A high-throughput co-culture CRISPR array for rapid fitness hypothesis testing and growth engineering in CHO cells Authors: Jannis Peter Marzluf, Lea Huckfeldt, Dominik Haidas, Christoph Zehe, Ann-Cathrin Leroux Published: 2026 ABSTRACT SummaryCell growth and fitness shape mammalian biomanufacturing performance, but subtle genetic effects are difficult to resolve with conventional growth assays. We developed a high-throughput competitive co-culture CRISPR array in Chinese hamster ovary (CHO) suspension cells that pairs a CRISPR-competent population with a CRISPR-inactive green fluorescent protein (GFP) reference pool and tracks GFP-ratio dynamics by flow cytometry. Guided by a prior genome-wide CRISPR screen, we tested 235 single-gene knockouts and 155 multiplex gene sets across seven batches and >100 days of cultivation. The assay detected expected lethal and growth-inhibitory perturbations and identified 28 growth-enhancing knockouts or gene sets that are poorly resolved by viable-cell-concentration measurements. Slope-based modeling converted GFP dynamics into inferred growth rate changes, and combinatorial exclusion with bulk transcriptomics linked multiplex phenotypes to candidate driver genes. The approach enables scalable fitness hypothesis testing and prioritization of CHO growth-engineering targets for downstream host-cell development. Introduction Cell growth and cellular fitness are fundamental determinants of performance in mammalian cell culture systems, influencing not only experimental reproducibility but also long-term stability, scalability, and economic feasibility in bioprocessing applications.1,2 In industrial cell line development, growth rate directly impacts timelines for clone generation, seed train expansion, and production scale-up, while subtle fitness defects can accumulate over extended culture periods and compromise productivity or product quality.3,4 Despite its importance, cell growth remains a complex, polygenic trait whose mechanistic and genetic principles are only partially understood, particularly under conditions optimized for high-density suspension culture.5,6 Chinese hamster ovary (CHO) cells represent the dominant host system for the production of recombinant therapeutic proteins, owing to their adaptability to suspension growth in chemically defined media and their capacity for human-like post-translational modifications.7,8 Although decades of optimization have yielded CHO production platforms achieving high titers and good product quality, these advances have largely relied on empirical process development and targeted genetic interventions rather than a systematic understanding of cellular functions.2,9,10 Growth rate, in particular, remains a frequent bottleneck during cell line development, with slow proliferation leading to long development timelines.11,12 Proteomic and transcriptomic profiling studies have delineated molecular programs associated with favorable growth phenotypes in CHO cells. High-density or fast-growing clones tend to upregulate DNA replication, mitotic progression, and ribosome biogenesis, while downregulating stress-responsive and oxidative metabolic pathways.13,14 These observations suggest that growth capacity is constrained by coordinated regulation across multiple cellular networks rather than single rate-limiting genes. Nevertheless, most engineering efforts to date have focused on modifying individual metabolic pathways, particularly those associated with the accumulation of inhibitory byproducts such as lactate or ammonia.15,16 While these interventions can improve growth under bioprocess conditions, they do not address the broader question of which genes fundamentally determine proliferative fitness under optimal growth environments. The existence of genetically encoded differences in intrinsic growth capacity is further supported by recent reports describing newly derived or re-isolated hamster cell lines with substantially faster doubling times than conventional CHO production hosts.17,18 These findings indicate that current industrial CHO lineages may carry growth-limiting genetic configurations that are not strictly required for survival or productivity. Identifying such dispensable or conditionally disadvantageous genes requires accurate maps of gene essentiality and fitness contribution in the relevant cellular context.19,20 Genome-wide knockout screens have enabled the identification of essential genes, synthetic lethal interactions, and condition-specific vulnerabilities across diverse cell types.19,20,21,22,23,24 In CHO cells, these tools have been adapted to study metabolic dependencies, stress responses, and productivity-related phenotypes.25,26 However, few studies have focused explicitly on growth rate as a primary phenotype, despite its relevance for both basic cell biology and industrial applications.21 One key challenge in studying cell growth is the accurate detection of subtle proliferative differences. Traditional growth assays, including direct cell counting, metabolic readouts such as ATP or tetrazolium reduction, and endpoint viability staining, are well suited for detecting large growth differences but lack the sensitivity and reproducibility required to resolve small differences.27,28 These limitations are exacerbated in high-throughput or pooled screening formats, where technical noise, seeding variability, and passage-to-passage fluctuations can obscure biologically meaningful effects.29,30,31,32 Comparative growth measurements are particularly sensitive to small deviations in initial cell seeding number or handling, which can introduce systematic biases when differences in proliferation rate are modest.27 This is problematic in functional genomics screens, where many perturbations are expected to exert weak but cumulative fitness effects rather than acute cytotoxicity.33,34 Co-culture-based competitive growth assays offer an alternative strategy for quantifying relative fitness by directly comparing two-cell populations within the same environment. By monitoring changes in population ratios over time, these approaches effectively integrate small growth differences across multiple cell divisions, increasing sensitivity while controlling for external variability.31,32 Fluorescent protein labeling, in particular the stable expression of green fluorescent protein (GFP) in one competing population, provides a non-invasive, scalable readout that allows continuous tracking of relative population frequencies by flow cytometry without requiring cell lysis or cell treatment. Competitive assays have been applied successfully in yeast and mammalian systems to study fitness effects of genetic perturbations, particularly when absolute growth measurements are noisy or poorly reproducible.24,33 In mammalian cells, early implementations of competitive co-culture assays relied primarily on viral transduction and constitutive marker expression, which introduced confounding factors related to integration site variability, expression burden, and non-isogenic controls.31,32 Moreover, most studies were performed in adherent cancer cell lines, limiting their relevance for suspension-adapted production hosts such as CHO cells. These limitations underscore the need for co-culture strategies that are compatible with suspension growth, scalable cultivation formats, and