AAV Gene Therapy Drug Development and Translation of Engineered Ocular and Neurotropic Capsids: A Systematic Review Using Natural Language Processing

ABSTRACT

Natural AAV serotypes often lack specificity and efficiency, leading to off‐target effects and a low therapeutic index. To overcome these limitations of naturally occurring serotypes, there has been a keen interest in the field to engineer novel capsids to enhance tissue and cell‐specific targeting, resulting in a high number of published literature reports over the past few years. To ensure a systematic review and illustrate advances in engineered capsids that enhance specificity and efficiency, we used Natural Language Processing with Linguamatics i2E to identify neurotropic and ocular AAV capsids tested in non‐human primates. By querying PubMed abstracts for specific mentions of AAVs, administration routes, and organ/tissue/species, we obtained 5907 hits, refined through an optimized process to 36 relevant and unique abstracts. Notable findings include numerous novel capsids summarized by route of administration: (1) systemic administration, targeting the central nervous system (e.g., AAV‐PHP.eB, AAV X1.1, and AAV.PAL2), (2) direct central nervous system injection (e.g., AAV2.Retro, Olig001, and AAV2.1A), and (3) ocular administration (e.g., AAV.44.9 (E531D), rAAV2tYF, and Anc80L65). Such engineered capsids exhibit enhanced tissue specificity, improved pharmacokinetics and pharmacodynamics, or reduced off‐target effects compared to the parent serotypes. Our study provides insight into state‐of‐the‐art translational and drug‐development considerations for engineered neurotropic and ocular capsids. We also highlight the effectiveness of Natural Language Processing and Large Language Models as tools in identifying and characterizing engineered neurotropic and ocular AAV capsids to summarize this rapidly growing class of drugs and area of therapeutics.

Study Highlights

• What Is The Current Knowledge on the Topic?

  • ○
    There is an emerging effort to engineer AAV capsids to improve specificity for the central nervous system (CNS) and the eye. Non‐human primates (NHPs) are commonly used to assess translational potential, yet evidence is scattered across studies, routes of administration (ROA), and naming conventions, making data synthesis and comparison difficult.
• What Question Did This Study Address?

  • ○
    We asked whether natural language processing (NLP) and large language models (LLMs) workflows could systematically identify and organize the published literature on engineered neuro‐ and ocular‐tropic AAVs in NHPs across administration routes (e.g., IV, intra‐CNS, and ocular), while capturing descriptors relevant for translation.
• What Does This Study Add to Our Knowledge?

  • ○
    The study uses a structured query with predefined dictionaries, in which it consolidates NHP data on engineered capsids into ROAs for neuro and ocular indications. The engineered capsids demonstrate enhanced tissue specificity, improved pharmacokinetics and pharmacodynamics, and reduced off‐target effects, and showcase the limitations of naturally occurring AAV serotypes. Such a workflow can improve reproducibility, reduce manual curation time, and support data‐driven decisions in vector engineering and translational pharmacology.
• How Might This Change Clinical Pharmacology or Translational Science?

  • ○
    By providing a summary of engineered capsids evaluated in NHPs, this work supports translation by informing capsid selection for tissue/organ, ROA, and frames NHP dose ranges to guide dose optimization. The standardized summaries help teams rapidly compare options and move more efficiently from literature to preclinical study design and early clinical planning for CNS and ocular programs.

1. Introduction

Gene therapeutics (GTx) employ vectors to deliver therapeutic genes into patients' cells and effectively repair, supplement, or counteract the effects of faulty genes or proteins responsible for disease. Promising gene delivery vectors, such as adeno‐associated virus (AAV), can carry replacement genes, RNA interference molecules, antisense oligonucleotides, and gene‐editing components, making them highly versatile cargo carriers [1]. Unlike small molecules and biologics that require repeated dosing, AAV GTx can potentially provide a long‐lasting (even lifelong) effect from a single treatment as it enables cells to continuously produce the required protein. AAV vectors are favored for their low immunogenicity, ability to target diverse tissues (tropism), and long‐term gene expression with minimal host genome integration.

AAV GTx have unique pharmacokinetics/pharmacodynamics (PK/PD) challenges due to complex processes they undergo within the body, from biodistribution, cellular entry, to transgene expression [2, 3, 4]. Targeting the CNS tissues is particularly challenging due to several barriers, such as blood–brain barrier (BBB) and blood cerebrospinal fluid barrier (BCSFB) that the capsids face following intravenous (i.v.) administration. The eye, while more accessible and relatively immune‐privileged, allows local gene delivery via intravitreal (IVT), subretinal (SR), or suprachoroidal routes. In practice, IVT is often preferred because it is less invasive, but the inner limiting membrane (ILM) may hinder vector movement to the outer retina, so transduction skews toward inner retinal cells (e.g., retinal ganglion cells). SR injection (SRI), while highly effective for targeted RPE/photoreceptor transduction, is invasive and typically limited to the bleb area, constraining geographic coverage [5].

In the brain as well as the eye, natural viral serotypes often lack the specificity and efficiency needed for effective and safe gene delivery, leading to the need for engineered capsids that can overcome these challenges [6]. Engineering AAV capsids with enhanced tropism can significantly improve biodistribution, allowing for targeted delivery to tissues. One way of doing so is by altering the surface proteins on AAV capsids to preferentially bind to receptors present on the target cells, thereby enhancing vector uptake and transgene expression in the desired tissues while minimizing off‐target effects. By refining the biodistribution profile of AAV vectors through capsid engineering, GTx can achieve precise delivery, reduce systemic exposure, and improve therapeutic outcomes. This targeted approach addresses the limitations in the therapeutic index of traditional AAVs, paving the way for more effective treatments, especially in tissues where specificity is crucial. This approach also provides an opportunity to engineer a capsid in a way to minimize immunogenicity, improving the safety and efficacy of this therapeutic approach [7].

The expanding body of literature in the current era of novel therapeutic modalities makes it increasingly challenging to keep track of relevant research and insights. Natural Language Processing (NLP) is a branch of artificial intelligence (AI) that deals with the interaction between computers and humans through natural language, with the ultimate goal of interpreting, comprehending, and processing human language in a meaningful manner [8]. It offers a powerful tool for harnessing this vast and growing body of scientific knowledge. NLP has historically been used in numerous disciplines, including biomedical research, to extract information from vast amounts of text data. This includes, but is not limited to, literature mining and data extraction, Electronic Health Records analysis, clinical trial matching, drug discovery and development, genomic sequence analysis, gene‐disease association, and others [8, 9, 10, 11, 12, 13].

The goal of this work was to use NLP to screen engineered AAV capsids targeting the CNS and eye, with a specific focus on non‐human primates (NHP), which are widely considered the most relevant species for human translation, to accelerate and enhance the translatability of GTx from preclinical models to human applications. Using this method allows us to systematically comb through vast amounts of research publications from PubMed and to extract data specifically related to engineered AAV capsids targeting the CNS/eye. Through an iterative process, we developed the NLP and Large Language Model (LLM) screening method to generate the most relevant results. Engineered capsids are then compared and sorted by transduction efficiency (e.g., the percentage of cells that have been successfully genetically modified by a viral vector) and tropism, promoter, species, dose levels, cell/organ specificity, etc. This system helps prioritize novel neurotropic and ocular capsids that have shown promising results in NHPs and may facilitate gene therapy drug development and translation.

2. Methods

2.1. NLP Software: Linguamatics I2E

The Linguamatics Interactive Information Extraction (i2e) NLP platform [i2e version 6.12R4] is a text mining tool that enables the efficient evidence curation from free‐text articles using NLP technologies [11, 12, 13]. Using the I2E Software, we designed a query script and created an algorithm to extract evidence related to engineered AAV GTx drug development and translation from PubMed abstracts. In this task, several biological ontologies or hand‐crafted dictionaries related to AAV, disease and disorder, organ, species, tissues, route of administration, and dosage were leveraged to identify biological terminologies from scientific literature. With the technology of named entity recognition in I2E, evidence with predefined patterns within a 150‐word or five‐sentence limit is retrieved. Specifically, we varied the extraction window size in increments from 100 to 200 words (by 25 words) within a span of five sentences to determine the segment length that best captured relevant scientific evidence while minimizing irrelevant context. For each configuration, we qualitatively assessed retrieval performance by manually reviewing a representative sample of extracted passages to evaluate their recall and precision in conveying the queried concepts. Through this iterative process, a window size of approximately 150 words (or five sentences) was found to provide the best balance between contextual completeness and specificity and was therefore adopted for the final extraction [12]. This word limit was used to ensure the information remains concise, focused, and digestible for efficient analysis or downstream processing. The structure and logic of the I2E query are summarized in Table S1, which includes representative query strings, ontologies, and inclusion/exclusion criteria. Each query was designed to capture abstracts containing specific combinations of AAV serotype, route of administration, and organ or tissue terms. Capsids were classified as “engineered” if the publication explicitly described capsid modifications (e.g., directed evolution, peptide insertion, rational design) or used a naming convention associated with engineered variants. Verification of target organ and tissue was performed by cross‐referencing extracted keywords with ontology‐based terms (e.g., NCI and MeSH dictionaries) and validated through manual review of the abstract context.

