Toward a 4D genome annotation of CHO cells for biomanufacturing

Abstract

Biotechnology and biopharma rely on detailed genome annotations for cell-line engineering, yet most production hosts are non-model organisms with limited resources linking sequence to physiology across space and time. A complete Chinese hamster ovary (CHO) 4D genome annotation that links sequence, networks, spatial constraints, process state, and passaging history does not yet exist; current efforts instead provide partial layers that must be connected into an actionable framework. Here we illustrate this framework for CHO cells, the dominant platform for recombinant protein biologics. Building such resources on a genomic foundation could reduce trial-and-error in biologics manufacturing by making cell-line and bioprocess design more predictable, transparent, and reproducible.

Biopharma Needs a Four-Dimensional Genome Annotation

Genomics and genome annotation have transformed modern biomanufacturing by turning genome sequences into important resources to support engineering efforts. High-quality reference assemblies, refined gene models, and functional annotations provide the coordinates and biological context needed to design genome edits and screening libraries [1,2], evaluate knock-ins and knockouts, select promoters and regulatory elements [3,4], resolve transgene integration sites and locus structures [5,6], compare sequence, structural, and copy-number variation (CNV) (see Glossary) across lineages [7,8], and connect genes to metabolic pathway models that are used for media and process optimization [9,10].

Yet many biomanufacturing phenotypes are not determined by sequence features alone. Productivity, stability, metabolism, and product quality depend on how genomic features interact with biological networks, spatial organization, process conditions, and time. To capture these dependencies, genome annotation must move beyond traditional one-dimensional resources toward multidimensional frameworks for biotechnology-relevant organisms. The first dimension [11] captures the linear genome and its catalog of genes and genomic features, including isoforms, accurate transcription start sites (TSSs) and promoters, enhancers, super-enhancers, repeats, variants, and well-annotated vector integration sites. The second dimension connects the functions of all genes and genomic features, i.e., their interactions and topological organization that connect genotype to phenotype. This includes gene-regulatory and signaling networks, and maps of diverse cell processes, such as metabolic, secretory, and glycosylation pathways. Important to biomanufacturing, models built from the second dimension annotation can link genes and pathways to cell growth and recombinant protein production. The third dimension of genome annotation accounts for the effects of genome to proteome structure on cell performance and bioproduction, including (i) chromatin architecture that govern promoter-enhancer wiring and position effects and gene expression, (ii) protein structures, experimental or predicted, that contextualize variants, active-site chemistry, allostery, post-translational modification (PTM) accessibility, and druggability for structure-aware editing and pathway engineering, and (iii) spatial organization and size of macromolecules and organelles in the cell, which impacts cell capacity and product quality. Finally, the fourth dimension captures dynamics at two scales: distal and proximal. Proximal dynamics describe rapid, reversible changes in the activity of genome-encoded enzymes and components following shifts in metabolic state (e.g., exponential or stationary phase) and/or process perturbations (e.g., temperature or pH shifts), affecting their expression, regulation, function, or transport and thereby altering the activity of broader metabolic, signaling, stress-response, and secretory programs during a production run. Distal dynamics encompass clonal evolution and drift (e.g., copy-number and structural variation, loss of heterozygosity (LOH), karyotypic changes, and persistent shifts in DNA methylation, chromatin accessibility, and post-transcriptional programs), together linking mechanistic readouts to run-to-run stability, productivity, and quality outcomes (Figure 1, Key figure).

Figure 1. Key figure. Overview of the 4D genome-annotation framework.

1D: The first dimension catalogs DNA, RNA, and protein sequences and their features (genes/isoforms, promoters/enhancers, repeats, small/structural variants, mitochondrial variants, and resolved vector-integration loci). 2D: The second dimension assembles interaction maps and quantitative network models (metabolism, secretion/glycosylation, regulation, signalling, protein complexes) to contextualize multi-omics and predict responses. 3D: The third dimension adds constraints from chromatin architecture, molecular structure of proteins and complexes, and cellular ultrastructure (ER-Golgi trafficking, mitochondria) that set capacity limits on folding, modification, and transport. 4D: The fourth dimension captures time- and condition-dependence across the bioprocess. In the 4D panel, the upper inset illustrates distal dynamics during cell passage and evolution, including clonal drift and genomic changes that can affect long-term stability. The lower inset illustrates proximal dynamics within a production run, in which process conditions alter biological networks and time-resolved culture trajectories. Together, these dynamic readouts link genome-encoded mechanisms to context-specific digital twins that can be used to interpret and predict changes in productivity, stability, and product quality over time. Abbreviations: CNV, copy-number variation; ER, endoplasmic reticulum; LOH, loss of heterozygosity; SV, structural variation; TF, transcription factor; TSS, transcription start site.