precise genome editing technologies. The integration of CRISPR-based perturbation methods with competitive co-culture readouts provides a powerful framework for dissecting growth-related gene function. CRISPR/Cas systems enable highly specific, programmable genomic modifications, allowing clean separation between genetic perturbation and phenotypic readout.35,36 When combined with co-culture GFP measurements over time, this approach is particularly well suited for identifying genes whose loss confers small but consistent growth advantages or disadvantages that only become apparent over extended cultivation periods.30,37 In CHO cells, such strategies are especially relevant given the polygenic nature of growth and the high degree of redundancy within mammalian regulatory networks.5,6 Growth-enhancing perturbations are unlikely to produce dramatic phenotypes in short-term assays but may nevertheless yield meaningful benefits during prolonged cultivation runs. Systematic identification of these genetic determinants requires experimental designs that preserve population diversity, minimize clone-specific artifacts, and enable sensitive detection of cumulative fitness effects.11,38,39,40 In this study, we employ a high-throughput co-culture CHO-optimized functional genomics system to systematically investigate genetic determinants of cell growth and cell fitness. By combining previous genome-scale CRISPR screening results21 with competitive co-culture-based growth assessment in suspension culture, we aim to resolve subtle proliferative phenotypes that are difficult to capture with conventional assays. This approach enables the identification and experimental validation of individual genes and gene combinations that modulate cell growth and fitness, thereby providing a more resolved view of the genetic architecture underlying CHO cell proliferation. In parallel, it establishes a scalable, high-throughput co-culture CRISPR array for rapid fitness hypothesis testing and growth engineering in CHO cells, creating a quantitative framework for rational host cell optimization with an emphasis on sustained growth performance under relevant conditions. Results Culture dynamics provides a cumulative readout of cell fitness To assess whether population dynamics provides a more reliable measure of cell fitness than viable cell concentration (VCC) readouts, model genes with established phenotypes in CHO cells were selected. EEF1A1 and PGK1 served as lethal targets confirmed in a previous genome-wide CRISPR screen21 (CHO_EG2025), PRDX1 was included as a growth-inhibiting gene based on prior studies, and FN1, LPL, and FUT8 were chosen as neutral controls (Table S1). Lethal and growth-inhibitory targets were deliberately included as positive controls because their knockout is expected to deplete the edited subpopulation over time, they provide an internal benchmark for the sensitivity and dynamic range of the competitive assay, against which neutral controls and candidate growth-enhancing perturbations can be calibrated. Gene editing was performed using two delivery formats: a stable Cas9-expressing CHO cell line and ribonucleoprotein (RNP) transfection. KO scores measured on day 2 indicated high-editing efficiencies across all targets, ranging from ∼50% to 90% (Figure 1A). The CHOCas9+ cell line achieved slightly lower KO scores than RNP delivery, with median values of 65.0% and 80.5, respectively (Figures 1B and 1C). Variation in sgRNA dose between 7.5 and 20 pmol did not significantly affect editing outcomes (Figure 1B). Although the CHOCas9+ cell line performed about 15 KO-score percentage points lower than RNP, it enables sgRNA-only delivery, reducing cost and complexity for high-throughput applications. Edited pools were cultivated for 16 days and sub-cultivated every 2–4 days. VCC normalized to wild-type (WT) controls showed transient reductions for EEF1A1 and PGK1 knockouts between days 6 and 11 in CHOCas9+ pools and days 9 and 11 in RNP pools (Figures 1D and 1E). These differences disappeared after day 11, and VCC values converged across all variants. PRDX1 did not exhibit significant growth inhibition in Cas9 pools, in RNP conditions, VCC appeared lower but without statistical significance. Overall, VCC-based measurements displayed high variability, limiting their ability to resolve moderate fitness effects as expected with PRDX1. The large WT-normalized VCC error bars in Figures 1D and 1E, which for several conditions span roughly 0.5–1.5 in magnitude, are themselves a direct manifestation of VCC assay variability, and they are the limitation that motivates an improved readout. In contrast, population dynamics, operationalized here as the temporal change in the fraction of functionally edited cells within the mixed populations provided a clearer and cumulative readout of fitness. Rather than measuring absolute cell numbers, this approach tracks the relative competitive advantage or disadvantage of a perturbed subpopulation over successive cell divisions, integrating even modest fitness differences into a detectable signal over time. KO-score trajectories quantified functional knockout fractions at days 2, 7, and 16 (Figures 1F and 1G). Both editing formats showed complete dropout of lethal knockouts EEF1A1 and PGK1 over time, while neutral targets remained stable. ΔKO scores confirmed significant depletion of lethal variants (Figures 1H and 1I), while PRDX1 did not show depletion under these conditions. The rapid loss of EEF1A1 and PGK1 KO cells by around day 5 (Figures 1F and 1G) is the expected kinetic signature of lethal perturbations and is consistent with the transient WT-normalized VCC reductions in Figures 1D and 1E, which resolve once the surviving non-edited cells dominate the pool; the growth effect is therefore attributable to the knockout during the early window in which edited cells are still present, even though the functional KO score is depleted thereafter. The PRDX1 KO score is maintained over time, indicating that PRDX1 loss is tolerated, yet VCC shows no statistically significant growth difference; its mild growth-inhibitory effect falls below the resolution of VCC in this monoculture format. These findings demonstrate how growth-based metrics are transient and insensitive to measure subtle differences, whereas population dynamics integrate effects over time, reduce variance, and eliminate the need for metabolic assays.Figure 1Comparative evaluation of knockout efficiency and growth impact using CHOCas9+ and RNP delivery systems(A) KO scores for six model genes (EEF1A1, PGK1, PRDX1, FN1, LPL, and FUT8) measured at day 2 post-editing using two delivery formats: CHOCas9+ (black bars) and RNP (orange bars). KO scores represent the proportion of functional knockouts determined by sequencing analysis.(B) Effect of sgRNA dose on editing efficiency in CHOCas9+ cells (7.5 pmol vs. 20 pmol) compared with RNP delivery (7.5 pmol).(C) Overall KO score distribution for CHOCas9+ and RNP conditions.(D and E) WT-normalized viable cell concentration (VCC) over 16 days for each knockout versus wild-type control under CHOCas9+ (D) and RNP conditions (E). Transient reductions were observed for EEF1A1 and PGK1 early in cultivation, followed by normalization after day 11; PRDX1 showed no significant effect. Statistical significance assessed by one-way ANOVA with Dunnett’s post-hoc test (∗p < 0.05) versus WT as neutral control, evaluated separately at each time point (“split-wise”).