Multilayered control of retinal stem/progenitor cell fate in the single-cell and organoid era: Developmental blueprints and regenerative opportunities

Abstract

Irreversible degeneration of retinal neurons and the retinal pigment epithelium is a major cause of vision loss, and current pharmacological or surgical treatments seldom rebuild lost tissue, placing stem cell-based regeneration at the center of therapeutic exploration. Retinal progenitor cells, Müller glia (MG)-derived progenitors, pluripotent stem cell-derived retinal pigment epithelium and photoreceptors, and emerging human retinal stem cell candidates together provide a diverse cellular repertoire whose behavior is governed by tightly coordinated fate-control mechanisms. Enabled by single-cell and spatial multi-omics in developing human retina and retinal organoids, these mechanisms can now be mapped at unprecedented resolution, revealing how distinct lineage trajectories and molecular states arise. This review synthesizes a multilayered framework of fate regulation encompassing the diversity and plasticity of embryonic progenitors, MG-derived progenitors, ciliary margin-like cells and putative adult retinal stem cells, and examines how transcription factor hierarchies, epigenetic landscapes, and non-coding RNAs interact with translational, metabolic and inflammatory cues to shape competence windows and photoreceptor vs inner retinal fates in development and disease. These insights are then connected to next-generation regenerative strategies, including engineered retinal organoids and sheets, MG reprogramming, and rational combinations of gene and cell therapies designed to precisely steer cell identity, maturation and circuit integration. By framing retinal regeneration within this multilayered control paradigm, we highlight key challenges for clinical translation and outline how targeted manipulation of fate-regulatory networks may accelerate the development of safe and effective stem cell therapies for blinding retinal disorders.

Key Words: Epigenetic competence; Human retinal organoids; Metabolic state control; Müller glia; Retinal degeneration; Retinal regeneration; Single-cell multi-omics; Spatial transcriptomics

Core Tip: Retinal degeneration is a structural failure that cannot be reversed by pathway modulation alone. Leveraging single-cell and spatial multi-omics from human retina and organoids, this review defines a multilayer fate-control paradigm - transcriptional instruction, epigenetic permission, and metabolic/inflammatory execution - that governs competence windows and photoreceptor vs inner-retinal outcomes. We compare major regenerative cell sources (retinal progenitors, Müller glia, ciliary marginal zone-like candidates, and human pluripotent stem cell-derived retinal pigment epithelium/ photoreceptors) and translate atlas insights into actionable engineering strategies, including organoid benchmarking, staged reprogramming, and rational gene + cell combinations to enhance maturation and circuit integration.

INTRODUCTION

The human retina depends on precisely wired microcircuits spanning photoreceptors, interneurons, and retinal ganglion cells (RGCs). Across major retinal degenerations, the dominant pathology extends beyond cellular dysfunction to progressive, often irreversible cell attrition accompanied by circuit disassembly[1,2]. This creates a central translational bottleneck: Although upstream molecular drivers can be modulated, most current therapies do not rebuild the tissue architecture and connectivity required for durable functional recovery. Because leading causes of blindness converge on cumulative neuronal and/or retinal pigment epithelium (RPE) loss, there is an urgent need for regenerative strategies that replace missing cells and reconstitute circuitry rather than merely slowing decline[3].

A practical way to define regenerative targets is to align disease classes with dominant missing cell types, which dictate cell source selection, delivery plane, and integration requirements. Outer-retina disorders such as age-related macular degeneration (AMD) and inherited retinal degenerations are characterized by photoreceptor dysfunction and/or RPE failure, whereas optic neuropathies exemplified by glaucoma center on RGC loss with the additional requirement of long-range axonal reconnection to central targets[4,5]. Together, these targets define two translationally distinct agendas in regenerative ophthalmology: Photoreceptor/RPE replacement for outer-retina failure, and RGC replacement coupled to connectivity engineering for optic neuropathy[6].

The current therapeutic landscape excels in mechanism-targeted disease modification but lags in reconstruction. In geographic atrophy, complement inhibition can slow lesion expansion without restoring atrophic tissue[7]. Anti-vascular endothelial growth factor therapy has transformed neovascular AMD management yet does not address irreversible cellular loss and requires sustained treatment to maintain benefit[8]. Gene therapy provides definitive proof-of-concept for molecular correction but is fundamentally constrained by the requirement for a viable target cell population[9]. Consequently, scalable solutions to the combined “tissue plus circuit” problem remain elusive, particularly for RGC disorders in which structural replacement is inseparable from long-distance connectivity[6].

Here, we posit that effective retinal regeneration will require developmentally informed and controllable fate engineering, defined as the ability to: (1) Specify lineage with high fidelity; (2) Achieve appropriate maturation states; and (3) Stabilize identity and function within the host microenvironment through multi-layer control spanning transcriptional, epigenetic, metabolic, and niche cues. High-resolution multi-omics atlases of human retinal development now provide quantitative benchmarks for competence windows, lineage trajectories, and maturation programs[10,11]. In parallel, emerging clinical and preclinical evidence supports the feasibility of cell replacement while underscoring key barriers, including safety, durability, and functional integration - ranging from stem cell-based RPE replacement efforts [e.g., retinal pigment epithelium stem cell-derived retinal pigment epithelium progeny at 4 weeks of differentiation (RPESC-RPE-4W); NCT04627428] to experimental strategies for RGC replacement and reprogramming[12-15]. Accordingly, this review will: (1) Summarize developmental principles that define competence windows; (2) Synthesize multi-layer control points required for stable specification; and (3) Evaluate strategies to enhance safety and functional circuit integration.

A LINEAGE LIBRARY FOR RETINAL REGENERATION: SOURCES, PLASTICITY, AND CONTROLLABILITY DIFFERENCES

Embryonic/developmental retinal progenitor cells: Lineage trajectories and competence windows

Embryonic retinal progenitor cells (RPCs) are the “reference progenitors” for retinal fate engineering because they natively execute the full developmental program that generates all retinal neurons and Müller glia (MG) in a temporally ordered manner. Over the last five years, single-cell atlases and especially multi-omic datasets (transcriptome coupled to chromatin accessibility) have moved the field from marker-based stage definitions to gene-regulatory-network (GRN) - level descriptions of competence windows - discrete periods when RPCs are permissive for producing particular lineages and subtypes[10,16,17]. A dual-omic atlas spanning human fetal macular and peripheral retina provides a particularly useful blueprint because it links fate choices to region-specific regulatory logic (macula vs periphery) and to chromatin constraints that are not obvious from RNA alone[11].

Mechanistically, “temporal identity” is increasingly viewed as an active, regulatable state rather than a passive readout of developmental time. For example, polycomb-associated regulation has been implicated in the early-to-late transition: Jarid2 facilitates progression of RPC temporal identity by repressing early temporal programs (e.g., Foxp1-linked competence), thereby reshaping what the progenitor can still generate at later stages[18]. In parallel, Ikaros-family transcription factors (TFs) act as early temporal competence regulators; recent work shows redundancy between Ikzf1 and Ikzf4 in shaping early-stage outcomes (including cone-biased competence) and repressing late fate programs[19].

From a translational perspective, authentic fetal RPCs are not a scalable clinical source, but their value is instructional: They define which TF modules, chromatin states, and signaling contexts must be recreated (or bypassed) when programming human pluripotent stem cell (hPSC)-derived progenitors, retinal organoids, or reprogrammed MG[20]. Thus, developmental RPC biology primarily improves “controllability” indirectly - by turning fate engineering into a blueprint-driven design problem rather than trial-and-error differentiation[21,22].

MG and MG-derived progenitors: Zebrafish robustness vs mammalian restriction

MG are the dominant endogenous regeneration substrate in zebrafish: After injury, they de-differentiate, re-enter the cell cycle, generate MG-derived progenitor cells, and repopulate multiple neuronal classes. In mammals, MG are broadly competent to mount gliosis and metabolic support but show limited spontaneous neurogenesis, making them an attractive yet challenging target for in situ regeneration. Comparative work over the past five years has sharpened the distinction between: (1) Injury-responsive reprogramming circuits that are readily engaged in fish; and (2) Stability mechanisms that constrain mammalian MG to a supportive glial identity[23].

On the “where we are” side, several proof-of-principle studies show that mammalian MG can be pushed toward neuronal programs using developmental TF logic. A prominent example is TF-driven reprogramming that induces MG to generate RGC-like neurons in adult mouse retina, supported by single-cell RNA/assay for transposase-accessible chromatin (ATAC) profiling and functional readouts[14]. Importantly, the same conceptual strategy is beginning to be tested in human MG contexts: ASCL1-driven neurogenesis has been demonstrated in human MG derived from fetal retina and retinal organoids, suggesting at least partial conservation of reprogrammability if the correct starting state and environment are provided[22].

On the “what’s missing” side, two classes of barriers dominate. First are cell-intrinsic locks (epigenetic and transcriptional): Mammalian MG often fail to sustain a stable neuronal trajectory or to mature into subtype-appropriate phenotypes without additional interventions. Second are microenvironmental constraints [inflammation, extracellular matrix (ECM), and circuit context] that can prevent integration even when fate conversion signatures appear. A recent Nature Communications’s study highlights a concrete, mechanistic “lock”: Prospero homeobox 1 (Prox1) transfer-linked mechanisms can suppress MG regenerative potential in adult mouse retina, and manipulating this axis partially restores regeneration-associated behavior[24].

A critical methodological caveat for the MG field is “apparent conversion” vs true lineage reprogramming. Recent analyses emphasize that viral tools and lineage strategies can produce misleading signals (e.g., reporter leakage, cell labeling ambiguity), motivating stricter fate mapping, orthogonal validation, and single-cell lineage-aware designs[25].

Overall, MG offer the most anatomically native route to regeneration (cells are already in the right layer), but they currently have the greatest controllability gap: Inducing the correct fate is feasible in principle, yet ensuring robust efficiency, subtype precision, maturation, and circuit integration remains the central challenge - especially for long-range RGC repair.

Ciliary margin zone-like cells and “adult retinal stem cell candidates”: Evidence boundaries and controversies

Unlike fish and amphibians, where a lifelong ciliary marginal zone (CMZ) fuels retinal growth, adult mammals do not exhibit a well-established CMZ stem-cell niche capable of robust in vivo neurogenesis after injury, and “adult retinal stem cell” claims in the ciliary margin/ciliary epithelium remain controversial[26]. Recent human single-cell and spatial studies instead identify transient progenitor-like states at the ciliary margin during fetal development that decline with maturation, supporting a developmental-edge program rather than a durable adult stem pool[27]. Retinal organoids can recapitulate putative CMZ-like states, which is valuable for hypothesis generation and for extracting developmental design rules, but does not constitute evidence of an adult human CMZ regenerative resource[21]. Therefore, we frame CMZ/adult-candidate populations strictly as hypothesis-generating: They may illuminate peripheral developmental programs and reactivation triggers, yet there is currently insufficient in vivo evidence in adult human retina to justify them as a near-term clinical cell source[27,28].