Several practical challenges emerged when balancing search sensitivity with relevance. Initial iterations often returned a number of abstracts referencing unrelated contexts (e.g., AAV as an autoimmune disease “ANCA‐associated vasculitis”) or mentions of natural serotypes without engineered variants. Using an iterative approach, we adjusted regular expressions, expanded the AAV‐specific dictionary, and maximized true positive capture while at the same time limiting redundant or incomplete records. Another key challenge involved the inconsistent naming of engineered AAV variants across publications. To address this, we used Genentech's internal GPT‐4O Large Language Model (LLM) to validate ambiguous cases and cross‐check context (e.g., determining whether a variant was an engineered or hybrid capsid). Manual curation was performed to correct misclassified entries. This iterative workflow combining automated NLP extraction with human and LLM‐assisted review proved important to maintain data quality and reproducibility.

The extracted data included: AAV serotypes, payload type, payload detail, disease and disorders, dose, route of administration, organ/tissue, and animal species. Our search query mandated the inclusion of AAV serotype, route of administration, and organ/tissue as essential fields. The additional categories were captured and populated by the NLP software when available. To enhance search accuracy, the mandated organ/tissue column was updated to include organ/tissue/species. For example, this change caused hits like “AAVPhP.eB was injected into cynomolgus macaques” to be included in the results rather than their previous exclusion as a result of not mentioning the organ/tissue. Thus, abstracts would be identified as a hit if they contained a mention of AAV, a route of administration, and a tissue or organ or species within the specified range.

Since there is no comprehensive public dictionary for all AAV serotypes, we created a custom dictionary. This dictionary was developed by identifying terms that contain “AAV” and capturing any word or phrase that immediately follows it (Table S1). The route of administration and organ/tissue/species columns were populated using a dictionary of terms relating to the aforementioned categories.

2.2. Data Processing

The previous data generation workflow yielded 5907 PubMed abstract hits that included mentions of an AAV serotype, a route of administration, and a tissue, organ, or species. These abstracts were then filtered by PubMed Identifier (PMID) to generate 5358 unique abstracts containing the necessary information (Figure 1). Filtration of NLP‐generated data, focused on mentions of NHP species, identified 458 NHP abstracts. This set was further refined to instances mentioning engineered AAVs, resulting in 37 abstracts. Finally, filtering by neurotropic and ocular indications narrowed the selection to 33 abstracts meeting the specified criteria. Mentions of AAVs in the literature are often inconsistently formatted, using variations such as dashes, spaces, or intervening phrases (e.g., ‘AAV‐PhP.eB’, ‘AAV PhP eB’, ‘AAV capsids such as PhP.eB’). This inconsistency reduces the effectiveness of the I2E software's pattern matching. Additionally, the lack of a comprehensive dictionary for all AAV serotypes (natural and variants) limits accurate identification. For example, the phrase “We identified a novel AAV‐based variant, XYZ, that shows an increase in neuronal transduction” would be excluded, as the term ‘AAV’ does not immediately precede ‘XYZ’. In such cases, the software defaults to identifying it as ‘AAV’. For this reason, NHP abstracts containing a non‐specified AAV serotype were manually screened to include abstracts that may have been excluded from the engineered AAV criteria. This process identified an additional 28 abstracts. These 28, in addition to the original 37 identified by the NLP software, were manually verified to ensure that the identified engineered AAV capsids were actively injected into NHPs, rather than just being mentioned. Further filtering ensured the NLP‐identified tissue or organ was genuinely neurotropic or ocular. The refined workflow identified 36 relevant articles (Figure 1). Within this “manual” workflow, LLMs provided significant help. Genentech's internal GPT‐4O model was employed to validate manual findings, distinguish engineered AAVs from naturally occurring variants, and confirm the use of the NHPs and mention of the injected organ. After relevant articles were identified, each was manually screened to confirm NLP‐generated information. Information summarizing the parent capsid, tissue tropism, cell specificity, routes of administration, and dose ranges used was collected per capsid.

FIGURE 1.

Summarized NLP workflow. NLP, natural language processing, LLM, large language model (1) NLP Criteria—The NLP software searches PubMed abstracts with the following criteria: [AAV] AND [Route of Administration] AND [Organ OR Tissue OR Species]. (2) This workflow produced 5907 hits, where each hit is an instance of the previous criteria within 150 words and/or 5 sentences. AAVs mentioned in the same abstract were typically grouped. (3) The data was further filtered into unique abstracts by removing duplicates, and then filtered by any mentions of an NHP species in its respective column, yielding 458 unique abstracts. (4a) The NLP‐generated data were filtered by researcher‐identified engineered AAV capsids. (5a) Remaining abstracts were finally filtered by NLP‐generated data in the tissue column, producing a total of 33 abstracts fitting the filtration criteria. (4b) The NHP abstracts were revisited, and the exact engineered AAV capsids mentioned in the abstracts were manually identified, producing 28 abstracts in addition to those from 4A. (5b) Manually identified abstracts and NLP‐generated data abstracts were combined into 65 total engineered capsid abstracts relating to NHPs. (6) The full‐text papers from all 65 engineered capsid abstracts were manually screened to identify papers that targeted the CNS and the eye. (7) These papers were finally filtered to ensure that they injected engineered capsids into NHPs and to remove cases where NHPs were simply mentioned. Genentech's internal LLM software aided in producing the data in steps 4B and 7 through the identification of engineered capsids and the validation of papers fitting all the criteria.

All of the capsids included in the NLP results are engineered capsids; however, in this work, mentions of “engineered” capsids will refer to those that are genetically modified with mutated sequences of a parent capsid, whether done through rational design, directed evolution, computer‐guided design, or a combination of these [7]. Hybrid AAV capsids refer to engineered capsids that contain a mixture of naturally occurring AAV serotypes, including those created through (1) transcapsidation, (2) adsorption modification, (3) mosaic, and (4) chimeric vectors [14]. Retrograde AAV capsids are engineered capsids that have the capacity to enable efficient retrograde transport of genetic material from axon terminals to neuronal cell bodies. This capability is useful for tracing neural circuits and targeting specific neuronal populations [15, 16]. Thus, in the NLP‐generated output, capsids will be classified as either (1) engineered, (2) hybrid, or (3) retrograde.

2.3. Application of Large Language Models

In the workflow, Genentech's internal GPT‐4O model, referred to as LLM, was used to facilitate the manual examination of papers that passed all filtration stages. For identifying engineered capsids, techniques such as recursion of thought, clarification prompting, and open‐ended prompting were employed to enhance the output's validity [10, 17]. At the end of the workflow to validate that all experimentally tested capsids injected into NHPs were captured, methods like chain of thought prompting were also utilized.

3. Results

The NLP workflow generated a total of 36 articles containing relevant information on engineered AAV capsids used to target the brain or eye in NHPs. These 36 articles contained 42 distinct engineered AAV capsids identified through the NLP search. Distinct capsids were divided into three categories by the route of administration: (1) Intravenously (IV) administered capsids that target the CNS, (2) capsids that were injected directly into the CNS, and (3) capsids that were injected into the eye (Figure 2). The full text of each literature source was reviewed to extract key information, including the number of unique capsids identified per category. These capsids were further classified into three categories: Engineered, Hybrid, and Retrograde capsids.

FIGURE 2.

NLP generated results.

3.1. Neurotropic Engineered AAV Capsids Following Intravenous Administration

The NLP workflow identified nine articles on intravenously (i.v.) delivered engineered AAV capsids targeting the CNS, revealing 13 novel engineered capsids. Examples include AAV.CAP‐B10 [18], AAV.Cap‐Mac [19], and AAV X1.1 [20] which decreased liver tropism and increased brain tropism compared to their parent capsids (Table 1). All capsids within this category used ubiquitous promoters (DNA sequence that promotes gene expression in an abundance of cell types and tissues). Across various studies, differential levels of brain transduction and peripheral organ detargeting (reducing AAV tropism to a specific tissue/cell type) were observed. These variances were influenced by the specific NHP species and the capsid modifications.