(F and G) KO score trajectories at days 2, 7, and 16 for CHOCas9+ (F) and RNP conditions (G). Lethal knockouts (EEF1A1 and PGK1) exhibited complete dropout over time, while neutral targets remained stable.(H and I) ΔKO scores calculated as the difference between initial and final KO scores for each gene under CHOCas9+ (H) and RNP conditions (I). Negative ΔKO values confirm depletion of lethal variants, whereas neutral targets show minimal change. All data are shown as mean ± SD; n = 8 (CHOCas9+); n = 4 (RNP). LPL∗ amplicon PCR failed to amplify for downstream sequencing-based KO-score determination; LPL ΔKO scores are therefore not reported in (F)–(I). Comparative evaluation of knockout efficiency and growth impact using CHOCas9+ and RNP delivery systems (A) KO scores for six model genes (EEF1A1, PGK1, PRDX1, FN1, LPL, and FUT8) measured at day 2 post-editing using two delivery formats: CHOCas9+ (black bars) and RNP (orange bars). KO scores represent the proportion of functional knockouts determined by sequencing analysis. (B) Effect of sgRNA dose on editing efficiency in CHOCas9+ cells (7.5 pmol vs. 20 pmol) compared with RNP delivery (7.5 pmol). (C) Overall KO score distribution for CHOCas9+ and RNP conditions. (D and E) WT-normalized viable cell concentration (VCC) over 16 days for each knockout versus wild-type control under CHOCas9+ (D) and RNP conditions (E). Transient reductions were observed for EEF1A1 and PGK1 early in cultivation, followed by normalization after day 11; PRDX1 showed no significant effect. Statistical significance assessed by one-way ANOVA with Dunnett’s post-hoc test (∗p < 0.05) versus WT as neutral control, evaluated separately at each time point (“split-wise”). (F and G) KO score trajectories at days 2, 7, and 16 for CHOCas9+ (F) and RNP conditions (G). Lethal knockouts (EEF1A1 and PGK1) exhibited complete dropout over time, while neutral targets remained stable. (H and I) ΔKO scores calculated as the difference between initial and final KO scores for each gene under CHOCas9+ (H) and RNP conditions (I). Negative ΔKO values confirm depletion of lethal variants, whereas neutral targets show minimal change. All data are shown as mean ± SD; n = 8 (CHOCas9+); n = 4 (RNP). LPL∗ amplicon PCR failed to amplify for downstream sequencing-based KO-score determination; LPL ΔKO scores are therefore not reported in (F)–(I). Co-culture enables single-parameter flow cytometric fitness readout The rationale for prioritizing co-culture-based validation over monoculture growth measurements is illustrated conceptually in Figure S3A, demonstrating how small growth-rate differences are poorly resolved in monoculture, but accumulate into a quantifiable signal in competitive co-culture format. To establish a high-throughput co-culture system capable of detecting changes in cellular fitness with a single-flow cytometric readout, we combined a CRISPR-competent CHOCas9+ cell line with a CRISPR-deficient GFP-expressing line. This design ensures that only half of the population is edited upon sgRNA delivery, while GFP expression enables rapid and cost-efficient monitoring of population dynamics by flow cytometry. Two prerequisites were verified: (1) comparable growth rates between GFP+ and CHOCas9+ cells and (2) stable GFP expression with clear separation from GFP-negative cells over extended cultivation. Rather than relying on monoclonal GFP+ reference clones, we deliberately employed a polyclonal reference pool. This choice minimizes bias introduced by clonal heterogeneity and avoids over-interpretation of fitness effects driven by clone-specific background variation. As discussed extensively for producer cell lines, pooled systems provide a more representative average phenotype and improve accuracy when subtle fitness differences are interrogated.40 Although the present co-culture assay is performed in a non-producer context, the same principles apply: the use of pools averages out clone-to-clone variability and better reflects population-level fitness, which is advantageous for sensitive, high-throughput screening applications. We acknowledge that pooling can, in principle, introduce additional intra-population heterogeneity because a pool comprises cells with differing baseline behaviors that may respond differently to a given perturbation; in the present design, this is mitigated by FACS-based enrichment of a stable reference pool and by control-normalization (ΔΔGFP), so that the net effect on the competitive readout is improved rather than degraded reproducibility. To generate GFP+ controls, a GFP-encoding plasmid was transfected into a mock CHO line followed by puromycin selection. Of nine transfected pools, three were pooled to yield three large, pooled pools (LPP1-3). These were monitored for GFP stability (Figure 2A). GFP levels declined after puromycin omission, reaching ∼85%–90% after 30 days. To improve signal stability, the pooled LPP1-3 population was FACS-sorted into four enriched pools (FELP1-4) based on GFP intensity, each containing 2E5 cells. Representative gating strategies and GFP intensity thresholds used to generate FELP1-4 are shown in Figures S1A–S1D. Mean fluorescence intensity (MFI) ranged from 4 × 105 to 6 × 105, comparable with a monoclonal GFP+ reference in the highest intensity sorted pool (Figure 2B). Monitoring of GFP signal intensity further demonstrated stable MFI profiles for all FELP populations over extended cultivation (>100 days), with FELP3 and FELP4 showing the highest long-term stability (Figures S1E and S2C). Next, FELP3 and FELP4 were mixed 1:1:2 with CHOCas9+ cells to create three co-culture pools (CC1-CC3), which were cultivated for 46 days. GFP ratios remained within ±10% of the starting point across all pools, confirming stability (Figure 2D). The flow cytometric gating strategy used for all co-culture experiments, including debris exclusion, singlet selection, and definition of the GFP-positive gate, as well as representative GFP fluorescence distributions for CC1-CC3, is shown in Figures S2A–S2F. Using this co-culture setup, we tested two model genes: EEF1A1 (lethal) and FN1 (neutral). VCC showed slight reductions at the first passage for both knockouts, but EEF1A1 remained suppressed at the second passage. All conditions converged by passage three (Figure 2E). Integrated VCC (IVCC) (Table 1) across the whole time period revealed minor, statistically insignificant reductions for EEF1A1, while FN1 remained comparable with unedited controls (Figure 2F). These changes were transient and difficult to quantify, highlighting the limitations of growth-based metrics. In contrast, GFP time-course analysis provided a clear signal (Figure 2G). Control cultures transfected for transient GFP expression declined from 100% to near zero by day 12, while EEF1A1 knockouts showed a strong enrichment of GFP+ cells to ∼80%, reflecting dropout of CRISPR-edited cells. Here, ΔGFP denotes the change in GFP+ cell fraction between day 0 and the final evaluation day within a given sample, serving as a direct measure of competitive fitness shift over the observation window. To account for spontaneous baseline drift in untreated control populations, ΔΔGFP further corrects ΔGFP against the mean ΔGFP of internal controls measured within the same experiment and at the same time point, thereby isolating perturbation-specific fitness effects from technical or biological noise inherent to the co-culture system (Table 2). ΔΔGFP calculations confirmed large shifts for EEF1A1 and negligible changes for FN1 (Figure 2H). KO-score trajectories mirrored this pattern, with EEF1A1 showing pronounced depletion of edited cells and FN1 remaining stable (Figure 2I). These results demonstrate that co-culture enables sensitive detection of gene essentiality and fitness changes using a single-flow cytometric readout, outperforming conventional growth-based measurements.Figure 2Development and validation of a co-culture system for flow cytometric fitness readout(A) GFP-positive fraction (%) over 30 days for three initial GFP-transfected pools (LPP1-LPP3) compared with GFP-positive and GFP-negative controls. Data represents SD; n = 3.