Defining the safety and efficacy of liver-directed AAV gene therapy using a human liver tissue-equivalent platform Authors: Ritu M Ramamurthy, Wen Ting Zheng, Sarah E Wachtman, Jonathan H Diaz, Sunil K George, Trang Simon, Yu Zhou, Meimei Wan, Baisong Lu, Stephen J Walker, Colin E Bishop, Christopher B Doering, H Trent Spencer, Anthony Atala, Christopher D Porada, Graça Almeida-Porada Published: 2026 ABSTRACT AbstractThis study reports on the use of human liver tissue equivalents (hLTEs) fabricated using major cell types at ratios that recapitulate native liver structure, physiology, and function, to investigate transduction efficiency, cellular tropism, functional impact, and genotoxicity of 2 adeno-associated virus (AAV) serotypes, AAV5 and AAV3b, encoding eGFP under the strong cytomegalovirus (CMV) promoter to ensure ubiquitous transgene expression in all cells. Additionally, AAV5 encoding eGFP or a bioengineered FVIII transgene (lcoET3) under the phosphoglycerate kinase (PGK) promotor was used to identify unique pathways specific to lcoET3 and/or the effects of different promoters. Overall, AAV5 yielded higher eGFP expression, both AAVs transduced endothelial and stellate cells more efficiently than hepatocytes and Kupffer cells, and both altered liver function biomarkers. Differential gene expression analysis showed that multiple genes involved in hepatotoxicity/inflammation were significantly dysregulated, with each serotype, promoter, and transgene producing a distinct pattern of transcriptional alterations. Integration site analysis identified AAV integrations throughout the human genome, with some in the vicinity of multiple genomic loci associated with oncogenesis. Interestingly, numerous integrations were also found within the human mitochondrial genome. Therefore, hLTEs are a valuable platform for human-relevant safety assessment and to gain critical insights for developing safer and more effective liver-directed AAV gene therapy.Keywords: AAV, gene therapy, safety, genotoxicity, liver tissue equivalents, organoids, 3D constructs hepatoxicity, hemophilia A, FVIII Introduction Hemophilia A (HA) is the most common X-linked bleeding disorder, has a prevalence at birth of 24.6/100,000 males, and affects over 1 million individuals worldwide.1 Persons with severe HA have <1% of the normal plasma levels of FVIII activity and present with life-threatening spontaneous bleeding, including connective tissue/muscle hematomas, internal bleeding, and hemarthroses leading to chronic debilitating arthropathies.2 Over the past several years, gene therapy (GT) clinical trials employing AAV vectors to treat adult HA patients have provided evidence of safety and efficacy and promised a “functional cure,” with a dramatic reduction in the annualized bleeding rate after just a single treatment.3,4,5 However, these trials have also highlighted unanticipated risks/hurdles, including hepatic inflammation, resulting in an increase in liver transaminases (alanine aminotransferase [ALT]/aspartate aminotransferase [AST]), which usually, but not always, respond to a course of steroid therapy, and a marked drop in FVIII plasma levels at roughly 3 years post-GT. This contrasts with results obtained with hemophilia B (HB) GT, in which most patients exhibited minimal toxicity and have enjoyed sustained correction for 13 years to date,6 with the exception of trials with fidanacogene elaparvovec7 and verbrinacogene setparvovec,8 in which the vectors encoded the FIX Padua variant, which has approximately 8 times the specific activity of native FIX. Transaminitis was frequently observed in these trials, but many patients exhibited supraphysiologic levels of FIX activity post-AAV GT, with levels in one patient reaching over 300% of the normal FIX levels and requiring anticoagulant therapy.8 Whether this aspect of the study may have contributed to the observed transaminitis is not clear. Nevertheless, the authors of both studies make a point of clearly stating that the transaminitis observed was generally mild, it was generally observed in patients in the higher/highest dose cohorts, and it was significantly less severe than that observed in similar GT trials treating HA. While HA and HB are clinically similar, differences in the affected gene size (F8 cDNA is 7.2 kb vs. 1.4 kb for F9), the native site of protein expression (hepatocytes for FIX vs. liver sinusoidal endothelial cells [LSECs] for FVIII),9,10 and mutation types render these diseases very different from a GT standpoint, stressing the importance of studying the effect of different transgenes on the intended target cells for each specific pathology. Furthermore, FVIII is a large, complex protein that exhibits a high propensity for retention, misfolding, and aggregation in the endoplasmic reticulum (ER), which can lead to ER stress, induction of the unfolded protein response (UPR), oxidative stress, and eventually cell death.11 These properties of FVIII make it far more challenging to deliver and express at high levels by GT. A wealth of preclinical work has also now demonstrated that AAV, which was long assumed to remain episomal, integrates into the genome of hepatocytes (in mice) at a frequency of approximately 0.1% per transduced cell.12,13,14,15 In addition, in GT for non-bleeding disorders, hepatocellular carcinoma (HCC) was reported in one liver-directed AAV-based preclinical murine study,12 and a larger study found that HCC risk depended on vector dose, promotor, enhancer, and the degree of hepatic cell division at vector administration.16 By contrast, a 10-year follow-up of AAV-mediated FVIII expression in adult HA dogs also detected >1,700 integration events and clonal expansion in liver tissue, but there was no evidence of tumorigenesis.17,18 A single study using mice, whose liver was repopulated with human hepatocytes, provided a first insight into the frequency with which AAV can integrate into the genome of human hepatocytes following transduction in vitro and in vivo.19 Recent biopsy data from 5 patients who received the only FDA-approved GT product for HA (valoctocogene roxaparvovec)20 documented the presence of hundreds to thousands of unique AAV genomic integrations in each recipient.21 There is, therefore, a need to better define the causes of liver inflammation and the potential for genotoxicity in novel AAV GT approaches. To improve GT safety, other questions also need to be answered, such as tissue tropism/cell specificity of AAV vectors of differing serotypes,22,23 as there are marked species-specific differences in AAV tissue and cellular tropism,24 as well as species-specific differences in post-entry events such as nuclear translocation or uncoating and conversion of the AAV genome to dsDNA.25,26 Clinical approaches to-date have relied upon tissue-specific promoters to restrict transgene expression to the desired cell type in vivo,27 and since multiple liver biopsies to define tropism/specificity in humans cannot be justified, thus far in humans, there is no definitive data concerning tropism/cell specificity for any of the serotypes of AAV vectors used in clinical trials. The use of a tissue-/cell-specific promoter does not limit transduction itself to a particular tissue or cell type; thus, such an approach continues to be inherently problematic, as off-target effects are possible, and the effective vector dose reaching the desired target cells is markedly reduced by binding and uptake of the vector by cells/tissues that cannot contribute to transgene expression. Collectively, these studies raise the critical question of the accuracy with which animal models will predict tropism/vector biodistribution, transduction efficiency, and eventual treatment success in humans. Organ tissue equivalents (OTEs) produced with normal primary human cells can be maintained and retain normal functionality for several months, have begun shifting the paradigm in drug testing, and have been shown to accurately predict the toxicities of former FDA-approved drugs that were subsequently recalled due to unpredicted toxicity.28,29 Here, we used human liver tissue equivalents (hLTEs), which are composed of all of the major cell types of the liver in physiologically relevant frequencies and are able to recapitulate the biology and function of the native human liver,28,29,30,31 to address multiple gaps in knowledge regarding the safety and efficacy of liver-targeted AAV GT, using two serotypes, AAV5 and AAV3b. Using a ubiquitous promoter to ensure transgene expression in all cells, we identified which cell types within the human liver, aside from hepatocytes, are transduced by each AAV serotype, and we evaluated the potential of these vectors to induce hepatocyte stress and innate immunity. We delineated the role ectopic expression of FVIII within hepatocytes (not the cell type responsible for FVIII production)10,11,32,33,34,35,36 plays in these processes, and we defined the potential for genotoxicity of AAV in the steady-state human liver. This work provides critical insights regarding the impact of liver-directed AAV-GT, and it validates the hLTE model as a valuable platform to bolster preclinical investigation for developing safer and more effective liver-directed AAV gene therapies. Results hLTE fabrication, cytotoxicity, and ultrastructural analysis of hLTEs post-AAV transduction hLTEs were fabricated using healthy primary adult human hepatocytes, hepatic stellate cells, Kupffer cells, and LSECs, at physiologically relevant ratios (75% hepatocytes, 10% Kupffer cells, 10% stellate cells, 5% endothelial cells), and they were transduced at a multiplicity of infection (MOI) of 105 viral genomes (vg)/cell, with two AAV serotypes encoding EGFP. EGFP was selected as a transgene to facilitate the quantification of transduction efficiency and the determination of cell tropism and effects of transduction on hLTE assembly. Transduction was started on the same day of fabrication to ensure the viral vectors had equal probability of transducing all cell types during the period of cellular self-aggregation. The experimental design and timeline of the study is shown in Figure S1, and it is described in the “materials and methods” section. To investigate whether transduction impacted hLTEs’ viability, and to ensure that hLTEs remained viable throughout the study, the percentage area of viable cells was quantified by imaging, after live/dead assays were performed (Figure 1A). Self-aggregation of hLTE-AAV5 (N = 3; n = 3), hLTE-AAV3b (N = 3; n = 2), and hLTE-NT (N = 3; n = 3) was completed within 4 days of fabrication, and all hLTEs remained viable for the 10-day period of evaluation, demonstrating that transduction did not impact viability after formation. At day 5, the percentage of viable cells was 99.3 ± 0.33, 99.3 ± 0.2, and 99 ± 0.4, and at day 10, it was 96.4 ± 1.1, 97.9 ± 0.6, and 98.1 ± 0.9 in hLTE-AAV5, hLTE-AAV3b, and hLTE-NT, respectively (Figure 1B), demonstrating that transduction with either AAV3b or AAV5 had no negative impact on viability of the hLTEs. Differences in morphology between AAV-transduced hLTE and hLTE-NT were also evaluated using H&E staining performed on day 10, and no significant histological differences were seen (N = 3, n = 3) (Figure 1C). To confirm the degree to which hLTE-NT and AAV-transduced hLTEs resembled the native human liver, and whether transduction caused ultrastructural changes, transmission electron microscopy (TEM) was performed at day 10 (N = 3; n = 3). TEM analysis revealed that all groups exhibited microvilli and bile canaliculus-like structures (Figure 1D), indicating the formation of a rudimentary biliary system, and the presence of tight junctions, confirming the establishment of appropriate cell-cell interactions (Figure 1E). Nevertheless, distinct ultrastructural changes were observed between the hLTE-NT and hLTE-AAV3b groups when compared with the hLTE-AAV5 group, with the latter exhibiting less-abundant mitochondria and ER (Figure S2A). Of note is that these data agree with previous studies showing that AAV transduction can induce