHGRNgroup IAAV2-I.4GDKVRMPgroup IAAV2-I.5EQGKHREgroup IAAV2-I.6DTKFTISgroup IAAV2-I.7VKDLRQPgroup IAAV2-I.8DRQEDRKgroup IAAV2-I.9KDLSRPSgroup IAAV2-I.10IPKERPAgroup IAAV2-I.11KDMQKPPgroup IAAV2-I.12DMSHDRRgroup IAAV2-I.13HNPDRVGgroup IAAV2-I.14DHTIIFSgroup IAAV2-I.15NVTGDSAgroup IAAV2-I.16GMVEQGMgroup IAAV2-I.17AKYQTDHgroup IAAV2-II.37VQDMTKGgroup IIAAV2-II.38SNDATKMgroup IIAAV2-II.39MSDHTKTgroup IIAAV2-II.40EMDITKDgroup IIAAV2-II.41DDSTKMQgroup IIAAV2-II.42PDATKQKgroup IIAAV2-II.50NHDATKAgroup IIAAV2-II.51LEDQTKSgroup IIAAV2-II.52NDATKWFgroup IIAAV2-II.53LVDDTKMgroup IIAAV2-II.54LHDTTKNgroup IIAAV2-II.55NHDTTKHgroup IIAAV2-II.56SYDTTKNgroup IIAAV2-II.57IQDTTKNgroup IIAAV2-II.58PDATKHGgroup IIAAV2-III.43EESRRREgroup IIIAAV2-III.44DDHKRREgroup IIIAAV2-III.45PDDERRKgroup IIIAAV2-III.46KDKLDIFgroup IIIAAV2-III.47AKDMRQPgroup IIIAAV2-III.48AMDMKSVgroup IIIAAV2-III.49REIIRPPgroup IIIAAV2-III.58PDAAKHGgroup IIIAAV2-III.59PHATKHGgroup IIIAAV2-III.60PDATNHGgroup IIIAAV2-IV.27DDATKQKgroup IVAAV2-IV.28GDKTKYPgroup IVAAV2-IV.29DRTKMHDgroup IVAAV2-IV.30DMTKWVQgroup IVAAV2-IV.31TSHDTTKgroup IVAAV2-IV.32DMTKVTFgroup IVAAV2-IV.33VDPTRKPgroup IVAAV2-IV.34LGDMTRVgroup IVAAV2-IV.35DDTTRIMgroup IVAAV2-IV.36QVDGTRVgroup IVAAV2-IV.37GDKTKYPgroup IVAAV2-IV.38GQDTTRFgroup IVAAV2-IV.61HADQTRVgroup IVAAV2-V.15AEKGKVPgroup VAAV2-V.16DQRMKPPgroup VAAV2-V.17QTDVTMVgroup VAAV2-V.18SQMMIKAgroup VAAV2-V.19NKMVPMPgroup VAAV2-V.20DKTHHRDgroup VAAV2-V.21GAYPKSPgroup VAAV2-V.22MDMKKPSgroup VAAV2-V.23KPIERQPgroup VAAV2-V.24DKSPGRVgroup VAAV2-V.25GVSQRAPgroup VAAV2-V.26FEKSRPAgroup VAAV2-V.62PTDKKMPgroup VAAV2-V.63DNLHRKEgroup VAAV2-I5.14MLDKTRSgroup I.5 Candidates’ selection and grouping for barcoded-AAV library screening campaigns Each capsid candidate was vectorized individually, packaging a uniquely barcoded eGFP reporter driven by a CMV promoter. Viral titers were quantified, and the candidates were distributed into six library pools (library 1–6; Figure S3C). Barcoded AAV2 and AAV2-7m8 were added as benchmarks to each library. Sequencing of the input libraries confirmed near-equal representation of all candidates, with only minor deviations (Figure S3C). These libraries were IVT-administered to NHP at a dose of 1.4 E + 12 vg/eye (equivalent to 1 E + 11 vg/variant). Eight weeks post-injection, central and peripheral retinal tissues were harvested for NGS analysis. To assess the performance of each AAV variant, the barcoded DNA and RNA reads were normalized to the respective barcode abundance in the input library. Four AAV2 variants exhibited ∼2-fold increase in transduction (i.e., AAV DNA delivery) relative to AAV2 in the peripheral (AAV2-I.1, AAV2-III.43, AAV2-III.44, and AAV2-IV.33) and five in the central retina (AAV2-I.1, AAV2-I.12, AAV2-III.43, AAV2-III.44, and AAV2-IV-33) (Figures 3D and 3E). AAV2-7m8 was the strongest overall transducer, achieving ∼5-fold higher delivery in the periphery and ∼19-fold higher delivery in the central retina compared with AAV2 (Figures 3D and 3E; Figures S4A–S4F). At the transcriptional level, several engineered capsids showed improved performance over AAV2. Seventeen variants demonstrated increased eGFP mRNA in the periphery, and 18 showed ≥2-fold increases in the central retina (Figures 3F and 3G). Importantly, a set of variants also outperformed AAV2-7m8 in the peripheral retina (Figure 3F), most notably AAV2-I.1 and AAV2-I.12, which displayed a broad transcriptional activity and mediated ∼2-fold higher transcriptional activity compared to AAV2-7m8. In the central retina, AAV2-III.43 and AAV2-III.44 were the most transcriptionally active novel variants, ranking immediately behind AAV2-7m8 (Figure 3G; Figure S4C). To further clarify the performance of our identified variants, we analyzed their transcriptional efficiency (log2(RNA/DNA)) as a measure of expression level per delivered genome. This analysis revealed that AAV2-7m8-driven eGFP expression was largely driven by its higher DNA delivery (Figures S4A–S4F) as several novel variants exhibited higher transcriptional efficiency. The top performers, AAV2-I.1 and AAV2-I.12, maintained strong transcriptional efficiency in both retinal regions (Figure S4A). Similarly, AAV2-I.13, AAV2-V.20, and AAV2-III.45 also showed consistently high efficiency across compartments (Figures S4C–S4E). Other variants displayed subregional specialization, such as AAV2-I.3, which showed enhanced transcriptional efficiency in the central retina (Figure S4A). Importantly, the analysis also revealed library/group-specific differences, as for example, all variants from library 2 performed poorly, with none exhibiting significant transcriptional activity comparable to AAV2 (Figures S4B, S4E, and S4F). In summary, all variants with enhanced transcriptional performance relative to AAV2 originated from groups I, III, and V (Figure 3C; Figures S4A, S4C, and S4E). In contrast, AAV2-I5.14 and variants from groups II and IV, i.e., variants that harbor a D.T[K/R] motif, generally underperformed relative to AAV2, with only two group IV members (AAV2-IV.29 and AAV2-IV.33) achieving transcriptional activity comparable to AAV2 (Figures S4B–S4D). snRNA-seq