(B) Mean fluorescence intensity (MFI) of GFP-positive cells in four FACS-enriched large pools (FELP1-FELP4). Data represent boxplot ± SD with min and max; n = 3, across 23 timepoints.(C) GFP-positive fraction (%) over >100 days for FELP1-FELP4, demonstrating long-term stability of FELP3 and FELP4. FELP1 and FELP2 cultures were terminated at ∼day 57. Data represents SD; n = 3.(D) GFP ratio over 46 days in three co-culture pools (CC1-CC3) mixed at 1:1 with CHOCas9+ cells; all remained within ±10% of baseline. Data represents SD; n = 3.(E) Viable cell concentration (VCC) over 16 days for EEF1A1 knockout, FN1 knockout, and wild-type controls in co-culture; transient reductions observed for EEF1A1 at early passages. Data represents SD; n = 3.(F) Integrated VCC (IVCC) for the same conditions; only minor differences detected. Data represents SD; n = 3.(G) GFP-positive fraction (%) over time for EEF1A1 and FN1 knockouts versus controls; EEF1A1 shows strong enrichment of GFP-positive cells (∼80%) indicating dropout of edited cells, while FN1 remains near baseline. Data represents SD; n = 3.(H) ΔΔGFP values calculated relative to initial state and wild-type controls; large shifts for EEF1A1, negligible changes for FN1. Data represents SD; n = 3.(I) ΔKO scores between initial and final time points confirm depletion of EEF1A1 knockouts and stability of FN1. Data shown as mean ± SD; n = 3.Table 1VCC calculations(1)Integrated viable cell concentration (IVCC) IVCC=∑k=1n−1VCC(tk)+VCC(tk+1)2·(tk+1−tk)where VCC(tk) and VCC(tk+1) denote the viable cell concentrations measured at consecutive culture days tk and tk+1, respectively, and n is the total number of VCC measurement time points within the interval of interest. IVCC represents the cumulative viable cell mass over time.Table 2Formulas used to quantify fitness effects from co-culture GFP measurements(2)Sample endpoint fitness effect (ΔGFP)ΔGFPi,b,d=GFPi,b,dend−GFPi,bDay0whereGFPi,b,dend is the final GFP-positive fraction observed for sample i in batch b at evaluation day d (Last Day of Cultivation) and GFPi,bDay0 is the corresponding batch-specific Day 0 GFP fraction(3)Batch-, control- and day-corrected fitness effect (ΔΔGFP)ΔΔGFPi,b,d=ΔGFPi,b,d−⟨ΔGFPINTb,d⟩where ⟨ΔGFPINT⟩b,d denotes the mean ΔGFP of INT-control samples measured in the same batch b and at the same evaluation day d as the sample of interest. Development and validation of a co-culture system for flow cytometric fitness readout (A) GFP-positive fraction (%) over 30 days for three initial GFP-transfected pools (LPP1-LPP3) compared with GFP-positive and GFP-negative controls. Data represents SD; n = 3. (B) Mean fluorescence intensity (MFI) of GFP-positive cells in four FACS-enriched large pools (FELP1-FELP4). Data represent boxplot ± SD with min and max; n = 3, across 23 timepoints. (C) GFP-positive fraction (%) over >100 days for FELP1-FELP4, demonstrating long-term stability of FELP3 and FELP4. FELP1 and FELP2 cultures were terminated at ∼day 57. Data represents SD; n = 3. (D) GFP ratio over 46 days in three co-culture pools (CC1-CC3) mixed at 1:1 with CHOCas9+ cells; all remained within ±10% of baseline. Data represents SD; n = 3. (E) Viable cell concentration (VCC) over 16 days for EEF1A1 knockout, FN1 knockout, and wild-type controls in co-culture; transient reductions observed for EEF1A1 at early passages. Data represents SD; n = 3. (F) Integrated VCC (IVCC) for the same conditions; only minor differences detected. Data represents SD; n = 3. (G) GFP-positive fraction (%) over time for EEF1A1 and FN1 knockouts versus controls; EEF1A1 shows strong enrichment of GFP-positive cells (∼80%) indicating dropout of edited cells, while FN1 remains near baseline. Data represents SD; n = 3. (H) ΔΔGFP values calculated relative to initial state and wild-type controls; large shifts for EEF1A1, negligible changes for FN1. Data represents SD; n = 3. (I) ΔKO scores between initial and final time points confirm depletion of EEF1A1 knockouts and stability of FN1. Data shown as mean ± SD; n = 3. Table 1VCC calculations(1)Integrated viable cell concentration (IVCC) IVCC=∑k=1n−1VCC(tk)+VCC(tk+1)2·(tk+1−tk)where VCC(tk) and VCC(tk+1) denote the viable cell concentrations measured at consecutive culture days tk and tk+1, respectively, and n is the total number of VCC measurement time points within the interval of interest. IVCC represents the cumulative viable cell mass over time. VCC calculations Table 2Formulas used to quantify fitness effects from co-culture GFP measurements(2)Sample endpoint fitness effect (ΔGFP)ΔGFPi,b,d=GFPi,b,dend−GFPi,bDay0whereGFPi,b,dend is the final GFP-positive fraction observed for sample i in batch b at evaluation day d (Last Day of Cultivation) and GFPi,bDay0 is the corresponding batch-specific Day 0 GFP fraction(3)Batch-, control- and day-corrected fitness effect (ΔΔGFP)ΔΔGFPi,b,d=ΔGFPi,b,d−⟨ΔGFPINTb,d⟩where ⟨ΔGFPINT⟩b,d denotes the mean ΔGFP of INT-control samples measured in the same batch b and at the same evaluation day d as the sample of interest. Formulas used to quantify fitness effects from co-culture GFP measurements Resolving subtle growth effects from a genome-wide CRISPR screen In a previous genome-wide CRISPR KO screen in CHO cells, 17,761 expressed genes were targeted using a pooled lentiviral sgRNA library. This screen identified 235 candidate genes associated with growth acceleration, which were grouped into 155 functional gene sets based on cellular pathways and gene annotations.21 All 235 candidate genes were individually validated in the singleplex co-culture assay, with each gene targeted by two dedicated sgRNAs in an independent competitive growth experiment (Figure 3A). The gene sets ranged from sizes of two genes up to 13 genes in a single set (Figure S3C). For each gene in singleplex, two sgRNAs with the highest enrichment scores from the original screen were included, and multiplexed gene sets were addressed using the top-ranked fold-change sgRNA per gene. The experimental design comprised seven batches, each consisting of three to four 96-well plates. Each batch contained internal controls, including GFP-transfected controls to monitor transfection efficiency, WT co-cultures without sgRNA substrate, intron-targeting (INT) controls, and non-targeting (NT) controls. In addition, five model genes from prior internal knockout studies were included: EEF1A1 and PGK1 as lethal targets, PRDX1 as a