In this opinion we demonstrate the value of multidimensional annotation using Chinese hamster ovary (CHO) cells as an example. CHO cells produce most biopharmaceuticals (95 of 107 approved mammalian-made products as of 2022; ~89%), given their favorable titers and quality attributes at commercial scale [12]. To move beyond trial-and-error cell line and process development, a CHO-centric ‘4D genome annotation’ can unify sequence features, interaction networks, spatial organization, and dynamics under bioprocess perturbations and clonal evolution.

The value of a 4D genome annotation is more evident when considering that individual methods provide incomplete information. For example, a dynamic flux model may predict metabolic flux variation related to changes in lactate secretion after a feed shift and may link those fluxes to the corresponding metabolic genes and reactions [10,13]. However, a 4D annotation would extend this interpretation by connecting the response to regulatory programs, chromatin state, and genetic drift over passage history, helping identify why one clone responds differently from another. Similarly, integration-site sequencing or chromatin profiling alone can identify where a transgene inserted or whether the locus is active [5,14], but a 4D annotation would connect the insertion site to local regulatory architecture, structural variation, epigenetic drift, and production stability over time. For product quality, glycosylation models can predict glycoform distributions [15,16], whereas a 4D annotation would place those predictions in the context of many other biological processes and experiments, linking glycosylation enzymes, nucleotide-sugar transporters, Golgi and secretory-pathway organization, media composition, and process-state trajectories to regulatory programs and clone-specific behavior over time. Indeed, the goal is to organize the outputs of diverse experiments and methods in the context of the host cell genome, which makes the mechanisms behind process and engineering outcomes easier to interpret, compare, and test, as it links cell performance to their physicochemical content.

One-Dimensional Annotation

A 1D annotation includes linear genome sequence and its feature catalog, including genes, isoforms, TSSs, promoters, enhancers, repeats, variants, and well-annotated vector integration sites. In CHO, the 1D layer is maturing through successive assemblies and resequencing efforts that progressively resolved gene models, variant landscapes, and insertional architectures relevant to biomanufacturing [1,7,17–19]. In contrast to the relatively stable human and mouse reference genomes, the genome sequences of CHO production lines exhibit substantial genomic plasticity that accumulates with passage and selection, impacting stability and productivity; consequently, cell line-specific assemblies and variant maps further refine precision engineering and host cell lineage characterization [20,21]. Thus, industrially relevant cell lines are being paired with publicly available reference assemblies, providing industrially-relevant sequence data for genome-wide screening and engineering [1,22].

The first public draft genomic sequence was of the ancestral CHO-K1 cell line, published in 2011. This assembly contained ~2.45 Gb of sequence and ~24k annotated genes, enabling initial study of bioprocess-relevant components (e.g., glycosylation) and providing candidate chromosomal scaffold assignments [17]. This resource established the CHO gene annotation, while subsequent work further demonstrated the feasibility of short-read based cataloging in production lines. A community portal, the Chinese hamster genome database (CHOGenome.org), was launched soon thereafter to centralize sequences and annotations [17,23,24].

A comprehensive reference became possible with the 2013 Chinese hamster (Cricetulus griseus) draft genome, based off the same hamster colony from which the original CHO cells were derived in 1957. This genome sequence reported 24,044 genes and, through resequencing of common CHO lineages, quantified >3.7 M SNPs, ~5.5×10⁵ indels, and 7,063 copy number variations (CNVs), connecting specific genomic differences to phenotypes relevant to cell line performance [7]. An independent chromosome-sorting strategy, published concurrently, reinforced assembly contiguity and assignment [25]. These resources established a species-level coordinate system for consistent CHO gene and variant annotation.

Subsequent cell-line-focused resequencing revealed distinctive karyotypes and regional instability. Notably, sequencing of the DHFR-negative CHO DXB11 line uncovered pronounced haploidy (15–20% of genes were found as single-copies) and cell line-specific CNV signatures, insights that explained ease of certain knockouts and can motivate locus selection for engineering [8].

Integration of long-read sequencing transformed the baseline reference. A 2018 hybrid assembly (PICR) merged PacBio SMRT and Illumina data, collapsing scaffold counts >28-fold and filling >95% of prior gaps, with most genes (and flanking regulatory sequence) now residing on single scaffolds [19]. Hi-C scaffolding (PICRH) polished this to chromosome-scale continuity, further enabling precise promoter/TSS placement, enhancer discovery, and cross-omic mapping [18]. Proteogenomic and Ribo-Seq methods further refined gene models, by more precisely identifying the translated regions of genes [26]. Complementary enhancer annotation has been expanded through high-throughput reporter assays in CHO, yielding a genome-scale catalog of functionally validated regulatory elements for expression engineering [3].

Given the pronounced divergence among major CHO lineages (e.g., K1, DG44, CHO-S) and the increasing accessibility of high-quality, long-read genome assemblies, CHO pangenomes built from multiple parental and engineered clone assemblies offer a framework to capture both core and accessory genetic diversity. These also reduce reference-mapping bias, standardize variant representation, and strengthen host-specific genotype-phenotype associations [7,27,28]. Such pangenome efforts including widely-used production clones, can use a variation-graph reference (e.g., on PICRH coordinates) would enable consistent mapping of annotations while resolving lineage-specific small and structural variants linked to metabolic and secretory phenotypes. In parallel, lineage-specific pangenomes could uncover clone- and derivation-dependent genetic features that contribute to phenotypic heterogeneity, stability, and engineering outcomes within a shared host background.

