2.1 hiPSC culture, the establishment of the hOPM cell line, RPE differentiation and induction of exon skipping using the CRISPR/Cas9 system

WTC11 cells¹⁶ were cultured according to the protocol provided at the following site: https://labs.gladstone.org/conklin/protocols.html. The establishment of the hOPM cell line and differentiation of hOPM into RPE were done following the previous report.¹⁵ After reaching confluence on a 12 well plate coated with Matrigel (Coring, 356231), hOPM cell lines were cultured with the RPEMM1¹⁵ without nicotinamide. After 7 weeks the culture medium was changed to X-VIVO 10 (Lonza, 04-380Q). We confirmed doxycycline-dependent differentiation of hOPM to RPE by morphological observation (Supplemental Figure S1A) and gene expression (Supplemental Figure S1B). To induce exon 6-skipping, hOPM RPE cells were transfected with the plasmid described below using the Neon Transfection System (ThermoFisher Scientific) and cultured for 7 weeks with X-VIVO 10. To induce oxidative stress, RPE cells were incubated with X-VIVO 10 containing 2 mm sodium iodate (Sigma Aldrich, S4007) for 1 day.

2.2 Genome edition using the CRISPR/Cas9 system and the establishment of clonal cell lines

The target site (Supplemental Table S1) of guide RNA (gRNA) was designed using CRISPRdirect (https://crispr.dbcls.jp/). Oligo DNAs were hybridized and inserted into the BbsI sites of pSpCas9(BB)-2A-GFP (PX458, Addgene, 48,138). To construct PX458 expressing the Cas9 VQR variant (D1135V/R1335Q/T1337R), each mutation was introduced by an inverse PCR-based mutagenesis kit (Toyobo, SMK-101). hOPM cells were transfected with the plasmids by the Neon Transfection System. After 2 days of culture, GFP-expressing cells were sorted using FACSAria II (BD). For the establishment of clonal cell lines, 5E3 cells were seeded on a six-well plate. Genomic PCR was carried out using PrimeSTAR GXL DNA polymerase (Takara, R050).

2.3 Reverse transcription and semi-quantitative or quantitative PCR (RT-qPCR), droplet digital PCR (ddPCR), immunocytochemistry

Total RNA was extracted using Sepasol RNA I Super G (Nacalai tesque, 09379) and cDNA was synthesized using ReverTra Ace (Toyobo, FSQ-301). Semi-quantitative PCR or qPCR was performed using PrimeSTAR GXL DNA polymerase or THUNDERBIRD SYBR qPCR Mix (Toyobo, QPS-101) and LightCycler 96 (Roche), respectively. The relative mRNA expression level was normalized by the expression level of ACTB and GAPDH. Quantitative ddPCR was carried out using the QX200 ddPCR system (BioRad) and ddPCR Supermix for Probes (BioRad, 1863024). The same amounts of cDNA were used for templates. Primer sequences are available in Supplemental Table S1.

Retinal pigment epithelium cells derived from hOPM cell lines were seeded on a chamber slide glass (Corning, 354118) coated with Matrigel. Immunostaining was done as previously described¹⁷ using antibodies indicated in Supplemental Table 2. Nuclei were visualized with 4′, 6-diamidino-2-phenylindole dihydrochloride. Fluorescence images were obtained using Zeiss Axio Imager M2 (ZEISS).

2.4 Phagocytosis assay

Porcine eyes were obtained from Tokyo Shibaura Zouki, and photoreceptor outer segment (POS) was prepared as previous reported.¹⁸ Fluoresbrite YG Microsphere latex beads (Polysciences, 17154–10) were mixed with 1.5 mg of POS in DMEM medium with 2.5% sucrose and rotated at room temperature for 1 h for opsonization of latex beads with POS to enhance the phagocytic activity.¹⁹ After washing with 0.9% NaCl solution, beads were resuspended with X-VIVO 10 containing 30% fetal bovine serum and incubated with RPE cells which were preincubated with X-VIVO 10 containing 30% fetal bovine serum for 1 h to prime cells.²⁰ After 24 h of culture, RPE cells were resuspended with X-VIVO 10 containing propidium iodate and analyzed by using FACSAria II.