growth-inhibiting target, and FN1 and FUT8 as neutral targets. Batches 1–6 were processed in parallel with a one-day offset, and batch 7 was executed at a later time point. Transfection efficiency was consistently high across all batches, with GFP-positive ratios approaching approximately 100% at the first measurement interval, 2–4 days post-transfection (Figure 3B). Knockout efficiency was quantified using the co-culture system, where approximately half of the population was CRISPR competent. Initial KO scores ranged from 20% to 30% for EEF1A1, FN1, PGK1, and PRDX1, while FUT8 reached approximately 50%, indicating near-complete editing in CRISPR-positive cells (Figure 3C). Time-course analysis revealed pronounced depletion of lethal and growth-inhibiting targets. Between days 4 and 25, ΔKO scores indicated near-complete loss of EEF1A1-, PGK1-, and PRDX1-edited cells, consistent with conditional lethality observed in previous studies and the co-culture proof of concept.21 FN1 and FUT8 exhibited moderate declines of less than an average of 10% points (Figures 3C and 3D). While all controls exhibited comparable long-term GFP ratio slopes, INT- and NT-controls showed a slight early increase in GFP+ fraction during the initial recovery phase after transfection, attributable to transient burden in the GFP-negative subpopulation: INT-control cells undergo active CRISPR editing, while NT-control cells metabolize sgRNA transfection substrate; WT-control cells, subjected only to electroporation, displayed the least transient increase. This transient effect resolved within the early cultivation window and did not confound steady-state fitness measurements (Figure 3E). Because all knockouts are carried in the GFP-negative (CRISPR-competent) cell population, the direction of the GFP+ fraction directly reports relative fitness: fitness-reducing perturbations and editing-associated burden deplete GFP-negative cells and thus enrich the GFP+ fraction, whereas growth-enhancing perturbations allow the GFP-negative population to outgrow the reference, depleting the GFP+ fraction. This system-inherent bias explains the decline in KO score for neutral targets FN1 and FUT8, with the slope of ∼0.5%/day approximating the ∼10% loss of KOs across 21 days of cultivation (Figure 3D). In contrast, EEF1A1 and PGK1 showed steep GFP enrichment, indicating rapid clearance of CRISPR-competent cells following KO lethality. PRDX1 displayed a gradual GFP increase, consistent with prior observations where single-cell PRDX1 KOs were viable but exhibited severely impaired growth (Figure 3E) (clone data not shown). To quantitatively capture these behaviors, two complementary fitness metrics are assessed: the sample-intrinsic change in GFP fraction over time (ΔGFP) and the control-normalized differential fitness effect at matched time points (ΔΔGFP) (Figure S3B). Day 25 ΔΔGFP values confirmed strong fitness-reducing effects for EEF1A1, PGK1, and PRDX1, while FUT8 showed only a minor, statistically significant growth advantage of negligible biological relevance. Importantly, direct VCC measurements for these same model genes showed largely overlapping growth trajectories across targets and controls (Figure S3D). Again, VCC-based growth readouts lack the sensitivity to resolve fitness differences that are readily detected by cumulative GFP dynamics in co-culture, reinforcing the necessity of competitive assays for quantifying subtle growth effects (Figures S3A and S3D).Figure 3Validation of growth-enhancing gene knockouts in CHO cells using a high-throughput co-culture assay(A) Schematic overview of the experimental workflow combining genome-wide CRISPR screen candidates with co-culture validation. A lentiviral library targeting 17,761 expressed genes was transduced into a CRISPR competent CHOCas9+ cell line. The cell-based library was cultivated and passive fitness/growth selection applied. Differential sgRNA abundance allowed identification of essential and fitness/growth-enhancing gene KO candidates. The growth-enhancing gene KOs mapped to cellular functions like metabolism by cytochrome P450, cytokine interactions, extracellular matrix receptors (ECM), amino acid transport and fatty acid metabolism. Gene KO candidates were grouped based on molecular functions and pathways to derive 155 functional gene sets. The 235 gene KO candidates and 155 functional gene sets added up to 390 variants screened in the Co-Culture CRISPR array, where ΔΔGFP is the main readout.(B) GFP-positive cell fraction (%) measured 2–4 days post-transfection across seven experimental batches. Data are shown as mean ± SD; n = 6.(C) KO score trajectories for EEF1A1, PGK1, PRDX1, FN1, and FUT8 from days 4 to day 25. Data represent mean ± SD; n = 6.(D) ΔKO scores (days 4–25) for the same targets compared with initial state after transfection. Bars indicate mean ± SD; statistical significance assessed by one-way ANOVA with Dunnett’s post-hoc test (∗p < 0.05) versus FUT8 as neutral control; n = 6.(E) GFP ratio time-course over 116 days for INT-, NT-, and WT-controls and model gene knockouts. Data shown as mean in bold and each replicate as thin line; n = 6.(F) ΔΔGFP (%) values (days 0–25) comparing target gene knockouts to INT-controls. Data are shown as mean ± SD; n = 6. Statistical significance determined by one-way ANOVA with Dunnett’s post-hoc test (∗p < 0.05) versus INT control as the reference group, consistent with INT-control serving as the normalization baseline for ΔΔGFP (Figure 3A created in BioRender, https://BioRender.com/bccphjn). Validation of growth-enhancing gene knockouts in CHO cells using a high-throughput co-culture assay (A) Schematic overview of the experimental workflow combining genome-wide CRISPR screen candidates with co-culture validation. A lentiviral library targeting 17,761 expressed genes was transduced into a CRISPR competent CHOCas9+ cell line. The cell-based library was cultivated and passive fitness/growth selection applied. Differential sgRNA abundance allowed identification of essential and fitness/growth-enhancing gene KO candidates. The growth-enhancing gene KOs mapped to cellular functions like metabolism by cytochrome P450, cytokine interactions, extracellular matrix receptors (ECM), amino acid transport and fatty acid metabolism. Gene KO candidates were grouped based on molecular functions and pathways to derive 155 functional gene sets. The 235 gene KO candidates and 155 functional gene sets added up to 390 variants screened in the Co-Culture CRISPR array, where ΔΔGFP is the main readout. (B) GFP-positive cell fraction (%) measured 2–4 days post-transfection across seven experimental batches. Data are shown as mean ± SD; n = 6. (C) KO score trajectories for EEF1A1, PGK1, PRDX1, FN1, and FUT8 from days 4 to day 25. Data represent mean ± SD; n = 6. (D) ΔKO scores (days 4–25) for the same targets compared with initial state after transfection. Bars indicate mean ± SD; statistical significance assessed by one-way ANOVA with Dunnett’s post-hoc test (∗p < 0.05) versus FUT8 as neutral control; n = 6. (E) GFP ratio time-course over 116 days for INT-, NT-, and WT-controls and model gene knockouts. Data shown as mean in bold and each replicate as thin line; n = 6. (F) ΔΔGFP (%) values (days 0–25) comparing target gene knockouts to INT-controls. Data are shown as mean ± SD; n = 6. Statistical