mitophagy, reducing the cellular content of mitochondria specifically within hepatocytes.37 They also support the studies showing that AAV5, after receptor mediated-endocytosis, undergoes trafficking through the endocytic compartments followed by retrograde transport to the Golgi or the ER38 and that accumulation of AAV within the ER, combined with high-level of expression of the EGFP transgene, can lead to induction of the UPR, which in turn can trigger ER “self-destruction.39 In hLTEs treated with 20 mM acetaminophen (APAP) (N = 3; n = 3) (APAP-treated hLTE-NT, positive control), TEM readily detected necrotic areas (Figure 1F). Cytotoxicity and ultrastructural analysis of hLTEs post-AAV transduction (A) Viability of hLTEs determined by live/dead staining; NT, non-transduced. Leica TCS LSI Macro Confocal, single macro zoom objective, scale bars, 150 μm. (B) % of viable area of each group in hLTE-AAV5, (N = 3; n = 3), hLTE-AAV3b (N = 3; n = 2), and hLTE-NT (N = 3; n = 3). Data are represented as mean ± SEM, and statistical analysis was performed using one-way ANOVA followed by Tukey’s multiple comparisons test. ns, non-significant; (C) H&E staining of hLTEs to determine morphological changes in NT and AAV-transduced hLTEs (N = 3, n = 3). Olympus BX63 using a 40× objective Scale bars, 20 μm. (D) Representative transmission electron microscopy (TEM) images of microvilli (solid arrows) and bile canaliculus-like structures (dashed arrows) in day 10 hLTEs (N = 3, n = 3). (E) Representative TEM images of tight junctions (open double-line arrows) in day 10 hLTEs. (F) Comparison of negative (NT) control and positive control for liver damage (hLTEs treated with 20 mM acetaminophen [APAP]) (N = 3; n = 3). FEI Tecnai BioTwin Transmission Electron Microscope (120 keV); white scale bars, 2 μm; black scale bars, 500 nm. AAV transduction impairs self-aggregation of hLTEs To investigate whether AAV disrupted and/or impaired hLTE self-aggregation, non-transduced (NT) hLTEs and hLTEs transduced with either AAV3b-CMV-EGFP or AAV5-CMV-EGFP were imaged and tracked using the IncuCyte Live-Cell Imaging System (Figures S2B and S2C). Visualization of hLTEs at day 4 after fabrication/transduction established that 66 ± 4.1 hLTEs had been formed in the NT group, while only 59.6 ± 2.4 and 32.6 ± 3.5 hLTEs were generated in the AAV5 and AAV3b groups, respectively, demonstrating that the AAV-transduced groups formed significantly less hLTEs than the NT controls (N = 10) (p < 0.05). In addition, for the remainder of the monitored period, the AAV-transduced groups continued to display significantly fewer organoids compared with the NT control (Figure 2A; Figure S2D). Self-aggregation, GFP expression, and transduced hLTE cell types post-AAV transduction (A) AAV transduction impairs self-aggregation of hLTEs. Mean hLTE counts in the 3 different groups (N = 10). (B) Average GFP integrated intensity/hLTE; NT, non-transduced (background) (N = 10). Pink ∗p ≤ 0.05, hLTE-NT vs. hLTE-AAV5; teal ∗p ≤ 0.05, hLTE-NT vs. hLTE-AAV3b; and blue ∗p ≤ 0.05, hLTE-AAV5 vs. hLTE-AAV3b. Mixed-effects model with Geisser-Greenhouse correction and Tukey’s multiple comparisons test was used for statistical analysis (α = 0.05). (C) Fold change in GFP expression measured using RT-qPCR, normalized to human GAPDH (n = 3) on day 10 hLTEs’ RNA (n = 749 ± 38, 575 ± 24, and 246 ± 40 hLTEs for NT, AAV5, and AAV3b groups, respectively). Statistical analysis: one-way ANOVA and Tukey’s multiple comparisons test. ∗∗∗p ≤ 0.001, ∗∗∗∗p ≤ 0.0001. (D) Representative immunofluorescence images of hLTE cell types, DAPI, and GFP. Scale bars, 20 μm. Percentage of area transduced in each of the four cell types comprising hLTEs determined using AI-based detection in hLTE-AAV5 (E, N = 3; n = 5) and hLTE-AAV3b (F, N = 4; n = 5). One-way ANOVA and Tukey’s multiple comparisons test were used to determine statistical significance. ns, p > 0.05; ∗p ≤ 0.05, ∗∗p ≤ 0.01, ∗∗∗p ≤ 0.001. Images were acquired using an IncuCyte S3 Live-Cell Analysis System, 4× objective, and an Olympus BX63 fluorescent microscope equipped with X-Cite 120×LED Boost, 40× objective. eGFP protein expression is higher in AAV5-transduced hLTEs eGFP protein expression was quantified in hLTEs transduced with AAV5 and AAV3b, from days 1–7 using the IncuCyte Live-Cell Imaging System after setting the fluorescence intensity baseline/background fluorescence on the imaging system using NT hLTE (Figure 2B). A significant increase in eGFP expression was observed in the AAV5-hLTE group from day 4–7, and in the hLTE-AAV3b group from day 5–7, when compared with the hLTE-NT control (Figure 2B) (N = 10; p < 0.05). Furthermore, from day 4–7, eGFP expression was significantly higher in hLTE-AAV5 than hLTE-AAV3b (N = 10; p < 0.05). In order to evaluate EGFP mRNA expression, RNA was extracted at day 10 (N = 10) from hLTEs, consisting of a total of 749 ± 38, 575 ± 24, and 246 ± 40 hLTEs for NT, AAV5, and AAV3b groups, respectively, and quantitative reverse-transcription PCR (RT-qPCR) was performed. In agreement with the eGFP protein expression data obtained by IncuCyte, the AAV5-hLTE group had significantly higher levels of EGFP mRNA transcripts than hLTE-AAV3b (N = 10; n = 3; p < 0.05) (Figure 2C). AAV5 and AAV3b have higher cell tropism for LSECs and stellate cells in hLTEs To determine which specific cell types within the human liver are more efficiently transduced by AAV5 and AAV3b, sections of AAV5-hLTE and hLTE-AAV3b were stained simultaneously with antibodies to liver-resident cell-specific markers and GFP. A representative image of stained AAV5-hLTE can be seen in Figure 2D demonstrating liver cell-specific staining and co-expression of GFP. The mean percentage area positive for GFP and/or the distinct markers identifying the different transduced cell types was measured using an AI model developed on the Aiforia platform, capable of identifying and quantifying different fluorescent stains in immunohistochemistry (IHC) images. Data show that AAV5 transduced liver endothelial cells and stellate cells more efficiently than Kupffer cells and hepatocytes (N = 3; n = 5) (p < 0.05) Figure 2E. Even though AAV3b transduced significantly more endothelial cells and stellate cells than Kupffer cells, and significantly more endothelial cells than hepatocytes, the difference in transduction between stellate cells and hepatocytes did not reach significance (N = 4, n = 5) Figure 2F. Of note, is that AAV3b transduced significantly more hepatocytes than AAV5 (Figure S3A), while no significant differences were seen in the transduction efficiency of any of the other liver cell types (Figure S3A). Impact of AAV transduction on liver metabolism and function To establish the impact of AAV transduction (N = 5) on disruption of hepatocellular metabolism due to vector capsid and genome exposure, liver function biomarkers were measured in the media of hLTEs from day 6–10. Even though viability remained comparable between groups at all points, standardization for the number of viable hLTEs formed in each group was required. Thus, the concentration of urea, a highly specific indicator of hepatocyte maturity and functionality of liver organoids, was used to normalize the raw concentrations of the other biomarkers (Figures S3B and S3C). Interestingly, we observed that urea levels decreased over time, an observation for which we do not have a definitive explanation. However, this decrease could be related to the absence of flow in our 3D spheroids, as prior studies have shown that flow and shear stress are important regulators of urea production by hepatocytes in culture, with both markedly enhancing urea production/secretion.40,41,42,43 In the AAV5-transduced group (N = 5, n = 374 ± 19), ALT was significantly elevated on days 6–8 post-transduction compared with the NT control and AST was elevated at all time points of analysis. In the AAV3b-transduced group (N = 5, n = 123 ± 20), AST was even more markedly elevated during that same time period (Figures 3A and 3B) but ALT was only significantly higher at day 7. Of note is that the proportion of AST to ALT in hepatocytes is about 2.5:1; thus the greater concentrations of AST identified by ELISA (Figures 3A and 3B) may simply mirror the higher relative abundance of AST found in hepatocytes. In hLTEs transduced with AAV-5, both ALT and AST returned to levels equal to those of the hLTE-NT control at days 9–10, while in hLTEs transduced with AAV-3b, both transaminases decreased when compared with hLTE-NT control. Production of albumin and alpha-1 antitrypsin (A1AT) were also evaluated in all groups. At all time points evaluated, A1AT levels were decreased in both AAV5-hLTE and hLTE-AAV3b when compared with hLTE-NT (p < 0.05) (Figure 3C), while in AAV-transduced hLTE, albumin production was lower at days 6 and 7 (p < 0.05), but these levels then normalized and surpassed hLTE-NT’s albumin levels by days 9–10 post-transduction (Figure 3D). Hepatotoxicity post-AAV transduction Liver function biomarker levels (normalized to urea) measured in media from days 6–10 (A–D) and day 10 (E–H) compared with acetaminophen (APAP)-treated hLTEs. NT, non-transduced. Experiments were performed in duplicates using media from 5 wells (N = 5; n = 374 ± 19, 288 ± 12, and 123 ± 20 hLTEs for NT/24h-APAP/48hAPAP, AAV5, and AAV3b groups, respectively) and analyzed using ANOVA/Dunnett’s/Tukey’s test (α = 0.05). (A–D) Pink ∗p ≤ 0.05, hLTE-NT vs. hLTE-AAV5, and teal ∗p ≤ 0.05, hLTE-NT vs. hLTE-AAV3b. (E–H) ns, p > 0.05; ∗p ≤ 0.05, ∗∗p ≤ 0.01, ∗∗∗p ≤ 0.001, ∗∗∗∗, p ≤ 0.0001. (I) Cytochrome-P450 (CyP450) 3A4 activity represented as concentration of luminescent product generated (RLU, relative luminescence unit), which is directly proportional to hLTE drug-metabolizing capacity. Three independent experiments (N = 3) were assayed in triplicate and analyzed using ANOVA/Tukey’s test. Each N consisted of n = 75 ± 4, 58 ± 2, and 25 ± 4 hLTEs for NT, AAV5, AAV3b groups, respectively. ns, p > 0.05. Significantly dysregulated genes in hLTE-AAV5 (J and L) and hLTE-AAV3b (K and M) related to hepatotoxicity (J and K, respectively) and NF-κB-signaling pathway/inflammation (L and M, respectively) determined in day 5 RNA from 10 wells (N = 10) consisting of n = 749 ± 38, 575 ± 24, and 246 ± 40 for NT, AAV5, and AAV3b groups, respectively. Experiments were repeated 3 times and analyzed using RT2 Profiler software. To confirm that the changes in biomarkers following transduction of the hLTEs with AAV resulted from hepatic damage, APAP, a known inducer of hepatoxicity was used as a positive control. hLTE-NT were treated for 48 h with 20 mM APAP (N = 5, n = 374 ± 19) starting on day 8, and Figures 3E and 3F show the levels of ALT and AST, respectively, in media collected from APAP-treated hLTE-N at days 9 and 10, demonstrating an increase in transaminases following APAP-induced hepatoxicity. Figure 3G shows that the levels of A1AT are decreased followed APAP treatment, while Figure 3H shows that albumin levels rise and fall during this period. To investigate whether AAV transduction impairs hLTE’s drug-metabolizing capacity, cytochrome P450 (CyP450) 3A4 activity was induced (Figure S3D) and quantitated in NT controls and in hLTEs transduced with AAV3b-CMV-EGFP or AAV5-CMV-EGFP. No significant changes (p > 0.05) in CyP450 3A4 activity were observed post-AAV transduction (Figure 3I; AAV5 group: N = 5, n = 58 ± 2; NT control: N = 5, n = 75 ± 4; AAV3b group: N = 5, n = 25 ± 4). Investigating gene expression involved in liver hepatoxicity and inflammation in response to AAV transduction Since protein biomarkers of AAV-induced hepatotoxicity were readily detected at day 6 post-transduction, two different RT-qPCR panels, each targeting 84 genes involved in hepatotoxicity and inflammation, respectively, were performed on RNA extracted from the hLTEs at day 5 post-transduction (N = 10; hLTE-NT n = 749 ± 38, hLTE-AAV5 n = 575 ± 24, hLTE-AAV3b n = 246 ± 40) to assess dysregulation of these genes. The hepatotoxicity panel (Figures S3E and S3F) demonstrated significant (p ≤ 0.05) dysregulation (>1.5-fold) of 