reveals broad and potent retinal cell transduction by novel AAV capsid variants To resolve the cell-type-specific transgene expression within the retina, we performed single-nucleus RNA sequencing (snRNA-seq) on peripheral retina samples from NHPs injected with AAV libraries 1 and 3, selected based on their strong performance in bulk analyses (Figure 3D; Figures S4A and S4B). Retinal tissues were dissociated into single-nucleus suspensions, followed by cDNA library preparation using the 10× genomics platform. Sequencing was achieved at an average depth of 25,000 reads per cell, with a median detectable gene count between 1,557 and 2,047 genes per cell and a mean of 19,579 genes detected per sample. Quality metrics confirmed robust data: 95.7% of reads were confidently mapped to the genome, and 41.1% to the transcriptome, consistent with expectations for snRNA-seq. Notably, recovery of AAV barcodes from the libraries was highly efficient, with 42–136 million barcode-specific reads per sample. Cell type annotation was accomplished by integrating known retinal cell markers and published NHP single-cell atlas references,42 resulting in the identification of all major retinal populations, including cones, rods, various subtypes of bipolar cells, amacrine cells, horizontal cells, Müller glia, immune cells, and rare types like retinal ganglion and horizontal cells (Figures 4A and 4B). Annotation quality was validated using established marker genes, with cell type abundances largely matching expectations except for a minor under-representation of ganglion cells in one sample (Figure S5A).Figure 4Single-nuclei RNA sequencing-based annotation of cell types and quantification of AAV-infected peripheral retinal cells in NHP(A) UMAP of clustered peripheral retinal cells in NHP, with similar cells grouped together and each cluster indicated by a different color. (B) Heatmap showing marker genes for each cluster, with dot size representing the percentage of cells expressing the marker and color indicating average expression levels. (C) Relative barcoded AAV mRNA expression compared to AAV2. Fold change is shown in color, and dot size represents the percentage of cells with at least one count of the respective barcode. Relative barcoded AAV cargo expression normalized to AAV2 in (D) all cells of sample, (E) RGCs, (F) Mueller Glial, (G) cones, and (H) rods. n = 2–8. Data are represented as mean ± SEM. Single-nuclei RNA sequencing-based annotation of cell types and quantification of AAV-infected peripheral retinal cells in NHP (A) UMAP of clustered peripheral retinal cells in NHP, with similar cells grouped together and each cluster indicated by a different color. (B) Heatmap showing marker genes for each cluster, with dot size representing the percentage of cells expressing the marker and color indicating average expression levels. (C) Relative barcoded AAV mRNA expression compared to AAV2. Fold change is shown in color, and dot size represents the percentage of cells with at least one count of the respective barcode. Relative barcoded AAV cargo expression normalized to AAV2 in (D) all cells of sample, (E) RGCs, (F) Mueller Glial, (G) cones, and (H) rods. n = 2–8. Data are represented as mean ± SEM. Analysis of barcoded mRNA expression confirmed bulk RNA-seq findings (Figures 3D and 3E), identifying AAV2-I.1, AAV2-I.12, AAV2-I.8, AAV2-III.43, AAV2-III.44, and AAV2-7m8 as the most potent vectors in the retina periphery. These variants achieved at least a 10-fold increase in eGFP expression within transduced cells compared to parental AAV2 (Figures 4C and 4D). Importantly, these top candidates showed broad tropism, driving expression in both inner and outer retinal layers, without marked preference for specific cell types (Figures 4C–4G; Figures S5B–S5H). Variants AAV2-I.12, AAV2-I.1, AAV2-7m8, AAV2-III.43, and AAV2-III.44 exhibited particularly uniform barcode expression, reaching at least 50% of cells across multiple retinal populations (Figure 4C). In summary, snRNA-seq analysis demonstrated that most novel AAV capsid variants mediate robust and widespread gene delivery across all principal retinal cell types in NHPs, following IVT injection. While minor differences in cell type-specific efficiency were observed, no variant displayed exclusivity for a single cell type. Discussion Directed evolution represents a powerful strategy for engineering AAV capsids with tailored properties, particularly through peptide insertions at loop VIII residue 588 of AAV2.19,22,27,28,34,43,44,45,46,47 In the present study, we performed two individual selection campaigns, where AAV2 peptide-display libraries were iteratively screened in either mouse or NHP retinae following IVT delivery. Our workflow integrated three rounds of in vivo selection, followed by a characterization phase in which the top-performing variants were benchmarked using a barcoded library approach combined with snRNA-seq to resolve capsid tropism and activity across retinal cell types. At each selection round, NGS was employed to quantify variant enrichment and prioritize candidates for subsequent characterization. This analytical