significance determined by one-way ANOVA with Dunnett’s post-hoc test (∗p < 0.05) versus INT control as the reference group, consistent with INT-control serving as the normalization baseline for ΔΔGFP (Figure 3A created in BioRender, https://BioRender.com/bccphjn). Across batches 1–6, 218 single-gene knockouts and 155 multiplex gene sets were screened, all derived from our previous study (Figure 4A). An intermediate analysis during the runtime of these batches identified 10 single-gene candidates as preliminary hits based on ΔΔGFP metrics. These genes were subsequently included in batch 7 as part of a 10x_Ultra_Gene_Set, designed to test potential synergistic or additive effects of combined validated KOs. This multiplex set was generated in 24 replicates to ensure sufficient representation of cells carrying simultaneous KOs of multiple target genes within the transfected pool. Batch 7 also included the 17 single-gene targets omitted from batches 1–6 (Figure 4A) and all relevant controls. Ranking all variants by gene-specific endpoint ΔΔGFP values, the eight top hits comprised six multiplex gene sets and two single-gene knockouts (NDRG1 and BRD9), all showing significant depletion of GFP-positive cells relative to INT-, NT-, and WT-controls (Figure 4B). Multiplex gene sets exhibited an initial increase in GFP-positive ratios, reflecting transient editing-associated toxicity and cell-cycle arrest during repair of multiple double-strand breaks. This effect was most pronounced in the 10x_Ultra_Gene_Set, where GFP-positive ratios reached ∼85%–90% within 10 days before declining at differing slopes after day 40 across all 24 replicates. BRD9 knockout displayed a consistent selective advantage throughout the cultivation period, whereas NDRG1 enrichment became evident only after day 46. Notably, this coincided with an unplanned 10-h interruption of orbital shaking, resulting in transient oxygen limitation. Given NDRG1’s role in oxidative metabolism, its knockout may confer a selective advantage under hypoxic conditions.41,42 To assess system performance and define statistical thresholds, ΔΔGFP distributions were compared against negative controls (EEF1A1, PGK1, and PRDX1) and INT-controls. On day 30, INT-controls remained tightly clustered with a 2 × SD threshold of ±5.4%, enabling clear separation to negative controls (EEF1A1, PGK1, and PRDX1) (Figure 4C). By the end of cultivation, variance increased and INT-controls drifted toward negative controls, expanding the 2 × SD threshold to ±18.3% (Figure 4D). While this reduced signal-to-noise for positive GFP shifts, it markedly improved resolution for negative GFP changes, favoring detection of growth and fitness-enhancing knockouts at later time points. Importantly, hit calls were highly consistent regardless of whether INT-, NT-, or WT-controls were used for ranking, with only minor differences in rank order (Figure 4E). Because INT-controls most closely mimic the physiological state of target cells undergoing DNA damage, statistical testing was performed against this baseline (Figure 4F). Using one-sided testing with an adjusted p value cutoff of 0.05 and a ΔΔGFP threshold of −18.3% (derived from INT-control variability), we identified 28 significant hits, including 15 multiplex gene sets and 13 single-gene knockouts that conferred improved growth or fitness (Figure 4F). Additionally, the ad-hoc validated 10x_Ultra_Gene_Set also showed an additive effect size compared with its constituent KOs. Outcomes and computed values are shown in Tables S2 and S3. The top 20 hits are shown in Figure 4G, with mean ΔΔGFP values and ±SD across replicates. Although the 10x_Ultra_Gene_Set exhibited the steepest GFP decline, it did not rank highest by ΔΔGFP due to batch 7 being concluded before INT-control drift had elevated sufficiently the GFP-positive ratios. The co-culture system enabled sensitive detection of subtle fitness differences across a total of 390 single-gene and multiplex gene sets derived from a previous genome-wide CRISPR screen, identifying 28 statistically significant hits including multi-gene knockout combinations and single-gene knockouts such as BRD9 and NDRG1. The array’s cumulative readout over extended cultivation improved resolution for growth-enhancing variants, while early time points remained sufficient to identify lethal or strongly deleterious KOs. These findings establish co-culture as a scalable platform for functional genomics in CHO cells, enabling rapid hypothesis testing for cell fitness and growth rate engineering.Figure 4Large-scale co-culture CRISPR screening and hit identification(A) Overview of experimental timeline and variant distribution. Batches 1–6 included 218 single-gene knockouts and 155 multiplex gene sets. Batch 7 contained 17 additional single-gene targets and one higher-order multiplex set (10x_Ultra_Gene_Set) tested in 24 replicates.(B) GFP-positive cell fraction over time for representative top 8 ΔΔGFP hits. Curves show INT-control, NT-control, WT-control, and target variant gene knockouts. Data represents means for controls and target with all target replicates being plotted individually. All variants are shown with their batch-corrected control means; n = 6 (Singleplex Gene Targets and controls); n = 4 (Multiplex Gene Sets and controls).(C) Distribution of ΔΔGFP values at day 30 for all variants compared with INT-controls. Dashed lines indicate ±2 × SD threshold (±5.4%).(D) Distribution of ΔΔGFP values at the end of cultivation (mostly ∼day 116). End of cultivation is always calculated relative to INT-1 at the day a culture was terminated. Dashed lines indicate ±2 × SD threshold (±18.3%).(E) Ranked mean endpoint ΔΔGFP (%) for all variants relative to INT-, NT-, and WT-controls.(F) Statistical significance assessed using one-sided Student’s t tests (alternative = “less”) comparing each gene or gene set against INT-controls, with p values combined if appropriate across batches using Fisher’s method and adjusted for multiple comparisons by Benjamini-Hochberg FDR correction (cutoff: q < 0.05). Hits were additionally filtered by ΔΔGFP threshold (−18.3%). Data represent batch- and end-day corrected means; n = 6 (Singleplex Gene Targets and controls); n = 4 (Multiplex Gene Sets and controls).