Beyond the nuclear genome, ultra-deep sequencing of the Chinese hamster mitochondrial genome and diverse CHO clones revealed widespread heteroplasmy with clone-specific variant spectra [29]. These mitochondrial variants can modulate respiratory capacity and metabolic phenotypes, contributing to clone-to-clone variability in production; newer single-cell mtDNA studies are beginning to resolve intercellular heteroplasmy [29,30].

Finally, 1D maps can include information on vector integration sites and regions of indispensable genomic sequence and essential genes. Long-read and targeted sequencing can resolve complete transgene locus structures-concatemers, breakpoints, and host-vector fusions, within their genomic context [5,6]. Together with ChIP-guided safe harbor locus candidates, these datasets are cataloging integration-sites and helping with root-cause analysis of genetic instability, further helping deploy site-specific integration strategies[6,14]. In parallel, large-scale CRISPR-based deletions and pooled deletion screens are delineating essential versus non-essential regions of the CHO genome. These efforts are advancing toward streamlined cell lines, where non-essential DNA can be removed without compromising productivity or stability [2,31]. Indeed, 1D annotation maps will be invaluable for contextualizing the genetic features of individual platform cell lines, facilitating cell host engineering and improving targeted integration.

Two-Dimensional Annotation

The second dimension connects genotype to phenotype by linking functions of all genes, genomic features, gene products, metabolites and macromolecular complexes, etc., including their interactions and topological organization. This includes gene-regulatory and signaling networks, protein-protein interaction (PPI) networks, and maps of cell processes, like metabolism and the secretory pathway. Important to biomanufacturing, models built from the second-dimension annotations can link genes and pathways to growth and protein production.

One of the most well-known examples is a genome-scale metabolic network reconstruction, a curated knowledge base of biochemical reactions linked to catalyzing proteins and participating metabolites [32]. Release of the Chinese hamster genome [7,25], enabled a community effort that produced the first CHO genome-scale metabolic reconstruction in 2016, which quickly became a resource to interpret CHO metabolism and identify bottlenecks relevant to cell engineering [9]. Subsequent expansions increased reaction and enzyme coverage [33] and incorporated enzyme kinetics and capacity constraints, deepening annotation and improving quantitative predictions [34]. Most recently, a new community-consensus reconstruction for the Chinese hamster considerably expands the reaction network with mapped protein structures, enabling structural systems-biology analyses and improved contextualization of complex biological phenotypes [35]. Together, these resources enable new approaches to media design, by-product minimization, and hypothesis generation directly relevant to bioprocess optimization [32,36].

Beyond metabolism, 2D annotation systematically maps additional cellular processes central to recombinant protein production [37]. Protein secretion, spanning synthesis, translocation, folding, post-translational modifications, and transport, can be captured in pathway reconstructions that layer functions, subcellular localization, and complex assembly. Early mammalian secretory-pathway networks were adapted from mouse to CHO by protein homolog mapping [38], establishing a CHO-specific map of the secretion machinery and enabling omics overlays to nominate engineering targets. Building on this foundation, a recent mammalian-wide secretory-pathway reconstruction provides a more comprehensive cross-species framework that identifies mechanisms regulating antibody production and offers broadly transferable annotations for CHO contextualization [39]. By further annotating the mechanistic function of each component in the network, the reconstruction of the secretory pathway can be connected to genome-scale metabolic models (GEMs), to quantify protein-specific energetic and substrate costs, reveal pathway bottlenecks, and predict trade-offs that limit specific productivity (q_p) [40].

Annotation of glycosylation pathways provides a genetic basis for common critical quality attributes (CQAs) of CHO-derived therapeutic proteins. These networks can be layered onto secretory reconstructions and GEMs to connect metabolic state and secretory capacity to glycosylation function. CHO-specific data have helped define a comprehensive glycosylation reaction network (CHOGlycoNET)[15] and mechanistic/machine learning (ML) models [16] that predict glycoform distributions and upstream precursor demands. Together, these tools support glycoengineering and bioprocess optimization by helping guide media and feed interventions to achieve target glycan CQAs.