framework, here established to guide AAV capsid development in complex ocular tissues, ensured rigorous, data-driven progression through the selection process and can therefore also be transferred to screening campaigns in other target organs.33,48 To validate capsid performance, we employed barcoded vector libraries, which enabled internally controlled comparisons of transcriptional efficiency against both AAV2 and the benchmark capsid AAV2-7m8. This strategy revealed a correlation between barcoded library performance and individual capsid validation in mice, providing confidence to extend the platform to a large-scale in vivo campaign in NHPs. One insight from NGS data was the enrichment of variants containing the D.T[K/R] motif, across both mouse and NHP retinal screens. However, while variants with this motif were consistently enriched at the viral genome level, subsequent functional assays using barcoded AAV libraries revealed limited to no transgene expression. These finding highlights some of the key limitations of screening complex AAV peptide libraries, which also leads to the enrichment of candidates, which have been enriched in the respective target tissue but do not mediate transgene expression.49,50 At this stage of investigation, we can only speculate that the consistent enrichment of D.T[K/R] motif variants with reduced transcriptional performance may be a result of increased retention in the extracellular space and/or of non-productive cell entry and intracellular processing.51 Alternative selection strategies, such as Cre-recombination-based AAV targeted evolution (CREATE) and multiplexed (M)-CREATE,52,53,54 overcome AAV engineering limitations by using Cre-loxP recombination in transgenic mice, where Cre in target cells recombines loxP-flanked cassettes in infecting AAV capsids to enable PCR amplification and sequencing of successful variants for positive/negative selection, but require such Cre-expressing mice and implemented in NHP screenings. A complementary strategy is the expression-based screening of target-individualized peptide-display libraries. In this approach, each variant expresses its specific CAP gene under the control of ubiquitous or cell type-specific promoters, enabling the selection of functionally active candidates. This method enriches for variants with activity in the intended cellular population without requiring transgenic animal models.33,55,56,57,58 Another important finding of this work was the species dependence of AAV capsid performance following IVT administration. In mice, the variant AAV2-GAYPKSP mediated robust and uniform retinal transduction, outperforming the benchmark AAV2-7m8 in both barcoded library and individual validation assays. AAV2-GAYPKSP has since become a routinely used internal-research tool for efficient IVT gene delivery, and its strong performance has been reproduced across mouse strains with varying genetic background, such as DBA/2 J mice (data not shown). This indicates that its potency is not restricted to specific mouse strains; such strain-specificity was reported for other AAV capsids,59,60,61 and further supports the broad pre-clinical applicability of AAV2-GAYPKSP. As a research tool, the vector also demonstrated strong activity in human iPSC-derived ROs. However, the failure to transduce the NHP retina prevents future clinical applications.39,62,63 Notably, while species-specific differences have been reported for systemically delivered AAVs,59,64,65 AAV2-GAYPKSP represents to our knowledge the first published example failing to translate from rodents to NHPs after IVT administration. It also contrasts with other clinically established IVT capsids that were discovered in rodents and translated with good performance into higher species.22,28 In the retina, several factors may contribute to species-specific differences, including ILM thickness, extracellular matrix composition, and receptor expression profiles.7 The example of AAV2-GAYPKSP further underscores the importance of implementing AAV capsid screening pipelines that prioritize higher species models, rather than relying solely on rodents or organoid systems that lack key anatomical barriers. Accordingly, we implemented our workflow directly in NHPs with a focus on identifying variants that mediate efficient retinal transduction after IVT delivery. In NHPs, AAV2 typically restricts transduction to foveal ganglion cells, whereas second-generation capsids display broader distribution patterns including the fovea and retinal periphery.17,22,28 After three rounds of screening of the AAV2 peptide display library in NHP, we nominated 70 variants for validation using the barcoded-library approach. Variant selection was guided by screening campaign outcomes, prioritizing progressive enrichment in retina and/or RPE across rounds R1-R3, and the presence or absence of enriched motifs, rather than solely the most abundant variants. The variants were grouped as: (1) the MLDKTRS variant, which was enriched in both retina and RPE and contains the D.T[K/R] motif (group I.5) (2) variants enriched in both retina and RPE (group I); (3) retina-enriched variants