TABLE 1.

Summary of AAV capsids engineered for CNS tropism following systemic administration. All listed novel engineered capsids were delivered using ubiquitous and constitutive promoters ^a .

Novel AAV capsid [Parent capsid]	Capsid classification	Promoters, species	Tissue tropism and key findings	Cell specificity	Dose range	References
AAV‐PHP.B [AAV9]	Engineered	CB7, CAG, CBh, Rhesus Macaques, Marmosets, Cynomolgous Macaques	CNS tropism found in mice is not replicated in NHP; Little transgene expression [protein] in CNS, comparable with AAV9 [CBh, CB7, Marmosets, Rhesus Macaques]; Toxicity at higher doses [CBh, CB7, Cynomolgus Macaques, Rhesus Macaques]	None	2 × 1012−1 × 1014 vg/kg	[21, 22, 23]
AAV.PHP.eB [AAV9]	Engineered	CAG, Marmosets	Increased brain transgene RNA expression and decreased liver transgene mRNA expression via pooled delivery; no change in neuronal transgene expression [protein] with single variant injection [CAG, Marmosets]	None	7 × 1013 vg/kg	[18]
AAV.CAP‐B10 [AAV‐PhP.eB/AAV9]	Engineered	CAG, Marmosets	Strong, broad transgene expression [protein] in the brain with neuronal specificity, 17‐fold fewer transgene‐positive liver cells compared to AAV9 [CAG, Marmosets]	Increased neuronal specificity compared to AAV9: 4‐fold higher transgene‐positive neurons	7 × 1013 vg/kg	[18]
AAV.CAP‐B22 [AAV‐PhP.eB/AAV9]	Engineered	CAG, Marmosets	12‐fold increased transgene RNA levels in the brain compared to AAV9 [CAG, Marmosets]	Qualitative increase in transgene‐positive astrocytes compared to AAV9	7 × 1013 vg/kg
AAV X1.1 [AAVX1/AAV9]	Engineered	CAG, Marmosets, Rhesus Macaques	Enhanced CNS targeting over AAV9, with increased transgene mRNA transcripts in the cortex, cerebellum, hippocampus, and lateral geniculate nucleus via pooled delivery; reduced transgene DNA and RNA levels in peripheral tissues [CAG, Rhesus Macaques]	45× higher transgene‐positive neurons than AAV9, with 98% specificity in rhesus macaques; strong transduction of human brain endothelial cells in vitro	1–2.5 × 1013 vg/kg	[20]
AAV‐MaCPNS1 [AAV9]	Engineered	CAG, Marmosets, Rhesus Macaques	Overall neuron‐biased transduction profile; Strong CNS & PNS transduction with no decrease in liver transduction compared to AAV9; Strong transduction of neurons and astrocytes in the cortex and thalamus, but no transduction of oligodendrocytes and endothelial cells [CAG, Marmoset, Rhesus Macaques]	Neuron‐biased transduction in the PNS and CNS: 6‐fold increase in transgene‐positive cortical neuronal compared to AAV9; 37‐fold and 22‐fold increase in transgene‐positive neurons in the lumbar and thoracic DRG, respectively, compared to AAV9	2.5 × 1013 vg/kg	[24]
AAV‐MaCPNS2 [AAV9]	Engineered	CAG, Marmosets, Rhesus Macaques	Strong CNS & PNS transduction with no decrease in liver transduction compared to AAV9; Strong transduction of neurons and astrocytes and astrocytes in the cortex and thalamus, but not oligodendrocytes and endothelial cells [CAG, Marmoset, Rhesus Macaques]	44‐fold increase in transgene‐positive astrocytes and 11‐fold increase in transgene‐positive neurons in the cortex compared to AAV9; 77‐fold increase in transgene‐positive neurons with 94% specificity, 44‐fold increase in transgene‐positive neurons in lumbar and thoracic DRG, respectively, compared to AAV9	2.5 × 1013 vg/kg
AAV.CAP‐Mac [AAV9]	Engineered	CAG, Marmosets, Rhesus Macaques, African Green Monkeys	Generally increased brain targeting with increased transduction [protein] compared to AAV9, decreased viral DNA and RNA in the liver compared to AAV9 [CAG, Marmoset, Rhesus Macaque, African Green Monkey]; tropism variation by NHP species & age	Cell Tropism varied by NHP species & age	2–7.6 × 1013 vg/kg	[19]
AAVPAL1A‐D [AAV9]	Engineered	CBh, Cynomolgus Macaques	PAL variants 1A‐1C detargeted from DRG and significantly increased transgene mRNA expression in all brain lobes, thalamus, corpus callosum, and midbrain compared to AAV9; All 4 variants significantly decreased transgene mRNA expression and vector genomes present from most peripheral organs, including the liver and kidneys [CBh, Cynomolgus Macaques]	Not reported	1–3 × 1013 vg/kg	[25]
AAVPAL2 [AAV9]	Engineered	CBh, Cynomolgus Macaques	2.7–6‐fold increase in transgene mRNA expression in the cerebrum and corpus callosum, and decreased liver transduction compared to AAV9, increased transgene RNA levels in DRG compared to AAV9 [CBh, Cynomolgus Macaques]	Similar transduction profile as AAV9 (neuron leaning) with weaker astrocyte transduction in the Thalamus	3–6 × 1013 vg/kg

Abbreviation: DRG, Dorsal Root Ganglion.

^a Data within the tables reflects information extracted from the full text manuscripts identified from NLP search.

While the dose range varied slightly, most capsids were relatively well tolerated after systemic administration. However, one study reported acute toxicities, including one animal fatality, liver damage, and abnormal coagulation, following high doses (≥ 2 × 10¹³ vg/kg) of AAV.PHP.B, leading to early study termination [21]. The lack of PHP.B (an AAV9 variant) BBB‐penetrance beyond certain strains of mice likely precludes clinical advancement of this capsid, at least for CNS‐targeting diseases [22, 23]. However, these adverse events underscore the need for a better understanding of safety signals (Section 4.2).

3.2. Neurotropic Engineered AAV Capsids Following Direct Administration to the CNS

Our NLP search results identified 13 articles featuring 9 CNS‐administered, engineered AAV capsids, summarized in Table 2. The search revealed various routes of administration, with 4 vectors injected via intra‐CSF delivery and 5 vectors injected into the brain parenchyma. Contrary to the search results from systemically administered vectors, which contained solely those that were genetically engineered, our search for CNS‐administered vectors also identified hybrid and retrograde capsids. Specifically, the search yielded 3 engineered capsids, 4 hybrid capsids, and 2 retrograde capsids.

TABLE 2.

Summary of AAV capsids engineered for CNS tropism following the intra‐CNS route of administration ^a .