2.5 Building of neural network and training

We made a model based on VGG16,²¹ which is a convolutional neural network model containing five blocks, the convolutional layer, the ReLU activation function and the Max pooling layer.²² We used the initial parameters of the five blocks from results obtained by ImageNet training. Then, stacked information in the blocks was further stacked to two dimensions. Finally, the probability of each class was returned by the softmax formula.

2.6 Construction of an automated model to choose differentiated regions

We compared total non-pigmented (undiff) (1,847 blocks), mixed (1,227 blocks) and pigmented (diff) blocks (1,387 blocks) using the model, and cross-segregation of the matrix was performed. We found that undiff and diff were identified with about 80% accuracy, but lower accuracy was observed with mixed blocks, which was predictable because the mix contains mingled diff and undiff regions (Supplemental Figure 2A). Only less than 50% of accuracy was obtained with the mixed category (Supplemental Figure 2A), which is predictable since the category contains both undiff and diff regions. Based on the results, we then generated a model to classify the blocks to diff and undiff automatically. We used undiff (322 blocks) and diff (358 blocks) for building a model for automated classification of diff and undiff blocks. Since we used the softmax formula, the output of the neural network model is total 1 probability, i.e. {Diff:0;9, Undiff:0.1}. The differentiation classification model was applied to a total of 1,907 blocks, and output is shown in Supplemental Figure 2B. Then, based on the probability of judgment as “diff”, we adjusted the threshold to make the Youden index the maximum. The judgment of either “mix” or “undiff” was also made by using the same method. The identification of “undiff”, “mix”, and “diff” patches was used for the judgment of the maturation of hOPM RPE cells, and mixed populations were used in the experiments for the evaluation of the effects of a CHM mutation and exon-skipping.

2.7 In vitro prenylation assay and western blotting

The in vitro prenylation assay was performed following previously reports.²³ In brief, RPE cells were resuspended with cold prenylation buffer (25 mm HEPES pH 7.2, 50 mm NaCl, 2 mm MgCl₂, 2 mm DTE, 20 μm GDP) containing a protease inhibitor cocktail (Nacalai tesque, 25,955). After cellular debris was removed, the supernatant was ultracentrifuged at 80,000 rpm for 60 min at 4°C. The supernatant was concentrated using Amicon Ultra (Millipore, UFC200324) and used for the in vitro prenylation assay or as a cytoplasm fraction. The supernatant was incubated with 1 μm RabGGTase (JenaBioscience, PR-103), 1 μm REP-1 (JenaBioscience, PR-105) and 5 μm biotin-geranyl pyrophosphate (JenaBioscience, LI-105) for 6 h at room temperature. Biotin-labeled RAB protein was detected using streptavidin-HRP (Invitrogen, 434323) and Amersham Imager 600 (Cytiva). In western blotting analysis of the cytoplasm fraction, mouse IgG anti-ACTIN (Sigma Aldrich, A4700), anti-RAB27A (Santa Cruz, sc-81,924), anti-RAB38 (Santa Cruz, sc-390,176), anti-REP-1 (Santa Cruz, sc-23,905) and HRP-linked anti-mouse IgG antibody (Cytiva, NA-9310) were used. The signal quantification was carried out by using ImageQuant TL (Cytiva).

2.8 Statistical analysis

Statistical analysis was done using Python Scipy (https://scipy.org/). The p-values were calculated Tukey's test as indicated in the figure legends with R software (The R Foundation for Statistical Computing, Vienna, Austria).

3.1 AI-based evaluation of cell differentiation into RPE

We established the hOPM cell line by introduction of the plasmids as previously described.¹⁵ In the hOPM cell line, RPE differentiation is not homogeneous; pigmented and non-pigmented regions are intermingled (Figure 1A). We first confirmed that OTX2 and MITF were expressed only in pigmented cells (Figure 1B), and the pigmented and non-pigmented regions are referred as “diff” and “undiff”, respectively. We explored whether deep learning could be used to identify the developmental transition into hiPSC-derived RPE. Photographs were taken using a conventional microscope (Nikon Eclipse, TE300) and camera (Nikon, D3300) (Figure 1A), and then divided into small squares; an expert separated the pigmented and non-pigmented areas. The former were dark, the cells were hexagonal, and the cell edges were visible (Figure 1A, C). Each “diff” block contains seven or eight hexagonal cells (Figure 1D, E). Blocks with both “undiff” and “diff” regions were classified as mixed (Figure 1D). The population of “diff” blocks increased linearly for 7 weeks, and was then maintained at about 30% of the total (Figure 1F).