3/84 key genes in the hLTE-AAV5 group (Figure 3J) and 9/84 genes in the hLTE-AAV3b group (Figure 3K). Analysis of the 84 key genes involved in inflammation, specifically the NF-κB signaling pathway (Figures S3G and S3H), which is a primary regulator of the innate immune response, revealed that significant (p ≤ 0.05) dysregulation (>1.5-fold) was observed in 5/84 key genes in the hLTE-AAV5 group (Figure 3L) and 6/84 key genes in the hLTE-AAV3b group (Figure 3M). To define the repertoire of differentially expressed genes (DEGs) in processes known to impact GT as a result of transducing hLTEs with AAV vectors, we used the NanoString nCounter Gene Therapy Optimization panel to analyze RNA isolated, on day 5 post-transduction, from NT hLTEs, hLTEs transduced with AAV3b-CMV-EGFP, and hLTEs transduced with AAV5-CMV-EGFP. Figures 4A and 4B show volcano plots and heatmaps, respectively, and Table S1 provides complete lists of all significantly (raw p ≤ 0.065) DEGs in the hLTE-AAV3b group vs. hLTE-NT (113 genes; 98 upregulated and 15 downregulated), the hLTE-AAV5 vs. hLTE-NT (44 genes; 17 upregulated and 27 downregulated), and hLTE-AAV3b vs. hLTE-AAV5 (170 genes; 138 upregulated and 32 downregulated). Significantly (p < 0.065) DEGs in hLTEs transduced with AAV3b-CMV-EGFP and AAV5-CMV-EGFP Total RNA extracted on day 5 (N = 10 wells; n = 749 ± 38, 575 ± 24, and 246 ± 40 for hLTE-NT, hLTE-AAV5, and hLTE-AAV3b, respectively) was analyzed using NanoString nCounter Gene Therapy Optimization panel and ROSALIND data analysis platform (N = 3). (A) Volcano plots displaying significantly (p < 0.065) DEGs when comparing (a) AAV3b-CMV-eGFP-transduced hLTEs with non-transduced (NT) control hLTEs, (b) AAV5-CMV-eGFP-transduced hLTEs with non-transduced (NT) control hLTEs, and (c) AAV3b-CMV-eGFP-transduced hLTEs with AAV5-CMV-eGFP-transduced hLTEs. (B) Heatmaps displaying significantly (p < 0.065) DEGs (red is upregulated, blue is downregulated) when comparing (a) AAV3b-CMV-eGFP-transduced hLTEs with non-transduced (NT) control hLTEs, (b) AAV5-CMV-eGFP-transduced hLTEs with non-transduced (NT) control hLTEs, and (c) AAV3b-CMV-eGFP-transduced hLTEs with AAV5-CMV-eGFP-transduced hLTEs. As can be seen in Table S2, when compared with NT control hLTEs, transduction with AAV3b-CMV-EGFP induced the upregulation of a large number of genes in many of the nCounter Gene Therapy Optimization pathways studied, while transduction with AAV5-CMV-EGFP led to the downregulation of genes in most of these same pathways. Directly comparing DEGs between hLTE-AAV3b and hLTE-AAV5 demonstrated that transduction with an identical vector exerted a far more dramatic effect on the transcriptome of hLTEs when the AAV3b serotype was used than did the AAV5 serotype, with the AAV3b vector significantly upregulating multiple genes in pathways related to systemic immune response (interleukin and chemokine signaling, as well as MHC class I and II antigen presentation) and innate immune response (TLR and type I IFN signaling and cellular sensing of RNA and DNA), which would collectively be expected to promote an immune response to the vector and transgene product. AAV transduction does not induce fibrosis in hLTEs during the 10-day period of study To assess whether AAV transduction triggered hepatic fibrosis, two methods were employed: RT-qPCR and histological analysis. RNA isolated from hLTE collected at day 10 post-transduction, the last day of evaluation, was analyzed by RT-qPCR to assess whether AAV transduction led to dysregulation of genes involved in fibrosis (N = 10; hLTE-NT n = 749 ± 38, hLTE-AAV5 n = 575 ± 24, hLTE-AAV3b n = 246 ± 40) (Figures 5A–5C). No significant (p > 0.05) dysregulation (>1.5-fold) of any fibrosis-related genes was observed in the hLTEs transduced with AAV5, while 4 genes (JUN, PDGFβ, TGFβ2, and PLAU/uPA) normally associated with a pro-fibrotic environment44,45,46,47 were significantly (>1.5-fold; p < 0.05) downregulated in the hLTEs transduced with AAV3b. These data support the conclusion that transduction with AAV3b or AAV5 does not overtly activate pathways traditionally implicated in the induction and progression of fibrosis within the liver. Day-10 hLTEs were also fixed in paraffin, sectioned, and subjected to Masson trichrome and Sirius red staining and quantification (Figures 5D and 5F). As in the preceding studies, APAP-treated hLTEs served as positive controls. Both methods stain for collagen fibers within the hLTEs, which are an indicator of onset of fibrosis. The acquired bright-field images were analyzed using ImageJ to quantify the fibrotic area percentage (Figures 5E and 5G). No significant increase (p > 0.05) in the percentage of fibrotic area was observed in the AAV-transduced groups compared with control hLTE-NT. Collectively, the results of these RT-qPCR and histological analyses suggest that AAV transduction does not promote fibrosis within the hLTEs, at least within the 10-day period of examination. Assessment of the potential for AAV transduction to induce fibrosis in hLTEs Scatterplot of expression of genes involved in fibrosis in hLTE-AAV5 (A) and hLTE-AAV3b (B). Significantly over-expressed genes are depicted in red, while under-expressed genes are depicted in green. (C) Under-expressed genes (fold change > −1.5) post-AAV5 transduction. RNA was extracted on day 10 post-transduction (n = 749 ± 38, 575 ± 24, and 246 ± 40 for hLTE-NT, hLTE-AAV5, and hLTE-AAV3b, respectively), and experiments were repeated three times. Data were analyzed on the RT2 Profiler Data Analysis online platform. (D) Representative images of Masson trichrome staining. Blue, collagen fibers. Scale bars, 20 μm. (E) Fibrotic area percentage calculated using ImageJ (n = 13, 12, 9, 19, 10 for hLTE-NT, hLTE-AAV5, hLTE-AAV3b, hLTE-24 h APAP, and hLTE-48 h APAP, respectively). (F) Representative images of Sirius red/Alcian blue staining. Bright red, collagen fibers. Scale bars, 20 μm. (G) Fibrotic area percentage calculated using ImageJ N = 3 (n = 10, 10, 11, 14, 11 for hLTE-NT, hLTE-AAV5, hLTE-AAV3b, hLTE-24 h APAP, and hLTE-48 h APAP, respectively). Statistical analysis was performed using ANOVA and post hoc Tukey’s test. ∗ Represents statistical significance (p ≤ 0.05). ∗p ≤ 0.05, ∗∗p ≤ 0.01, ∗∗∗p ≤ 0.001. Images were acquired using an Olympus BX63 using a 40× objective. Integration site analysis post-AAV5 and post-AAV3b transduction To define the total number and genomic location of AAV vector integration48 following transduction of the human liver, integration site (IS) analysis using long-read sequencing was performed on DNA extracted from AAV-transduced hLTEs, using pooled samples of 5 experiments from a single donor (N = 5; n = 75, 58, and 25 hLTEs for NT, AAV5, AAV3b groups, respectively). Figure 6A shows the total number of probable ISs, and Figure 6B shows the total number of probable ISs with the distance in base pairs (bp) from transcription start sites (TSSs) after AAV5 transduction. Chromosome ideograms of data from hLTE-AAV5 can be seen in Figure 6C. Blue lines represent each IS location, and the length of the line correlates to the number of supporting reads for each IS. Of note is that multiple ISs were observed in the vicinity of genomic loci that have been associated with oncogenesis and, somewhat surprisingly, in mitochondrial DNA (Figure 6A; Figure S4). The total number of probable ISs, and the total number of probable ISs with the distance in base pairs (bp) from TSS after AAV3b transduction can be seen in Figures 6D and 6E. Chromosome ideograms of data from hLTE-AAV3 can be seen in Figure 6F. Blue lines represent each IS location, and the length of the line correlates to the number of supporting reads for each IS. Importantly, there was no significant difference between AAV5 and AAV3b with respect to the percentage of ISs that fell within genes (as opposed to intergenic regions), with 53.1% of AAV5 ISs falling within genes and 53.5% of AAV3b ISs falling within genes. While a higher absolute number of ISs in/near oncogenic loci was observed in hLTEs transduced with AAV3b-CMV-EGFP than those transduced with AAV5-CMV-EGFP (Figure 6; Figure S4), there was no significant difference between the two serotypes with respect to the percentage of ISs that fell within oncogenic regions, once values were adjusted to account for the total number of reads for ISs with AAV5 vs. AAV3b. Integration site analysis of hLTEs post-AAV transduction Total number of integration sites (ISs) with ≥1 supporting read post-AAV5 (A) and post-AAV3b (D) transduction. Chromosome ideograms of data from hLTEs transduced with AAV5 (C) and AAV3b (F). The length of the blue line is proportional to the number of supporting reads. The number of ISs in proximity to transcription start sites (TSSs) categorized based on relative position (in base pairs [bp]) to TSSs post-transduction with AAV5 (B) and AAV3b (E). Integration site analysis was performed with pooled samples of 5 experiments from a single donor (N = 5; n = 75, 58, and 25 hLTEs for NT, AAV5, AAV3b groups, respectively). hLTEs are efficiently transduced by an AAV5 vector encoding a liver codon-optimized bioengineered-FVIII transgene without impairing viability or metabolic activity After establishing that transducing hLTEs with the AAV5-based vector caused minimal impairment to hLTE aggregation, far less gene dysregulation, and fewer integrated vector genomes, the next set of studies investigated the effects of AAV5-mediated delivery of a liver codon-optimized bioengineered-F8 transgene (lcoET3) on hLTEs. Moreover, to determine if delivering different transgenes to hLTEs, using AAV5 as a vector, would result in different outcomes, hLTEs were transduced at an MOI of 105 vg/cell, with either AAV5-PGK-lcoET3 or AAV5-PGK-EGFP at the same MOI, and compared with NT negative control hLTEs. The AAV5-PGK-lcoET3-transduced hLTEs maintained similar viability and metabolic activity to that of the other groups, as assessed on days 5 and 10 post-transduction, using live/dead staining (N = 3; n = 3) and ATP assays (N = 3; n = 4) (Figures 7A–7E). Transduction efficiency in the AAV-transduced hLTE groups (N = 3; n = 96) was established by determining vector copy number (VCN)/cell (Figure 7F), transgene mRNA expression levels (Figure 7G), eGFP expression by IncuCyte-Live-Cell Imaging Figure 7H), and FVIII production by ELISA (Figure 7I). Although VCNs were similar between groups transduced with AAV5-PGK-EGFP and AAV5-PGK-lcoET3, mRNA expression of lcoET3 was significantly higher than eGFP when normalized to the house keeping gene (HKG) RPL13A (Figure 7G). Although different methods were used to quantify the 2 proteins, both eGFP and lcoET3/FVIII proteins were detectable and significantly higher than those detected in LTE-NT. Viability, metabolic activity, gene dysregulation, and drug-metabolizing capacity post-transduction with AAV5 encoding EGFP or a bioengineered F8 transgene (lcoET3) (A) Representative images of live/dead assay performed three independent times (N = 3), and 3 images (n = 3) acquired per N. NT, non-transduced. Images were acquired using Leica TCS LSI Macro Confocal using a single macro zoom objective. ImageJ analysis of live/dead images to determine % viable area on day 5 (B) and day 10 (C). ATP assay to assess metabolic activity represented by RLU (relative luminescence unit) on day 5 (D) and day 10 (E). Three independent experiments (N = 3) were performed, and per N; ATP levels of four hLTEs were measured (n = 4). One-way ANOVA and Tukey’s multiple comparisons test was used for statistical analysis of live/dead and ATP assay (α = 0.05). ns, p > 0.05. Vector copy number (VCN) (F) and transgene mRNA expression (G) measured on day 10 of three biological replicates/from 3 separate donors (N = 3), each consisting of 96 hLTEs (n = 96). One-way ANOVA and Tukey’s test were used to perform statistical analysis (α = 0.05). NT, non-transduced. ns, p > 0.05; ∗p ≤ 0.05. (H) IncuCyte analysis of NT and AAV5-PGK-eGFP hLTEs on day 10 of 3 independent experiments (N = 3), each containing 48 hLTEs per group (n = 48). Student’s t