Novel AAV capsid [Parent capsid]	Capsid classification	Promoters, species	Brain tropism and key findings	Cell specificity	Route of administration	Dose range	References
AAV2.Retro [AAV2]	Retrograde	CAG, CMV, hSyn, Rhesus Macaques, Cynomolgus Macaques	Higher transgene expression [protein] in extra striatal regions than AAV2 [CMV, Rhesus Macaques], more retrograde labeling with constitutive promoter than neuron‐specific promoter [CAG, hSyn, Rhesus Macaques]; retrograde transduction of projections into the striatum [CMV, hSyn, Rhesus Macaques]	Most transgene‐positive cells were neurons	Intrastriatal	5.4 × 109–6.6 × 1011 vg	[16, 26, 27, 28]
AAV‐PhP.eB [AAV‐PhP.B/AAV9]	Engineered	CAG, Rhesus Macaques	Increased transgene expression levels broadly in the brain compared to PhP.B [protein], especially in deeper brain tissues such as putamen, globus pallidus, caudate nucleus; weak transgene expression [protein] in liver cells [CAG, Rhesus Macaque]	Putamen: 80%–90% of transgene‐positive cells were neurons and oligodendrocytes Hippocampus: Neurons	IT	1.5 × 1011 vg/kg	[29]
		Promoter not reported, Rhesus Macaques	22.2% of radiolabeled capsid found in the nervous system, with ~12% distributing into the liver [Rhesus Macaque]	Not reported	ICM	2.5 × 1012 gc	[30]
AAV2.1A [90% AAV2 & 10% AAV1]	Hybrid—mosaic	CMV, hSyn, Rhesus Macaques, Cynomolgus Macaques, Japanese Macaques	Increased localized transgene expression intensity & number compared to AAV2 [CMV, hSyn, Rhesus Macaques, Cynomolgus Macaques, Japanese Macaques]	Increase number of transgene‐positive neurons with similar neuron specificity to AAV2 using ubiquitous promoter, higher neuronal specificity than AAV9	Cortical injections	5 × 109–5 × 1010 gc	[31]
AAV2.1B [50% AAV2 & 50% AAV1]			Increased localized transgene expression intensity & number compared to AAV2; transduction of both neurons and astrocytes [CMV, hSyn, Rhesus Macaques, Cynomolgus Macaques, Japanese Macaques]	Increased number of transgene‐positive astrocytes with ubiquitous promoter compared to AAV2, similar glial transduction to AAV1
AAV‐PhP.B [AAV9]	Engineered	CAG, Rhesus Macaques	Weaker transgene protein expression levels broadly in the brain compared to PhP.eB, especially in deeper brain tissues (putamen, globus pallidus, caudate nucleus); weak transgene expression [protein] in the liver [CAG, Rhesus Macaques]	Putamen—Neuronal + Oligodendroglial Hippocampus—Neuronal	IT	1.5 × 1011 vg/kg	[29]
		CAG, Rhesus Macaques	High transduction [protein] in all cortical layers, weak in putamen, other subcortical regions, transduction of neurons, astrocytes, oligodendrocytes [CAG, Rhesus Macaques]	None	ICM	1 × 1012 vg/kg	[32]
		SYN1, Rhesus Macaques	Widespread transgene‐positive neurons across various brain regions, including throughout the cortex and in subcortical structures [Syn1, Rhesus Macaques]	Neurons (Syn1 promoter)	ICV	1.87–3.95 × 1012 gc/kg	[33]
AAV‐F [AAV9]	Engineered	CBA, Cynomolgus Macaques	Transduction [protein] in the cerebellum, brain stem, and hippocampus; Decrease in vg/ug of DNA in the spinal cord and in peripheral organs compared to AAV9 [CBA, Cynomolgus Macaques]	Slightly more transgene‐positive neurons in the spinal cord than AAV9	IT	4.17–4.47 × 1012 vg	[34]
Olig001 [AAV 1, 2, 6, 8, 9]	Hybrid—chimeric	CBh, rhesus macaques, Cynomolgus Macaques	Strong, widespread transgene‐positive oligodendrocytes [protein] in the brain [CBh, Rhesus Macaques, Cynomolgus Macaques]	Oligodendrocytes (90%–94%) in striatum & corpus callosum [CBh, Rhesus, Macaques]	Intraputaminal	3–3.75 × 1011 vg	[35, 36]
					Intrastriatal
scAAV2.5 [AAV1 & AAV]	Hybrid—Chimieric	Promoter not reported, Cynomolgus Macaques	Similar transduction to AAV9 with 100‐fold lower transduction in the spleen [vector DNA/diploid NHP genome]: Low levels of transgene‐positive cells [protein] in widespread brain and spinal cord regions [Cynomolgus Macaques]	Not reported	ICM	2 × 1012 vg	[37]
AAVDJ8R [AAVDJ8]	Retrograde	CamKIIa, CAG, hSyn, Cynomolgus Macaques	High number of transgene‐positive neurons [protein] retrograde labeled in the cortex to striatum pathway [CamKIIa, Cynomolgus Macaques]	Neurons	Intrastriatal	2 × 1011 vg	[15]

Abbreviations: ICM, Intra‐ cisterna magna; ICV, Intracerebroventricular; IT, Intrathecal.

^a Data within the tables reflects information extracted from the full‐text manuscripts of associated articles.

The five capsids were administered via intraparenchymal injections, typically targeting striatal structures, except for AAV2.1A, which was injected into various cortical locations. Among the four capsids injected via intra‐CSF methods, at least one novel engineered capsid was identified for each of the commonly used intra‐CSF routes: intrathecal (IT), intracisterna magna (ICM), and intracerebroventricular (ICV). To note, these capsids generally required a lower vector dose compared to intravenously injected AAVs, likely due to the more targeted nature of the injection sites.

3.3. Ocular‐Tropic AAV Capsids

For over 20 years, AAV gene therapy has been explored for ocular indications. During this time, scientists have continuously worked to enhance tissue specificity and safety profiles [38]. It is unsurprising that our NLP search results yielded numerous unique capsids in this category. We identified 14 articles reporting on engineered ocular capsids delivered to NHPs. From these articles, we discovered 20 novel engineered capsids fitting the criteria: 16 were genetically engineered, 3 were hybrids of two naturally occurring serotypes, and 1 was an engineered version of a hybrid vector (Table 3).

TABLE 3.

Summary of engineered AAV capsids for ocular indications ^a .