Beyond metabolism and secretion, two-dimensional annotation expands into additional layers that capture how cells sense, decide, and execute production programs. Gene regulatory networks [41] connect transcription factors (TFs) to their targets inferred from transcriptomic and perturbation data, while co-expression networks [42] identify coordinated gene modules associated with growth, proteostasis, secretory load management, and productivity; together, these regulatory layers frequently intersect with the unfolded protein response (UPR) and stress-adaptive transcriptional programs that shape robustness and recombinant protein output [43,44]. Complementing these networks, cis-regulatory maps of TSSs, promoters, and enhancers in CHO provide a framework for TF-target inference and promoter engineering [4,45]. Signaling networks further connect extracellular inputs such as nutrients, cytokines, and osmolality to cell cycle control, apoptosis, translational capacity, and glycosylation precursor supply, linking process conditions to phenotype [46]. Although comprehensive CHO-specific signaling reconstructions are still limited, studies of pathways such as mTORC1, JAK-STAT, and stress MAPKs highlight their importance for growth and recombinant protein production [47–50]. At the protein level, interaction networks and complex assemblies define secretory bottlenecks, as multi-subunit machines such as signal peptidase, PDIs/ERO1, OST, and COPII/COPI can constrain throughput at high q_p. While a curated CHO-wide interactome is not yet available, investigators typically rely on orthology-transferred PPI resources (e.g., STRING [51] for Cricetulus griseus) and CHO spatial/compartmental proteomics to contextualize complex stoichiometries [52]; alternatively, the recent mammalian secretory-pathway reconstruction compiles complexes that can be directly contextualized for CHO [39]. Finally, CRISPR screens, ranging from pathway-focused to genome-scale studies, now provide direct experimental validations by linking genetic perturbations to growth and productivity phenotypes, thereby strengthening regulatory, secretory, and metabolic network models [53–57].

Three-Dimensional Annotation

The third dimension of genome annotation captures how 3D structures of the nucleome, macromolecules, and organelles shape gene regulation, pathway flux, and ultimately bioproduction performance. In biomanufacturing, 3D features determine where and how genetic programs are executed (chromatin architecture), how proteins function and respond to edits (molecular structure), and how physical capacity constrains throughput (cellular ultrastructure) (Figure 2). Together, these annotations complement 1D (sequence) and 2D (network) maps by adding geometric and spatial constraints.

Figure 2. Three structure types for 3D genome annotation.

(A) Chromatin architecture captures the spatial organization of chromosomes, including topologically associating domains (TADs), that position promoters with distal enhancers and place loci within active or repressed chromatin environments. Active TADs are generally associated with accessible chromatin and transcriptionally permissive states, whereas repressed TADs are associated with reduced accessibility and gene silencing. These features inform landing-site selection, expression stability, and position effects by indicating regions that are more (or less) likely to support durable, high transcription. (B) Molecular structure links genes to the 3D conformations of their protein products. After translation and folding, proteins adopt structures with catalytic residues, cofactor/ligand pockets, interfaces, transmembrane segments, and PTM-accessible surfaces. Understanding these features guides structure-aware engineering to improve function, reduce undesired allostery, maintain multimeric contacts, and preserve desired glycosylation or stability. (C) Cellular ultrastructure reflects how proteins and macromolecules assemble into compartments (e.g., ER, ERGIC, Golgi) that set physical capacity limits on throughput. Organelle size, copy number, and localization dictate feasible flux and quality attributes; these compartments are plastic and remodel under load (e.g., expanded ER/Golgi to cope with increased secretory demand). Abbreviations: ER, endoplasmic reticulum; ERGIC, ER-Golgi intermediate compartment; PTM, post-translational modification; TAD, topologically associating domain.

Chromatin Architecture

At the level of genome organization, chromatin folding influences which genes can be accessed together and how regulatory elements interact with their targets [58]. In producer cells, this spatial layer is relevant to questions of transgene position effects, clone-to-clone variability, epigenetic silencing, and the long-term stability of productive states[5,59,60]. Although CHO-specific 3D genome resources remain limited compared with those available for human and mouse systems, growing access to chromosome-scale assemblies and chromatin profiling now help researchers study how local and higher-order genome structure shape transcriptional robustness and transgene behavior in bioprocess contexts. For CHO engineering, the value of this layer is not in reproducing a full human nucleome framework, but in identifying genome environments that support durable expression, favorable regulation, and resistance to drift over passaging.

Molecular Structure

A second component of 3D annotation is structural biology, which places proteins and macromolecular assemblies into a spatial framework that can be linked to function, spanning resources that contain experimentally determined structures, like the Worldwide Protein Data Bank (wwPDB) [61], to high-accuracy predictions from tools like AlphaFold [62]. Beyond single-chain models, complex-aware predictions (e.g., AlphaFold-Multimer/AF3) [63,64] provide interface-resolved hypotheses for PPIs, helping to identify catalytically competent assemblies, localize variants and PTMs to binding surfaces, and refine subunit stoichiometry for additional constraints in GEM simulations and contextualization of PPIs [65]. In CHO, structure-aware annotation can help in studies of metabolic enzymes and secretory machinery by facilitating analysis and engineering of their catalytic properties, interaction interfaces, and regulation. Integrating structural information with 2D reconstructions can make these models more mechanistically informative. In iCHO3K, this approach linked altered central-carbon flux states in a lactate-free CHO cell line to a plausible structure-based regulatory mechanism, providing a rationale for reduced glycolytic throughput and glucose uptake [35].