test was used to determine statistical significance (α = 0.05). ∗p ≤ 0.05. (I) FVIII ELISA using protein extracted from hLTEs on day 10 from 96 hLTEs (n = 96) per group, run in triplicate and at 2 different dilutions. ∗∗∗∗p ≤ 0.0001. Evaluation of liver function biomarkers post-transduction with AAV5 encoding lcoET3 To establish the impact of AAV transduction on disruption of hepatocellular metabolism due to vector capsid and transgene exposure, liver function biomarkers were measured by ELISA in the media of hLTEs (N = 3; n = 96/group) from days 5 to 10 post-transduction. Interestingly neither AAV5-PGK-EGFP nor AAV5-PGK-lcoET3 transduction significantly impacted the levels of ALT (Figure 8A) or AST (Figure 8B) when compared with NT hLTEs. Albumin was elevated only in the AAV5-PGK-lcoET3-transduced group at day 5 post-transduction and on days 8 and 9 (Figure 8C). However, A1AT was significantly decreased in both transduced groups throughout the duration of the study period (Figure 8D). Transduction verification and changes in liver function biomarkers post-transduction with AAV5 encoding bioengineered F8 transgene (A–D) Biomarkers in media collected from days 5–10 (N = 3, n = 96/group). Two-way ANOVA followed by Dunnett’s test was used to determine significant differences (α = 0.05). Pink ∗p ≤ 0.05, NT vs. AAV5-PGK-EGFP, and purple ∗p ≤ 0.05, NT vs. AAV5-PGK-lcoET3. (E) CyP450 activity (RLU, relative luminescence unit) measured in uninduced (−IND) and induced (+IND) NT, AAV5-PGK-eEGFP and AAV5-PGK-lcoET3, as well as NT (20 mM APAP with induction). One-way ANOVA followed by Šídák’s multiple comparisons test to perform statistical analysis (p ≤ 0.05 was considered significant). ns, p > 0.05; ∗∗∗∗p ≤ 0.0001. (F) Gene Therapy Optimization panel run using NanoString nCounter System to determine dysregulation of genes post-AAV transduction; 3 biological donors (N = 3), 96 hLTEs (n = 96). Volcano plots displaying significantly (p < 0.065) DEGs when comparing AAV5-PGK-EGFP-transduced hLTEs with hLTE-NT control, AAV5-PGK-lcoET3-transduced hLTEs with hLTE-NT control, and AAV5-PGK-lcoET3-transduced hLTEs with AAV5-PGK-EGFP-transduced hLTEs. (G) Gene Therapy Optimization panel run using NanoString nCounter System to determine dysregulation of genes post-AAV transduction. Heatmaps displaying significantly (p < 0.065) DEGs when comparing AAV5-PGK-EGFP-transduced hLTEs with hLTE-NT control, AAV5-PGK-lcoET3-transduced hLTEs with hLTE-NT control, and AAV5-PGK-lcoET3-transduced hLTEs with AAV5-PGK-EGFP-transduced hLTEs. RNA from 3 biological donors (N = 3), 96 hLTEs (n = 96). (H) Volcano plot displaying significantly (p < 0.05) DEGs when comparing AAV5-CMV-EGFP-transduced hLTEs with AAV5-PGK-EGFP-transduced hLTEs. (I) Heatmap displaying significantly (p < 0.05) DEGs when comparing AAV5-CMV-EGFP-transduced hLTEs with AAV5-PGK-EGFP-transduced hLTEs. To investigate whether AAV transduction and/or the encoded transgene impaired hLTEs’ drug-metabolizing capacity, CyP450 3A4 activity was induced and measured in control NT hLTEs and hLTEs transduced with AAV5-PGK-EGFP or AAV5-PGK-lcoET3. No significant changes (p ≤ 0.05) in CyP450 3A4 activity were observed post-transduction with either AAV5 vector when compared with NT-hLTEs, demonstrating that neither AAV itself nor expression of an EGFP or F8/lcoET3 transgene altered drug metabolism. However, when induced hLTEs were treated with 20 mM APAP (positive control for hepatotoxicity), a significant reduction was seen in drug-metabolizing capacity (p ≤ 0.0001) (Figure 8E). Investigating gene expression involved in liver hepatoxicity and inflammation in response to AAV transduction In order to investigate alterations in gene expression caused by AAV transduction and assess the impact of the specific transgene being expressed, we employed the NanoString nCounter Gene Therapy Optimization panel to profile 800 genes across 40 different pathways related to systemic immune response, innate immune response, cell state, host/vector interactions, and toxicity. Figure 8F shows volcano plots, Figure 8G shows heatmaps, and Table S3 provides complete lists of all significantly (raw p ≤ 0.065) DEGs on day 5 post-transduction in the group transduced with AAV5-PGK-EGFP vs. control NT-LTEs (19 genes; 2 upregulated and 17 downregulated) and the group transduced with AAV5-PGK-lcoET3 vs. control hLTEs-NT (12 genes; 8 upregulated and 4 downregulated) and directly comparing the group transduced with AAV5-PGK-lcoET3 vs. the group transduced with AAV5-PGK-EGFP (24 genes; 20 upregulated and 4 downregulated). Table S4 provides a list of all nCounter pathways with DEGs for each of these 3 comparisons. It is worth noting that, as can be seen in Table S4, the majority of upregulated DEGs when hLTEs transduced with AAV5-PGK-lcoET3 were directly compared with those transduced with AAV5-PGK-EGFP fell within pathways identified by nCounter as being related to systemic immune response and innate immunity. These data provide some of the first direct evidence that the specific transgene being encoded by an identical AAV vector can differentially impact the repertoire of DEGs in the human liver following transduction. The selective upregulation of multiple genes related to immune response by the AAV5 vector encoding a fVIII transgene compared with the same vector encoding EGFP could have important implications for the use of AAV vectors to treat HA. Impact of promoter driving transgene expression on differential expression of genes involved in human liver hepatoxicity after AAV transduction Given the marked effect the transgene exerted on gene expression in transduced hLTEs, we next explored whether the promoter driving transgene expression may also exert an impact on the response of the human liver to transduction with AAV vectors by directly comparing the DEGs in hLTEs transduced with AAV5-CMV-EGFP with those transduced with AAV5-PGK-EGFP. This comparison revealed 204 DEGs, with 172 genes upregulated and 32 genes downregulated in hLTEs transduced with an AAV5 vector in which EGFP was driven by the CMV promoter, when compared with those transduced with an AAV5 vector in which EGFP was driven by the PGK promoter. A volcano plot of this comparison appears in Figure 8H, a heatmap of this comparison appears in Figure 8I, and a complete list of the DEGs (p < 0.05) is provided in Table S5. It is intriguing to note that many of the DEGs that are upregulated in the hLTEs transduced with the AAV vector in which EGFP is driven by the CMV promoter compared with those transduced with an identical vector in which EGFP is under the control of the endogenous human PGK promoter are assigned by nCounter to pathways associated with Systemic Immune Response (genes in 16 different pathways upregulated), Innate Immunity (genes in 11 different pathways upregulated), and Toxicity (genes in 4 different pathways upregulated). The number of DEGs in each of the nCounter-assigned pathways between these two groups of AAV5-transduced hLTEs is shown in Table S6. Discussion Currently, there are multiple gaps in knowledge regarding the safety and efficacy of liver-directed AAV vector-based therapeutics.49,50 Despite success in some clinical settings, these novel therapies are still plagued by issues related to hepatic reactions, host immunity, and the potential for genotoxicity.17,18,19,51 Animal models have proved to be vital for preclinical evaluation of AAV-based treatments, but failure to accurately recapitulate the multitude of human responses to GT, has often led to unexpected outcomes, regarding both efficacy and safety, during clinical trials.51 These shortcomings highlight the need to develop novel preclinical models that can fill specific translational gaps by providing information of direct human relevance, when moving from preclinical animal studies to humans. This study reports on the use of physiologically relevant hLTEs that are able to recapitulate the biology and function of the native human liver28,29,30,31 to investigate the efficiency of transgene expression, cellular tropism, functional impact, and genotoxicity of 2 AAV serotypes—AAV5 and AAV3b—encoding EGFP under the strong CMV promoter to ensure ubiquitous transgene expression in all cells that are transduced within the hLTEs. Data show that hLTEs remained viable throughout the study, and that AAV5 yielded higher eGFP expression than AAV3b, while causing minimal impairment to the self-aggregation/assembly of the hLTEs. It is possible, nonetheless, that the levels of transduction mediated by AAV3b would be sufficient to produce a therapeutic benefit in, e.g., HA. Of note is that other groups demonstrated that AAV3-based vectors perform better in the liver than those based upon AAV5,52,53,54 but important differences between our studies and these prior publications make a direct comparison difficult. The first and foremost is that we utilized AAV3b, not AAV3-ST, and these two serotypes have 4 differing amino acids in the VP1 protein of their capsid leading to substantial differences in hepatocytic transduction. Indeed, one of the studies using an engineered AAV3 capsid showed much higher transduction of human hepatocytes within the liver of “humanized” mice with this engineered AAV3 capsid than with AAV5.54 A second factor that confounds direct comparison of our data with the use of “humanized liver” mice is that while this murine model provides an elegant means of assessing the affinity of various AAV serotypes for human hepatocytes, no other human hepatic cells (i.e., Kupffer cells, stellate cells, endothelial cells) are present.54 In addition, the apparent tropism of each AAV serotype for human hepatocytes is likely impacted by the relative ability of each AAV serotype to bind to murine cells present within the liver and all other tissues throughout the body. For example, as AAV5 can transduce non-hepatic tissue at low-to-moderate levels,55,56,57,58 a fraction of the AAV5 vector administered systemically to these mice may have been bound by other tissues and never reached the liver, giving the appearance of a greatly reduced ability of this serotype to transduce hepatocytes. As the only cells present in our hLTEs are human hepatic cells, such competition with other murine tissues/non-hepatic cell types does not occur, providing another possible explanation for the differing results between the two studies. Another study53 starts with the initial premise that AAV5 does not transduce hepatocytes efficiently54 when compared with AAV3, and the authors then perform directed evolution of the AAV5 capsid to generate novel AAV5 variants that have improved hepatocyte transduction efficiency compared with the initial AAV5. Data from a study52 that perfused cadaveric human liver with AAV vectors of multiple serotypes show that AAV-LK03 yields much higher transduction than does AAV5. Unfortunately, as AAV-LK03 differs from the AAV3b serotype we employed, a direct comparison cannot be made. It is also important to note that the authors used doses nearly 2 logs lower than the effective dose used in the HA clinical GT trials employing AAV-LK03 and AAV5.59,60 While all these findings are extremely important and can guide future therapies, human clinical data in which patients have been successfully treated for HA seem to suggest that AAV5 can yield appreciable transduction of the human liver, even if the doses employed with AAV559 were substantially higher than those with AAV3-LK03.60 Thus, for all of these patients to be phenotypically corrected and exhibit therapeutic plasma levels of FVIII, appreciable transduction of their liver must have taken place following systemic administration of an AAV5 vector in which transgene expression was driven by a hepatocyte-specific promoter. Similarly, the authors of a recently published study comparing various preclinical models for assessing AAV-based GT61 concluded that “based on the fact that each model brings a unique perspective that adds to the overall functional evaluation of AAV vectors, we propose that multiple models should be used to paint a more complete