containing a D.T[K/R] motif (group II); (4) retina-enriched variants lacking this motif (group III); (5) RPE-enriched variants containing the D.T[K/R] motif (group IV); (6) RPE-enriched variants lacking the motif (group V).This identified potent capsids such as several candidates, particularly those from groups I and III, exhibited robust expression across both the central and peripheral retina. These variants were not the most abundant retina-captured AAVs in the peptide screening campaign, highlighting the importance of multi-round enrichment dynamics over raw abundance alone in identifying viable retina-tropic. Within group I, AAV2-I.1 and AAV2-I.12 demonstrated broad pan-retinal transduction and transcriptional activity, driving higher eGFP expression in the periphery compared to AAV2-7m8. Further analysis of AAV2-I.1 and AAV2-I.12 revealed their enhanced transcriptional efficiency (log2(RNA/DNA)) in both retina center and periphery in comparison to AAV2.7m8. In the future, variants with higher transcriptional efficiency (i.e., higher expression per viral genome) could allow for lower dosing to achieve therapeutic benefit, thereby reducing the risk of dose-limiting toxicity adverse events or inflammation. Notably, at the current stage of capsid characterization, we cannot rule out competition for shared cell-entry receptors between variants and benchmarks. Given this caveat, we cannot exclude the possibility that the performance of individual variants may be over- or underestimated. Future head-to-head studies evaluating AAV2-I.1 and AAV2-I.12, alongside other top-performing variants, as single AAV productions against AAV2-7m8 and additional retina-tropic vectors such as AAV2.NN, AAV2.GL, and NHP#9 are needed to determine the value of the capsids identified here for future clinical applications.28,66 Quantitative comparison of capsid-mediated cell type specificity formerly required large animal cohorts, which is particularly challenging in NHPs due to cost, ethical considerations, and biological variability. The integration of barcoded library screening with single-cell transcriptomics now enables efficient, accurate, and direct head-to-head evaluation of multiple vectors across all retinal cell types within the same animals.30,67 In the present study, snRNA-seq was applied to group I and III variants, confirming the strong transcriptional activity comparable to AAV2-7m8 mediated by those new variants (Figure 3F; Figures S4A and S4C). Notably, AAV2-I.12, AAV2-I.1, AAV2-III.43, and AAV2-III.44 also achieved expression profiles across cell types similar to AAV2-7m8 (Figure 4C), for which a broad cell type tropism in human retinal explants was reported previously.68 This suggests that additional engineering strategies, such as mutations across multiple variable regions (e.g., VR-I, VR-IV, VR-V, and VR-VI) and incorporation of stuffer sequences, may be necessary to generate more cell type-restricted capsids. Recent studies following these strategies have yielded promising new vectors for the NHP retina, including AAV2-P2-V2 and AAV2-P2-V3.31 Alternatively, enhanced specificity might be achieved by screening peptide-display libraries cloned onto different AAV serotype backbones in combination with cell-sorting strategies tailored to the desired target cell type. In summary, the present study combines iterative in vivo selection, large-scale barcoded library screening, and single-cell transcriptomics to identify novel AAV2 variants with enhanced retinal tropism, particularly in NHPs. We identify two AAV2 variants, AAV2-I.1 and AAV2-I.12, driving broad pan-retinal cargo expression with enhanced transcriptional efficiency. We also highlight specific pitfalls in the discovery process, such as misleading motif enrichment signatures and the species-specific failure of AAV2-GAYPKSP, both of which provide valuable lessons for the capsid engineering community. Our findings underscore the importance of higher species selection pipelines, coupled with high-resolution single-cell readouts, in directing the design of next-generation capsids for translational retinal gene therapy. Material and methods Animals For mouse experiments, 10–12 weeks-old female mice (18–22 g; C57BL/6NCrl) were purchased from Charles River Laboratory. Mice were kept in a 12 h light/dark cycle with free access to food and water. Before dosing, the mice were housed for 1 week. All procedures were approved by Boehringer Ingelheim and the animal welfare institution (Regierungspräsidium; Tübingen, Germany). All procedures were performed in compliance with Directive 2010/63/EU governing the use of animals in scientific research. For NHP studies, purpose-bred male cynomolgus monkeys (2.5–4.2 kg; Macaca fascicularis) of Mauritian origin were selected, aged 2–4 years at the beginning of pre-dosing phase. As part of the breeder’s health procedure, all animals were tested for tuberculosis and given prophylactic treatments. Upon arrival, the animals were acclimated to study procedures for at least 2 weeks, during which they were inspected clinically and tested for tuberculosis. Water was provided at ad libitum and routinely