Novel AAV capsid [Parent capsid]	Capsid classification	Promoters, species	Eye tropism & key takeaways	Cell specificity	Route of administration	Dose range	References
AAV.SPR/AAV44.9 (E531D) [AAV44.9]	Engineered	hGRK1, Cynomolgus Macaques	Highly efficient localized photoreceptor transduction with some lateral spread, transgene‐positive para and perifoveal cones [protein] after extrafoveal SRI, superior overall performance compared to AAV44.9 [hGRK1, Cynomolgus Macaques]	Highly specific transgene‐positive cell expression in photoreceptors	SRI (peripheral & extrafoveal)	9 × 1010 vg	[39]
rAAV2tYF [AAV2]	Engineered	CB, Cynomolgus Macaques	Well tolerated, high vector DNA levels in anterior segment, vitreous, and retina with limited peripheral transduction; Transduction [protein] in ganglion cell layer, inner nuclear layer, inner plexiform layer [CB, Cynomolgus Macaques]	None	IVT	4 × 1010–4 × 1011 vg/eye	[40]
		PR1.7, Cynomolgus Macaques	Vector DNA primarily in the retina, with some DNA in the vitreous, and less DNA in the lens; potential concern of severe ocular inflammation which led to euthanasia of ⅙ animals [PR1.7, Cynomolgus Macaques]	Not reported	SRI	1.2 × 1011–1.2 × 1012 vg/eye	[40]
		CMV, Cynomolgus Macaques	Removal of empty capsids to decrease total capsid dose resulted in increased transduction [protein] and decreased inflammation [CMV, Cynomolgus Macaques]	Not reported	IVT	3.4 × 109–7.6 × 1011 capsids/eye	[41]
AAVP2‐V1 [AAV2]	Engineered	CBA, Cynomolgus Macaques	Increases fluorescent signal intensity 2‐fold and 1.5‐fold increase in transduction positive area in foveal ring [protein]; transgene‐positive retinal ganglion cells, photoreceptors, axons innervating optic nerve [CBA, Cynomolgus Macaques]; enhanced ability to avoid Anti‐AAV2 nAbs in vitro [human vitreous samples]	Transgene expression is primarily limited to foveal ring RGCs and photoreceptors	IVT	1 × 1010–1 × 1011 vg/eye	[42]
AAVP2‐V2 [AAV2]		CBA, Cynomolgus Macaques	Tropism largely resembled that of AAV2, with increased transgene‐positive cell number and area in foveal ring; High dose—3‐fold higher fluorescent signal intensity & 4‐fold higher transduction area [protein] than AAV2 [CBA, Cynomolgus Macaques]	Low Dose—primarily transgene‐positive retinal ganglion cells in foveal ring High dose—transduction primarily confined to the retinal ganglion cell layer adjacent to foveal pit within the foveal ring
AAVP2‐V3 [AAV2]		CBA, Cynomolgus Macaques	Tropism largely resembled that of AAV2, with increased transgene‐positive cell number and area in foveal ring; 6‐fold higher transduction area [protein] at high dose than AAV2 [CBA, Cynomolgus Macaques]
AAVDGEDF [AAV2]		CBA, Cynomolgus Macaques	Tropism largely resembled that of AAV2, with increased transgene‐positive cell number and area in foveal ring; 5× fluorescent signal intensity at high dose [CBA, Cynomolgus Macaques]
AAV2.NN [AAV2]	Engineered	CMV, Cynomolgus Macaques	NHP: Penetrance into all retinal layers, transduction in the fovea [protein]; Human Retinal Explant: broad transduction across the explant with most prominent transduction in photoreceptors	Average of 54.4% transgene‐positive photoreceptors in the fovea centralis in NHPs	IVT, Human Retinal Explant	1 × 1012 vg/eye	[39]
AAV2.GL [AAV2]	Engineered
AAV2/8 (Y733F) [AAV2]	Hybrid Engineered	CAG, Rhesus Macaque Explants	Transgene expression [protein] primarily in the periphery of fovea explants, with higher transduction in outer nuclear layer from foveal explants	Primarily transgene‐positive cones	NHP Retinal Explants	1 × 1010 vg	[43]
		CAG, Human	Transduction [protein] primarily in inner nuclear layer	Transgene‐positive cells primarily within inner nuclear layer	Human Retinal Explant
AAV2/2 (quadY‐F) [AAV2]	Engineered	CAG, Rhesus Macaque Explants	Transgene expression [protein] primarily in the periphery of fovea explants, with higher transduction in outer nuclear layer from foveal explants	Primarily transgene‐positive cones	NHP Retinal Explants	1 × 1010 vg
		CAG, Human	Extensive transduction [protein] in all retinal layers	None	Human Retinal Explant
rAAV2(4pMut)ΔHS [AAV2]	Engineered	CBA, hGRK1, Cynomolgus Macaques	Strong localized transgene‐positive photoreceptors with a degree of lateral spread [CBA, hGRK1, Cynomolgus Macaques]; Substantial transgene expression [protein] in RPE, specifically in the macula [hGRK1, Cynomolgus Macaques]	Robust photoreceptor transduction with stronger transduction in cones than in rods	SRI (submacular, peripheral, extrafoveal)	9 × 1010–1.2 × 1011 vg	[44]
AAV2 Tyrosine 4yF [AAV2]	Engineered	CMV, Rhesus Macaques	This work elucidates the effect of nAbs on natural and mutant capsid transduction		IVT	Not reported	[45]
AAV2.7m8 [AAV2]	Engineered	CAG, Rhesus Macaques	Transgene expression [protein] primarily in the periphery of fovea explant, with higher transduction in the outer nuclear layer & centrally towards foveola	Primarily transgene‐positive cones	NHP Retinal Explants	1 × 1010 vg	[43]
				None	Human Retinal Explant
		CAG, Human	Extensive transgene expression in all retinal layers
		CMV, Rhesus Macaques	This paper elucidates the effect of nAbs on natural and mutant capsid transduction		IVT	Not reported	[45]
		CBA, Cynomolgus Macaques	Strong localized transduction with some vector spread; transgene‐positive retinal ganglion cells, spread to inner nuclear layer with increased dose; transduction of axons in optic nerve at lowest dose & increasing with dose, no Anterior Segment transduction [CBA, Cynomolgus Macaques]	Transgene‐positive cells in RPE and rod photoreceptors, with cone transduction at higher doses	SRI	1 × 1010–1 × 1012 vg/eye	[46]
			Transduction of both the anterior segment and posterior pole of the eye at doses ≥ 1 × 1010 vg, with deeper penetration of the posterior pole at higher doses; transgene‐positive retinal ganglion cells, Müller cells, bipolar cells, & axons of optic nerve [CBA, Cynomolgus Macaques]	None	IVT	1 × 108–1 × 1012 vg/eye
AAV8BP2 [AAV8]	Engineered	CBA, Cynomolgus Macaques	Transgene expression [protein] in RPE and cones, ganglion cells in the optic nerve; spread to inner nuclear layer cells with increased dose, weak transduction of bipolar cells at higher dose [CBA, Cynomolgus Macaques]	None	SRI	1 × 1010–1 × 1012 vg/eye	[46]
			No transduction [protein] in retina or anterior segment at lower doses; Transduction [protein] in inner nuclear layer, ganglion layer, optic nerve, ciliary body, and cornea at higher doses; Few transgene‐positive bipolar cells; Limited transduction deeper retinal regions [CBA, Cynomolgus Macaques]	Transduction of retinal ganglion cells & optic nerve, limited transduction of deeper retinal cells	IVT	1 × 109–1 × 1012 vg/eye
scAAV2(Y444 + 500 + 730F) [AAV2]	Engineered	CBA, Cynomolgus Macaques	Internal Limiting Membrane peeling prior to AAV administration resulted in increased number of transgene‐positive cells without retinal damage [CBA, Cynomolgus Macaques]	Primarily transgene‐positive Müller cells	IVT	9.5 × 1011 vg	[47]
Anc80L65 [Anc80]	Engineered	CMV, Rhesus Macaques	Anc80L65 had faster onset of strong expression compared to AAVs 2, 5, 8, 9; Increased cone transduction compared to AAV8; Strong transduction intensity [protein] in the outer nuclear layer and RPE in NHP [CMV, Rhesus Macaques]	Stronger transduction of rods than cones	SRI	1 × 1010 gc	[48]
AAV2/4 [AAV2]	Hybrid	TetOn, Macaques	AAV 4 capsid proteins can be detected in the Outer Plexiform Layer up to 2.5 years after SRI; Vector particles only found in cells not displaying transgene expression [TetOn, Macaques]	None	SRI	1 × 1010 vg	[49]
rAAV2/6 [AAV6]	Hybrid	Promoter not reported; African Green Monkeys	Intracameral injection caused self‐resolving transient inflammation that did not alter corneal or retinal thickness; Transgene expression [protein] in Iridocorneal angle, trabecular meshwork, ciliary body and iris, none in cornea or retina with high dose, trace transduction found in Iris for low dose	Transduction of cells surrounding iridocorneal angle and iris	Intracameral	1 × 1011–1 × 1012 vg/eye	[50]
rAAV2/9 [AAV9]	Hybrid	Promoter not reported; African Green Monkeys	Intracameral injection caused self‐resolving transient inflammation that did not alter corneal or retinal thickness; Transgene expression [protein] in Iridocorneal angle, trabecular meshwork, ciliary body and iris, none in cornea or retina with high dose, trace transduction in Iris for low dose
rAAV2/2 (MAX) [AAV2]	Engineered	Promoter not reported; African Green Monkeys	Intracameral injection caused self‐resolving transient inflammation that did not alter corneal or retinal thickness. Highest number of transgene‐positive cells in iridocorneal angle of vectors tested; transgene Iridocorneal angle, trabecular meshwork, ciliary body and iris; trace transduction in Iris for low dose

Abbreviations: IVT, intravitreal; RPE, retinal pigment epithelium; SRI, subretinal injection.

^a Data within the tables reflect information extracted from the full text manuscripts of associated articles.

Of the 20 capsids identified, 7 were tested via subretinal injection, 11 via intravitreal injection (IVT), 3 were delivered via intracameral injection, and 5 were tested in vitro in NHP or human retinal explants. Route of administration plays a significant role for ocular indications, as there can be limited distribution to other regions of the eye depending on the injection site. While lower doses were typically used for these capsids, several groups tested higher doses comparable to those in the direct CNS‐injected group to determine the safety profile of the tested capsids [40, 46, 50].

4. Discussion

The field of AAV gene therapy is rapidly expanding, with scientists continuously publishing new insights into engineered AAV capsids. These insights cover a range of topics, including immunogenicity management, species translatability, and enhancing various aspects of natural AAV serotype performance. The use of NLP software and LLMs has played a crucial role in aggregating this extensive body of literature and streamlining the search process.

Our findings are, however, constrained by the (1) literature available at the time of this research and (2) database used. Since the time of the search, additional studies have identified new engineered capsids potentially offering substantial advancements in the field [51]. A recent example is VCAP‐102 [52]. This capsid demonstrates CNS targeting capabilities across multiple species, including cynomolgus macaques, African green monkeys, marmosets, and mice. VCAP‐102 achieves transgene expression levels that are 20‐ to 400‐fold higher than those of AAV9 across various brain regions. This enhanced CNS tropism is attributed to its direct interaction with alkaline phosphatase (ALPL), a conserved membrane‐associated protein present on brain vasculature [52]. Another limitation of our work may also be the databases used. Our NLP software searched through PubMed due to the number of journals cited within PubMed and the availability of abstracts compared to full‐text manuscripts. Work published in journals that are not represented on PubMed was likely missed.

As mentioned, Genentech's internal GPT‐4O model was used within this workflow to support manual filtration of articles. Namely, LLMs coupled with manual filtration were used to confirm the identity of engineered capsids and to validate key information extracted from articles. While GPT‐4O was largely effective in summarizing the key findings and identifying whether a capsid was engineered or not, it was less effective in extracting engineered capsids from the articles, and iterative manual curation was used throughout the process.