The controversy centers on translating “progenitor-like signatures” into bona fide stem cell claims. Sphere formation or progenitor marker expression in vitro has historically been used to argue for adult retinal stem cells in the ciliary epithelium, yet in vivo neurogenesis and durable, functionally integrating neuronal output remain inconsistent[29-31]. This gap reflects both biological limitations (true scarcity or quiescence of such cells in adults) and methodological constraints (sampling bias, culture-induced plasticity, and lineage ambiguity)[32,33]. Recent multi-omic work in Science Translational Medicine proposes a stem cell population with regenerative potential when integrating single-cell multiomics across fetal retinas and retinal organoids, but the decisive test remains whether such candidates can reproducibly generate clinically relevant retinal neurons with integration-level function in appropriate in vivo settings[34].

Practically, CMZ-like and adult-candidate populations are best treated as hypothesis-generating reservoirs: They may illuminate edge-region developmental programs and reactivation triggers, but they currently rank lower than hPSC-derived products and MG-based approaches in translational controllability.

hPSC-derived cells: HPSC-RPE, rod/cone precursors, and organoid-derived cell populations

hPSC-derived products represent the most scalable and manufacturing-controllable lineage library for regenerative ophthalmology. The past five years have shifted the conversation from “can we make retinal cells?” to “can we make the right cells, at the right maturity, with measurable comparability to human tissue, and with clinically acceptable safety and delivery performance?” Multimodal benchmarking of human retinal organoid development against primary human retinal tissue provides quantitative standards for maturation trajectories and molecular congruence - an essential step for defining release criteria beyond marker snapshots[10].

RPE remains the most clinically advanced hPSC-aligned lineage because it is a monolayer epithelium with well-defined identity markers, functional assays (phagocytosis, polarized secretion), and relatively tractable integration requirements compared with neurons. Notably, a 2025 Cell Stem Cell clinical report describes low-dose outcomes of RPESC-RPE-4W transplantation in geographic atrophy, providing contemporary evidence primarily on safety/tolerability and feasibility while underscoring the importance of product identity and release criteria in human studies[12]. Consistent with general principles of early clinical development, efficacy readouts in such early cohorts should be interpreted as exploratory rather than confirmatory, given typical limitations in sample size, controls, and endpoint variability[10,35].

Photoreceptor precursors (rods/cones) are more complex: They require neuronal maturation, synaptic incorporation, and alignment with host bipolar circuitry, and they must be evaluated with assays that distinguish true integration from material transfer or transient donor-host exchange. Recent transplantation work in a cone-dominant ground squirrel model demonstrates survival and characterization of transplanted human induced pluripotent stem cell (iPSC)-derived photoreceptors, illustrating progress toward anatomically relevant large-eye settings[36]. Meanwhile, organoid engineering is being pushed toward functional rescue signals; for example, transplantation of genome-edited retinal organoids has been reported to restore selected physiological functions coordinated with severely degenerated host retinas, underscoring the potential - and the need for rigorous interpretation of “function” and connectivity[37].

Finally, organoid-derived mixed populations are increasingly viewed as a feature rather than a flaw when the goal is tissue-like architecture, but this comes at a controllability cost (heterogeneity, off-target cell types, batch effects). Recent methods address this by improving scalability, purification, and release-readiness - such as scalable organoid generation platforms and label-free enrichment approaches that increase photoreceptor purity[38].

In aggregate, hPSC-derived RPE currently offers the strongest combination of scalability and controllability, while hPSC-derived photoreceptor strategies are rapidly maturing but remain integration-limited. Organoid benchmarking and multi-omic comparability frameworks are now the enabling infrastructure that connects developmental blueprints to clinically auditable fate engineering (Figure 1; Table 1).

Figure 1 Retinal regenerative lineage library: Sources vs controllability/scalability/translational fit. A: Lineage/cell source library: A categorization of the major regenerative substrates discussed in this review. Embryonic retinal progenitor cells (RPCs) serve primarily as developmental blueprints (reference only). Among adult-associated substrates, Müller glia (MG)-based approaches and human pluripotent stem cell-derived products represent the most tractable therapeutic routes, whereas ciliary marginal zone-like/adult-candidate populations are presented here as hypothesis-generating developmental-edge programs with currently limited evidence as a near-term clinical cell source; B: Decision matrix & indication stage: A multi-axis evaluation of these sources based on three translational criteria: Controllability: The precision of fate specification and stability of the maturation state [high in RPCs/retinal pigment epithelium (RPE); lower in MG/ciliary marginal zone]. Scalability: Good Manufacturing Practice manufacturing readiness (highest in human pluripotent stem cell-derived RPE). Integration burden: The complexity of host connectivity required for function (highest for photoreceptors/RPCs; lowest for RPE). The bottom timeline maps these sources to their “best-fit” disease stages, proposing MG reprogramming for early-stage intervention and cell replacement (RPE/photoreceptors) for intermediate-to-end-stage degeneration. Critical cautions: Key risks are highlighted, including adeno-associated virus promoter leakage artifacts in MG studies, the confounding of material transfer vs synaptic integration in photoreceptor grafts, and surgical complexity for RPE patches. RPCs: Retinal progenitor cells; Nuclear factor I; Prox1: Prospero homeobox 1; AAV: Adeno-associated virus; CMZ: Ciliary marginal zone; hPSC: Human pluripotent stem cell; RPE: Retinal pigment epithelium; GMP: Good Manufacturing Practice; PVR: Proliferative vitreoretinopathy; QC: Quality control.

Table 1 Retinal regenerative lineage library - comparative controllability, scalability, and translational fit.

Lineage/source class	Representative product/strategy in manuscript	Primary strengths	Dominant controllability limits	Key safety risks	Best-fit indication stage	Translational role in your framework
Embryonic/developmental RPCs	“Reference progenitors” defining competence windows	Gold-standard developmental program; informs TF/enhancer logic	Not scalable; not a clinical source	Ethical/availability constraints	N/A (instructional)	Blueprint to recreate/bypass competence windows
MG in situ	In vivo reprogramming	Anatomically native; correct laminar position	Competence locked by Notch/NFI/Prox1; prone to reactive reversion	Off-target AAV expression; scarring; uncontrolled proliferation	Early-intermediate, niche-permissive	“Gene-only” regeneration if evidence-standard met
CMZ-like/adult stem cell candidates	Hypothesis-generating reservoirs	May reveal edge-niche triggers	In vivo neurogenesis inconsistent; culture-induced artifacts	Lineage ambiguity	Low priority	Mechanistic inspiration rather than a product source
hPSC-RPE (suspension)	Subretinal injection; RPESC-RPE-4W	Mature monolayer identity; imaging-friendly endpoints	Repolarization on diseased Bruch’s; heterogeneous distribution	Proliferation/tumorigenicity monitoring needed	Intermediate-advanced GA/AMD	Clinically advanced “vanguard” replacement
hPSC-RPE (patch/scaffold)	CPCB-RPE1, sheets/patches	Pre-polarized monolayer; surrogate substrate	Higher surgical burden; complication spectrum	PVR/retinal detachment risk	Advanced structural loss	Durability-first strategy
Photoreceptor precursors	“Goldilocks zone” post-mitotic precursors	Potential for vision restoration via synaptic integration	Integration vs material transfer confound; maturity tuning required	Ectopic differentiation; limited connectivity proof	End-stage vs mid-stage stratification needed	Replace + require rigorous mechanism-of-benefit parsing
Organoid-derived laminated outputs	Engineered scaffolds; ecosystem completion	Tissue-like architecture; testbed for causality + product	Heterogeneity, batch effects; microenvironment missingness	Off-target tissues; stress programs	Preclinical → translational	Programmable developmental proxy + manufacturing challenge

RPCs: Retinal progenitor cells; TF: Transcription factor; N/A: Not applicable; MG: Müller glia; NFI: Nuclear factor I; Prox1: Prospero homeobox 1; AAV: Adeno-associated virus; CMZ: Ciliary marginal zone; hPSC-RPE: Human pluripotent stem cell-derived retinal pigment epithelium; RPESC: Retinal pigment epithelium stem cell; RPESC-RPE-4W: Retinal pigment epithelium stem cell-derived retinal pigment epithelium progeny at 4 weeks of differentiation; GA: Geographic atrophy; AMD: Age-related macular degeneration; CPCB-RPE1: California Project to Cure Blindness-retinal pigment epithelium 1; PVR: Proliferative vitreoretinopathy; RPE: Retinal pigment epithelium.

SINGLE-CELL AND SPATIAL MULTI-OMICS: HOW DEVELOPMENTAL BLUEPRINTS QUANTIFY FATE TRAJECTORIES

Milestone progress in retinal development and disease single-cell genomics

Over the past five years, retinal single-cell genomics has shifted from “cell-type catalogs” to quantitative developmental blueprints that can be operationalized for fate engineering[39]. A pivotal advance has been the emergence of integrated multi-omic reference atlases that jointly capture transcriptomic states and regulatory potential (chromatin accessibility) at scale, enabling cell-type–resolved TF logic, cis-regulatory element (CRE) inference, and disease-variant interpretation[40,41]. Large-scale human retina multi-omics resources integrating millions of nuclei and hundreds of thousands of single nucleus ATAC profiles now approach near-complete coverage of major retinal classes and many rare subtypes, offering a reference frame for mapping new datasets, benchmarking organoids, and contextualizing disease states[42].

On the developmental side, the field has moved toward time-resolved, region-aware atlases that explicitly model macular vs peripheral programs and provide GRN hypotheses for lineage decisions. Dual-omic profiling across human post-conception windows has quantified trajectories for dozens of cell classes and nominated candidate TF-CRE drivers that differ between macula and periphery, directly supporting “competence window” concepts in humans rather than extrapolations from rodents[11]. A complementary milestone is the integration of single-cell RNA sequencing with spatial transcriptomics in developing human eye tissues to resolve transient progenitor states (including ciliary-margin-localized progenitors) and to localize differentiation waves across the tissue-critical when fate decisions are spatially gated[27].