tested for micro-organisms and environmental contaminants. Animals were fed a certified laboratory diet (LabDiet) twice daily, supplemented with fresh fruit and vegetables. The monkeys were housed in a climate-controlled room with a minimum of eight air changes per hour, maintaining temperatures between 19°C and 25°C and relative humidity between 40% and 70%. Humidity levels were returned to normal within 24 h. Artificial lighting was controlled automatically to provide a 12 h light/dark cycle. All procedures were approved by the LAVE (Landesamt für Verbraucherschutz und Ernährung) in Nordrhein-Westfalen (NRW), Germany. AAV library generation and virus production To generate the AAV2-7mer peptide display library, random (NNK)7 sequences were introduced in the SfiI restriction site of the ITR-flanked cap2 gene at loop VIII residue 588. Double stranded (NNK)7 were ligated with SfiI digested backbone at 1:10 molar ratio and were transformed by electroporation in ElectroMAX DH5α-E electrocompetent E. coli (Thermo Scientific, Germany). Plasmid library was subsequently purified using NucleoBond Xtra Maxi EF (Macherey-Nagel, Germany) following an overnight incubation at 30°C and 225 rpm. Virus production was performed using HEK293T cells maintained in Dulbecco’s modified Eagle’s medium (DMEM) with GlutaMAX, 10% fetal bovine serum (Thermo Scientific, Germany) according to a standardized protocol.69 Briefly, 6 × 107 cells were seeded in 16-Layer CELLdiscs (Thermo Scientific, Germany) for 72 h at 37°C. Calcium chloride transfection was performed using pHelper and 0.5 μg library DNA per disc. We have tested different pHelper: plasmid library ratios and found that this condition (multiplicity of transfection [MOT] 250), allowing the optimal virus yields, library diversity and genotype-phenotype correlation. Cells and culture medium were harvested 72 h after transfection using 0.5 M EDTA. Pelleted cells were lysed by resuspension in lysis buffer (1 M NaCl, 50 mM Tris, 10 mM MgCl2, 0.001% Pluronic F-68; (Thermo Scientific, Germany)) in the presence of salt active nuclease (Serva electrophoresis GmbH, Germany). AAV-containing supernatant was precipitated on ice with 40% PEG-8000 (25%; %v:v, Thermo Scientific, Germany) followed by centrifugation. The PEG-pellet was resolved in resuspension buffer (1 M NaCl, 50 mM Tris, 0.001% Pluronic F-68, pH 8.0) and fractionated on iodixanol gradient. Buffer exchange was performed on an Amicon Ultra-15 MWCO 100,000 ultrafiltration tube (Merck Millipore) with AAV formulation buffer (1 × PBS, 1 mM MgCl2, 2.5 mM KCl, 10% glycerol, 0.001% Pluronic F-68, pH 7.4, sterile-filtered; Sigma-Aldrich, Germany), and the final virus samples were analyzed by droplet digital PCR (QIAGEN, Germany) using an ITR-specific probe set. In vitro neutralizing antibody assay Neutralizing antibodies (NAbs) in NHP serum against AAV2 were quantified in vitro using a transduction inhibition assay as described previously.40,41 Prior to infection, 2.5 E + 04 HEK293H cells/well were seeded in 96-well plates. Samples were serially diluted ×2 fold (1:2–1:1,024) in serum-free DMEM (Thermo Scientific, Germany) and pre-incubated with AAV vectors encapsulating anti-fluorescein isothiocyanate (FITC) reporter for 30 min at 37°C. Then vectors (2.5 E + 03 vg/cell; n = 3) were added to the cells 24 h post-seeding and incubated for 72 h at 37°C. Supernatants were collected, cleared by centrifugation (200× g, 5 min). To quantify reporter expression, cleared supernatants were analyzed using a Meso Scale Discovery (MSD) immunoassay (Meso Scale Diagnostics, USA) according to manufacturer’s procedure. Transduction efficiency was expressed as relative signal normalized to virus-only controls. NAb titers were defined as the reciprocal serum dilution causing a 50% reduction in transduction (ID50). Animals with ≥50% inhibition at 1:5 serum dilution were classified as seropositive. In vivo selection and quantification in mice and NHP In C57BL/6NCrl mice, AAV2 peptide display libraries were IVT administered at a dose of 4.2 E + 9 vg/eye. Individual AAV variants were injected at 1.0 E + 9 vg/eye. Animals were euthanized at 21, 28, or 107 dpi, depending on experimental requirements. Retina and RPE were immediately collected, snap-frozen in liquid nitrogen, and stored at −80°C. For NHP studies, the AAV2 peptide display library was administered intravitreally at 2.0 E + 11 vg/eye. AAV2-7m8 and AAV2-GAYPKSP were IVT administered at doses of 1 E + 11 and 1 E + 12 vg/eye. According to experimental requirements, animals were euthanized at 8 dpi or at 8- or 12-week endpoints following IVT administration. DNA was extracted from all animal samples using RLT reagent and AllPrep DNA/RNA kit (QIAGEN, Germany). Capsid variants genomes were amplified by PCR using primers flanking the 7-mer peptide inserts; forward (5′-GGT TCT CAT CTT TGG GAA GCA AG-3′) and reverse (5′-TGA TGA GAA TCT GTG GAG GAG-3′) using the Phusion High Fidelity DNA polymerase (Thermo Scientific, Germany). Following a second nested PCR amplification using Cap2-specific forward primer (5′-GCA GTA TGG TTC TGT ATC TAC CAA CC-3′) and reverse primer (5′-GCC