The upcoming sections will focus on the NLP results and delve into critical aspects of engineered capsids that enhance AAV performance, focusing on tissue tropism, biodistribution, species translatability, and immunogenicity. These topics are highlighted by our NLP‐driven research, and we will discuss the associated challenges and their implications for advancing drug development for AAV GTx.

4.1. Engineered Capsid Tropism and Biodistribution

4.1.1. Tissue Tropism

A primary objective in capsid engineering is the development of strategies to selectively target specific tissues or even distinct cell types. However, while targeting a particular tissue and avoiding others presents a challenge, pinpointing a specific cell type within that tissue can be even more difficult. Our findings indicate that AAV transduction strength and biodistribution are influenced by various factors. These include the route of administration, capsid characteristics (such as tissue tropism and detargeting capabilities), promoter specificity, neutralization by anti‐AAV or transgene product antibodies, and receptor expression levels across different species. The selection of the administration route requires a careful balance between optimizing tissue targeting, maximizing the therapeutic index, and ensuring patient acceptability [53]. Interactions between the viral capsid and the promoter can significantly affect both the localization and efficiency of transduction. For instance, a study comparing vectors with either a full‐length Chicken β‐Actin (CBA) promoter or a truncated CBA hybrid (CBh) promoter in rats illustrates how promoter modifications impact viral vector performance. Using an AAV2 capsid, both vectors showed similar transduction levels and neuronal specificity. However, with an AAV9 capsid, the truncated CBh promoter caused a 42% decrease in neuronal transduction and a 2% increase in oligodendrocyte transduction compared to its AAV9, CBA counterpart [54]. In another example, examining the Frontal Eye Field (FEF) of non‐human primates, researchers compared the constitutive CAG promoter to the neuron‐specific hSYN promoter. Using rAAV2.retro, this group observed that the two promoters provided vastly different transduction patterns, with only the vector using the CAG promoter being able to retrogradely transduce the ipsilateral claustrum and the contralateral FEF [26]. These findings highlight the critical role of promoter‐capsid interactions in optimizing gene delivery, underscoring the need for precise promoter selection to enhance gene therapy efficacy. Other aspects influencing tissue tropism will be covered in the following sections.

4.1.1.1. Biodistribution by IV Administration Into the CNS

The effectiveness of intravenous delivery is contingent upon the ability of AAV serotypes to bypass the blood–brain barrier (BBB). A study assessed the biodistribution of natural AAV serotypes 1–9 after intravenous injection in mice. They found that AAV9, and to a lesser extent AAV8, were the only capsids capable of transducing the brain, while other serotypes primarily targeted major peripheral organs [55]. AAV9 is commonly chosen as a parent vector for capsid engineering targeting various tissues when delivering AAV intravenously (most comparisons in Table 1 are made to AAV9).

Some capsids identified from our NLP findings demonstrated their enhanced ability to cross the BBB and transduce the brain, while also detargeting peripheral organs [18, 19, 20]. One example identified in our results, AAV.CAP‐Mac is an AAV9‐based vector that shows a ninefold increase in DNA delivered to the brain and a concurrent 50% decrease in liver DNA expression compared to AAV9 in newborn rhesus macaques [19]. Other unique capsids like AAVPAL2 demonstrated increased transgene mRNA expression (HA‐tagged human frataxin) by 4–6 fold in the cerebrum, increased mRNA expression by 2.7‐fold in the corpus callosum, and decreased liver mRNA expression by 50% as well [25]. AAVPAL2 was also found to increase transgene mRNA expression in the neuroretina compared to AAV9, thus providing therapeutic potential for diseases affecting both the brain and the retina [25]. Engineered capsids like these allow efficient target tissue transduction following systemic delivery, while minimizing off‐target effects and safety concerns.

4.1.1.2. Biodistribution Following Direct CNS Administration

While it may be thought that various types of intra‐CSF injections typically lead to minimal peripheral exposure, some papers suggest that this is not always the case. Due to the clinical relevance of intrathecal (IT) and intracisterna magna (ICM) injections, Rosenberg et al. [30] tracked the biodistribution of capsids via Positron Emission Tomography (PET) scanning after intra CSF administration in NHPs. Capsids were tested across doses, capsid types, routes of administration, and vector immune status to determine how each of these factors could contribute to biodistribution. Observing native (AAV9, AAVrh.10, and AAVrh91) and engineered capsids (AAVPhP.eB and AAVhu68), their findings suggest that across capsids, a large percentage of the administered dose was found to spread to peripheral organs. Of the peripheral organs tested, the liver typically showed the highest percent biodistribution with both ICM and IT injections [30, 56].

Although intra‐CSF administration has more peripheral exposure than initially thought, this route of administration can still provide benefit through the strength of transduction. For example, AAV.PhP.B. is an AAV9‐based engineered capsid shown to increase CNS transduction compared to AAV9 after IV administration in mice. This was not evident in the NHP species [22, 23], indicating a lack of species translation when it comes to transduction. Liguore et al. found that AAVPhP.B, whose transduction does not translate to higher species via IV administration, does show CNS transduction following direct administration routes in NHPs. Specifically, PhP.B‐mediated CNS transduction (vg/cell) was significantly higher in the frontal, motor, temporal, and cerebellar cortices following intra‐cisterna magna (ICM) injection compared to carotid artery injection. Similarly, eGFP protein expression was notably higher in the orbitofrontal and dorsal premotor cortices in animals injected via ICM compared to those receiving carotid artery injections [32], indicating that direct administration may provide higher transduction for certain capsids.

Intraparenchymal injection is thought to be an ideal route of administration for targeting unimodal CNS indications, as it directly targets the tissue, has minimal peripheral exposure, and allows for the usage of lower doses [53, 57, 58]. While there are several clinical trials currently ongoing with this route of administration [59], several challenges persist depending on the indication. Some of the inherent challenges caused by intraparenchymal injection are the invasive nature and safety concerns, as it carries the risk of causing intracerebral hemorrhages and other surgical complications [58]. Additionally, due to the localized nature of this administration method, intraparenchymal injections typically show limited vector distribution, which may prevent multifocal brain indications from effectively being treated through this route. However, one intraparenchymally injected capsid AAVDJ8R, a variant of the engineered AAVDJ8 parent capsid, was shown to efficiently transduce cortical neuronal projections following intrastriatal delivery [15], indicating that altering capsid properties may enable broader dissemination even after focal delivery. AAVDJ8R also allowed for functional manipulation of these cells in NHPs [15]. While there have been other efforts to address the limited vector distribution [60, 61], engineered retrograde capsids like this provide a method to address this issue that may also address additional concerns such as cell specificity, depending on the indication.

4.1.1.3. Biodistribution Following Ocular Administration

Biodistribution in the eye is largely dependent on the route of administration. Following SRI, the majority of capsids identified by the NLP search did not show transduction in the anterior segment, and only one, rAAV2tYF, identified transduction within the vitreous [40]. Following intracameral injection, none of the capsids identified reported transduction of the retina or vitreous [50]. It was only after IVT injections that some of the capsids identified in our findings, like AAV2.7m8 and rAAV2tYF [40, 46] were capable of consistently transducing multiple regions of the eye. Thus, indications requiring more spread between regions may benefit more from a route of administration like IVT, or from capsids with the ability to transduce multiple areas like rAAV2tYF. Indications requiring more localized retinal transduction will then benefit from more focused injection methods like SRI. Localized retinal transduction using AAV provides targeted therapeutic benefits for outer (posterior) retinal conditions, while the intravitreal route is advantageous for inner (anterior) retinal targeting due to its less invasive nature and broader distribution in the inner retina [62].

4.1.2. Cellular Tropism

4.1.2.1. Neuronal and Glial Cell Tropism

The NLP search yielded several hits of vectors demonstrating highly efficient cell targeting and transduction via various administration routes. Neurons were the most commonly targeted neural cell type, with several natural serotypes exhibiting this tropism in vivo [63, 64]. Neuronal tropism typically occurs following intraparenchymal injection, which limits peripheral exposure and facilitates more localized transduction. AAVX1.1, an engineered vector based on the AAVPhP.eB backbone, was reported to target brain endothelial cells in mice following systemic administration [20]. The efficiency of this capsid for transducing the rhesus macaque brain via intravenous delivery was also investigated, and using a ubiquitous promoter, they observed not only brain endothelial cell targeting, but also a marked increase in neuronal specificity. Approximately 98% of the transduced cells in the cortex were neurons, a 45‐fold increase compared to the parent AAV9 [20]. An instance of robust neuronal transduction not picked up by our NLP search, due to its absence in peer‐reviewed PubMed publications, is STAC‐BBB. The capsid, developed by Sangamo Therapeutics, showed a 700‐fold increase in neuronal mRNA expression after intravenous injection into cynomolgus macaques. It is also designed to avoid targeting the dorsal root ganglion (DRG) and peripheral tissues [65]. Capsids with such high neuronal specificity in NHPs are rare and challenging to develop, so it is crucial to generate workflows that can help identify these capsids.