picture and help us make the most informed decision as to which vector should be used in each clinical application.” Importantly, a recently published study61 using liver ductal progenitor cell organoids, demonstrated the AAV5 serotype to be most effective at entering human liver cells, but the efficient entry of AAV5 did not lead to efficient transgene expression. Interestingly, these authors also found that AAV-LK03 and AAV3b were more efficient at driving transgene expression in the organoids, despite less-efficient entry and the reporter EGFP transgene being driven by the identical liver-restrictive ApoE/hAAT promoter. Our studies used primary adult liver cells and a EGFP transgene driven by the constitutive PGK promoter, rather than a liver-restrictive promoter, making a direct comparison between our studies and this prior publication difficult, but these data highlight the complex role played by the capsid in governing not only transduction but also subsequent transgene expression. The studies reported here also investigated tropism of these two AAV serotypes within human liver cells to investigate which of the serotypes transduced cell types within the liver besides hepatocytes. In hLTEs, both AAV5 and AAV3b transduced LSECs and stellate cells at much higher efficiency than hepatocytes or Kupffer cells. The low level at which Kupffer cells were transduced is encouraging from a safety standpoint, as the uptake of various viral vectors by Kupffer cells has been shown to trigger a robust inflammatory response and reduce hepatic gene transfer.62,63,64 When activated, Kupffer cells upregulate cell surface markers including CD163, CD33, and CD68, and they express multiple pro-inflammatory genes, such as TNF-α, IL-1β, IL-6, IL-18, and inducible nitric oxide synthase (iNOS).65,66 Importantly, our DEG analysis revealed that none of these genes were upregulated following transduction of the hLTEs with either AAV5-CMV-EGFP or AAV3b-CMV-EGFP, thus supporting our contention that the negligible transduction observed in Kupffer cells is encouraging from a safety standpoint. While the lower efficiency of hepatocyte transduction is surprising and concerning, the observation that these serotypes of AAV both efficiently transduce human LSECs is encouraging, as these cells are the native site of FVIII production in the liver and, as such, possess the necessary machinery to synthesize and secrete this complex protein without cellular stress/toxicity. Additionally, LSECs act as key regulators of immune balance in the liver,67 and their accessibility through the sinusoids significantly increases the likelihood of both successful gene delivery and subsequent secretion of the transgene product into the circulation. Of note is that these substantial advantages will never be realized with AAV vectors in current clinical use to treat HA, given their utilization of hepatocyte-specific promoters, based on the assumption that both AAV5 and AAV3b exhibit a pronounced tropism for human hepatocytes,19,68 and the rationale that the slow turnover of hepatocytes in the adult liver69 should prolong AAV’s episomal therapeutic gene expression. Moreover, activated stellate cells can produce pro-inflammatory factors and interact with leukocytes and Kupffer cells to initiate an immune response70,71,72,73,74; thus it is of concern that these AAV vectors transduce human stellate cells at significant levels. These studies also demonstrated the impact of AAV5 or AAV3b transduction on the disruption of hepatocellular metabolism. At the earliest time points post-transduction analyzed, secreted levels of albumin were lower in transduced hLTEs when compared with hLTE-NT controls. These levels then normalized and surpassed those of hLTE-NT by the end of the study. In contrast, during this same time period, A1AT levels were consistently decreased and remained significantly lower post-transduction, when compared with hLTE-NT controls, with both AAV serotypes. Although, to our knowledge, a decrease in A1AT has not been reported in patients receiving AAV-GT to date, it would be interesting to know whether A1AT levels were monitored in these patients, as a persistent reduction in A1AT could lead to long-term liver damage. For reasons that are not well understood, most subjects in clinical AAV GT trials for HA have experienced hepatotoxicity, as evidenced by the elevated circulating levels of liver transaminases.27,75 Transduction of hLTEs with AAV5-CMV-EGFP or AAV3b-CMV-EGFP resulted in elevated AST with both AAV serotypes, but ALT was only elevated post-transduction with AAV5. Moreover, ALT and AST levels returned to control levels by day 10 post-transduction of the hLTEs, mirroring the transient nature of the transaminitis reported in AAV GT clinical trials.59,76,77 These data demonstrate that the hLTEs are able to accurately model not only stressor-induced damage and decreased functionality but also the liver’s ability to recover following injury. Interestingly, the elevation in transaminases in human patients has been suggested to be due, at least in part, to the transduced hepatocytes being attacked by the recipient’s adaptive immune system.5 As no components of the adaptive immune system are present in the hLTEs, our findings suggest that there are other local factors that could contribute to the elevation in transaminases, such as direct toxicity due to transduction itself or the high-level expression of the transgene driven by the CMV promoter. Since the hLTEs contain Kupffer cells, the elevated transaminases could also be driven by inflammation due to an innate immune response.78 Analysis of key genes involved in hepatotoxicity and inflammation showed that multiple genes were significantly differentially expressed (>1.5-fold) in the hLTEs transduced with either AAV5-CMV-EGFP or AAV3b-CMV-EGFP when compared with hLTE-NT, with each serotype producing a distinct pattern of transcriptional alterations. Of note, as early as 4 days post-transduction, both serotypes induced upregulation of genes known to play a role in the innate immune response, demonstrating the utility of hLTEs as a model to study the immune aspects of AAV GT. Furthermore, S100A8 was significantly upregulated in the hLTEs transduced with AAV3b as early as day 5. This observation further strengthens our conclusion that hLTEs effectively capture the innate immune response of the human liver to AAV3b, as S100A8 has been reported to act as a danger-associated molecular pattern (DAMP) molecule and stimulate innate immune cells via binding to pattern recognition receptors such as TLR4 and AGER, which then activates the MAP-kinase and NF-κB signaling pathways, thereby amplifying the pro-inflammatory cascade.79,80,81,82 While these analyses, did not provide compelling evidence of AAV-induced fibrosis, especially in the AAV5 group, AAV3b surprisingly resulted in significant down-regulation of four genes (JUN, PDGFβ, TGFβ2, and PLAU/uPA) that have traditionally been associated with the induction and progression of liver fibrosis. As such, transduction with these serotypes of AAV does not appear to directly activate pathways leading to a pro-fibrotic milieu within the human liver. An important caveat to these studies, however, is that their duration was likely too short to enable development and observation of bona fide fibrosis. Directly comparing DEGs between hLTE-AAV3b and hLTE-AAV5 demonstrated that transduction with an identical vector exerted a far more dramatic effect on the transcriptome of hLTEs when the AAV3b serotype was used than when the AAV5 serotype was used, with the AAV3b vector significantly upregulating multiple genes in pathways related to systemic immune response (interleukin and chemokine signaling, as well as MHC class I and II antigen presentation) and innate immune response (TLR and type I IFN signaling and cellular sensing of RNA and DNA), which would collectively be expected to promote an immune response to the vector and transgene product. While the mechanism responsible for the markedly different cellular response to these two different serotypes of AAV is not known, these data suggest that caution should be exercised when choosing which serotype of AAV to use in clinical studies. While extensive data are now available on genomic integration in mice and dogs that received AAV vectors, very limited such data exist for the human liver.13,17,19,21,48,77,83 We performed IS analysis following transduction of hLTEs to define the frequency of insertions of AAV into the human genome and to identify the precise loci at which these insertions occurred. These analyses identified myriad ISs throughout the human genome, but no specific pattern emerged to explain the mechanism by which sites were selected for integration. Both vectors did, however, integrate in the vicinity of multiple genomic loci that have been associated with oncogenesis, suggesting caution for AAV therapeutic products. It is currently not clear why AAV3b appears to exhibit a greater tendency for integration near such loci than AAV5, at least in the hLTE platform. As both vectors harbor the inverted terminal repeats (ITRs) from AAV2, one would imagine their genomic integration profile would be very similar as the selection of sites for integration is thought to be driven by the degree of sequence homology between the ITR and the genomic loci under consideration and to preferentially occur within genomic sites harboring double-strand breaks.84 Equally noteworthy is the identification of numerous AAV integrations within the human mitochondrial genome, as one would expect mitochondrial mutations to severely compromise cell health/viability. To our knowledge, this phenomenon has only been reported once before,85 and questions were raised as to whether the reported integrations may have been an artifact of library preparation.86 This observation highlights just how much we still have to learn about this vector that is in widespread clinical use. While AAV GT has the potential to provide long-term correction for HA, a decline in FVIII expression over time has been observed in individuals who have received these therapies.77,87 Although there are a multitude of molecular mechanisms that could explain the observed decline in circulating FVIII levels, a leading theory holds that this loss is driven, at least in part, by activation of the UPR and cell stress/toxicity88 resulting from FVIII’s ectopic expression in hepatocytes, which are not the native site of synthesis.11,32,33,34,35,36 From the limited clinical data available on liver biopsies collected from HA subjects treated with AAV GT, no overt evidence of induction of the UPR has been seen.5 To address this important issue, we performed studies to understand the transgene-specific response to AAV5-mediated gene transfer by investigating differential gene expression between hLTEs transduced with AAV5-PGK-EGFP vs. hLTEs transduced with AAV5-PGK-lcoET3 to identify pathways specific to a bioengineered FVIII transgene. In contrast to current clinical vectors that employ a hepatocyte-specific promoter to drive FVIII expression, we used the constitutive PGK promoter to ensure any cell types within the hLTEs that were transduced by AAV5 would express the vector-encoded EGFP or lcoET3 transgene, thereby allowing us to determine which cell types may be at risk if FVIII expression exceeds the cell’s intrinsic processing/secretory threshold. Interestingly, mRNA for lcoET3 was present in much higher copies than that for EGFP despite the two groups of hLTEs having been transduced at the same vector MOI. While we do not have a definitive explanation for this finding, it is well known that different transgenes can yield varying levels of mRNA despite being driven by the same promoter due to inherent sequence-specific differences in mRNA stability and as a result of differing post-transcriptional processing. EGFP mRNA has been reported to have a fairly short half-life of just 1–4 h, which could provide a possible explanation for the lower levels seen for this mRNA compared with those for