(G) Top 20 hits ranked by mean endpoint ΔΔGFP (%) with error bars representing ±SD across replicates. Data represent batch- and end-day corrected means; n = 6 (Singleplex Gene Targets and controls); n = 4 (Multiplex Gene Sets and controls). Large-scale co-culture CRISPR screening and hit identification (A) Overview of experimental timeline and variant distribution. Batches 1–6 included 218 single-gene knockouts and 155 multiplex gene sets. Batch 7 contained 17 additional single-gene targets and one higher-order multiplex set (10x_Ultra_Gene_Set) tested in 24 replicates. (B) GFP-positive cell fraction over time for representative top 8 ΔΔGFP hits. Curves show INT-control, NT-control, WT-control, and target variant gene knockouts. Data represents means for controls and target with all target replicates being plotted individually. All variants are shown with their batch-corrected control means; n = 6 (Singleplex Gene Targets and controls); n = 4 (Multiplex Gene Sets and controls). (C) Distribution of ΔΔGFP values at day 30 for all variants compared with INT-controls. Dashed lines indicate ±2 × SD threshold (±5.4%). (D) Distribution of ΔΔGFP values at the end of cultivation (mostly ∼day 116). End of cultivation is always calculated relative to INT-1 at the day a culture was terminated. Dashed lines indicate ±2 × SD threshold (±18.3%). (E) Ranked mean endpoint ΔΔGFP (%) for all variants relative to INT-, NT-, and WT-controls. (F) Statistical significance assessed using one-sided Student’s t tests (alternative = “less”) comparing each gene or gene set against INT-controls, with p values combined if appropriate across batches using Fisher’s method and adjusted for multiple comparisons by Benjamini-Hochberg FDR correction (cutoff: q < 0.05). Hits were additionally filtered by ΔΔGFP threshold (−18.3%). Data represent batch- and end-day corrected means; n = 6 (Singleplex Gene Targets and controls); n = 4 (Multiplex Gene Sets and controls). (G) Top 20 hits ranked by mean endpoint ΔΔGFP (%) with error bars representing ±SD across replicates. Data represent batch- and end-day corrected means; n = 6 (Singleplex Gene Targets and controls); n = 4 (Multiplex Gene Sets and controls). Growth modeling quantifies fitness effects and transcriptomics resolves multiplex gene sets To quantify fitness effects beyond endpoint ΔΔGFP values, we implemented a slope-based growth model in which sliding-window linear regression was applied to time-resolved GFP trajectories to identify the peak rate of GFP-positive fraction decline for each target gene (Figure S4B). The resulting slope was corrected against the INT-control slope estimated over the identical time window and batch, yielding a control-corrected metric (Δs). Δs values were subsequently converted to inferred growth rate changes (Δμ) via a logit-based transformation anchored to a fixed INT-control reference growth rate of 0.9241 days−1 (18 h doubling time), enabling direct estimation of target-specific doubling times (Table 3; full model specification in STAR Methods section). This model assumes a linear relationship between the control-corrected peak slope (Δs), defined as the difference between the most negative local slope of the target gene’s GFP+ trajectory and the corresponding INT-control slope estimated over the identical time window, and the inferred relative growth rate change (Δμ), within a valid range of −3% to +3% GFP per day, corresponding to growth rate changes of approximately −15% to +15% (Figure 5A). Back-testing with simulated starting GFP ratios of 50% and slope intervals from −5%/day to +5%/day over 10 days confirmed that the approximation is accurate for moderate changes, although extreme slopes introduce error as linearity is lost (Figure 5A). Despite these caveats, the model provides a practical estimate of growth advantages or disadvantages in co-culture. Slope fitting was applied to all significant hits, and inferred Δμ values were plotted in Figure 5B. Examples for the top 15 hits by slope fit are shown in Figure S4. The highest inferred Δμ was observed for the 10x_Ultra_Gene_Set, followed by singleplex and multiplex gene sets with effects ranging from 1% to 10%. For the next validation step, previously banked cell pools were thawed and expanded before growth rate measurements. For batches 1–6, all biological replicates of each singleplex or multiplex gene set were pooled prior to cell banking. Upon thaw, each pooled cell bank was re-subdivided into six replicate cultures. Batch 7 knockout pools were not pooled but instead maintained as individual replicates carried forward directly from the original transfections. Due to throughput limitations associated with handling thousands of individual cultures, this pooling strategy was applied exclusively to batches 1–6, and these constraints should be considered when interpreting downstream analyses. To provide an experimental benchmark for the model-based predictions, we extended cultivation for approximately 30 days beyond the main co-culture screen and directly measured growth rates for selected hits and controls (Figures 5C and 5D). Several caveats apply: first, not all singleplex or multiplex KO variants had reached complete GFP-negative status at the time of inoculating the 30 days validation experiment, such that residual GFP-positive cells reduced the apparent growth rate of edited populations. Second, cultures had undergone 80–120 days of prior cultivation, and long-term stability of the observed advantages cannot be assumed, particularly for knockouts and multiplex gene sets affecting stress response or other important regulatory pathways. As a result, the measured growth rates should be interpreted as directional validation of growth and fitness effects rather than precise quantitative estimates. The primary aim of this study is to establish and characterize the co-culture screening system as a scalable fitness prioritization tool; definitive assessment of growth-enhancing perturbations will require isolation and timely characterization of edited clones in follow-up studies. Despite these limitations, measured growth rates for top hits were broadly consistent and order of magnitude with inferred Δμ values. Variants such as 101_TGF-beta, 087_GrowthM, and the 10x_Ultra_Gene_Set as well as individual knockouts including ESX1 and XPNPEP1 exhibited measurable growth improvements consistent with model predictions. Together, these results demonstrate that competitive co-culture combined with slope-based modeling can reliably resolve subtle fitness differences and provide a useful approximation of their relative impact, supporting its utility for prioritizing targets for CHO cell growth engineering.Table 3Formulas to calculate slope and infer changes in growth rates based on changes in GFP-positive ratio(4)Linear model for local GFP dynamics GFP(t) = m·t+bwhere GFP(t) is the mean GFP-positive fraction (%) at aligned culture day t, m is the local slope of the GFP trajectory (% GFP per day), and b is the intercept of the fitted linear model.(5)Control-corrected peak slopeΔs = mtarget-mINTwhere mtarget is the most negative slope identified for the target gene across all eligible sliding windows, and mINT is the corresponding INT-control slope estimated over the same time window and within the same batch. Negative values of Δs indicate a stronger decline in GFP-positive fraction relative to INT-control dynamics.(6)Estimated endpoint GFP fractionfend=clip(50+Δs100)where fend is the estimated GFP-positive fraction after one day, Δs is expressed in % GFP per day, and clip(·)constrains values to the open interval (0,1)to ensure numerical stability and biological plausibility. The starting fraction is fixed as fstart = 0.5.(7)Inferred change in growth rateΔμ=ln⁡(fstart1−fstart)−ln⁡(fend1−fend)where Δμ is the inferred change in relative growth rate (day−1), derived using a logit transformation of GFP fractions.(8)Inferred target growth rateμtarget = μINT1+Δμwhere μINT = 0.9241 days−1 is the fixed reference growth rate of INT-controls (corresponding to an 18 h doubling time), and μtarget is the inferred growth rate of the target population.(9)Inferred doubling timeDTtarget=ln2μtarget×24hwhere DTtarget is the inferred doubling time of the target population in hours.Figure 5Growth modeling/validation and gene set deconvolution using slope-based inference and transcriptomics(A) Relationship between GFP slope (%/day) and inferred relative growth rate (Δμ), modeled using a logit transformation anchored to a reference growth rate of 0.9241 days−1. Slopes between −3%/day and +3%/day correspond to Δμ of approximately −15% to +15%, assuming a starting GFP fraction of 50%.