Our search produced very few capsids specifically targeting glial cells, though some vectors demonstrated enhanced astrocyte transduction compared to their parent capsid. One noteworthy example is AAVMaCPNS2, an AAV9 variant that exhibits robust CNS and PNS transduction following systemic administration. In addition to its strong neuronal transduction in the cortex, AAVMaCPNS2 reported a 44‐fold increase in astrocyte transduction compared to AAV9 [24]. Additionally, although AAV9 and AAV.PhP.B [66] can transduce astrocytes following intravenous administration, AAV.CAP‐B22 showed a qualitative increase in astrocyte transduction compared to parent capsid AAV9 [18]. AAV2.1B was another capsid identified from the NLP search that was shown to increase astrocyte production by nearly 10% compared to parent AAV2 [31].

Beyond astrocyte targeting, the NLP helped identify AAV capsids with tropism to other glial cells, namely, oligodendrocytes. Strong oligodendrocyte cell tropism was demonstrated for the Olig001 vector, which showed 90%–94% oligodendrocyte specificity following intrastriatal injection [35]. Some groups also showed tropism towards oligodendrocytes for AAVPhP.B and PhP.eB following intrathecal injection [29].

4.1.2.2. Tropism Towards Retinal Tissue

Targeting specific retinal cell types has been a longstanding aim of AAV gene therapy for ocular diseases. The eye's relative immune privilege, the variety of administration routes, and the low dose required make the use of AAVs an appealing therapeutic approach for treating ocular conditions [67]. However, mutations in specific cell types and layers, such as photoreceptors and the retinal pigment epithelium (RPE), are at the root cause of various ocular diseases, making it crucial to develop cell‐specific GTx to minimize off‐target effects [68]. Voretigene neparvovec‐rzyl (Luxturna) is the first approved AAV gene therapy program targeting inherited retinal diseases caused by biallelic mutations in the RPE65 gene [69] which was monumental for the field of gene therapy. Since this point, efforts targeting increased transduction of the RPE layer have continued, as an increasing number of engineered capsids have been identified. Our search led us to several vectors capable of widespread transduction of the RPE through SRI, including AAV2.7m8 [46], rAAV2(4pMut)ΔHS [44], and AAV8BP2 [44].

Mutations, dysfunction, or neuronal degeneration of photoreceptors remain one of the major causes of permanent vision loss [70, 71]. Because of this, it is important to find ways to target these cells without incurring off‐target effects. Our NLP results led us to a photoreceptor targeting capsid called AAV44.9(E531D) that displayed highly efficient, localized photoreceptor transduction following SRI in NHPs [39]. This capsid was found to transduce up to 98% of foveal cones in the foveal pit and 100% of central rods, showing strong lateral spread and cone specificity outside of the SRI bleb [39]. AAV44.9(E531D) also showed the unique ability to target parafoveal and perifoveal cones. This is promising, as groups have long struggled to specifically transduce cones in these areas [72]. At the time of writing, this AAV is currently being used in Phase I/II clinical trials to treat X‐linked retinoschisis [73].

4.2. Overcoming Immunogenicity and Safety Concerns With Capsid Engineering

Immunogenicity concerns remain a major challenge of AAV gene therapy, in some cases leading to dose‐limiting toxicities, as well as reductions in efficacy and durability [74, 75, 76]. Immunogenicity can arise in response to various components of the AAV drug product, including the capsid, DNA payload, and the expressed RNA or protein product, triggering both innate and adaptive immune responses. These responses can, among other things, lead to reduced transduction (efficacy risk) [77, 78], clearance of transduced cells (pharmacological and toxicity risk), and other adverse events, such as systemic hypersensitivity reactions, inflammation, hepatotoxicity, and other target organ‐related toxicities. While all these factors contribute to the overall immune profile of AAV gene therapy, capsid‐related immunogenicity and toxicities present unique challenges that warrant specific engineering strategies to mitigate immune recognition, enhance targeting to desired cell types, and improve therapeutic outcomes.

Innate immune activation following AAV can impact clearance [79] while also triggering inflammatory toxicities such as infusion‐related reactions, complement activation, and coagulation disorders, and hepatotoxicity [80]. The AAV capsid itself is a key driver of Toll‐like receptor 2 (TLR2) activation and other pattern recognition receptors contributing to innate immune recognition. To mitigate these effects, some capsid engineering strategies, such as those described in Bentler et al. [81] have focused on modifying surface‐exposed capsid residues to dampen TLR‐MyD88 signaling while preserving transduction. Since innate immunity not only drives acute inflammatory responses but also influences downstream cellular (e.g., priming cell responses to AAV capsids) and humoral (e.g., immune complex formation) adaptive responses, strategies to minimize innate activation may improve both the tolerability and long‐term efficacy of AAV gene therapies.

One of the earliest hurdles to effective AAV gene transfer is humoral immunity, as pre‐existing neutralizing antibodies (nAbs) against the capsid (from natural exposure) can limit vector transduction and impact safety. Some studies estimate 30%–60% of the global population has anti‐AAV nAbs, with seroprevalence varying by geographical location and serotype [77]. Screening and exclusion of subjects based on seropositivity is common practice for investigational preclinical and clinical studies, as well as approved products, particularly systemically administered AAVs. This strategy aims to mitigate the risk of reduced efficacy and adverse events. Several approaches to minimize exclusion of patients in need are under active investigation, including plasmapheresis, use of antibody‐degrading enzymes, as well as leveraging novel capsids that may have reduced cross‐reactivity to pre‐existing antibodies, and serotype switching, which may support re‐dosing in patients [82].

Neutralizing antibodies also have potential for cross‐reactivity to engineered capsids based on natural serotypes [45]. To address some of these concerns, researchers have employed capsid engineering to incorporate mutations that confer resistance to nAbs and mitigate immune responses [83, 84]. Mainly, AAVP2‐V1, an engineered vector based on AAV2, has been shown to enhance retinal transduction in cynomolgus macaques as compared to the parent capsid [42] harbors a mutation that showed increased ability to avoid nAbs in human vitreous humor samples containing anti‐AAV2 antibodies [42]. While in vivo nAb avoidance was not tested in NHPs, capsids like AAVP2‐V1, which contain mutations known to avoid nAbs response, strongly demonstrate the potential of capsid engineering and its benefit to the field. Another example includes chimeric mutant AAV‐DJ, which significantly reduces in vivo seroreactivity compared to its wild‐type counterparts (AAV2/8/9), demonstrating enhanced immune evasion and transduction efficiency in mice [85]. A valuable approach was reported for liver transduction as well—novel liver‐tropic AAV variants, such as NP40 and NP59, developed through directed evolution and capsid shuffling, demonstrated significantly improved human hepatocyte transduction in vivo (mouse xenografts and primary organoids) compared to wild‐type counterparts, while also exhibiting reduced seroreactivity in a small cohort of human patient sera—a valuable approach that can inform the development of enhanced AAV vectors for CNS and eye gene therapy applications [86].

A unique engineering strategy to avoid nAbs and allow for repeated AAV administration is a method called serotype switching. Indeed, Chen et al. reports serotype switching of the X1 peptide from AAV9 into AAV1 allowed repeat administration of AAV in mice [20]. While there is a degree of seroreactivity between AAV serotypes [84, 87], the ability to increase CNS transduction while using a different AAV backbone poses great therapeutic potential. Not only does this avoid the nAb response, it also has the potential to expand the patient selection criteria, as patients with a high nAb titer against one serotype may be able to receive treatment due to the use of a different AAV backbone.