lcoET3. In addition, the lcoET3 transgene was carefully codon optimized based on transcripts known to be stable and expressed at high levels within the human liver, while the EGFP transgene was not. The lack of liver-specific codon optimization for EGFP may have resulted in the presence of codons that are “rare” in liver cells, thereby slowing translation, triggering translation-coupled decay, and reducing mRNA stability. Another factor that may contribute to the higher levels of lcoET3 mRNA compared with those for EGFP is that the lcoET3 cassette includes the polyadenylation signal (PAS) derived from the 3′-flanking region of the bovine growth hormone (BGH) gene. The BGH-PAS is widely used due to the higher levels of transgene expression it yields as a result of markedly increasing the stability of the encoded mRNA. As the EGFP cassette does not possess the BGH-PAS, this may also be a factor in the lower levels of EGFP mRNA we observed, despite similar vector copy numbers. Nevertheless, viability, metabolic activity, and drug-metabolizing capacity were all largely unaffected by AAV5 transduction, irrespective of the transgene encoded by the vector. However, in contrast to the results of the initial study in which EGFP expression was driven by the strong viral CMV promoter with reporter vectors, transduction of the hLTEs with an AAV5 vector in which EGFP (or lcoET3) was under the control of the endogenous PGK promoter induced only slight alterations in the levels of transaminases and albumin. This difference is likely due to the much less robust expression of the EGFP and lcoET3 transgenes afforded by the PGK promoter. Despite the use of this endogenous promoter to achieve far more physiologically relevant levels of transgene expression, a decrease in A1AT levels continued to be consistently observed in hLTEs transduced with either AAV5-PGK-EGFP or AAV5-PGK-lcoET3. Interestingly, comparing the functional impact of transducing the hLTEs with AAV5 encoding EGFP vs. lcoET3, no major differences were observed, with the exception of the hLTE group transduced with AAV5-PGK-lcoET3 consistently exhibiting elevated albumin levels. Transduction of the hLTEs with AAV5 encoding the F8 transgene resulted in upregulation of a larger number of genes than the identical AAV5 vector encoding EGFP, which demonstrates that there is certainly a transgene-specific component to the hLTE response to AAV transduction. MRE11A and ULK2 were upregulated in hLTEs transduced with either AAV5-PGK-EGFP or AAV5-PGK-lcoET3, but significance was only achieved in the hLTEs transduced with AAV5-PGK-lcoET3. The nuclease MRE11A has been shown to localize to AAV ITRs and prevent second strand synthesis.89 The more pronounced upregulation of this gene in the hLTEs transduced with AAV5-PGK-lcoET3 suggests a heightened response of the human liver to the F8 transgene. ULK2 is a marker for autophagy; as such, its upregulation in both the hLTEs transduced with AAV5-PGK-lcoET3 and the hLTE group transduced with AAV5-PGK-EGFP is consistent with a previous report that AAV induces autophagy.37 This cell response may actually be beneficial in the context of GT, since autophagy has been shown to positively impact the efficiency of the intracellular processing of AAV in hepatocytes.37 Of note, IL5 was significantly upregulated in the hLTE group transduced with AAV5-PGK-lcoET3, indicating heightened inflammation or cellular stress. Finally, comparing the DEGs in hLTEs transduced with AAV5-CMV-EGFP with those transduced with AAV5-PGK-EGFP to determine the impact of the type of promoter driving the transgene expression revealed that 172 genes were upregulated and 32 genes downregulated in hLTEs transduced with an AAV5 vector in which EGFP was driven by the CMV promoter, when compared with those transduced with an AAV5 vector in which EGFP was driven by the PGK promoter. The majority of these DEGs were assigned to pathways associated with systemic immune response, innate immunity, and toxicity, demonstrating a far more nefarious effect when a CMV promoter was used. In conclusion, we successfully used hLTEs as a model system to define the human liver’s response to two serotypes, AAV5 and AAV3b, effectively showcasing the capacity of the hLTEs to recapitulate the biology and function of native human liver and to reproduce some of the clinical observations reported in patients who received AAV GT. This platform also provided unique insights regarding the impact of AAV serotype and promotor selection, as well as the potential for integrations into genomic loci associated with oncogenesis and into the mitochondrial genome. In their current stage, hLTEs are not without limitations, as they lack vascularization, adaptive immune elements, and plasma proteins, each of which is likely important to fully reproduce the physiopathology of AAV transduction in vivo. In particular, with the model presented herein, AAV transduction was performed at the early stage of hLTE formation to ensure the viral vectors had equal probability of transducing all cell types in the organoid. The rationale for choosing to perform the transduction at this time is that the early stage of hLTE formation would more accurately mimic the conditions of in vivo AAV delivery and transduction. This is because the organizational structure of one-cell-thick hepatocytic cell plates covered by fenestrated endothelium lacking a solid basement membrane facilitates the direct contact with the sinusoidal blood flow, resulting in efficient exposure of endothelial cells and hepatocytes to the viral vectors. In addition, Kupffer cells (located within the lumen of the liver sinusoids) and stellate cells (situated between the liver sinusoids and hepatocytes) are both readily exposed to the AAV vectors, allowing their transduction, assuming that the AAV vector has tropism for those cells. Because the hLTEs used were composed of mature primary cells (unlike OTEs fabricated with pluripotent stem cells), the hLTEs do not need to undergo differentiation and maturation to liver-specific cells. Thus, we do not foresee that the maturation state of the hepatocytes or the other liver cells that comprise the hLTEs had an impact on levels of AAV transduction. Importantly, this study design favored access of the AAV vectors to all liver cells, rather than transducing a fully formed hLTE with mature cell-to-cell adhesions, in which the cells on the more external layers (i.e., endothelial and stellate cells) would be exposed to a significantly higher MOI than hepatocytes. However, it is possible that establishment of cell-to-cell interactions could affect the expression of specific AAV receptors, which could in turn impact cell transduction. It is important to note that each 3D model is different and results obtained can differ depending upon whether primary, pluripotent, or immortalized cells are used and which extracellular matrix is employed or even as a result of donor-to-donor and batch-to-batch variability. Thus, it is possible that the outcomes seen regarding liver function, inflammation, and even AAV integration events would be different if cell types other than the four mature primary liver cell types we employed were used, the process for manufacturing the hLTEs were different, and if the hLTE had been transduced at a later stage of maturation once cell-to-cell adhesions were fully established. Nevertheless, with the passage of the FDA Modernization Act 2.0, which has recognized the limitations of current animal models and paved the path for the development and use of novel, all-human systems, we believe the data presented herein have increased our understanding of AAV liver transduction and complement the existing models61 as a promising platform to bridge existing translation gaps and thereby provide critical insights to bolster preclinical investigation for developing safer and more effective liver-directed AAV gene therapies. Materials and methods Study design 3D aggregates (hLTEs at early stage of maturation) containing 4 different types of liver cells were transduced with AAV5-CMV-EGFP and AAV3b-CMV-EGFP, at an MOI of 105 vg/cell. The early stage of hLTE maturation process was chosen for transduction to ensure the viral vectors had equal probability of transducing all cell types in the organoid. While this MOI can be considered high, it is directly relevant from a clinical standpoint, as it approximates the theoretical blood vector maximum concentration and effective dose of vector “seen” by hepatocytes in the lowest dose cohort (1.2 × 10ˆ13 vg/kg) in AAV5-hF8-SQ clinical trials. Since both vectors were used at the same MOI, and we were looking for potential toxicity, it seemed most appropriate to use the highest theoretical dose “seen” by hepatocytes and not lower doses as have been employed in other ex vivo studies. All vectors employed in these studies were single stranded to maintain consistency, as the fVIII transgene (unlike fIX) is far too large to package within an AAV vector using self-complementary design. hLTEs were allowed to self-aggregate for 3 days followed by media change. Experimental overview and time points for sample collection to determine transduction efficiency and cell tropism and for genomic and transcriptomic analysis were defined prior to conducting the experiments and can be seen in Figure S1. The minimum number of organoids required for each experiment was pre-determined by trial experiments and power analysis using G∗Power software. hLTEs were separated into three groups: NT negative control (hLTE-NT) (N = 86; n = 75), AAV5 transduced (hLTE-AAV5) (N = 62; n = 58), and AAV3b transduced (hLTE-AAV3b) (N = 46; n = 25), where “N” denotes the number of different experiments and “n” denotes the number of hLTEs (technical replicates) within each experiment. When positive controls for liver damage were required, hLTE-NT were treated with 20 mM APAP (N = 6; n = 75) for 24 and 48 h. To track hLTE assembly and to measure GFP expression post-transduction, IncuCyte real-time imaging of hLTEs was performed by an investigator blinded to sample identity (days 1–7). On day 5, live/dead assays were performed to establish viability and RNA was extracted to determine whether transduction resulted in dysregulation of genes involved in inflammation and hepatotoxicity. Urea, albumin, A1AT, and transaminases (ALT and AST) were quantified by ELISA in media collected from hLTEs on days 6–10. On day 7, hLTEs were fixed to stain for GFP. Drug metabolism/toxicity in hLTEs was assessed by measuring the CyP450 3A4 (CYP3A4) activity on day 9. On day 10, live/dead assays were repeated to confirm hLTE viability throughout the study, and hLTEs were fixed for TEM and histological analyses, which were then performed in a blinded fashion. Since fibrosis has delayed onset, RNA was also extracted on day 10 to determine whether transduction resulted in dysregulation of genes involved in fibrosis. To determine the specific response of hLTEs to the F8 transgene, the 10-day experiment described above was repeated, but using an identical AAV5 vector encoding either EGFP or F8 under the transcriptional control of the human PGK promoter (N = 3; n = 576). Live/dead and ATP assays were performed on day 5 and day 10 to assess viability and metabolic activity, respectively. Media was collected from days 5–10 to measure the levels of biomarkers of liver function. CYP3A4 activity was measured on day 10 to assess the drug-metabolizing capacity of hLTEs following transduction. NT hLTEs treated with 20 mM APAP for 48 h were included as a positive control for liver damage in some experiments. On day 10, DNA, RNA, and protein were extracted from hLTEs to measure VCN, transgene mRNA expression, and transgene protein levels, respectively. Fabrication of hLTEs and transduction with AAV hLTEs were fabricated by self-aggregation as described previously.29 hLTEs were fabricated by self-aggregation as described previously.27 In brief, primary human hepatocytes (Lonza, Walkersville, MD, USA), hepatic stellate cells (Lonza, Walkersville, MD), Kupffer cells (BioIVT, Westbury, NY, USA), and liver-derived endothelial cells (Lonza, Walkersville, MD, USA) were thawed in a 37°C water bath followed by cell count and viability determination using a NucleoCounter NC-200 (ChemoMetec, Denmark). The different cell types were combined in a 50-mL tube at a concentration of 7.9 × 105 cells/mL, at physiologically relevant ratios (75% hepatocytes, 10% Kupffer, 10% stellate, and 5% endothelial cells) in organoid-forming media consisting of complete hepatocyte growth medium (HCM) (Lonza, Walkersville, MD, USA) with 10% FBS (R&D Systems, Minneapolis, MN, USA), gentamicin (50 μg/mL; Gibco, Waltham, MA, USA), and 10 μg/mL rat tail collagen 1 (Corning, NY, USA), a low-density hydrogel-like matrix, providing support for cell aggregation. Ultra-low attachment 96-well, with 79 microwells/well, Elplasia (96 × 79) plates were then primed with 100 μL of organoid-forming media and centrifuged at 1,700 × g for 5 min at room temperature (RT). To generate 1,000-cell hLTEs, Elplasia (96 × 79) plates were seeded with 100 μL of organoid-forming media containing 7.9 × 104 cells per well and centrifuged at 200 × g for 3 min to promote rapid cell aggregation. AAV viral stocks were then thawed on ice and added to the 3D cell aggregates at an MOI of 105 vg/cell. hLTEs were thoroughly washed on day 4 to remove residual AAV particles, and HCM with gentamicin was used for the remainder of the time in culture. The 2,000-cell hLTEs were fabricated in ultra-low-attachment 96 well plates (Corning, NY, USA) in 150 μL of organoid-forming media and centrifuged at 200 × g for 3 min to promote rapid cell aggregation. 