For translational relevance, posterior-segment multi-omics has expanded beyond neural retina to include RPE/choroid ecosystems and AMD-relevant regulatory architecture. A recent single-cell multiome and enhancer “connectome” framework coupled single-cell RNA (scRNA)/single-cell ATAC with chromatin conformation and functional reporter assays to nominate AMD-linked noncoding mechanisms and prioritize causal variants in RPE/choroid contexts - an explicit bridge from atlas to mechanism[43]. Finally, organoid systems have entered a more quantitative era: Multimodal spatiotemporal phenotyping has paired organoid time courses with primary human tissue benchmarks, producing cross-scale standards for maturation trajectories and GRN inference rather than relying on marker-by-marker resemblance[10,44].

Spatial transcriptomics and in situ technologies: Returning “cell state” to lamination and microenvironment

Retinal biology is inseparable from architecture: Laminar placement constrains synaptic partners, while regional specializations (e.g., fovea/macula vs periphery; dorsal-ventral gradients) impose distinct transcriptional programs. Dissociation-based single-cell RNA sequencing necessarily discards this information; spatial transcriptomics and image-based in situ methods restore it by preserving layer, neighborhood, and lesion context.

High-resolution image-based spatial transcriptomics has begun to map subtype organization in situ at scale. MERFISH applied to mouse retina generated a single-cell spatial atlas that resolves major classes and many subtypes while uncovering spatially patterned gene expression and laminar organization that cannot be reconstructed from dissociated data alone[45]. In humans, spatial approaches are increasingly deployed to capture developmental and disease-relevant niches. Spatial transcriptomic resources for the developing human retina have been used to quantify spatiotemporal cell organization and infer cell-cell communication in context, providing a scaffold for interpreting competence windows as tissue-position-dependent rather than purely time-dependent[45]. Spatial methods are also being used to resolve regionally restricted stress and inflammatory programs in adult retina models, underscoring how degeneration is often geographically heterogeneous - a point that directly impacts where and how regenerative grafts should be deployed[46].

Importantly, spatial layers enable explicit modeling of the microenvironmental constraints that shape fate stability after transplantation or reprogramming - vascular proximity, immune infiltration, reactive gliosis gradients, and extracellular-matrix states[47,48]. These are precisely the variables that determine whether a cell’s transcriptomic identity will remain stable and whether integration is feasible in the outer vs inner retina[6,49].

Practical analytic frameworks for quantifying fate trajectories and interaction logic

Establishing a coherent analytic pipeline that links cellular states to developmental directions, interaction logic, and comparative benchmarks is essential for dissecting regenerative potential[50]. The standard analytical workflow typically begins with trajectory inference to order cells along continuous developmental progressions. Graph-based or manifold approaches are routinely employed to define lineage bifurcations, yet pseudotime must be interpreted as a conditional axis constrained by sampling windows and tissue regions[27,51,52]. To resolve the ambiguity of static snapshots, RNA velocity and dynamical modeling provide local directional vectors of state change, particularly valuable when transitions are rapid or progenitor compartments are heterogeneous. While recent advances have expanded velocity analysis from simple splicing kinetics to transcriptomic “vector fields” that capture complex flows, these methods remain sensitive to technical noise. Therefore, robust trajectory reconstruction requires triangulation - integrating velocity directionality with lineage marker kinetics and temporal experimental structures to ensure biological fidelity[53].

Once the developmental trajectory is established, the focus shifts to quantifying fate potentials and their regulatory drivers. Probabilistic fate-mapping tools, such as CellRank, combine transcriptomic similarity with directional velocity data to identify “decision regions” in which subtle regulatory shifts tip lineage outcomes. This probabilistic approach is particularly effective for comparing normal vs perturbed differentiation, such as following TF manipulation[54]. Crucially, these intrinsic fate decisions are often governed by extrinsic cues. Cell-cell communication inference frameworks (e.g., CellChat) formalize these interactions by modeling ligand-receptor probability networks[55]. In the context of the retina, quantifying signaling inputs from MG, microglia, and the vasculature is critical for decoding the niche environments that either promote regeneration or enforce homeostatic restriction.

Finally, ensuring the translational rigor of these findings requires valid benchmarking across datasets and species. Reference-mapping approaches have become indispensable for projecting new data - such as organoid-derived cells or disease samples - onto established healthy atlases without the need for de novo integration, facilitating standardized “disease-state positioning”[56]. Furthermore, bridging the gap between regenerative model organisms (e.g., zebrafish) and non-regenerative mammals relies on cross-species mapping. Advanced alignment tools now move beyond simple marker homology to align latent cellular states, enabling the precise identification of conserved vs divergent gene regulatory programs that underlie the loss of regenerative competence in the human retina[57].

From atlas to causality: Validating control nodes with perturbation experiments

Atlases generate hypotheses; fate engineering requires causal nodes - TFs, enhancers, signaling pathways, and metabolic constraints whose manipulation predictably shifts lineage outcomes. The strongest validation paradigm combines: (1) Perturbation; (2) Single-cell readouts; and (3) Quantitative comparison to developmental references.

CRISPR in retinal organoids has become a practical route to causality because it enables controlled genotype perturbation in a human developmental context with scalable phenotyping. For example, gene-edited human epithelium stem cell retinal organoids demonstrate that loss of neural retina leucine-zipper (NRL) redirects photoreceptor fate toward S-cone-like programs, offering a direct causal lever over rod-vs-cone identity[58]. More recent organoid-based CRISPR studies continue to nominate timing and fate regulators (e.g., developmental genes influencing photoreceptor specification dynamics), reinforcing organoids as a testbed for when and how strongly a node must be modulated to move fate[59].

Beyond coding genes, retina-relevant functional genomics is increasingly extending to noncoding control. In RPE/choroid contexts relevant to AMD, enhancer connectome strategies coupled with functional reporter assays (including allele-specific STARR-seq in RPE cells) provide a template for testing whether candidate regulatory variants or enhancers actually modulate disease-relevant programs - an approach that can be adapted to fate-specification enhancers in retinal lineages[43].

Finally, perturbations are not limited to CRISPR: Culture chemistry, morphogen timing, and niche engineering (ECM composition, stiffness, inflammatory cues) are often the most clinically actionable levers. The key methodological recommendation for the field is to treat these as quantitative perturbations and to score outcomes by reference mapping to developmental atlases plus trajectory/fate-probability metrics - so that “better differentiation” becomes a measured improvement in trajectory congruence, fate stability, and microenvironmental compatibility rather than a qualitative marker panel[10] (Figure 2; Table 2).

Figure 2 The “atlas-to-engineering” workflow: From quantitative blueprints to validated regenerative therapies. This pipeline illustrates how multi-omic data is operationalized for fate engineering: (1) Reference maps (input): The integration of single-cell transcriptomics (single-cell RNA sequencing), dual-omic profiling (single nucleus assay for transposase-accessible chromatin sequencing), and spatial transcriptomics creates quantitative developmental blueprints; (2) Trajectory & fate inference: Computational frameworks such as RNA velocity and CellRank use these blueprints to define lineage directionality and probabilistic fate decision regions; (3) Regulatory node nomination: Inference of gene regulatory networks identifies the specific transcription factor and cis-regulatory element logic driving lineage restrictions; (4) Perturbation testbed: Candidate control nodes are causally validated using CRISPR-Cas9 or small-molecule screens within human retinal organoids to test their ability to shift fate; (5) Benchmarking & quality control: Engineered cells are evaluated against primary tissue references using multimodal metrics to ensure they meet composition and maturation envelopes; and (6) Translation: Validated protocols are deployed for cell replacement (with structural/functional endpoints) or in situ reprogramming. Critical caution (bottom): A minimal evidence standard is required to distinguish true regeneration from artifacts, necessitating genetic lineage tracing, orthogonal identity validation, and physiological proof-of-function to rule out viral promoter leakage. QC: Quality control; scRNA: Single-cell RNA; snATAC: Single nucleus assay for transposase-accessible chromatin; 3D: Three-dimensional; TF: Transcription factor; cCRE: Cis-regulatory element; dCas9: Catalytically dead Cas9; RPE: Retinal pigment epithelium; OCT: Optical coherence tomography; FAF: Fundus autofluorescence; MG: Müller glia.

Table 2 Atlas-to-engineering toolkit - multi-omics reference types, inference outputs, and what they enable experimentally.

Reference/method class	What it quantifies (output)	“Control objects” you can engineer	Best validation experiment	Key pitfalls you should flag in text
scRNA/snRNA atlases	Cell states; trajectories; GRNs	TF modules; lineage branch points	Perturb TFs; scRNA readout + reference mapping	Marker mimicry; stress-induced pseudo-states
Multiome (RNA + ATAC)	State + chromatin accessibility	Competence windows; enhancer permission space	Time-gated TF pulses aligned to accessibility shifts	Accessibility ≠ activity; batch effects
3D genome/enhancer-promoter maps	Regulatory architecture	Cis-regulatory nodes; enhancer hubs	dCas9 recruitment/CRISPRi to specific enhancers	Context dependence; cell-type specificity required
Spatial multi-omics	Niche-positioned states	Layer-aware targets; microenvironment coupling	Perturb niche cues + spatial readouts	Resolution limits; deconvolution artifacts
CellRank/fate probability	Decision regions; fate bias	“Threshold tuning” at branchpoints	Perturb node then compare fate probabilities	Velocity assumptions; sampling density
CellChat/LR inference	Niche signaling network	Immune/niche gating; permissive vs restrictive cues	Ligand blockade/receptor editing + readouts	LR inference is probabilistic, not causal
Cross-species mapping	Conserved vs divergent programs	Identify why mammalian competence is lost	Match intervention nodes across species	Orthology mismatch; latent space alignment bias
Reference mapping benchmarks	Congruence to fetal tissue	Quantitative maturity score	Iterative differentiation optimization loop	Overfitting to reference; missing rare subtypes

scRNA: Single-cell RNA; snRNA: Single-nucleus RNA; GRNs: Gene regulatory networks; TF: Transcription factor; ATAC: Assay for transposase-accessible chromatin; 3D: Three-dimensional; dCas9: Catalytically dead Cas9; CRISPRi: CRISPR interference; LR: Ligand-receptor.

Multilayered fate-control networks: From TFs to metabolism and inflammation

Retinal regeneration is best conceptualized as a coupled control system rather than a single-factor event. Fate outcomes are dictated by the convergence of instruction (lineage-determining TFs), permission (epigenetic competence), and execution (post-transcriptional regulation, proteostasis, metabolism, and the injury-immune niche)[10,60]. Human single-nucleus dual-omic (RNA + ATAC) atlases now resolve these layers at cell-type and state resolution, providing a blueprint for mechanism-anchored, benchmarkable engineering designs in organoids and eventually in vivo[11,61] (Figure 3).