TGG AAG AAC GCC TTG TGT G-3′) amplicons containing the Illumina adapters and unique indices (NEB GmbH, Germany) were quantified using qubit, pooled at equimolar ratio, and sequenced on a Nextseq (Illumina Inc., USA). For the barcoded AAV variants testing experiments, top enriched capsid variants and benchmarks were vectorized to encapsulate a unique barcoded eGFP reporter driven by CMV promoter. Each variant was produced individually. In mice, barcoded libraries were delivered at 2.0 E + 9 vg/eye. In NHP barcoded libraries were injected IVT at 1.4 E + 12 vg/eye. Barcodes were used to quantify transgene expression after transduction by its corresponding capsid variant. An equal amount of the barcoded AAV variants was mixed with each other to prepare the viral library mix injected into the animals. Retinae were harvested from the animals and RLT reagent/AllPrep DNA/RNA kit (QIAGEN, Germany) were used to isolate DNA and RNA from the retinal samples. RNA samples were treated with dsDNase (Thermo Scientific, Germany), and cDNA synthesis was performed using Maxima H Minus cDNA synthesis master mix (Thermo Scientific, Germany). Barcoded eGFP transcripts were amplified from DNA and cDNA samples with primers binding to upstream (5′-ATC ACT CTC GGC ATG GAC GAG C-3′) and downstream (5′-GGC TGG CAA CTA GAA GGC ACA-3′) of the barcodes using the Phusion High Fidelity. Amplicons containing the Illumina adapters and unique indices (NEB, Germany) were quantified using qubit, pooled at equimolar ratio, and sequenced on a Nextseq (Illumina Inc., USA). For RT-qPCR-based quantification of relative eGFP mRNA expression in mouse retinae transduced with AAV2-7m8 or AAV2-GAYPKSP, viral transcripts were measured using TaqMan assays (Thermo Scientific, Germany). Total RNA was reverse-transcribed as described above, and cDNA was mixed with 2× TaqMan Gene Expression Master Mix, 20× eGFP primer–probe set (forward: 5′-CTGCTGCCCGACAACCA-3′; reverse: 5′-TGTGATCGCGCTTCTCGTT-3′; probe: 5′-TACCTGAGCACCCAGTCCGCCCT-3′) for viral transcript quantification, and 20× RNA polymerase II primer–probe set (forward: 5′-GCCAAAGACTCCTTCACTCACTGT-3′; reverse: 5′-TTCCAAGCGGCAAAGAATGT-3′; probe: 5′-TGGCTCTTTCAGCATCTCGTGCAGATT-3′) as the housekeeping control. All reactions were performed in triplicate using a QuantStudio 6 Flex system (Applied Biosystems, USA) under the following cycling conditions: 50°C for 2 min, 95°C for 10 min, followed by 40 cycles of 95°C for 15 s and 60°C for 1 min. Absolute transcript numbers for eGFP and RNA Pol II were derived from standard curves using plasmids encoding the respective genes. Relative expression levels were quantified compared to BSS-buffer-treated control retinae. IVT injections In mice, IVT injections were performed similar to procedures previously described.28 In brief, animals were anesthetized and Tropicamid eye drops (Mydriadicum Stulln) were applied to dilute pupils. Injections of 1 μL were performed with a 34-gauge beveled needle free-hand under visual control (OPMI Lumera 700, Zeiss). Special care was taken to avoid damage to the lens. NHPs were anesthetized with ketamine and medetomedine, mydriadicum and local anesthesia was instilled. With NHPs lying on the back, 5% povidine solution was instilled. The temporal injection site was kept as consistent as possible between eyes taking care not to touch the retina directly with the canula. 80–100 μL of AAV suspension was injected into the vitreous. After injection, dexpanthenole ointment was applied to the cornea. As we used research grade material AAVs only, special care was taken to avoid potential inflammation of the eye. Therefore, NHPs were treated with Pred forte (1% prednisolone), 3× daily in both eyes on days 1–8 and received in addition systemic immunosuppression (6.6 mg/kg Methylprednisolone intramuscular [i.m.], once a week after dosing). No severe inflammation or adverse events were detected during our studies. Ocular analysis Optical coherence tomography (OCT) and autofluorescence (AF-SLO) imaging were performed using a Heidelberg Spectralis Flex HRA+OCT (Multicolor, FLEX system, SN: SPEC-06627-S3610). AF-SLO imaging was performed to image eGFP expression in the mouse retina as previously described.28 For NHPs, wide-field autofluorescence images were taken in an overlapping 6-field pattern to capture a large area of the retina. The animals were anesthetized with ketamine and medetomidine, and atipamezole was used as an antidote at the end. Pupils were diluted with 0.5% tropicamide and 2.5% phenylephrine. Immediately before imaging, midazolam was given to keep the eyes centrally positioned. Histology and immunohistochemistry After fixation in 10% neutral buffered formalin (mice up to 48 h, NHP up to 96 h) the eye samples were processed with an automated tissue processor (Tissue-Tek VIP 6, Sakura), embedded in paraffin and cut into 3 μm sections. Staining with hematoxylin and eosin (H&E) was performed according to standard protocols. Immunohistochemical detection of eGFP-expression was performed on the Leica Bond platform (Leica Biosystems, Melbourne, Australia). Murine eyes were stained using a polyclonal rabbit