We examined whether “undiff”, “mixed” and “diff” blocks could be separated by AI. K-Fold cross validation (k = 5) was used to create a model based on the VGG16 convolutional neural network. Cross-segregation consistently afforded 70–80% accuracy in terms of “diff”/“undiff” separation at 4/5, 6/7 and 8/9 weeks (Table 1, 2, and 3). Artificial intelligence distinguished “diff” regions at these times with remarkable accuracy (> 99%) (Table 4).

3.2 AI-based identification of various developmental stages of hiPSC-derived RPE cells

Using a fresh culture, we took 10 photographs daily in weeks 1, 4, 8 and 13 after RPE differentiation commenced and prepared small blocks (14 × 21 blocks from one photo). “Diff”/“undiff” classification was based on the softmax model described in the Materials and Methods. Cross-validation revealed that the accuracy was 90–100% (data not shown). Then, we prepared a new series of photographs (the test data) and evaluated predictions of culture-weeks. The accuracy was very high in weeks 1, 4 and 8, but only about 50% in week 13 (Table 5), suggesting that later samples exhibited similar profiles. We used the system to confirm maturation of hiPSCs into RPE in the following experiments.

3.3 Construction of exon-skipping vectors and evaluation thereof in Y79 cells

We chose CHM as a model for evaluation of exon-skipping in hOPM cells. CHM mutations of the patients are found in most of coding regions of CHM (Supplemental Figure 3A). As the numbers of nucleotides of exons 6, 9, and 10 are multiples of 3, the exon 6, 9, or 10 skipping is in-frame. We designed gRNAs targeting the splice acceptor (SA) or splice donor (SD) sites and found that Cas9 cut close to the SD sites of exons 6, 9 and 10 (Supplemental Figure 3B). Before expressing the skipping vector in hOPM cells, we checked whether the vector worked in a conventional cell line Y79 retinoblastoma line²⁴ for the first screen in various ways. After electroporation of the vector into Y79 cells, we sorted GFP-positive cells. The RT-PCR and droplet digital PCR (ddPCR) results indicated reduced expression of CHM mRNA containing exon 6, 9 or 10 and also induction of CHM mRNA synthesis in the absence of exon 6, 9 or 10 (Supplemental Figure 4A, B); ddPCR quantitated the exon-skipping rate (Supplemental Figure 4B). The decomposition (TIDE) assay²⁵ predicted high indel efficiency around the SD site of exon 6, 9 or 10 explaining the high level of CHM exon 6, 9 or 10 skipping (Supplemental Figure 4C), meaning that the first screening using Y79 identified the promising vectors. We then focused on the hiPSC line.

3.4 CHM exon 6-skipping was associated with expression of truncated REP-1 in hiPSCs

Plasmids expressing Cas9, gRNA, and GFP were introduced to WTC11 cells, and RT-PCR of purified GFP-expressing cells revealed bands corresponding to CHM-encoding mRNA without the target exons (Figure 2A). We confirmed exon skipping by sequencing the RT-PCR products, and the genomic DNA regions around the target SD sites (Figure 2A, B, Supplemental Figure 3B). The molecular weights of the truncated REP-1 were as expected in all cases (Figure 2C). REP-1 lacking exon 9 or 10 exhibited faint bands, but the expression level of REP-1 lacking exon 6 was much stronger (Figure 2C). The result indicated that REP-1 lacking exon 6 is more stable than REP-1 lacking exon 9 or 10 and exon 6-skipping is more effective. Thus, we focus on CHM exon 6-skipping below.