Cellular immunogenicity, and particularly CD8+ T cell responses, have been attributed to the cytotoxicity of transduced cells, leading to durability loss in clinical trials [83, 88] and safety concerns [89]. One group leveraging capsid engineering to improve liver targeting and reduce immunogenicity identified a capsid gene with reduced presentation of capsid peptides by MHC class I and reduced killing of human hepatocytes [90]. Other groups hope to do this on a larger scale by leveraging machine learning to predict presented peptides: MHC complexes that will inform the design of AAV capsids to reduce the likelihood of generating T cell epitopes. This is done through the identification of cytotoxic CD8+ T‐cell epitopes within capsid proteins and evaluating their ability to activate T cell responses, in part using peptide: MHC prediction algorithms. Although in their infancy, these efforts present a great opportunity to deplete capsids of T‐cell‐activating peptides, thereby providing the potential to improve the durability and safety of future AAV products [91].

In the CNS, a common way to minimize the immune response to delivered capsids has been to use targeted CNS injections rather than systemic delivery. While it is thought that the CNS is not subject to immune response, this has been refuted due to the presence of specialized immune cells in the CNS and resident leukocytes [92]. Microglia are responsible for patrolling the CNS for pathogens and damage and initiating an immune response when necessary [93]. These responses, at times, can involve significant inflammation that may lead to neurodegeneration if left unchecked [92, 94]. Direct CNS administrations may also result in peripheral exposure of the capsid, specifically when delivered to CSF, resulting in a systemic immune response [58, 95]. Additionally, studies have identified that some safety concerns—specifically DRG and spinal cord toxicity—can occur whether delivered intravenously or directly into the CNS [96]. Finding strategies for mitigating immune response beyond minimizing the dose or using targeted routes of administration is needed.

With ocular indications, the route of administration choice may greatly influence the immune response. Some studies have shown that intravitreal (IVT) administration of AAV can induce local inflammation and increase humoral responses against the vector capsid, particularly compared to subretinal delivery [97, 98, 99], meanwhile suprachoroidal injection was linked to a reduced humoral response but increased infiltration of inflammatory cells in the retina and choroid [100]. Off‐target ocular tissue transduction has been linked to increased ocular inflammation, and choosing a capsid with a select target cell tropism may reduce these risks [98]. Increased immune response following IVT can not only be a safety concern but can also potentially limit administration in the contralateral eye for bilateral diseases if done sequentially. Capsid screening for improved vectors for IVT administration based on combined retinal transduction and reduced reactivity to anti‐AAV sera [101] is currently underway. Combining these efforts, along with reducing transduction in nontarget tissues, may improve the tolerability of ocular‐delivered AAV gene therapies.

As far as translational efforts to evaluate safety go, nonclinical species, including NHP, are largely under‐predictive of systemic toxicity liabilities that have been observed clinically with wild‐type capsids. In the absence of clinical data, attempts to leverage reverse translation to inform predictive safety liabilities are limited. In the interim, existing toxicity data from nonclinical species, in combination with the anticipated minimal efficacious dose, can inform whether there is a therapeutic index enabling progression of a given capsid/payload combination. Further, identification of early biomarkers (e.g., elevated transaminases, coagulation parameters, and complement activation) may help identify early risk factors and support improved interventional strategies.

4.3. Species Differences/Translation

Species‐specific differences in receptor expression, BBB composition, epigenetics, and other factors can largely affect the translatability of findings between species [102, 103, 104]. A well‐known example of this occurs with the translatability of the AAV‐PhP.B capsids following IV delivery. AAV‐PhP.B is well studied and is known to induce increased CNS tropism in mice compared to AAV9 [105]. However, many groups struggled to replicate this increased CNS tropism following IV delivery in other mouse strains, and higher order species like rhesus macaques and marmosets [22, 23]. It was discovered that the increased CNS tropism originally found in WT C57BL/6J mice after intravenous delivery was likely due to the increased prevalence of the Ly6A receptor in the CNS, and that this effect also extended to AAVPhP.eB as well [106, 107]. Huang et al. identified that strain‐specific genetic differences in the Ly6A receptor in mice play an important role in the increased CNS tropism of AAVPhP.B and PhP.eB binding. Later, they concluded that this receptor is necessary for the increased tropism following IV administration, indicating that the functional Ly6A receptor is upregulated in specific mice strains and that the absence of a homolog in primates may limit the immediate translatability of this discovery [106, 107].

There can also be species‐related differences between primates. For instance, after AAVX1.1 increased CNS and brain endothelial specificity in mice, its transduction capabilities were tested ex vivo in human, southern pig‐tailed, and rhesus macaque explants. Seeing a significant increase in CNS transduction compared to AAV9 in these explants, in vivo experiments were subsequently performed in marmosets and macaques. In marmosets, they did not observe a significant difference between AAV9 and AAVX1.1 transduction; however, in rhesus macaques, there was a significant increase in AAVX1.1 CNS tropism [20]. On the other hand, certain capsids identified that the increased CNS transduction in marmosets like AAVCap‐B10 and AAVCap‐B22 did not translate to both rhesus and cynomolgus macaques [19, 25].

This group also observed species‐specific variation in the discovery of AAV.CAP‐Mac. CAP‐Mac, which was originally identified in marmosets, has been shown to efficiently transduce neurons in the CNS of infant rhesus macaques as well. However, AAV.CAP‐Mac, despite showing increased CNS tropism compared to AAV9 in African green monkeys and rhesus macaques, shows a cell specificity governed by species and age. Another species‐specific feature is cell specificity—it displays tropism towards neurons in infant green monkeys and infant rhesus macaques, but shows astrocyte‐biased tropism in adult rhesus macaques, and brain vasculature‐biased tropism in marmosets [20].

Overall, heterogeneity seen in these examples highlights the importance of considering species‐specific differences and age‐related changes in BBB makeup when developing engineered capsids. Due to this variability, screening engineered capsids across multiple primate species, along with identifying the receptors essential for novel engineered capsid binding, can provide valuable insights into the translatability of a novel capsid to humans.

5. Conclusion and Future Directions

The advancement of AAV gene therapy has shown immense promise in treating genetic diseases, particularly through the precise genomic modifications enabled by natural and engineered AAV capsids. This field is rapidly expanding, as evidenced by the increasing number of AAV products receiving FDA approval, as well as a rapidly growing literature database. Ocular diseases have remained a prominent focus for gene therapy, as demonstrated by the increasing number of clinical trials targeting ocular conditions [108]. In contrast, the central nervous system (CNS) remains a challenging target from a clinical perspective.

Our study also highlights the critical role of NLP and LLMs in systematically identifying and characterizing neurotropic and ocular AAV capsids, specifically those tested in NHPs. NHPs are often considered the most translatable species for preclinical research, making them essential for evaluating the potential of novel capsids for human applications. The engineered capsids identified demonstrate enhanced tissue specificity, improved pharmacokinetics and pharmacodynamics, and reduced off‐target effects, thereby addressing the limitations associated with naturally occurring AAV serotypes. Despite these advancements, challenges remain, such as the prevalence of nAbs and species‐specific translation difficulties. Future research should focus on integrating advanced computational approaches, such as machine learning and artificial intelligence, to further refine the identification and optimization of AAV capsids. By addressing existing barriers and leveraging cutting‐edge computational tools, the field of AAV gene therapy can move closer to achieving its full clinical potential, ultimately improving patient outcomes across a range of genetic disorders.

Beyond systematic literature curation, the NLP and LLM‐assisted framework described here can be adapted for automated identification of emerging AAV capsid data, integration into searchable vector databases, and prioritization of preclinical candidates with optimal translational properties. Incorporating these tools into routine preclinical workflows could enhance reproducibility, reduce manual curation time, and support data‐driven decision‐making in vector engineering and translational pharmacology.

Author Contributions

All authors wrote the manuscript. M.M., S.S., and H.‐Y.W. designed the research; C.A., H.‐Y.W., and C.K. performed the research; C.A., H.‐Y.W., and M.M. analyzed the data. All authors interpreted the data and provided final approval to submit the manuscript.

Funding

This research was funded by Genentech Inc.

Conflicts of Interest

At the time of this research, M.W., C.K., E.M.‐R., S.S., and M.M. were salaried employees of Genentech Inc. (a member of the Roche Group) and were stockholders in Roche. C.A. declares no conflicts of interest.

Supporting information

Data S1: cts70428‐sup‐0001‐Supinfo.docx.

Agbim C., Wu H.‐Y., Kim C., Mutter‐Rottmayer E., Sadekar S., and Markovic M., “ AAV Gene Therapy Drug Development and Translation of Engineered Ocular and Neurotropic Capsids: A Systematic Review Using Natural Language Processing,” Clinical and Translational Science 18, no. 12 (2025): e70428, 10.1111/cts.70428.