2,000-cell hLTEs were transduced at the same MOI of 105 vg/cell and washed on day 3 to remove residual AAV particles. HCM with gentamicin was used for the remainder of the time in culture. AAV5-CMV-EGFP and AAV3b-CMV-EGFP were purchased from Vigene Biosciences Inc. (Rockville, MD, USA), while AAV5-PGK-FVIII (lcoET3)90 and AAV5-PGK-EGFP were produced by VectorBuilder, (Chicago, IL, USA). Although the vectors for the two major sets of experiments were purchased from two different vendors (Vigene vs. VectorBuilder), both companies use the same production protocol, i.e., helper-free transient triple transfection of HEK293T cells. Both companies employ SDS-PAGE to assess vector purity, and both have purity >80% as a release criteria. While the titers differed somewhat from vector to vector (as a function of both manufacturer and transgene, as the smaller EGFP transgene leads to more efficient packaging of vector genomes compared with the larger lcoET3 transgene), it is important to note that both companies perform ddPCR-based assays to the AAV ITR to quantitate titer in vector genomes. As such, the empty:full ratio (even if markedly different between vector preps) would not affect the experimental outcome, as all hLTEs were transduced at the same MOI, based on this droplet digital polymerase chain reaction (ddPCR)-titering, which would not detect empty capsids, which do not contain ITRs. Live/dead assay Viability of hLTEs post-AAV transduction was measured on days 5 and 10 using double-staining with calcein-AM (green) and ethidium homodimer-1 (red) (Invitrogen, Carlsbad, CA, USA) indicating live and dead cells, respectively. hLTEs were incubated for 30 min at RT in media with 2 μM calcein-AM and 4 μM ethidium homodimer-1. hLTEs were then washed twice with PBS, and imaged using a Leica TCS LSI Macro Confocal Microscope. To determine % viable area, images were analyzed using ImageJ. Histology hLTEs were collected on day 10, fixed for 1 h in 4% paraformaldehyde, and embedded in 2% agarose and paraffin. 5-μm sections were cut, baked for 1.5 h at 60°C, and deparaffinized. Hematoxylin and eosin (H&E), Masson trichrome, and Sirius red staining were performed according to standard protocols. Transmission electron microscopy On day 10, hLTEs were collected and fixed in 2.5% glutaraldehyde for 1 h followed by washing with PBS. They were processed for TEM and imaged with an FEI Tecnai BioTwin Transmission Electron Microscope (120 keV) by the Wake Forest University School of Medicine microscopy core. IncuCyte real-time imaging hLTEs were placed in an IncuCyte real-time imaging system (Sartorius Stedim Biotech GmbH, Göttingen, Germany) housed in a cell culture incubator. Phase-contrast and fluorescence images were automatically recorded at regular intervals from days 1 to 7. The negative control hLTE-NT group was used to set the threshold to exclude any minimal background fluorescence from the evaluation, and these settings were applied to all groups. Data on counts/well and fluorescence intensity were exported, and GFP expression post-AAV transduction in hLTEs was represented as green integrated intensity per well/green object count per well, i.e., per organoid (GCU × μm2). Immunohistochemistry IHC was used to identify transduced cells (GFP) and each different cell type within the hLTEs. The samples were split into two sets, due to the limited number of channels available for imaging. Set 1 was imaged for DAPI, GFP (FITC channel), hepatocytes (TxRed channel), and endothelial cells (Cy5 channel), and Set 2 was imaged for DAPI, GFP (FITC channel), stellate (TxRed channel), and Kupffer cells (Cy5 channel). Deparaffinized slides were treated with 0.3 M glycine in PBS for 30 min, followed by antigen retrieval with proteinase K for 5 min (Abcam, Cambridge, MA, USA), and blocked for 3 h with a solution of 0.2% Triton X-100, 1% DMSO, 1% BSA, and 1% serum in PBS. The slides were then incubated overnight with primary antibodies: Set 1—GFP (AAV transduction; ab13970, 1:200), CK8+18 (hepatocytes; ab17139, 1:75), and lymphatic vessel endothelial hyaluronan receptor-1 (LYVE-1; NB100-725, 1:200) and Set 2—GFP (AAV transduction; ab13970, 1:200), vimentin (stellate; ab8978 1:100), and CD68 (Kupffer; ab213363 1:100) in the same blocking buffer at 4°C in a humidified chamber. Antibodies whose catalog number begins with “ab” were purchased from Abcam (Cambridge, MA, USA), while those beginning with “NB” were purchased from Novus Biologicals (Centennial, CO, USA). Slides were washed the following day and incubated with secondary antibodies (ab150169 1:250, ab150114 1:500, and ab150075 1:500; Abcam, Cambridge, MA, USA) for 45 min at RT followed by counterstaining the nuclei with DAPI (1:1000). ProLong Gold (Thermo Fisher Scientific, Waltham, MA, USA) was used to mount the slides. After 24 h of curing at RT, slides were imaged using an Olympus BX63 fluorescence microscope equipped with X-Cite 120 LED Boost. A total of 15 independent sections (n = 15) were analyzed for AAV5, and n = 20 for AAV3b. Development of AI model and image analysis Digitized images (resolution of 0.22 μm/pixel) were uploaded to the Aiforia Cloud image processing and management platform (Aiforia Inc., Cambridge, MA, USA) for analysis with deep learning convolutional neural networks (CNNs) and supervised learning. AI model training A supervised CNN was trained on annotations from digitized hLTE samples to recognize areas of transduction and different cell types: transduction area designated as “Green,” hepatocytes and stellate cells as “Red,” and endothelial and Kupffer cells as “Magenta.” Both the Red & Magenta AI models were trained on the most diverse and representative 44 whole-slide images to create a generalizable AI model capable of accurately quantifying the percentage GFP/transduction-positive area (including total hLTE area, total area of different cell types [Magenta and Red], and area expressing GFP-marker [Green]). Diverse training data were included to capture the variability in image/staining quality. We also chose slides with known artifacts (lint, bubbles, etc.) and trained the AI model to consider these artifacts as background and exclude them from analysis. AI model design and ground-truth definitions The AI model consisted of multiple feature layers, with each feature layer containing unique classes that were annotated for CNN input training data. The AI model consisted of 3 feature layers: (1) total hLTE was annotated using semantic segmentation to distinguish the total hLTE area from background and artifacts, (2) total cell area (stained with either Magenta or Red) was annotated using semantic segmentation to distinguish total cell area from non-cell positive organoid, and (3) total area of Green-positive staining was also annotated using semantic segmentation to distinguish green-positive stained area from non-green positive cells. The total hLTE, Cells, and Green-positive layers were then merged into a chained analysis pipeline, where segmentation results from the first layer (total hLTE) were used as a clipping mask in the next layer (Cells), and so on, to quantify total Green-positive staining per Magenta- or Red--positive cell area within total organoid area. Additional AI model training parameters Individual CNNs were trained for each layer using the image augmentation parameters, perceptive view (field of view), and level complexity summarized in Tables S7 and S8. AI model performance verification Magenta AI model The total hLTE layer was trained for 7,500 iterations on approximately 0.761 mmˆ2, and the resulting AI model performed at a total verification error rate of approximately 0.27% (0.08% false-positive, 0.18% false-negative). The Magenta cells layer was trained for 10,000 iterations on 0.168 mmˆ2, and the resulting AI model performed at a total verification error rate of 0.20% (0.16% false-positive, 0.04% false-negative) compared to the input training annotations. Last, the Green-positive stain was trained on 0.154 mmˆ2, and the resulting AI model performed at a total verification error rate of 0.32% (0.27% false-positive, 0.04% false-negative) compared with the input training annotations. Red AI model Layers 1 and 3 (total hLTE and Green) were copied from the Magenta AI model, and Layer 2 was retrained on annotations optimized to detect total Red-cell-stained area, resulting in an AI model capable of detecting total hLTE area, total Red-cell-stained area, and total Green-cell-stained area. The Red cells layer was trained for 9,628 iterations on 0.344 mmˆ2 and resulted in 0.45% (0.41% false-positive, 0.04% false-negative) total verification error when compared to the input training annotations. Final trained AI models for both Magenta- and Red-stained cells had grayscale intensity activated for all three layers, and grayscale values for each channel were output as % max intensity in the analysis results. ELISA and chromogenic assays Culture media to measure albumin, urea, A1AT, and transaminases (ALT and AST) were collected on days 6–10, spun down, and immediately stored at −80°C. Urea was measured using QuantiChrom urea chromogenic assay kit (BioAssay Systems, Hayward, CA, USA), and the other biomarkers were measured using ELISAs according to the manufacturer’s instructions (Abcam, Cambridge, MA, USA). Samples were run in triplicate for A1AT and albumin, and in duplicate for urea, ALT, and AST. In FVIII studies, ELISAs and chromogenic assays were run in triplicate for each time point and each sample. Urea was used to normalize the values of other biomarkers. To quantify FVIII/lcoET3, hLTEs from day 10 were lysed and assayed using the VisuLize Factor VIII Antigen Plus Kit (Affinity Biologicals, Ancaster, ON, Canada) per the manufacturer’s instructions. Lysis was carried out using buffer PTR containing a protease inhibitor cocktail and 50× enhance solution (Abcam, Waltham, MA, USA). Samples were run in triplicate and at two dilutions with pooled samples of 96 hLTEs per group. Data are presented as IU/mL normalized to urea. Each replicate represents the FVIII level of 20 hLTEs. Cytochrome P450 assay To measure cytochrome P450 3A4 (CYP3A4) activity, hLTEs were incubated for 24 h with media containing rifampicin (25 μM), dexamethasone (50 μM), and phenobarbital (58 μg/mL) to induce enzymatic activity. The P450-Glo CYP3A4 Assay and Screening System was then used according to the manufacturer’s protocol to measure the conversion of a P450 3A4 substrate