Figure 3 The multilayered fate-control paradigm for retinal regeneration: From lineage sources to therapeutic outputs. This schematic illustrates the hierarchical framework governing retinal cell fate engineering. The lineage library: Distinct cell sources available for regeneration, ranging from developmental references (embryonic retinal progenitor cells) and endogenous candidates (Müller glia, ciliary marginal zone-like cells) to scalable manufacturing sources (human pluripotent stem cell-derived retinal pigment epithelium, organoids, and photoreceptors) (left). The multilayer control hierarchy: Successful fate engineering requires the convergence of three regulatory tiers: (1) Instruction (tier 1): Lineage-determining transcription factors (e.g., Atoh7, neural retina leucine-zipper, ASCL1) that provide combinatorial and time-gated programming logic; (2) Permission (tier 2): The epigenetic landscape (chromatin accessibility, histone modifications, 3D architecture) that defines competence windows and determines whether transcription factor instructions can access cis-regulatory elements; and (3) Execution (tier 3): The stabilization and maturation layer, comprising non-coding RNAs, translational/proteostatic control, metabolic conditioning (e.g., glycolysis/OXPHOS balance), and the immune microenvironment (e.g., inflammation phase control), which collectively ensure the functional completion of the transcriptional program (middle). Regenerative outputs & therapies: The application of this logic to specific clinical targets - retinal pigment epithelium repair, photoreceptor replacement, and retinal ganglion cell restoration - highlighting key translational bottlenecks such as synaptic integration, long-range axon guidance, and safety/purity release criteria (right). hPSC: Human pluripotent stem cell; RPE: Retinal pigment epithelium; RGC: Retinal ganglion cell; NRL: Neural retina leucine-zipper; ATAC: Assay for transposase-accessible chromatin; TFs: Transcription factors; miRNA: MicroRNA; lncRNA: Long non-coding RNA; circRNA: Circular RNA; ECM: Extracellular matrix; ER: Endoplasmic reticulum; CCR2+: C-C chemokine receptor type 2-positive; NF-κB: Nuclear factor kappa B; AAV: Adeno-associated virus; dCas9: Catalytically dead Cas9.

Transcription-factor layer: Lineage determinants and maturation drivers

The TF layer constitutes the retina’s core “instruction set”, orchestrating lineage branching, subtype specification, and maturation. A recurring principle is that TFs function as time-gated controllers whose outcomes depend on the accessible enhancer landscape and the developmental/injury state[11,62]. Accordingly, TF-based fate engineering is moving away from single “master regulators” toward staged, combinatorial logic that mirrors development: (1) Establishing competence; (2) Imposing lineage direction; (3) Refining subtype identity; and (4) Driving maturation.

Temporal logic in photoreceptor and inner neuron specification: Photoreceptor generation exemplifies temporal TF logic. Multi-omic benchmarking indicates that progression from RPCs to rods requires synchronized activation of TF modules alongside stage-specific enhancer opening[10]. This logic is validated by gene-edited organoids, where loss of NRL disrupts rod identity and biases cells toward an S-cone-like program, supporting a competence-dependent “fate-enforcing” role for lineage TFs[58]. For inner retinal neurons, enhancer logic is similarly decisive: Deletion of a remote Atoh7 enhancer disrupts RGC development and axonogenesis, demonstrating that fate depends on discrete cis-regulatory control rather than TF presence alone[63]. These findings collectively imply a practical engineering rule: A TF cocktail effective in one temporal window may fail if delivered outside the appropriate competence state or within a restrictive post-injury niche[64].

MG reprogramming - overcoming barriers via staged inputs: MG reprogramming presents a higher barrier because it must overcome reactive stabilization (gliosis), reopen closed neurogenic enhancers, and impose directional lineage programs. Accordingly, reported successes generally require staged TF combinations rather than single factors[65,66]. In vivo rodent studies suggest that pairing proneural competence drivers with RGC-associated TF programs can induce MG-derived cells showing RGC programming features at the marker/transcriptomic level[14]. Human-centric evidence is currently strongest in fetal-derived or organoid-derived MG contexts, where ASCL1-driven programming can induce neurogenic states[22]. Importantly, methodological artifacts [e.g., adeno-associated virus (AAV) promoter leakage and intermediate-state misinterpretation highlighted by polypyrimidine tract-binding protein 1 (PTBP1)-related controversies] mandate stringent validation; we therefore summarize an explicit evidence hierarchy and minimal validation criteria in Table 3[67,68]. High-throughput screening platforms may further optimize combinatorial timing by identifying small molecules that potentiate TF-driven reprogramming[69].

Table 3 Müller glia reprogramming “minimal experimental standard” - non-negotiable evidence hierarchy for conversion claims.

Evidence tier	What must be demonstrated	Minimum required controls	Readouts that count as “orthogonal”	Pass/fail interpretation rule	Typical artifact this prevents
Tier 1: Genetic lineage tracing	Converted neurons are MG-origin	MG-specific inducible CreERT2; quantify recombination efficiency	Reporter-independent validation	No tracing = claim not interpretable	Mis-assigned cellular origin
Tier 2: Vector specificity/promoter leakage control	Transgene expression is cell-type restricted	AAV-GFAP leakage tests; alternative promoters; no-virus controls	Spatial mapping of transgene vs cell identity	Leakage unresolved = conversion invalid	“Apparent conversion”
Tier 3: Single-cell identity triangulation	True fate switch vs stress mimicry	Injury-only vs intervention; batch controls	scRNA ± ATAC; stress signatures; GRN congruence	Marker-only = insufficient	Stress-induced pseudo-neurons
Tier 4: Morphology + protein-level confirmation	Neuronal morphology consistent	Blinded morphometrics; layer localization	Immunostaining + morphology metrics	Partial markers without morphology = weak	Marker contamination
Tier 5 Physiology + circuit-level function	Functional maturation & integration	Electrophysiology controls; synapse evidence	Patch clamp; stimulus responses; connectivity proxies	No function = not therapeutic	Immature “neuron-like” cells
Tier 6: Contextual reproducibility	Robust across injury paradigms	Excitotoxic vs mechanical injury	Same validation stack across models	Single-context only = fragile	Context-dependent artifacts

MG: Müller glia; CreERT2: Cre recombinase-estrogen receptor T2 fusion protein; AAV: Adeno-associated virus; GFAP: Glial fibrillary acidic protein; scRNA: Single-cell RNA; ATAC: Assay for transposase-accessible chromatin; GRN: Gene regulatory network.

Epigenetic layer: “Permission” and competence windows

The control object - from descriptive states to causal constraints: The epigenetic landscape - chromatin accessibility, histone modifications, DNA methylation, and three-dimensional (3D) genome architecture-functions as the molecular substrate of competence windows. It defines a “permission system” that regulates TF efficacy by controlling access to CREs and stabilizing promoter-enhancer communication[40,70]. Dual-omic atlases in human retina quantify the time-resolved coupling of expression and accessibility, explaining why identical TF inputs yield distinct lineage outputs depending on which enhancer repertoires are accessible and which repressive barriers are enforced[11]. By integrating RNA + ATAC with methylation and 3D genome data, the field increasingly resolves state-specific enhancer grammars and nominates epigenetic bottlenecks where lineage progression stalls despite inductive TF availability[71,72].

Mechanistic evidence - competence as an active gatekeeper: Accumulating evidence supports epigenetic competence as a causal gatekeeper rather than a correlative marker. Polycomb-associated repression (e.g., H3K27me3) and ATP-dependent chromatin remodeling enforce orderly progenitor progression; JARID2 and CHD4 exemplify regulators that prevent promiscuous or premature lineage activation[25]. DNA methylation can also be rate-limiting; in rod development, active demethylation upstream of canonical TF programs is required to unlock otherwise latent differentiation logic[73]. Together, these data justify treating competence windows as an engineerable control layer - barriers that must be actively lowered to permit differentiation - especially in adult/injured contexts where reactive programs can re-impose restrictive chromatin states[74,75].

From motif to mechanism - multi-omics bridges the TF-regulatory gap: Atlas resources enable a compact “motif-to-mechanism” workflow: (1) Nominate TFs and candidate CREs from trajectory-resolved maps; (2) Link elements to targets using multi-modal integration and 3D contacts; (3) Validate causality with perturbations (organoids and/or in vivo); and (4) Benchmark engineered states back to primary atlases to quantify convergence vs “maturation gaps”[40,74]. This enhancer-aware workflow converts noncoding features from descriptive annotations into actionable engineering logic and supports targeted epigenome editing strategies (e.g., locus-level priming) to enable TF execution with less global identity destabilization[76-79].

Noncoding RNA layer: Homeostasis buffers and switch-like regulators

The non-coding RNA (ncRNA) layer functions as a dual-purpose control system: It buffers transcriptional noise and defines switch thresholds for quiescence, stress responses, and neurogenic entry[80,81]. MicroRNAs (miRNAs) typically sharpen lineage boundaries by coordinating repression of cell-cycle and stress pathways[61,82]. Long ncRNAs can scaffold chromatin modifiers to regulate enhancer usage and competence[83]. Circular RNAs add translational utility because their stability supports sustained expression, making them plausible delivery chassis for durable trophic or differentiation cues[84,85].

From a translational standpoint, ncRNAs are most actionable as “trim controls” integrated with TF and epigenetic programming. Integrated small RNA-seq in human organoids indicates dynamic miRNA tuning in response to physiological stimuli[86]. In regenerative models, specific miRNAs (e.g., miR-18a) can modulate the gain of regenerative programs and the balance between progenitor amplification and photoreceptor differentiation[87]. Mechanistically, long ncRNAs such as Malat1 have been linked to regeneration-relevant signaling-epigenetic coupling (e.g., via Egr1)[88]. Engineered delivery platforms (including circular RNA-based expression) may provide durable execution-layer support when repeated protein dosing is impractical[89,90]. Because ncRNA perturbations can rewire broad networks, validation should combine cell-type-resolved transcriptomics with safety monitoring for unintended reactive states.

Translational control and proteostasis: The executive gatekeepers

The “execution” of fate - synaptic maturation, photoreceptor outer-segment maintenance, and long-range axon integrity - hinges on translational capacity and proteostasis, and can diverge from transcript-level identity[91]. Ribo-seq and integrative analyses demonstrate stage-specific uncoupling between mRNA abundance and translational efficiency across retinogenesis, implying that transcript-level markers can overestimate functional completion[92-95]. Executive-layer failure modes in degeneration include stress-triggered translational control (e.g., eIF2α phosphorylation with ATF4/CHOP activation) and downstream proteostasis collapse (autophagy/UPS/ER quality control), which can precede overt cell death in vulnerable lineages[96-100]. RNA-binding proteins provide specificity (e.g., Musashi family loss precipitating photoreceptor degeneration), while hubs such as mechanistic target of rapamycin complex 1 link nutrient sensing to differentiation timing, reinforcing the need for phase-appropriate translation scheduling[101,102].