Cellular Ultrastructure

Finally, the third dimension also encompasses intracellular organization, which imposes structural and biophysical constraints on the secretory capacity of the cell. Compartmental and spatial proteomics in CHO producer contexts, resolving proteins across ER, ERGIC, Golgi, endo-lysosomal, mitochondrial, and cytoskeletal fractions, show enrichment of the core secretory machinery and ER-Golgi transport components in high-secretion states, suggesting that secretion can become limited by the availability of functional machinery and by physical capacity, such as ER and Golgi membrane area and lumenal space, especially at high productivity[52]. More recently, proximity-labeling interactome methods have provided a complementary view by identifying host proteins that interact with the recombinant product and may shape secretory performance or reveal candidate bottlenecks for engineering [66,67]. Together, these approaches help define intracellular bottlenecks that govern protein folding, processing, and vesicle trafficking, supporting more mechanistic models and more targeted strategies to improve CHO productivity.

Overall, these 3D layers shift annotation from a parts list to a spatially informed design framework. For CHO bioengineering, this means moving beyond the question of which genes or pathways matter, toward how their spatial arrangement across the genome, proteome, and cell interior support or constrain productive phenotypes. As these resources mature, 3D annotation should help identify chromatin states compatible with stable expression, structural liabilities in key proteins and complexes, and ultrastructural bottlenecks that become limiting at high specific productivity.

Four-Dimensional Annotation

Four-dimensional (4D) genome annotation adds time and process context to the earlier layers. Sequence features (1D), interaction networks (2D), and cellular organization (3D) are interpreted in relation to how they change across bioprocess phases and as clones are expanded over successive passages (Figure 3). This time-aware perspective is essential for CHO bioprocessing because many behaviors impacting productivity and quality are not fixed properties of a clone. Instead, they emerge as cultures progress within a run (proximal dynamics) and as cell populations adapt and drift over extended passaging (distal dynamics).

Figure 3. Proximal and distal dynamics in the fourth dimension of genome annotation.

(A) Proximal dynamics capture fast, reversible changes in the activity and usage of genome-encoded components in response to perturbations such as feed/medium changes, temperature shifts, pH/osmolality changes, and nutrient availability. These responses are mediated by the regulation and activity of metabolic enzymes, transporters, signaling proteins, transcription factors, glycosylation machinery, and secretory-pathway proteins, and can be quantified using time-course extracellular metabolite profiles, 13C fluxomics, flux analysis, transcriptomics, proteomics, and reporters of regulatory or secretory state. (B) Distal dynamics capture slower, cumulative changes across selection and repeated passages, including copy-number and structural variation, loss of heterozygosity (LOH), and chromosomal rearrangements, and regulatory drift in DNA methylation, chromatin accessibility, and post-transcriptional programs. These longer-term changes impact the baseline network and cellular capacity and can shift how a clone responds to the same process interventions in subsequent runs, with consequences for growth, productivity, and product quality over time.

On short time scales, 4D annotation captures proximal dynamics by linking process perturbations to time-dependent changes in annotated genomic features and genome-encoded networks. Feeding, temperature shifts, pH control, osmolality changes, and transitions in growth state are not represented only as extracellular inputs or flux constraints; they are connected to the genes, regulatory elements, and network components that execute the cellular response. For example, changes in nutrient availability and pH can alter the activity of transporters and metabolic enzymes, while also engaging signaling and transcriptional programs that regulate stress adaptation, protein synthesis, glycosylation, and secretion [44,46]. In this view, time-course uptake/secretion profiles [68,69], ¹³C fluxomics [69], and in silico flux analysis [13,70] define when pathway usage changes; while mRNA/protein turnover [71,72], pulse-chase glycoproteomics [73] and secreted proteomics [74,75] identify which genome-encoded components drive or constrain those changes. A glucose bolus, for instance, may initially increase glycolytic flux and lactate secretion, but its downstream consequences can include greater base addition for pH control and altered osmolality [68]. These changes can activate osmosensing and stress-response pathways, alter transcription-factor activity, and remodel secretory and glycosylation programs [15,44,52]. Linking these responses to annotated genes, regulatory networks, transporters, enzymes, and secretory-pathway components distinguishes 4D genome annotation from a dynamic process model alone: the goal is not only to describe that the culture state changed, but to identify which genome-linked mechanisms explain the change and which annotated features could be engineered or monitored in future runs.