3.5 CHM exon 6-skipping was induced in RPE cells with a frameshift mutation in that exon

We established a clonal hOPM cell line (CHM#1) with a 17 nt deletion in exon 6 and a premature stop codon (Figure 3A). Plasmids expressing Cas9 and gRNA targeting the CHM exon 6 SD site (gCHM_Ex6SD) or control luciferase (gLuc) were introduced into CHM#1 RPE cells via electroporation. Immunocytochemistry confirmed the RPE differentiation; OTX2 and PAX6 were expressed in the nuclei, and ZO-1 in the cell membranes, of all three samples (Figure 3B). We used the AI-based staging method to evaluate whether the mutation affected the maturation of RPE cells over 14 weeks. Artificial intelligence judged that the cells were at 13 weeks of culture period among the 1, 4, 8 and 13 week options with >90% prediction in all three samples (Figure 3C). This suggested that neither the frameshift mutation in CHM exon 6 nor the CHM exon 6-skipping adversely affected differentiation into an RPE.

Exon 6-skipping was confirmed in CHM#1_sk6 RPE cells (Figure 3D). We found that the CHM mRNA level was reduced (compared with that in WT RPE cells), probably owing to nonsense mediated decay in CHM#1 RPE cells, but this deficit was rescued in CHM#1_sk6 RPE cells (Figure 3D). Finally, we examined REP-1 expression levels using western blotting. In CHM#1 RPE cells, REP-1 expression was minimal (Figure 3E); however, we observed truncated REP-1 expression in CHM#1_sk6 RPE cells (Figure 3E). Because the REP-1 expression level in CHM#1_sk6 was not restored as the CHM mRNA level, REP-1 lacking exon 6 was probably unstable compared with full-length REP-1.

3.6 CHM exon 6-skipping rescued the reduced phagocytic activity of CHM#1 RPE cells under oxidative stress

We then evaluated phagocytic activity via flow cytometry. We observed no decrease in the percentage of FITC-positive cells among CHM#1 RPE cells (Supplemental Figure 5). In addition, the cell death levels of CHM#1 RPE and control cells were comparable (Supplemental Figure 5). Since CHM dysfunction is associated with high levels of oxidative stress, we employed sodium iodate, which induces oxidative stress to RPE cells.^{26, 27} In WT RPE cells, treatment with sodium iodate for 1 day decreased phagocytic activity and increased cell death (Supplemental Figure 5 and Figure 4A, B); both phenomena were accelerated in CHM#1 RPE cells (Figure 4A, B). In CHM#1_sk6 RPE cells, although the level of (accelerated) cell death was the same as that of as CHM#1 RPE cells (Figure 4B), the phagocytic activity was significantly restored (Figure 4A). This suggested that REP-1 lacking exon 6 was partially functional and contributed to the phagocytic activity of RPE cells under oxidative stress conditions.

3.7 REP-1 lacking exon 6 is likely to contribute to prenylation of RAB38 but not RAB27A

REP-1 controls cellular RAB localizations via prenylation; we thus examined the prenylation status of RABs in RPE cells under oxidative stress using an in vitro prenylation assay, which labels unprenylated RABs with biotin. CHM#1 RPE cells showed significantly higher levels of unprenylated RABs than did WT RPE cells (Figure 4C). The high levels of unprenylated RABs were maintained in CHM#1_sk6 RPE cells (Figure 4C). Most RABs are prenylated and then localized to cellular membranes; unprenylated RABs reside in the cytosol. Unprenylated RAB27A accumulates in choroideremia patients; the cytosolic levels of RAB27A increases in RPE cells differentiated from hiPSCs derived from a choroideremia patient.^{28, 29} In addition to RAB27A, RAB38 has been suggested to contribute to choroideremia.³⁰ The cytosolic levels of RAB27A and RAB38 were significantly increased in CHM#1 RPE cells, probably reflecting REP-1 dysfunction (Figure 4D). Although high RAB27A levels were also observed in CHM#1_sk6 RPE cells, the RAB38 level was significantly decreased (compared with that in CHM#1 RPE cells; Figure 4D). Thus, REP-1 lacking exon 6 may regulate the prenylation status of certain RABs, including RAB38, and the phagocytic activity of RPE cells.