to a luminescent product (luciferin), which is directly proportional to the activity of CYP3A4. The raw result ([RLU], relative luminescent unit), after subtraction of the blank, was compared with a standard curve generated with luciferin to represent data as concentration of luciferin generated. Samples were run in triplicate. ATP assay Metabolic viability of hLTEs for the FVIII studies was measured by quantifying ATP levels with the CellTiter-Glo Luminescent Cell Viability Assay (Promega, Madison, WI). hLTEs were suspended in 100 μL of media, equilibrated to RT for 30 min, followed by addition of 100 μL of CellTiter-Glo reagent and 2 min of mixing on an orbital shaker, followed by an incubation for 10 min at RT. Luminescence (integration time, 0.25 s/well) was recorded using a GloMax Navigator Microplate Luminometer (Promega, Madison, WI). Luminescence from spent media, representing background, was subtracted from all samples, and the data were represented as RLU. RNA extraction and RT-qPCR RNA was extracted from the hLTEs (pooling hLTEs from 10 wells/experiments for each group—for the AAV3b-CMV-EGFP vs. AAV5-CMV-EGFP experiments, NT = 750 hLTEs, AAV3b-transduced = 250 hLTEs, and AAV5-transduced = 580 hLTEs; for the AAV5-PGK-EGFP vs. AAV5-lcoET3 experiments, 96 hLTEs were pooled per group, and the experiment was repeated three times with cells from 3 different human donors) on days 5 and day 10 using the AllPrep DNA/RNA/miRNA Universal Kit (QIAGEN, Valencia, CA, USA) according to the manufacturer’s protocol. The concentration of RNA was measured using a NanoDrop 2000 (Thermo Fisher Scientific, Wilmington, DE, USA), and the integrity of RNA was verified using an RNA 6000 Nano kit and a 2100 Bioanalyzer (Agilent, Santa Clara, CA, USA). An RT2 Profiler PCR Array kit and QuantStudio 3 Real-Time PCR System (Applied Biosystems, Waltham, MA) were used to determine whether transduction resulted in the dysregulation of 84 key genes involved in hepatotoxicity (PAHS-093Z), inflammation (PAHS-025Y), and fibrosis (PAHS-120Z) between the control group and the groups transduced with each serotype of AAV (Agilent, Santa Clara, CA, USA). Human GAPDH was used as the housekeeping gene for normalization. Data were analyzed on the RT2 Profiler Data Analysis online platform, and a Student t test was used to perform statistical analysis (α = 0.05). EGFP expression was also determined using RNA from day 10 which was converted to cDNA (200 ng per 20 μL reaction) using an Omniscript RT kit (QIAGEN, Valencia, CA, USA) followed by RT-qPCR using the PowerUp SYBR Master Mix (Applied Biosystems, Waltham, MA), primers specific for EGFP and human GAPDH (200 nm final concentration), and a QuantStudio 6 Flex Real-Time PCR (Applied Biosystems, Waltham, MA). The cycle conditions were as follows: 1 cycle of UDG activation at 50°C for 2 min, 1 cycle of inactivation at 95°C for 2 min, 40 cycles of denaturation at 95°C for 15 s and anneal/extend at 60°C for 1 min, and standard melt curve analysis. RT-qPCR was repeated 3 times. GFP primers Forward 5′- CCACCCTGACCTACGGCGTG-3′ Reverse 5′- CCTGGACGTAGCCTTCGGGC-3′ Human GAPDH primers Forward 5′-CACTGCTGGGGAGTCCCTGC-3′ Reverse 5′-GCACAGGGTACTTTATTGATGG-3′ NanoString nCounter and Rosalind data analysis Total RNA was extracted on day 5. Once quality was verified, 50 ng of total RNA was used to run the nCounter Gene Therapy Optimization Panel (NanoString Technologies), as per the manufacturer’s protocol. Gene expression data were analyzed using ROSALIND (https://rosalind.bio/), with a HyperScale architecture developed by ROSALIND, Inc. (San Diego, CA). Read distribution percentages, violin plots, identity heatmaps, and sample multi-dimensional scaling (MDS) plots were generated as part of the QC step. Normalization, fold changes, and p values were calculated using criteria provided by NanoString. ROSALIND follows the nCounter Advanced Analysis protocol of dividing counts within a lane by the geometric mean of the normalizer probes from the same lane. Housekeeping probes to be used for normalization are selected based on the geNorm algorithm as implemented in the NormqPCR R library.91 Fold changes and p values are calculated using the fast method as described in the nCounter Advanced Analysis 2.0 user manual. p value adjustment is performed using the Benjamini-Hochberg method of estimating false discovery rates . Clustering of genes for the final heatmap of DEGs was done using the PAM (Partitioning Around Medoids) method using the FPC R library92 that takes into consideration the direction and type of all signals pathway, the position, role and type of every gene. Hypergeometric distribution was used to analyze the enrichment of pathways, Gene Ontology, domain structure, and other ontologies. The fold change and p values were extracted and sorted into different themes in the Gene Therapy Optimization panel. Given that multiple hLTEs (96–750, depending on the specific experiment) in each experimental group were pooled for RNA extraction, the resultant number of “replicates” of each condition was artificially small. As such, we considered a raw p value of ≤0.065 and a fold change of 1.2 to be significant for identifying DEGs between the various experimental groups. Digital PCR for VCN Digital PCR (dPCR) to determine VCN was carried out using 5 ng of DNA and a QuantStudio Absolute Q dPCR System, using the manufacturer’s recommended cycle setting and primers specific for the AAV2 ITR (Absolute Q Viral Titer dPCR Assay, Assay ID: DTH49PU, Thermo Fisher Scientific, Waltham, MA) and the housekeeping gene RPLP0 (Assay ID: Hs02992885_s1, Thermo Fisher Scientific, Waltham, MA), which was used to normalize the data to VCN/cell. For each independent experiment, DNA was extracted from 96 hLTEs to perform dPCR. RT-digital PCR for transgene mRNA expression RNA from day 10 hLTEs was extracted as described above. One-step RT-dPCR was carried out using total RNA and a QuantStudio Absolute Q Digital PCR system (Thermo Fisher Scientific, Waltham, MA) with primers and probes specific for EGFP, F8 (lcoET3), and RPL13A (housekeeping gene). Data are presented as copies of target mRNA/1,000 copies of RPL13A. Total RNA was extracted from 96 hLTEs (n = 96) to perform dPCR for each independent experiment. eGFP primers and probe Forward: 5′-TAAACGGCCACAAGTTCAGC-3′ Reverse: 5′-GAACTTCAGGGTCAGCTTGC-3′ Probe: 5′-/56-FAM/TGGCATCGCCCTCGCCCTCG/36-TAMSp/. lcoET3 primers and probe Forward: 5′-CAGGCTTCTAGGCCCTACTC-3′ Reverse: 5′-GTGATGCTGCACCTTCCAAA-3′ Probe: 5′-/56-FAM/TGCCTAGGTTCAGCTCCCTGCTCC/36-TAMSp/. RPL13A primers and probe Assay ID: Hs04194366_g1, Thermo Fisher Scientific, Waltham, MA. Integration site analysis DNA was extracted from hLTEs 1 week post-transduction using the QIAamp DNA Micro kit (QIAGEN, Valencia, CA). DNA concentration was measured using a Qubit 3 Fluorometer (Invitrogen, Carlsbad, CA). Custom PacBio Integration Site Analysis was carried out by Azenta Life Sciences (South Plainfield, NJ, USA) to determine the loci and frequency of genomic integration post-AAV transduction, and the proximity of these ISs to genomic loci associated with oncogenesis was conducted using the COSMIC database for annotation. Library preparation and sequencing DNA from hLTEs was used to prepare ISA libraries using the IDT xGen enrichment. DNA of samples was mechanically fragmented, followed by pre-capture library preparation, and then hybridization capture with baits corresponding to the ITR in the viral vector. An Illumina Nextera XT library preparation kit was used to construct final Illumina sequencing ready libraries from captured fragments. Libraries were sequenced on a NovaSeq with a target depth of 40M read pairs per sample in 2 × 150-bp configuration. Data analysis Raw BCL files generated by the sequencer were converted to raw fastq files for each sample using bcl2fastq v.2.20. Fastq reads were trimmed with adapter sequences and aligned to the human reference genome using bwa v.0.7.4. The bam file was subsequently processed with Samtools v.0.1.16. Finally, ISs were identified and clustered by GENE-IS v.1.0. Statistical analysis GraphPad PRISM was used for statistical analysis, and specific tests used are indicated in the respective figure legend or “materials and methods” section, including sample size and replicates. Tests used included Student’s t test or analysis of variance (ANOVA) with post hoc Tukey’s, Dunnett’s, or Šidák multiple comparisons test, as appropriate, as well as mixed analysis for some experiments. A confidence interval of 95% was considered, and α was set to 0.05. Data are presented as mean ± standard error of mean (SEM). Data and code availability All data are available in the main text or the supplemental information. Acknowledgments This study was supported by National Institutes of Health, USA grant R01 HL166462 (G.A.-P. and C.D.P.) and Pfizer Global Medical Grants System, USA grant #61589649 (G.A.-P. and C.D.P.). Author contributions G.A.-P., C.D.P., and A.A. conceptualized the study; G.A.-P., C.D.P., R.M.R., S.J.W., C.B.D., H.T.S., C.E.B., and B.L. developed methodology; R.M.R., W.T.Z., G.A.-P., C.D.P., J.H.D., S.E.W., S.K.G., T.S., Y.Z., M.W., and S.K.G. performed experiments; G.A.-P. and C.D.P. secured funding; R.M.R., C.D.P., and G.A.-P. drafted the manuscript. Declaration of interests The authors declare that this study received partial funding from Pfizer Global Medical Grants System, grant #61589649. The funder was not involved in the study design, collection, analysis, interpretation of data, the writing of this article, or the decision to submit it for publication. C.B.D. and H.T.S. are co-founders of Expression Therapeutics, Inc. and own equity in the company. The terms of this arrangement have been reviewed and approved by Emory University in accordance with its conflict-of-interest policies. Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.omta.2026.201772. Supplemental information Document S1. Figures S1–S4 and Tables S7 and S8 Table S1. DEGs in AAV3b-CMV-eGFP vs. AAV5-CMV-eGFP Please see Excel spreadsheet entitled Table S1 with 3 different tabs, included as a supplemental information Table S2. Pathways of DEGs for AAV3b-CMV-eGFP vs. AAV5-CMV-eGFP Please see Excel spreadsheet entitled Table S2, included as a supplemental information Table S3. DEGs in AAV5-PGK-lcoET3 vs. AAV5-PGK-eGFP Please see Excel spreadsheet entitled Table S3 with 3 different tabs, included as a supplemental information Table S4. Pathways of DEGs in AAV5-PGK-lcoET3 vs. AAV5-PGK-eGFP Please see Excel spreadsheet entitled Table S4 included as a supplemental information Table S5. Pathways of DEGs in AAV5-CMV-eGFP vs. AAV5-PGK-eGFP Please see Excel spreadsheet entitled Table S5 included as a supplemental information Table S6. Pathways of DEGs in CMV vs. PGK promoter Please see Excel spreadsheet entitled Table S6 included as a supplemental information Document S2. Article plus supplemental information References Associated Data Supplementary Materials Document S1. Figures S1–S4 and Tables S7 and S8 Table S1. DEGs in AAV3b-CMV-eGFP vs. AAV5-CMV-eGFP Please see Excel spreadsheet entitled Table S1 with 3 different tabs, included as a supplemental information Table S2. Pathways of DEGs for AAV3b-CMV-eGFP vs. AAV5-CMV-eGFP Please see Excel spreadsheet entitled Table S2, included as a supplemental information Table S3. DEGs in AAV5-PGK-lcoET3 vs. AAV5-PGK-eGFP Please see Excel spreadsheet entitled Table S3 with 3 different tabs, included as a supplemental information Table S4. Pathways of DEGs in AAV5-PGK-lcoET3 vs. AAV5-PGK-eGFP Please see Excel spreadsheet entitled Table S4 included as a supplemental information Table S5. Pathways of DEGs in AAV5-CMV-eGFP vs. AAV5-PGK-eGFP Please see Excel spreadsheet entitled Table S5 included as a supplemental information Table S6. Pathways of DEGs in CMV vs. PGK promoter Please see Excel spreadsheet entitled Table S6 included as a supplemental information Document S2. Article plus supplemental information Data Availability Statement All data are available in the main text or the supplemental information.