anti-GFP antibody (ab290, Abcam, 1:1,500) after enzyme-induced epitope retrieval with the Bond Enzyme Pretreatment Kit (Enzyme 1, Leica Biosystems, Newcastle, United Kingdom) for 5 min at 37°C. NHP eyes were stained with a rabbit monoclonal anti-GFP antibody (clone EPR14104-89, ab183735, Abcam, 1:250) after heat-induced antigen retrieval (95°C, 20 min, Bond Epitope Retrieval Solution 2, pH9, Leica Biosystems). Visualization was performed using the Bond Polymer Refine Detection System (3,3′ Diaminobenzidine (DAB) as chromogen, Leica Biosystems, Newcastle, United Kingdom). Slides were digitized using an Axio Scan.Z1 slide scanner (20× objective, Carl Zeiss Microscopy GmbH, Jena, Germany). Human RoC model Differentiation and culture of hiPSC ROs, derivation of RPE cells as a product from RO differentiation, assembly and culture of the retina-on-chip platform was conducted as previously described by Achberger et al.38,39 At day of transduction (day 0), RO had an age of 273 days, containing precursors of rod, cones, horizontal, amacrine, Müller glial, and bipolar cells; at this time point no retinal ganglion cells (RGCs) are present). AAV2-7m8 and AAV2-GAYPKSP (MOI = 2.0 E + 05) were injected into the apical compartment above the RO, mimicking sub-retinal delivery. Transduction efficacy, kinetics, and cell tropism were assessed by eGFP imaging on day 3 and 7 post-transduction using the Zeiss Observer 7 (5× objective; Carl Zeiss, Germany) and Zen Software 3.3 (Carl Zeiss, Germany). For immunocytochemistry, RoC were fixed and RO extracted, embedded, cryosectioned and immune-stained as previously described.39 Immunostainings were performed using the following antibodies: DKK3 (Rb, 1:500, PA5-21325, Thermo Scientific, Germany), GNAT1 (Rb, 1:500, 55167-1-AP, Proteintech, USA), ARR3 (Gt, 1:50, sc54355, Santa Cruz, USA), GFP (Ch, 1:1500, GFP1010, Aves Labs, USA). The anti-GFP antibody was used to enhance the native eGFP signal. Images were taken at 20× with a spinning disc confocal microscope (Cell Observer SD, Carl Zeiss Microscopy, Germany) using the Zen Blue Software v.2.6. Maximum intensity projections were used. Nuclei isolation from NHP retina Nuclei were isolated from snap-frozen retina by discontinuous iodixanol density gradient centrifugation (OptiPrep; Serumwerk Bernburg, Germany), on ice throughout. The following buffer were prepared for nuclei isolation Solution B (gradient buffer: 150 mM Tricine, 25 mM KCl, 5 mM MgCl2, pH 7.8), Solution D (homogenization buffer: 20 mM Tricine, 25 mM KCl, 5 mM MgCl2 in 0.34 M sucrose, pH 7.8), Complete Homogenization Buffer (HB: Solution D supplemented fresh with spermidine, spermine, DTT, complete protease inhibitor (Roche, Switzerland), and 0.2 U/μL Protector RNase Inhibitor (Roche, Switzerland)). Tissue was homogenized in Complete HB using a 2 mL Dounce homogenizer (pestle A and B) with 0.15% IGEPAL CA-630 (v/v). After centrifugation at 1,000× g (10 min, 4°C), nuclear pellets were resuspended in Complete HB/IGEPAL, combined with an equal volume of Working Solution C (∼50% iodixanol in Solution B), filtered through a 40 μm membrane, and under layered with 30% iodixanol gradient solution (ρ ≈ 1.175 g/mL) for isopycnic centrifugation at 10,000× g (25 min, 4°C). Pellets were resuspended in Minute Anti-Clumping Nuclei Storage Buffer (Invent Biotechnologies, Inc., USA; 0.2 U/μL RNase inhibitor), diluted 1:2 in 2% BSA/PBS with RNase inhibitor, and passed through a second 40 μm filter to remove aggregates. Nuclei were counted by Hoechst 33342 staining and hemocytometer. snRNA-seq snRNA-seq libraries were prepared using the Chromium Next GEM Single Cell 3′ v.3.1 Dual Index kit (10× Genomics, USA; protocol CG000315 Rev C). Nuclei were diluted to 1,000/μL in 2% BSA/PBS with 0.2 U/μL RNase inhibitor, targeting 10,000 recovered nuclei per sample. GEM generation, reverse transcription, post-GEM cleanup, cDNA amplification, and library construction were performed per the manufacturer’s instructions, with PCR cycle numbers adjusted to Qubit-measured cDNA input. Sample indexing used the Dual Index TT Set A (PN-3000431); sequences are given in Table S1 (oligos P13–P21). AAV-BC enrichment libraries were generated from 10× Genomics cDNA by a two-step PCR. Because 10× cDNA carries a partial Read1 sequence at its 5′ end, primer P02 (Table S1) replaces the gene-specific reverse primer used in bulk applications. PCR-1 (enrichment): 100 ng cDNA input; NEBNext HotStart Q5 2× Master Mix (New England Biolabs, USA) with EvaGreen (Biotium, USA); primers P01 (AAV-BC Specific Fwd) and P02; 98°C 2 min then cycles of 98°C 10 s/69°C 30 s (depending on yield, per-sample). Replicates pooled; purified with NucleoSpin PCR Clean-Up (Macherey-Nagel); right-sided SPRIselect size selection. PCR-2 (sample indexing): 10 ng PCR-1 product; NEBNext Q5 Master Mix with EvaGreen; sample-specific i5 forward primers P03–P11 (unique 8 nt i5, P5 adapter, Read1 site) and shared i7 reverse primer P12 (Custom BC i7 Rev; fixed index, P7 adapter, Read2 site nested within PCR-1 amplicon for added specificity); all at 200 nM. Program: 98°C 2 min then 7 cycles of 98°C 10 s/65°C 30 s. Purified with