Over longer time scales, distal dynamics reflect the cumulative effects of selection and passage. CHO cell lines can diverge through copy-number and structural variation, regional differences in genomic stability and haploidy, LOH, and chromosomal rearrangements including translocations [8,76,77]. Gene expression also changes over time, including shifts in DNA methylation at regulatory regions and changes in chromatin accessibility [78,79]. These slow variables create memory effects that bias how a given clone responds to standard process interventions and can help explain run-to-run differences in stability, metabolism, and productivity [80,81]. Distal dynamics, framed from a modeling and engineering perspective, treat genomic and epigenomic state as slowly-varying factors that shape faster responses within bioreactors. Furthermore, it prioritizes clone development and intervention strategies that provide more stable genomic contexts, while also considering how regulation and secretory capacity may adapt during extended operation. Emerging single-cell and lineage-resolved methods can quantify how population heterogeneity evolves over extended culture, helping to connect distal dynamics to production instability and clonal behavior in CHO at the single-cell level [82,83]. These approaches include genomic barcoding to trace CHO lineages and quantify clonal dynamics during selection and routine cultivation [84], as well as microfluidic single-cell cultivation with live-cell imaging to monitor time-resolved single-cell behavior under controlled conditions [85]. When aligned to the bioprocess timeline, such datasets can provide practical indicators of drift and stability and define response signatures to common perturbations such as temperature shifts and feeding changes.

Framed this way, the fourth dimension contextualizes all scales of CHO annotation, spanning genotype, networks, and structure. These are treated as functions of time and condition, converting static catalogs into predictive trajectories for clone selection and process design that jointly optimize productivity, stability, and product quality.

Functionalizing 4D annotation with digital twins and AI virtual cells

Genome annotation lays the groundwork for simulating cells and processes in silico, e.g., using a digital twin. These are working, testable model surrogates that explain how specific inputs (e.g., environmental changes) produce phenotypes (e.g., growth, productivity) under defined biological and chemical constraints. They allow one to run in silico experiments with tailored mechanistic and statistical models and return predictions with calibrated uncertainty [36]. In practice, one can develop digital twins with multiple different processes or modeling frameworks. Such hybrid models can leverage data across biological scales and time- varying conditions [36,86,87].

A broader vision is the AI Virtual Cell (AIVC) [88], in which a unified cell-state representation is learned from multi-omics and process measurements. Rather than targeting a single clone and process, the AIVC aims to learn a transferable representation that can be queried through standardized computational assays. In a CHO context, these would correspond to in silico assays that mirror common bioprocess workflows, applying defined perturbations (e.g., feed changes, temperature shifts, pH, perfusion settings, or engineering modifications) and returning outputs such as growth, byproduct formation, q_p, and CQAs. This framing should support consistent evaluation across defined perturbation studies, allowing different models to be compared on the same clone- and process-matched datasets.

However, modeling efforts remain limited by data availability and structure. Training will require open, well-annotated data, transparent reporting across multiple performance metrics, and biologically interpretable outputs that make key constraints and trade-offs explicit [89]. A 4D genome annotation can help address this bottleneck by providing a structured reference that links sequence variation, regulatory and interaction context, spatial and structural constraints, and temporal dynamics to the specific states and transitions that models attempt to predict. Importantly, not every application needs every layer. Predicting growth or metabolite byproducts may rely on biomass composition, exchange fluxes, and metabolic network structure, but these inputs can be connected to a wealth of data, if scaffolded upon a genomic foundation. In this context, it can be interpreted through transporter and metabolic gene-protein-reaction rules, isoenzyme usage, enzyme-capacity constraints, and regulatory programs that determine which proteins are available or active in a given clone and process state. In contrast, predicting productivity loss or altered product quality may require additional layers, such as transgene integration site, copy-number variation, chromatin accessibility, secretory-pathway capacity, and glycosylation genes. A comprehensive 4D annotation would make these task-specific inputs retrievable with their genomic origin, provenance, uncertainty, and process context, allowing digital twins and AIVC-style models to distinguish whether a predicted phenotype is driven by culture conditions, clone-specific genome structure, regulatory state, pathway capacity, or their interaction. Existing methods usually cannot do this in isolation, but together can connect process behavior to molecular mechanisms in a way that remains auditable, comparable across clones and studies, and extensible as new measurements become available.

Concluding remarks and future perspectives

Building 4D annotation frameworks that support CHO and other host cell lines for biomanufacturing requires a shift in emphasis: we need to move from collecting disconnected datasets and collaborate on coordinated resources designed for integration and testing against real process outcomes. The goal is not only to describe how CHO cells behave across time and passages, but to support better decisions about productivity, stability, and CQAs under defined operating conditions. This requires shared reference data, clearer reporting of bioprocesses, and community-agreement on how to evaluate if technologies improve production. To make this framework actionable, it is useful to distinguish the minimum content of each annotation layer from the added value that emerges when those layers are connected across clone, process, and time (Box 1). Unresolved challenges and priorities are summarized in the Outstanding Questions (see Outstanding questions).

Box 1. Minimum actionable layers for CHO 4D genome annotation.

A complete CHO 4D genome annotation does not require that every possible assay be collected for every clone or process. Instead, it requires that each data layer be properly annotated, e.g., with clone information, genomic coordinate system, defined biological networks, and process context. Each data type can be useful on its own, but its value increases as additional annotations are added and connected across dimensions.

1D sequence and feature layer.