Engineering levers include integrated stress response (ISR) tuning (e.g., ISRIB) and proteostasis reinforcement, but this layer is also a common source of artifactual “conversion” claims. The PTBP1 controversy illustrates why fate claims must rely on rigorous genetic lineage tracing and orthogonal identity validation rather than viral-promoter proxies alone[67,68]. We therefore recommend a minimal validation standard that pairs genetic fate mapping with multimodal identity readouts (transcriptomic/proteomic plus morphology/physiology) and explicit monitoring of execution-layer states (ISR/proteostatic flux).

Metabolic cues: Bioenergetic state as a fate and maturation driver

The control object - metabolism as an upstream determinant: Metabolism is not merely downstream support but an upstream constraint on fate fidelity and maturation because it sets redox tone, metabolite-sensitive signaling/chromatin modifications, and energetic capacity for high-cost outputs such as outer-segment biogenesis and axon growth[103-105].

From correlation to causality - glycolytic flux and mitochondrial remodeling: Causal perturbation studies in organoids indicate that early glycolytic flux can act as a required signaling node in retinogenesis, with lactate functioning as a regulatory signal rather than a waste product[106]. Flux reprogramming can also bias progenitors toward rod fates via pathway coupling (e.g., Wnt-linked coordination of proliferation timing and maturation outputs)[107]. As lineages progress, mitochondrial quality control becomes central; PINK1/PRKN-dependent remodeling has been reported as a scheduled event relevant to RGC differentiation[108].

Maturation and microenvironment: “Maturation gaps” in engineered systems often reflect environmental deficits rather than transcriptional failures. Physiological oxygen gradients and appropriate ECM cues can improve inner/outer retinal phenotypes and outer-segment features in organoid systems[109,110]. Single-cell metabolic analyses further suggest that mature cell types occupy distinct metabolic regimes, reinforcing the need for lineage-appropriate metabolic environments[111].

Engineering levers - metabolic conditioning for regeneration: A practical design rule is that transcriptional programming initiates identity, but bioenergetic engineering enables execution and durability. Metabolic conditioning (oxygen/ECM/flux shaping) should therefore be staged alongside TF and epigenetic interventions[107,110,112,113]. In injury contexts, local glycolysis can be required for axonal regrowth, and NAD-centered support is mechanistically aligned with RGC vulnerability and salvage-pathway enrichment[114,115].

Inflammatory and injury-microenvironment layer: Fate bias under damage

The control object - injury as a restrictive state space: In adult retina, injury rapidly reconfigures tissue into an inflammatory control state that biases fate toward reactive stabilization (gliosis, scar-like remodeling, immune recruitment) and actively suppresses productive neurogenesis. This restrictive state space reflects interacting variables: Reactive MG programs, innate immune composition (microglia vs infiltrating monocytes), and the cytokine/ECM milieu.

Mechanistic resolution - from signaling hubs to temporal paradoxes: Single-cell ligand-receptor analyses identify nuclear factor kappa B (NF-κB) as a key glia-centered hub activated by microglia-derived signals to enforce a reactive, non-neurogenic MG state; inhibiting this axis can unlock neurogenic potential in TF-driven paradigms, establishing causality[116]. Comparative and cross-stage omics further indicate a temporal paradox: Inflammatory signaling may be permissive for early plasticity/proliferation yet restrictive for neurogenic completion, motivating sequential regimens (priming → resolution) rather than blanket immunosuppression[117,118]. Niche composition also matters; C-C chemokine receptor type 2-positive monocyte infiltration can suppress MG-derived neurogenesis, supporting “niche editing” as a practical lever[119].

Engineering levers - microenvironment calibration and phase control: These insights support microenvironment calibration: Target fate-restrictive hubs (e.g., NF-κB/transforming growth factor-β/Notch) to prevent terminal reactive locking while implementing phase control that permits early activation and fosters later maturation[116,117]. Success should be judged by fate completion and integration-level readouts, not by suppression of damage markers alone.

Integrated synthesis: A coupled multilayer model and the “minimal controllable node set”

The convergence of instruction, permission, and execution: Across multi-omics and regeneration studies, a convergent design rule is that durable fate change is most reproducible when interventions are combinatorial, time-gated, and microenvironment-aware[120,121]. TF instruction is effective only within an epigenetic permission space and must be sustained by execution-layer stability (translation-proteostasis, metabolism, and immune niche).

Empirical validation - lessons from reprogramming and immune gating: Empirical data from in vivo reprogramming and immune gating support this coupled logic; to avoid redundancy, we consolidate the inflammation timing and niche-composition rules in “Inflammatory and injury-microenvironment layer: Fate bias under damage” and Table 3.

The “minimal controllable node set” - evidence base, validity envelope, and insufficiency conditions: To translate these biological insights into an implementable design scaffold for retinal regeneration, we propose a context-dependent “minimal controllable node set” (Table 4). Importantly, we do not present this as a universal recipe or a sufficient condition for success. Rather, it is an operational starting panel motivated by convergent evidence across three coupled layers - instruction, permission, and execution - and intended to reduce design degrees-of-freedom and improve interpretability. In this section, we explicitly distinguish descriptive evidence (single-cell/spatial atlases that nominate candidate nodes) from causal evidence (perturbation studies demonstrating necessity/sufficiency in defined contexts). Recent human retina multiome resources quantify state-linked chromatin accessibility constraints and nominate cell-type-specific regulatory nodes, supporting the measurability of “competence windows” but not, by themselves, proving causal controllability (e.g., adult retina scRNA + ATAC atlases; developing dual-omic trajectories)[11,60].

Table 4 “Minimal controllable node set” as an operational scaffold for more reproducible, mechanism-anchored fate engineering - modules, targets, tools, and success metrics.

Module	Engineering objective	Representative control nodes mentioned	Implementable tools (examples you already cite)	Timing logic	Primary success metrics	Typical failure modes
TF intent	Specify lineage/subtype + maturation	Combinatorial TF designs; staged programming	AAV timed expression; programmable delivery platforms	Competence induction → commitment → maturation	Fate fraction + subtype markers + functional readiness	Single-factor insufficiency; wrong temporal window
Epigenetic permission	Open required enhancer repertoire	Enhancer priming; cis-regulatory logic; 3D genome nodes	dCas9-based recruitment/CRISPRi; enhancer-first interventions	Must precede/overlap TF pulses	ATAC congruence; motif availability; reference-mapped maturity	Global de-repression; non-specific dedifferentiation
Microenvironment calibration	Prevent reversion; support integration	NF-κB; monocyte infiltration (CCR2+); metabolic/ECM/oxygen tuning	Immune phase control; niche editing; metabolic conditioning	Permissive inflammation early → resolution late	Stability over time; reduced gliosis; integration-level readouts	Reactive reversion; inflammatory bottleneck; stress collapse

TF: Transcription factor; AAV: Adeno-associated virus; 3D: Three-dimensional; dCas9: Catalytically dead Cas9; CRISPRi: CRISPR interference; ATAC: Assay for transposase-accessible chromatin; NF-κB: Nuclear factor kappa B; CCR2+: C-C chemokine receptor type 2-positive; ECM: Extracellular matrix.

Evidence anchor for module 1 - TF intent (instruction): Causal perturbations in human systems and in vivo models support that lineage/subtype specification requires explicit TF logic and timing. For example, CRISPR-edited NRL^-/- human retinal organoids fail to establish rod identity and default toward S-cone-like photoreceptors, demonstrating that TF intent is a necessary controller of fate direction in a human-relevant system[58].

Evidence anchor for module 2 - epigenetic permission (competence/CRE accessibility): Multiple experimental lines show that TF intent is constrained by cis-regulatory architecture and competence barriers: (1) A remote enhancer deletion at the Atoh7 locus phenocopies key RGC developmental defects, underscoring that discrete cis-regulatory modules can be rate-limiting for executing a neuronal program[122]; (2) Human retina 3D genome/topology maps reveal retina-specific enhancer-promoter wiring and super-enhancer interactions, reinforcing that controllable nodes often reside in cis-regulatory/3D architecture rather than in TF abundance alone[11]; and (3) In adult mammalian MG, combined inhibition of Notch signaling and nuclear factor I (NFI) factors converts nearly all MG into neurons, providing strong causal evidence that “permission” gates can dominate fate outcomes even without changing the target tissue identity program exogenously[123].

Evidence anchors for module 3 - execution and stabilization (microenvironment/metabolism/translation): Fate initiation does not guarantee fate completion or durability. In vivo, inflammatory signaling can act as a suppressive execution-layer gate: NF-κB inhibition after damage enhances Ascl1-mediated MG-to-neuron reprogramming, establishing that immune-state structuring is a causal lever for conversion efficiency and stability[116]. In engineered human systems, microenvironmental and metabolic conditioning can accelerate or enable maturation outputs: Hyaluronan, a native interphotoreceptor-matrix component, improves photoreceptor maturation/outer-segment ultrastructure in human retinal organoids, and developmental studies show that glycolytic flux can be instructive for photoreceptor fate/maturation via signaling coupling (e.g., glycolysis-Wnt axis)[107,110].

Validity envelope (when this scaffold is expected to be useful): We expect the “minimal node set” to be most informative when: (1) The target cell population is within a definable competence state (developmental/organoid trajectories or early injury windows); (2) The intended outcome is fate conversion and/or maturation that can be benchmarked by multimodal identity and function readouts; and (3) The host niche is either permissive or explicitly engineered (immune-state and metabolic/ECM conditioning included).

Insufficiency conditions (where additional nodes are required): This scaffold is necessary but often insufficient for: (1) Severely remodeled adult human degenerative niches (e.g., chronic gliosis/scarring and immune remodeling); and (2) Restoration of long-range circuitry, particularly RGC replacement, which additionally requires axon guidance/target selection and circuit-level reconnection beyond cell-intrinsic fate programming. Experimental regeneration studies in the adult visual system show that manipulating axon guidance mechanisms (e.g., Slit/Robo) is required to achieve correct target innervation during regeneration, illustrating that circuit repair adds control requirements outside the minimal fate-control set[124]. Accordingly, Table 4 should be read as a mechanism-anchored starting panel that organizes causal levers by layer and timing, while remaining empirically falsifiable in each target lineage, disease stage, and host microenvironment.