NucleoSpin PCR Clean-Up. Libraries were quantified by Qubit dsDNA HS assay (Thermo Fisher Scientific, Germany) and sized on an Agilent 2100 Bioanalyzer with the High Sensitivity DNA kit (Agilent Technologies, USA). Pooling was proportional to target sequencing depth: 26,100 reads/cell for GEX libraries and 10,000 reads/cell for AAV-BC libraries. Sequencing was performed on an Illumina NovaSeq 6000 (Illumina, Inc., USA). Data and statistical analysis All processing was performed on a Linux high performance computing cluster. For both the peptide display and barcode libraries, the “cutadapt” tool (v.1.17)70 was used to trim the flanking primer sequences leaving only the peptide-coding or barcode DNA sequences. These were then quantified using the Linux “uniq-c” command. Subsequent processing was done in R (v.4.0.2) with the Bioconductor package infrastructure (v.3.11).71 Raw counts were normalized to counts per 10 million reads. The DNA sequences were translated to their corresponding peptide sequences using the “translate” function of the “Biostrings” R Bioconductor package. Motifs were identified by hierarchically clustering the peptides based on their BLOSUM62 alignment distances (Biostrings package), partitioning into clusters with the “cutreeDynamic” function of the “dynamicTreeCut” package (v.1.63) using minimal cluster size = 20, and calculating the cluster consensus matrices and corresponding consensus peptide sequences (Biostrings package). Motif position weight matrices were generated from the consensus matrices using the Bioconductor “motifStack” package. Enrichment across rounds was calculated as the slope of the normalized counts going from the input library to the round 3 samples. For snRNA-sequencing, data were demultiplexed using bcl2fastq2, followed by trimming, mapping, counting, and cell calling with CellRanger v.7.2. Count matrices were pre-filtered to exclude low-quality nuclei, barcode-swapped, and chimeric fragments using the DropletUtils R package. Doublets and aggregates were identified with scDoubletFinder, and ambient transcripts were removed using DecontX for AAV barcode data and SoupX for gene expression data. Cells with mitochondrial transcript content over 0.5% were excluded, and a gene detection range of 600–4,000 per cell was set. Filtered count data were normalized with SCTtransform v.2 and integrated in Seurat v.5, followed by linear dimensional reduction, uniform manifold approximation and projection (UMAP) embedding, and shared nearest neighbor (SNN) graph construction for clustering and visualization. Cell types were identified by integrating with a down-sampled single-cell atlas dataset from Peng et al.,42 and the resulting annotations were validated with a predefined marker gene set. AAV expression data were pseudo-bulked using AggregateExpression in Seurat v.5 and normalized to barcode frequencies in the original AAV input libraries, then further normalized to the benchmark variant AAV2 and AAV2-7m8 to estimate relative performance and account for sample variability. For assessing cell type tropism of single variants, a joint cell type dataset (CTD) was created from the merged Seurat object with expression weighted celltype enrichment (EWCE), and the specificity matrix was computed. Quantification and statistical analysis were conducted using Graphpad Prism v.10 (GraphPad Software). Data are presented as mean ± SEM. Datasets with unequal variance between groups were log-transformed before analysis, as noted in the figure legends. Equal variance and normality were assumed for all other datasets. Statistical significance was assessed using one-way ANOVA for comparisons between two or more groups, as indicated. The sample size (n) for each dataset is specified in the figure legends. Data and code availability The supporting data of this study can be obtained from the corresponding author upon reasonable request. Acknowledgments The authors extend their gratitude to Phillipp Doll, Isabel Lang, Gudrun Zimmermann, Christine Mayer, Nadine Rehm, and Martina Steinrock for their valuable technical assistance and scientific insights. This study was funded by the Eye Health & Research Beyond Borders (EH&RBB), Boehringer Ingelheim Pharma GmbH & Co. KG, Biberach, Germany, and support from the 10.13039/100008349Boehringer Ingelheim Discovery Research Postdoc program. The graphics and the graphical abstract were generated using biorender.com. Author contributions Conceptualization, D.B., C.S., and S.M.; investigation, Y.F., D.B., T.S., M.S., C.S., and S.M.; data curation, M.O. and M.S.; formal analysis, Y.F., M.O., D.B., M.S., C.S., and S.M.; performance, evaluation and analysis of ophthalmological data, F.F.R. and I.P.S.; writing – original draft, Y.F., C.S., and S.M.; all authors contributed to reviewing and editing. Declaration of interests Y.F., D.B., M.O., R.B., B.S., T.S., U.M., C.S., S.M., report a relationship with Boehringer Ingelheim Pharma GmbH & Co. KG that includes employment. B.K. reports a relationship with Labcorp Early Development Services GmbH. M. S. and S.W. report a relationship with Systasy Bioscience GmbH that includes employment.