Minimum content includes a clone- or lineage-relevant reference genome, sequence variants, and mapped transgene integration sites. This layer supports genome editing, screening-library design, integration-site analysis, and clone comparison. Added regulatory elements, mitochondrial variants, essential or non-essential regions, and pangenome-aware annotations improve interpretation of recombinant-protein expression, production stability, and differences across CHO lineages.

2D functional-network layer.

Minimum content includes genome-linked metabolic, secretory, and glycosylation networks that represent the core machinery for recombinant-protein production, with explicit provenance and confidence. This layer supports pathway interpretation, flux analysis, and bottleneck identification. Added signaling, gene-regulatory, and protein-interaction networks help explain how this machinery is controlled. Its value further increases when connected to clone-specific variants from the 1D layer and spatial constraints from the 3D layer.

3D spatial and structural layer.

Minimum content includes chromatin context around genes and integration sites and predicted or experimental protein structures for key pathway components. This layer can already support landing-site interpretation and structure-aware engineering. Its value increases as additional annotations are added, including protein complexes, interaction interfaces, PTM-accessible sites, and quantitative ER/Golgi capacity constraints that help explain why the same gene, pathway, or product may behave differently across genomic or cellular contexts.

4D temporal and process layer.

Minimum content includes time-resolved process metadata, uptake and secretion profiles, growth and productivity measurements, product-quality readouts, and passage history. This layer defines when culture behavior changes across process phases and clone expansion. A time-course dataset becomes part of a 4D genome annotation only when these dynamic measurements are linked back to genome, network, and spatial layers.

These layers can be leveraged individually, but the main value of 4D annotation emerges when they are connected. Static resources, such as reference genomes, regulatory maps, network reconstructions, chromatin profiles, or protein structures, provide the coordinate systems and mechanisms needed to interpret dynamic process data. Conversely, time-resolved measurements without standardized genome, network, or spatial annotation may describe process behavior but do not by themselves constitute a 4D genome annotation.

Outstanding Questions.

• What is the minimal clone-matched 4D reference dataset needed to make genome annotation predictive for bioprocess decisions?

• How should CHO pangenomes and cell-line-specific assemblies be organized so that annotations remain comparable across major production lineages?

• Which CHO regulatory, interaction, chromatin, structural, and ultrastructural maps are most critical to replace cross-species transfer?

• How can process metadata, sampling timelines, and product quality readouts be standardized so datasets from different laboratories and companies can be integrated?

• Which measurements best capture the time-dependent activity of genome-encoded regulatory, metabolic, transport, and secretory programs during a run, and which measurements best capture distal drift across passages?

• How should uncertainty be represented when integrated annotations combine direct CHO evidence with transferred knowledge from other species?

• What community benchmarks are needed to test whether 4D annotations improve prediction across clones, media, and operating modes?

• What subset of 4D annotation is sufficient for specific applications such as clone selection, media optimization, stability prediction, or digital-twin development?

• How can federated learning be implemented so as to elevate all industry players, while protecting individual company IP?

A first priority is clone-matched reference programs. Existing resources span multiple CHO backgrounds and process modes, but we need some synchronization within the same clone so measurements and annotations are directly comparable rather than confounded by differences in lineage, media, or operating mode. Community efforts, including the Advanced Mammalian Biomanufacturing Innovation Center (AMBIC) and the National Institute of Standards and Technology CHO reference cell line (NISTCHO), are establishing openly accessible, industry-relevant reference cell lines and standardized cultivation platforms, with traceable provenance and detailed characterization of growth, productivity, and product quality attributes [90,91]. These references can help generate a synchronized public 4D reference dataset for single clones under well-defined processes, connected by a shared set of analytical readouts and validated through targeted perturbations with blinded replicates and cross-site runs. Until those foundations exist, comparability, traceability, and true cross-study synthesis will remain limited.

A second priority is standardized data collection of bioprocess context and sampling timelines so that datasets from different studies can be meaningfully compared and combined. In practice, integration often breaks down because bioprocess datasets are generated under different operating conditions, measured on different schedules, and reported with inconsistent detail. In addition, a substantial fraction of industrially-generated data remains inaccessible due to confidentiality, IP, and internal reporting constraints, which further limits cross-study synthesis. At a minimum, CHO datasets could be accompanied by key process readouts and time-resolved measurements, including major nutrient and byproduct concentrations, temperature and pH trajectories, osmolality, viable cell density and viability, and productivity metrics. On the analytics side, reporting should clearly state which CQAs were assessed, the analytical methods used, and how values were summarized across the run. Adopting lightweight “Minimum Information” reporting standards with accompanying CHO templates (e.g., modeled after MIAME [92] or MIAPE [93]), would immediately increase the value of both academic and industrial datasets by enabling more reliable alignment and interpretation across studies [89]. To share industry data while managing IP concerns, federated learning models could be deployed, but they are just beginning to be explored [94,95].