REGENERATIVE STRATEGY I: PROGRAMMABLE DEVELOPMENT IN THE ORGANOID ERA AND TISSUE-ENGINEERING OUTPUTS

Organoids as a scalable developmental proxy and the quality control imperative

Retinal organoids have transformed regenerative ophthalmology by converting inaccessible human developmental programs into a manipulable, scalable system where lineage decisions and failure modes can be interrogated at cell-type resolution. This platform offers a distinct value proposition over rodent models by capturing human-specific developmental logic - particularly macular-leaning programs - while remaining compatible with manufacturing-style process control and gene editing. However, the utility of organoids as “experimentable carriers” is constrained by systematic biases, including self-organization-driven heterogeneity, the absence of perfusion or immune ecology, and late-stage maturation bottlenecks[125]. Accordingly, organoid platforms are increasingly treated as Quality-by-Design systems in which multi-omics defines measurable critical quality attributes (CQAs) and guides process optimization rather than serving only as descriptive endpoints.

To bridge the gap between stochastic self-organization and reproducible manufacturing, we operationalize multi-omics as a Quality-by-Design tool rather than an academic endpoint. Recent retinal organoid single-cell studies show that organoid-to-organoid heterogeneity and off-target tissue contamination are pervasive and can be quantified at scale, enabling “composition envelopes” to detect drift and to rationally optimize patterning conditions[126]. Beyond superficial marker checks, defensible quality control (QC) requires quantifying: (1) On-target cell-type stoichiometry; (2) Developmental congruence; and (3) Execution-layer liabilities (stress/reactive programs) in a manner that can be translated into CQAs.

Two-tier, omics-enabled QC/chemistry, manufacturing, and controls (CMC) bridge. We distinguish deep multi-omics for process development and periodic lot qualification from minimal targeted assays for routine Good Manufacturing Practice (GMP) lot release. In practice, the essential omics criteria for clinical translation are those needed to define and periodically re-qualify CQAs: (1) scRNA-based cell-type composition and off-target detection (composition envelopes); (2) Reference-mapping maturity/trajectory scores against primary human retina atlases to quantify developmental congruence; and (3) Stress/reactive signatures (e.g., hypoxia/ISR/gliosis) that predict execution-layer failure. Optional/adjunct omics (e.g., single-cell ATAC/multiome, spatial transcriptomics, proteomics/metabolomics) can refine mechanism and comparability but are not required for every release lot[125]. Routine release testing should therefore focus on validated identity/purity/potency/safety assays derived from the above omics learnings and aligned with regulatory expectations (Table 3). Within this framework, electrophysiological platforms and light-response assays are best positioned as potency assay development/validation and periodic qualification tools, while routine release can rely on a reduced, validated surrogate potency panel that is mechanistically anchored and practical for manufacturing-scale deployment[127].

Closing the loop: From stochastic self-organization to pathway control

The maturation of organoid technology enables a paradigm shift from “optimization by intuition” to control-by-measurement. Developmental signaling pathways (e.g., Wnt, bone morphogenic protein, Notch, Shh) act as programmable inputs that can be systematically perturbed to steer lineage outcomes. This “perturbation-readout loop” allows for the precise shaping of cell yields; for instance, high-throughput analysis has demonstrated that early, timed Wnt activation significantly enhances specific retinal cell-class yields weeks later[128]. Similarly, rigorous bone morphogenic protein-timed patterning has been employed to stabilize retinal sheet outputs for transplantation[129]. The methodological implication is that organoid development should be treated as a deterministic system where timed pathway inputs are coupled with quantitative single-cell readouts, allowing for iterative, model-driven refinement that aligns biological programming with the reproducibility standards required for clinical translation.

Structural and micro-physiological engineering: Scaffolds, chips, and ecosystems

For therapeutic applications, the physical and physiological context of the graft is as critical as cellular identity. Engineering efforts are currently advancing along two synergistic fronts: Structural reconstruction and micro-physiological simulation.

Structural engineering: Recognizing that architecture is intrinsic to therapy, the field has evolved from simple cell suspensions to laminated sheet products and bio-scaffolded constructs. Recent innovations employ ultrathin, micromolded “ice-cube-tray” scaffolds to enforce high-density photoreceptor polarity and precise layer reconstruction, addressing the geometric limitations of organoid-derived grafts. Bioactive scaffold designs are further refining this approach by spatially controlling the placement of photoreceptors relative to supporting epithelium, thereby pre-assembling the necessary cellular adjacencies for integration[130,131].

Micro-physiological systems: To overcome the “microenvironmental missingness” of static culture, perfusion and organ-on-chip systems have been developed to impose controlled oxygen gradients and nutrient transport. These platforms directly mitigate hypoxia-linked stress programs and support long-term maturation, serving as critical conditioning tools prior to transplantation[109,132]. Simultaneously, the “cellular ecosystem” is being completed by incorporating vascular and immune components. The co-culture of microglia-like cells and endothelial networks within organoids not only models neuro-immune interactions relevant to degeneration but also generates “integration-permissive” grafts that better withstand the inflammatory host environment[125,133].

Ultimately, the metric of success for these engineered constructs is functional integration. The field is converging on a standardized evaluation stack that demands ultrastructural evidence of host-graft synaptic connectivity and circuit-level participation - such as graft-mediated light responses - rather than relying solely on survival or gene expression[134,135]. This integrates conceptually with corrective workflows, where genome-edited, patient-derived organoids serve as both the testbed for molecular rescue and the substrate for functional replacement[37].

REGENERATIVE STRATEGY II: CLINICAL PATHWAY AND EVIDENCE HIERARCHY FOR CELL-REPLACEMENT THERAPY

RPE replacement: Delivery modalities and the multimodal endpoint architecture

The RPE represents the vanguard of ocular cell therapy, owing to its well-defined monolayer identity and accessibility for longitudinal imaging. Over the past five years, the field has progressed from establishing safety in small, non-randomized cohorts to refining delivery modalities and standardizing endpoints. However, the clinical evidence base remains predominantly early-phase and feasibility-oriented; primary objectives are typically safety/tolerability and feasibility of administration, and efficacy inferences are therefore exploratory and limited[136] (Table 5).

Table 5 Clinical pathway + endpoint stack + chemistry, manufacturing, and controls interface for cell-replacement therapy.

Clinical modality	Delivery plane/format	Biological trade-off	Recommended endpoint stack (early phase)	CMC release criteria anchor (identity/purity/potency)	Major confounders to control	Evidence level framing in your text
RPE - cell suspension	Subretinal injection	Scalable + simpler surgery; risks non-uniform monolayer	OCT graft coverage; FAF atrophy expansion; microperimetry/dark adaptation	Identity: Polarity markers; purity: Residual pluripotency; potency: Phagocytosis + TEER	Bruch’s membrane integrity; atrophic niche variability	Level II-III feasibility; signals not definitive
RPE - patch/scaffold	Sheet/patch implant	Pre-polarized monolayer; higher surgical complexity	Same stack + implant positioning stability	Same axes + mechanical integrity (implant)	PVR/retinal detachment; immune response modulation by scaffold	Durability-oriented but complication-prone
RPE - strip (hybrid)	“Strip” transplantation	Balances handling vs structure	Same stack; add uniformity metrics	Same axes	Surgical learning curve; comparability issues	Emerging modality
Photoreceptor precursors	Subretinal; precursor “Goldilocks” maturity	Requires synaptic integration; risk of material transfer illusion	Structural survival + layer targeting; function with integration-sensitive assays	Potency: Light-response surrogates + integration proxies	Disease stage; host residual photoreceptors	First-in-human transition; mechanism-of-benefit controversy
Immunology-as-engineering adjunct	Immunosuppression; scaffold immune barrier; hypoimmune iPSC	Enable durable allografts; must avoid immune escape	Safety surveillance; inflammation markers; imaging	Purity + proliferation markers; in vivo tumor surveillance	Immune privilege is relative	“Universal donor” direction but needs safeguards

CMC: Chemistry, manufacturing, and controls; RPE: Retinal pigment epithelium; OCT: Optical coherence tomography; FAF: Fundus autofluorescence; TEER: Transepithelial electrical resistance; PVR: Proliferative vitreoretinopathy; iPSC: Induced pluripotent stem cell.

The delivery dilemma: Suspension vs scaffold. Clinical strategies have bifurcated into two primary approaches, each with distinct biological trade-offs. Cell suspension delivery (subretinal injection) offers surgical simplicity and scalability. However, it faces predictable biological constraints: Heterogeneous cell distribution, imperfect repolarization on diseased Bruch’s membrane, and variable survival within atrophic niches. The phase 1/2 RPESC-RPE-4W study illustrates this approach, reporting an acceptable safety/tolerability profile in early cohorts, while also highlighting that conclusions about comparative efficacy and durability remain constrained by cohort size and study design[12]. Conversely, scaffold-based implants (sheets/patches) aim to deliver a pre-polarized, mature monolayer, providing a surrogate substrate that bypasses the compromised host Bruch’s membrane. Available follow-up data support surgical feasibility and graft persistence, but the extent of sustained functional benefit remains to be established in adequately controlled settings; any observed immunomodulatory advantages of scaffolds should be regarded as preliminary and hypothesis-generating[137]. A third, hybrid modality - “strip” transplantation - has recently emerged using iPSC platforms, aiming to balance the surgical manageability of suspensions with the structural integrity of patches[138].

The endpoint stack. Given that early-phase trials are rarely powered for definitive visual acuity superiority, the field is converging on a multimodal endpoint architecture to explore potential treatment signals. This “stack” integrates structural metrics (optical coherence tomography-based integrity, graft coverage), atrophy dynamics (fundus autofluorescence-based expansion rates), and functional correlates (microperimetry, dark adaptation). This triangulation can support exploratory structure-function concordance analyses - such as graft persistence aligning with stabilized sensitivity trends - yet functional endpoints require cautious interpretation because psychophysical tests (including microperimetry) can show strategy heterogeneity and measurement variability, particularly across small early cohorts[12] (Figure 4).