A third priority is to make integration confidence-aware, especially when multiple layers rely on information transferred from other species. Orthology-based transfer and imported interaction resources remain valuable starting points, but uncertainty can accumulate when several transferred layers are combined, for example, protein interactions [96], metabolic networks, and secretory constraints. To keep models useful for engineering decisions, each integrated layer should carry an explicit level of confidence, with clear provenance distinguishing CHO evidence from transferred annotations. Where transfer uncertainty most strongly affects engineering conclusions, targeted CHO measurements should be prioritized [37].

A fourth priority is prospective benchmarking to test whether integrated 4D resources improve prediction under blinded, pre-defined conditions. Context-specific models have already shown that they can model cell performance in the condition in which they were trained [9,36,97], but the next step is to evaluate generalization across clones, media, and operating modes using agreed-upon metrics. This may require a multi-site effort spanning common CHO backgrounds in fed-batch and perfusion, where teams register model predictions in advance and then compare them to standardized experimental outcomes. Relevant process readouts include growth trajectories, lactate production and consumption dynamics, nutrient limitations, time-resolved uptake and secretion profiles, and targeted flux measurements where feasible. These benchmarks help rank models, and further identify which annotation layers and process descriptors are required for reliable prediction.

Finally, delivering the long term vision of a CHO “virtual cell” or generalizable digital twin will depend on closing the loop between data generation, model updating, and actionable engineering recommendations for cell line and process changes. For models to become dependable tools rather than conceptual descriptions, they must be reproducible and reusable across groups, with clearly stated assumptions and constraints. Releases must be tracked as resources evolve, inputs and outputs standardized, and workflows benchmarked to evaluate predictions against growth, productivity, and CQA readouts. Releasing community resources that can be directly applied in cell line and process development, including curated reconstructions, computable pathway networks, reference datasets, and benchmarking workflows, will accelerate adoption and reduce duplicated effort. If developed in a coordinated way, 4D genome annotations can evolve from organizing concepts into calibrated foundations for next-generation models that generalize across processes and shorten development cycles.

Highlights.

• Recent CHO genome assemblies, refined gene models, and regulatory maps are making genome annotation increasingly useful for cell-line engineering.

• Genome-scale metabolic, secretory, glycosylation, regulatory, and signaling models are expanding the ability to connect genes with bioprocess phenotypes.

• New structural and spatial approaches are helping link chromatin context, protein function and organelle capacity to production outcomes.

• Time-resolved process measurements and clone-passage data now motivate genome-grounded frameworks that connect culture behavior to annotated cellular mechanisms.

• A complete CHO 4D genome annotation does not yet exist, but current resources provide partial layers that can be connected into an actionable roadmap for predictive biomanufacturing.

Glossary

AI Virtual Cell (AIVC)

A generalizable computational framework that learns cell-state representations from multi-omics and process data and can be queried with standardized in silico assays.

Chromatin architecture

The three-dimensional folding of the genome, including loops and higher-order domains, that influences regulatory contact, gene expression, and transgene position effects.

Copy-number variation (CNV)

Gain or loss of genomic segments that changes gene dosage and can alter metabolism, growth, productivity, and long-term clone stability.

Critical quality attribute (CQA)

A measurable product property, such as glycosylation, purity, or charge profile, that affects therapeutic safety, efficacy, or manufacturability.

Digital twin

A calibrated computational surrogate of a biological system or process used to simulate perturbations, predict outcomes, and support engineering decisions.

ER-Golgi intermediate compartment (ERGIC)

Membrane trafficking compartment that mediates transport and sorting between the ER and Golgi apparatus.

Federated learning models

Machine-learning models trained across multiple institutions or datasets without requiring the underlying data to be centralized, allowing shared model development while helping protect proprietary or sensitive data.

Genome-scale metabolic model (GEM)

A network model linking genes, reactions, and metabolites to analyze cellular metabolism and predict flux redistribution under genetic or process perturbations.

Heteroplasmy

Coexistence of more than one mitochondrial DNA sequence within a cell or population, potentially contributing to metabolic heterogeneity.

Loss of heterozygosity (LOH)

Loss of one parental allele in a genomic region, which can unmask recessive variants and alter phenotype or stability.

Pangenome

A reference framework representing shared and lineage-specific genomic content across multiple related genomes, reducing bias from reliance on a single reference assembly.

Post-translational modification (PTM)

Chemical modification of a protein after translation, such as glycosylation or phosphorylation, that affects localization, activity, stability, or interactions.

Safe harbor locus

A genomic site expected to support durable transgene expression with minimal disruption of essential host-cell functions.

Secretory pathway

The integrated cellular machinery that synthesizes, folds, modifies, transports, and secretes recombinant proteins through the ER, Golgi, and vesicular system.

Specific productivity (q_p)

The rate of recombinant product formation per cell, commonly used to evaluate performance of production clones.

Vector integration site

The genomic location where an introduced DNA construct inserts, often influencing expression level, stability, and local genome integrity.