Figure 4 Regenerative ophthalmology strategies for retinal degeneration: Stage-matched, multimodal fate engineering. This schematic synthesizes the clinical pathways for retinal fate engineering across three distinct therapeutic lanes, aligned with disease progression. Outer retina replacement: Strategies for RPE and photoreceptor replacement (discussed in “Regenerative strategy II: Clinical pathway and evidence hierarchy for cell-replacement therapy”), detailing delivery modalities (suspension vs scaffold) and the requirement for a multimodal endpoint stack (optical coherence tomography, fundus autofluorescence, microperimetry) alongside rigorous chemistry, manufacturing, and controls release criteria (identity, purity, potency) to ensure graft function (lane 1). Inner retina reprogramming: In situ Müller glia reprogramming (discussed in the section of Regenerative strategy III: MG reprogramming - from “feasible” to controllable, reproducible, and translatable) is depicted as a staged control sequence: Unlocking competence (e.g., via Notch/nuclear factor I inhibition), identifying lineage instruction (ASCL1), and calibrating the immune niche. A critical warning highlights the “minimal experimental standard” required to rule out viral leakage artifacts (lane 2). Multimodal combinations: “Gene + cell” and circuit-bypass strategies (discussed in “Gene + cell and multimodal combination strategies: Turning fate-control networks into therapies”) that couple biological replacement with host microenvironment conditioning or optogenetic re-entry (lane 3). Patient stratification: The timeline illustrates the necessity of precision stage-matching - deploying reprogramming for early-stage rescue, cell replacement for intermediate degeneration, and circuit-bypass for end-stage atrophy - guided by artificial intelligence-driven progression prediction (bottom). RPE: Retinal pigment epithelium; OCT: Optical coherence tomography; FAF: Fundus autofluorescence; CMC: Chemistry, manufacturing, and controls; QC: Quality control; MG: Müller glia; TF: Transcription factor; NFI: Nuclear factor I; Prox1: Prospero homeobox 1; NF-κB: Nuclear factor kappa B; CCR2+: C-C chemokine receptor type 2-positive; RGC: Retinal ganglion cell; AI: Artificial intelligence.

Photoreceptor replacement: The maturity Goldilocks zone and integration controversy

Photoreceptor replacement is moving from advanced preclinical validation toward first-in-human safety/feasibility evaluation, including the 2024-2025 initiation of phase 1/2a studies (e.g., NCT06789445)[139]. As in other early-phase programs, efficacy readouts should be treated as exploratory. Unlike the RPE, photoreceptor benefit is expected to depend on synaptic incorporation, leading to a critical consensus on donor maturity: Post-mitotic precursors - rather than fully mature cells or multipotent progenitors - appear to occupy the “Goldilocks zone”, balancing survival, migratory capacity, and synaptic potential[140].

The mechanism-of-benefit controversy. A central translational challenge is distinguishing true synaptic integration from cytoplasmic material transfer. Recent murine studies have shown that donor-host material exchange can mimic “integration” signals (e.g., fluorescent reporter transfer) without establishing functional connectivity. This controversy necessitates a rigorous interpretation of early clinical readouts: Functional changes observed in early cohorts may reflect neurotrophic rescue, material transfer, or altered host physiology rather than circuit reconstruction[23]. Accordingly, clinical trial designs should stratify by disease stage and incorporate integration-sensitive endpoints. In end-stage disease (minimal host photoreceptors), functional gain is more plausibly donor-driven; in mid-stage disease, benefits are more likely confounded by paracrine rescue or material transfer. Preclinical models employing human cone transplantation in end-stage degeneration support that, when donor state and niche variables are well controlled, transplantation may yield measurable circuit-relevant signals, but these results should not be over-extrapolated to heterogeneous human disease settings[140].

Immunology as engineering: From privilege to universal donors

The concept of “immune privilege” in the subretinal space is relative, not absolute. Most allogeneic programs require calibrated immunosuppression, and the implant itself is increasingly viewed as an immune-engineering tool; evidence suggests that bioengineered scaffolds can function as physical barriers that limit antigen dispersion and modulate local immune responses[141].

Simultaneously, the field is advancing toward “universal” donor strategies via two parallel tracks: HLA-homozygous iPSC banking and gene-edited hypoimmune platforms. A 2024 non-human primate study reported long-term survival of hypoimmune iPSC-derived cells in immunocompetent recipients, supporting feasibility for “off-the-shelf” concepts[142]. Nevertheless, the clinical significance of these approaches must be established in human trials, and immune-evasion designs necessitate rigorous safeguards against tumorigenicity and unintended persistence. Accordingly, clinical programs increasingly incorporate explicit release criteria for proliferation/residual pluripotency markers and implement stringent in vivo surveillance protocols to detect ectopic differentiation or tumor formation[12].

Manufacturing is efficacy: The CMC-clinical interface

In cell therapy, manufacturing consistency is a primary determinant of clinical outcome. Variability in differentiation efficiency, purity, or cryopreservation viability can dominate efficacy signals as much as the biological target itself. Therefore, the regulatory pathway now treats GMP and release criteria as core components of the “efficacy” definition[136].

The release criteria framework. A defensible release strategy integrates three orthogonal axes: (1) Identity, ensuring marker profiles match the intended lineage (e.g., RPE polarity); (2) Purity, strictly limiting residual pluripotent cells and off-target lineages; and (3) Potency, utilizing mechanism-relevant bioassays - such as outer segment phagocytosis and transepithelial electrical resistance for RPE, or light-response surrogates for photoreceptors. As exemplified by the RPESC-RPE-4W trial, explicitly linking these QC outputs to clinical dosing is essential for interpretability[12]. Furthermore, because adverse events like fibrosis or ectopic proliferation may be delayed, regulatory consensus now mandates multi-year follow-up protocols governed by standardized safety oversight, ensuring that the long-term behavior of these living drugs is transparently mapped[136].

From atlas metrics to release specifications. Regulatory guidance consistently anchor release specifications in identity, purity, potency, sterility, viability and cell number, while allowing product-specific justification for additional CQAs. We therefore position multi-omics as a CQA-discovery and periodic qualification layer (especially to manage drift, donor/Line variability, and manufacturing changes), whereas routine lot release relies on a reduced, validated assay panel derived from those omics learnings. This lifecycle approach aligns with potency-focused guidance and comparability expectations when process changes occur[110,143-146] (Table 6).

Table 6 Two-tier, omics-enabled quality control framework linking multi-omics resources to Good Manufacturing Practice release criteria for retinal organoids and organoid-derived products.

QC domain (CQA)	Essential for clinical translation?	Essential multi-omics criteria (process dev/periodic lot qualification)	Minimal routine GMP lot-release panel (examples; targeted assays)	Engineering → CMC translation output	Key evidence/precedent	Ref.
Identity & composition (intended lineage; correct cell-type stoichiometry)	Yes (CQA definition + re-qualification)	scRNA-seq cell-type composition; “composition envelope” across lots; reference-mapping to human retina atlas (maturity/trajectory score)	Targeted marker panel (flow/qPCR/IF): Lineage markers + subtype markers; morphology metrics where applicable	Converts atlas cell states into measurable CQAs; derives reduced marker sets; flags drift early	Retinal organoids vs adult retina single-cell reference; organoid heterogeneity quantified at scale	[143]
Purity/off-target tissues (non-retinal CNS, mesenchymal, RPE contamination etc.)	Yes	scRNA-seq off-target fraction; stress/reactive state detection (hypoxia/ISR/gliosis modules)	Release: Residual pluripotency (OCT4/TRA-1-60/NANOG) negative; proliferation (Ki67) limits; off-target marker negatives; viability & total cell number	Sets acceptance criteria for “allowed impurities”; links drift to process parameters (media/patterning/selection)	Organoid variability across systems supports need for systematic QC	[144]
Maturity/developmental congruence (photoreceptor/RPE functional readiness)	Yes for qualification; not per-lot mandatory	Reference mapping to fetal/adult retina trajectories; optional scATAC/multiome for competence state; optional spatial for lamination	Release: Maturity-linked targeted markers (e.g., phototransduction/synaptic readiness proxies) + predefined in-process timepoints	Defines “Goldilocks” maturity window; prevents under-/over-mature lots	HA conditioning improves photoreceptor maturation and uniformity (supports measurable maturation CQAs)	[110]
Potency (mechanism-linked biological activity)	Yes	Omics used to select potency mechanisms (pathway engagement signatures) and to justify assay choice; optional proteomics/metabolomics to connect transcript → function	Release: Validated potency assay(s) aligned to mechanism (e.g., RPE phagocytosis/TEER; photoreceptor light-response surrogates + integration-sensitive proxies)	Bridges omics biomarkers → potency assay design; supports assay justification	FDA potency guidance emphasizes mechanism-linked potency tests; lifecycle potency assurance	[145]
Safety/genetic stability (tumorigenicity risk, genome integrity)	Yes	Genomic characterization strategy (karyotype/CNV; WCB/MCB characterization); optional WGS where justified	Release: Sterility/mycoplasma/endotoxin; viability; residual pluripotency negative; proliferation limits; stability post-thaw	Links bank characterization to release and long-term follow-up	Cell substrate characterization expectations; ATMP quality requirements in trials	[146]
Comparability (manufacturing changes)	Yes when changes occur	Re-map lots with scRNA composition + maturity score; stress signature comparison; optional multiome/spatial if MoA-critical	Release: Same validated panel + bridging study endpoints	Provides quantitative “sameness” evidence after changes	FDA comparability guidance for CGT products	[162]

QC: Quality control; CQA: Critical quality attribute(s); scRNA-seq: Single-cell RNA sequencing; GMP: Good Manufacturing Practice; CMC: Chemistry, manufacturing, and controls; qPCR: Quantitative polymerase chain reaction; IF: Immunofluorescence; ISR: Integrated stress response; OCT4 Octamer-binding transcription factor 4 (POU5F1); TRA-1-60: Tumor rejection antigen 1-60; Ki67: Marker of cellular proliferation (Ki-67 antigen); CNS: Central nervous system; RPE: Retinal pigment epithelium; scATAC: Single-cell assay for transposase-accessible chromatin; HA: Hyaluronic acid; TEER: Transepithelial electrical resistance; FDA: Food and Drug Administration; CNV: Copy number variation; WCB: Working cell bank; MCB: Master cell bank; WGS: Whole-genome sequencing; ATMP: Advanced therapy medicinal product; CGT: Cell and gene therapy; MoA: Mechanism of action.

Regenerative strategy III: MG reprogramming - from “feasible” to controllable, reproducible, and translatable

The barriers to regeneration - quiescence locks and the inflammatory microenvironment: While MG in regeneration-competent species like zebrafish readily re-enter the cell cycle to repopulate neuronal lineages, mammalian MG default to reactive gliosis - a stabilizing response that actively suppresses neurogenic competence. A crystallization of recent findings reveals that mammalian MG are not deficient in neurogenic potential but are “permission-restricted” by parallel inhibitory circuits. Genetic dissection has identified the Notch pathway and NFI family as primary gatekeepers; disabling these brakes can unlock broad MG-to-neuron conversion capacity even in the adult retina[123]. Furthermore, non-cell-autonomous mechanisms impose additional layers of repression. A particularly instructive example is Prox1, which accumulates in MG via intercellular transfer from neurons; disrupting this transfer restores regenerative potential and delays vision loss in retinitis pigmentosa models, reframing “incompetence” as a targetable, microenvironment-mediated barrier[24].

Critically, the injury microenvironment further biases fate by coupling MG state transitions to inflammatory signaling. NF-κB has emerged as a central reactivity hub that promotes gliosis while suppressing neurogenesis; its inhibition post-damage enhances Ascl1-mediated reprogramming, establishing a direct link between inflammation and fate restriction[116]. Similarly, the infiltration of C-C chemokine receptor type 2-positive monocytes into the injured retina acts as a potent negative regulator of regeneration[119]. Consequently, effective reprogramming strategies must be conceptualized as a coupled system: Co-designing “competence unlocking” (targeting Notch/NFI/Prox1) with microenvironment calibration (modulating NF-κB/monocytes).

Engineering fate: The logic of combinatorial and staged programming

Consistent with developmental principles, successful in vivo reprogramming demands a combinatorial and staged approach rather than a single-factor “switch”. The core control objective is tripartite: (1) Competence Induction to escape glial identity; (2) Lineage direction toward a specific neuronal class; and (3) Maturation stabilization.

In adult mouse retina, AAV-mediated delivery of proneural bHLH factor combinations - a “cocktail logic” - has proven superior to single factors in stimulating regenerative responses[65]. Building on this, strategies applying developmental TF logic have reported the generation of ganglion-like cells, illustrating that lineage direction is feasible but requires precise factor tuning[22]. However, a major advance toward controllability is the recognition that TF instruction fails if competence gates remain closed. The mechanistic template provided by the simultaneous disabling of Notch/NFI factors demonstrates that removing brakes first allows endogenous machinery to execute neuronal differentiation at scale[123]. This logic extends to human contexts, where ASCL1 can induce neurogenic programs in human MG, though maturation remains a bottleneck[22,147]. Thus, the prevailing design principle is hierarchical: Competence gates determine whether TF instruction is executable.

The PTBP1 warning and a minimal experimental standard

The PTBP1 controversy represents a methodological inflection point for the field. Initial reports suggesting that Ptbp1 knockdown could drive efficient glia-to-neuron conversion were subsequently overturned by rigorous genetic lineage tracing and loss-of-function studies, which demonstrated a lack of genuine fate switching[67]. This epistemological crisis, compounded by the identification of widespread promoter leakage in AAV-glial fibrillary acidic protein systems - where neuronal expression artifacts masquerade as conversion - has necessitated the establishment of a “minimal experimental standard” to govern future claims[148].

Validation must now proceed through a non-negotiable hierarchy of evidence. First, genetic lineage tracing using inducible systems (e.g., MG-specific CreERT2) with quantified recombination efficiency is required to unambiguously track origin. Second, this must be coupled with rigorous vector specificity controls to rule out transgene-dependent promoter leakage. Third, cellular identity must be triangulated via “orthogonal evidence” - combining single-cell multi-omics to exclude stress-marker mimicry with morphological and protein-level confirmation. Finally, phenotype validation must extend beyond markers to electrophysiology and circuit-level readouts matched to the target fate. Crucially, these findings must demonstrate contextual reproducibility across distinct injury paradigms (e.g., excitotoxic vs mechanical injury) to distinguish robust reprogramming from context-dependent artifacts (Table 3).

Translational bottlenecks: From efficiency to integration

Translating MG reprogramming into clinical reality requires solving five interlinked bottlenecks that span from cellular specification to systemic integration.

The cellular challenge - efficiency, precision, and stability: The first hurdle is balancing efficiency with bounded proliferation; while unlocking competence (e.g., via Notch/NFI modulation) drives high conversion rates, this expansion must be strictly controlled to prevent maladaptive scarring or tumorigenesis[123]. Simultaneously, cell-type precision remains an obstacle; generating bipolar/amacrine-like cells is relatively accessible, whereas directing RGC-like fates is constrained by the complex requirements for subtype specification and long-range axon guidance[22]. Furthermore, the long-term stability of these engineered fates is constantly opposed by reactive reversion programs, often necessitating active inhibition of inflammatory hubs such as NF-κB to prevent transcriptional drift[116].

The systemic challenge - integration and delivery: Beyond cellular identity, the ultimate clinical endpoint is circuit integration - restored vision requires not just the presence of neurons, but synaptogenesis and target engagement. This must be achieved via safe delivery platforms; AAV vectors require exquisite engineering to ensure specificity and avoid off-target immune activation[149-151].

Consequently, the field is converging on modular “closed-loop” designs that integrate competence unlocking, fate instruction, and niche calibration. Recent evidence implicating peripheral immune cells as gatekeepers further underscores that immunomodulation is not merely an adjunct but an enabling module for regenerative success[119].

GENE + CELL AND MULTIMODAL COMBINATION STRATEGIES: TURNING FATE-CONTROL NETWORKS INTO THERAPIES

Combination logic: Integrating specification, maturation, and niche calibration

A translational framework for retinal regeneration necessitates moving beyond isolated cellular programming to a tripartite engineering strategy that co-designs: (1) Lineage specification; (2) Physiological integration; and (3) Host-niche calibration. This implies that transcriptional instructions must be paired with epigenetic “permissions” and an execution layer capable of sustaining metabolic fitness and synaptic connectivity under stress. Recent advancements in “gene + cell” therapies underscore this modular interdependence; for instance, the transplantation of genome-edited human retinal organoids into degenerated retinae has demonstrated that combining intrinsic genetic programming with extrinsic environmental coupling can restore fundamental visual responses[37]. This reinforces the paradigm that neither vectors nor cells alone suffice in late-stage degeneration; rather, success depends on the synergistic coupling of engineered grafts with a receptive host environment.

Beyond biological replacement, multimodal combinations offer alternative restorative pathways. The pairing of mutation-independent optogenetic gene therapy with wearable stimulation devices has enabled partial functional recovery in retinitis pigmentosa, establishing a “circuit re-entry” principle[152]. This suggests that when native photoreceptors are irretrievably lost, regeneration can be effectively reframed as bypassing the damaged hardware to reconstitute upstream computations. Furthermore, host-niche calibration must be elevated from an auxiliary consideration to a core therapeutic module. Clinical successes in complement inhibition for geographic atrophy - such as pegcetacoplan and avacincaptad pegol - demonstrate the feasibility of slowing lesion expansion[7]. Conceptually, repurposing such anti-progression agents as pre-conditioning or co-therapies could dampen inflammatory barriers to graft survival, provided the immunomodulation is tuned to support integration. Ultimately, the implementability of these logic circuits relies on delivery innovations. Transient, non-integrating systems (e.g., virus-like particles for epigenome editing) are strategically vital, as they allow for the time-gated unlocking of regenerative competence without the risks associated with constitutive ectopic expression[153].

Personalization and stratification: Precision matching of patients to modalities

The clinical utility of regenerative therapies will depend on a rigorous stratification framework that integrates omics-defined molecular states with multimodal clinical phenotyping. Rather than a “one-size-fits-all” approach, interventions must be stage-matched: Early disease warrants rescue or reprogramming; intermediate stages may support partial integration; while advanced degeneration necessitates circuit-bypass strategies or intensive replacement coupled with rehabilitation. In the context of AMD/geographic atrophy, deep learning algorithms applied to optical coherence tomography imaging have matured sufficiently to predict individual progression trajectories and distinguish natural history from therapeutic effects[154]. Integrating these artificial intelligence-driven tools into trial design will be critical for enriching patient cohorts likely to demonstrate measurable benefit, thereby enhancing the statistical power to detect true biological restoration.

For inner-retinal disorders like glaucoma, stratification must advance beyond intraocular pressure to incorporate cell-type vulnerability and microenvironmental context. Emerging high-resolution atlases now link specific RGC subtypes to distinct spatial niches and degeneration resilience, providing a molecular basis for region-aware targeting[155]. These resources enable a precision medicine logic where outcomes are measured against the specific vulnerability profiles of local neuronal populations. On the manufacturing front, personalization should be interpreted as the stratified deployment of standardized products rather than bespoke autologous production. Hypoimmune iPSC platforms that evade both adaptive and innate immune recognition in immunocompetent primates offer a viable path toward “universal donor” tissues[142]. Such off-the-shelf products, when paired with precise patient stratification, could significantly broaden therapeutic access while minimizing the burden of chronic immunosuppression.

High-value technology directions for the next 5 years

The translational trajectory of the next half-decade will be defined by the convergence of high-resolution mapping, standardized manufacturing, and rigorous validation. First, the field is transitioning from cell atlases to spatial multi-omics-driven “target-to-circuit” mapping. By quantifying how microenvironmental niches modulate neuronal vulnerability and synaptic connectivity, these spatial references define the precise coordinates for circuit-aware interventions, nominating molecular control nodes constrained to relevant anatomical layers[11]. This spatial logic must be matched by standardized organoid QC frameworks. To overcome batch heterogeneity, QC must evolve from simple marker expression to multimodal spatiotemporal benchmarking against human retinal tissue, ensuring that engineered grafts possess the requisite synaptic readiness and metabolic maturity for functional integration[10].

Finally, the translation of in vivo reprogramming demands a stricter evidence hierarchy to preclude artifacts. Lessons from controversial PTBP1 studies underscore the necessity of a standardized validation stack: Genetic lineage tracing, orthogonal multi-omic identity confirmation, and physiological proof-of-function[156-158]. In parallel, delivery technologies must advance to support these complex regulatory logic sets. The development of time-gated, cell-type-specific, and reversible delivery vehicles - such as transient epigenome editors - will be paramount in implementing multilayer fate-control networks safely, distinguishing genuine regeneration from transient cellular stress responses[159-161].

CONCLUSION

Retinal degeneration is a terminal structural failure in which conventional pathway modulation rarely rebuilds lost tissue and circuitry. Thus, the field must move from single-factor serendipity toward more reproducible, mechanism-anchored fate engineering across three coupled layers: Instruction (time-gated TF programs), permission (competence windows and chromatin accessibility), and an execution layer (translation-proteostasis, metabolism, and the injury-immune niche). Single-cell/spatial multi-omics now quantify these layers, but clinical impact depends on operationalizing atlas outputs into CMC-ready QC (omics for process development/periodic qualification; reduced identity-purity-potency-safety panels for routine release) and enforcing strict evidence standards to separate true regeneration from artifacts. Near-term progress will be driven by multimodal convergence - linking programmable fate control with stage-matched immunometabolic niche calibration - to advance retinal reconstruction toward safe, reproducible clinical translation.

ACKNOWLEDGEMENTS

The authors gratefully acknowledge Zhejiang Luming Biotechnology Co., Ltd. (2^nd Floor, 113-1 to 113-5 Nanliu Road, Chashan Street, Ouhai District, Wenzhou, Zhejiang Province, China) for their technical support and assistance in scientific figure preparation, formatting optimization, and related manuscript support for this work. The authors are solely responsible for the scientific content, interpretations, and conclusions of this manuscript.