(B) Inferred Δμ (%) for the top 30 hits ranked by slope, derived from INT-control-corrected slopes. The 10x_Ultra_Gene_Set is highlighted in orange. Data represent mean; n = 6 (singleplex); n = 4 (multiplex).(C) Measured relative growth rates over ∼30 days for previously banked KO pools from batches 1–6 and batch 7 hits, compared against INT-controls. Color coding indicates statistical outcome: orange, significant increase; red, significant decrease; gray, not significant. Data represent means; n = 6 (all targets and controls); n = 24 (10x_Ultra_Gene_Set).(D) Mean measured growth rate differences (Δμ, %) relative to INT-controls for selected knockouts and gene sets after ∼30 days. Data are shown as mean ± SD; n = 6; n = 24 (10x_Ultra_Gene_Set). Statistical significance determined by Brown-Forsythe and Welch ANOVA with Benjamini-Hochberg FDR correction, p < 0.05.(E) Heatmap of gene set membership (rows) and constituent genes (columns). Gray, not present; blue, present in set; orange, significant individual hit (ΔΔGFP < −10%, q < 0.05). Sets shown in bold font were resolved by slope analysis; sets in gray font (101_TGF-beta, 122_Transcription, 152_Immune, 154_Biogenesis) were flagged for transcriptomic follow-up. Gene Set 100_DNAdam was included as a positive-control to validate that transcriptomic profiling correctly recovers known gene drivers.(F) Differential expression heatmaps for unresolved gene sets. Log2 fold-changes were calculated versus INT-controls from GeTMM-normalized counts; significance assessed by Welch’s t test. Blue, downregulation; red, upregulation; asterisks indicate significance tier (∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001). Data represent n = 4. Table 3Formulas to calculate slope and infer changes in growth rates based on changes in GFP-positive ratio(4)Linear model for local GFP dynamics GFP(t) = m·t+bwhere GFP(t) is the mean GFP-positive fraction (%) at aligned culture day t, m is the local slope of the GFP trajectory (% GFP per day), and b is the intercept of the fitted linear model.(5)Control-corrected peak slopeΔs = mtarget-mINTwhere mtarget is the most negative slope identified for the target gene across all eligible sliding windows, and mINT is the corresponding INT-control slope estimated over the same time window and within the same batch. Negative values of Δs indicate a stronger decline in GFP-positive fraction relative to INT-control dynamics.(6)Estimated endpoint GFP fractionfend=clip(50+Δs100)where fend is the estimated GFP-positive fraction after one day, Δs is expressed in % GFP per day, and clip(·)constrains values to the open interval (0,1)to ensure numerical stability and biological plausibility. The starting fraction is fixed as fstart = 0.5.(7)Inferred change in growth rateΔμ=ln⁡(fstart1−fstart)−ln⁡(fend1−fend)where Δμ is the inferred change in relative growth rate (day−1), derived using a logit transformation of GFP fractions.(8)Inferred target growth rateμtarget = μINT1+Δμwhere μINT = 0.9241 days−1 is the fixed reference growth rate of INT-controls (corresponding to an 18 h doubling time), and μtarget is the inferred growth rate of the target population.(9)Inferred doubling timeDTtarget=ln2μtarget×24hwhere DTtarget is the inferred doubling time of the target population in hours. Formulas to calculate slope and infer changes in growth rates based on changes in GFP-positive ratio Growth modeling/validation and gene set deconvolution using slope-based inference and transcriptomics (A) Relationship between GFP slope (%/day) and inferred relative growth rate (Δμ), modeled using a logit transformation anchored to a reference growth rate of 0.9241 days−1. Slopes between −3%/day and +3%/day correspond to Δμ of approximately −15% to +15%, assuming a starting GFP fraction of 50%. (B) Inferred Δμ (%) for the top 30 hits ranked by slope, derived from INT-control-corrected slopes. The 10x_Ultra_Gene_Set is highlighted in orange. Data represent mean; n = 6 (singleplex); n = 4 (multiplex). (C) Measured relative growth rates over ∼30 days for previously banked KO pools from batches 1–6 and batch 7 hits, compared against INT-controls. Color coding indicates statistical outcome: orange, significant increase; red, significant decrease; gray, not significant. Data represent means; n = 6 (all targets and controls); n = 24 (10x_Ultra_Gene_Set). (D) Mean measured growth rate differences (Δμ, %) relative to INT-controls for selected knockouts and gene sets after ∼30 days. Data are shown as mean ± SD; n = 6; n = 24 (10x_Ultra_Gene_Set). Statistical significance determined by Brown-Forsythe and Welch ANOVA with Benjamini-Hochberg FDR correction, p < 0.05. (E) Heatmap of gene set membership (rows) and constituent genes (columns). Gray, not present; blue, present in set; orange, significant individual hit (ΔΔGFP < −10%, q < 0.05). Sets shown in bold font were resolved by slope analysis; sets in gray font (101_TGF-beta, 122_Transcription, 152_Immune, 154_Biogenesis) were flagged for transcriptomic follow-up. Gene Set 100_DNAdam was included as a positive-control to validate that transcriptomic profiling correctly recovers known gene drivers. (F) Differential expression heatmaps for unresolved gene sets. Log2 fold-changes were calculated versus INT-controls from GeTMM-normalized counts; significance assessed by Welch’s t test. Blue, downregulation; red, upregulation; asterisks indicate significance tier (∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001). Data represent n = 4. To dissect which genes contributed to the multiplex phenotypes, we employed a two-step strategy: (1) combinatorial analysis comparing slopes of multiplex sets to those of their constituent single-gene knockouts. If the multiplex slope was fully explained by individual gene effects, the phenotype was attributed to those genes. (2) Differential transcriptomics at late screening stages to identify knockouts associated with nonsense-mediated decay or reduced expression of the target genes. We hypothesized that since starting populations harboring growth-enhancing variants are small, the culture dynamics would allow for selective co-enrichment of accelerating variants since these would be naturally selected for, if they had an effect. The exclusion principle applied in Figure 5E illustrates this approach. Most multiplex effects were explained by the presence of DDIT4, NDRG1, or ZEB1 within the set. For unresolved cases (sets 101, 122, 152, and 154), transcriptomic profiling was performed. Notably, set 152 was included despite not meeting formal significance thresholds because it exhibited late-onset GFP decline, suggesting enrichment of a small subpopulation. Transcriptomic analysis confirmed additive contributions from stress-response genes such as NDRG1, DDIT4, and PML in set 100_DNAdam, which served as a reference for interpretation. To characterize the downstream programs associated with the profiled gene sets, we additionally performed gene set enrichment analysis (GSEA) on the same late-stage transcriptomes (Figure S5). Enrichment patterns were set-specific and, in several cases, converged with the disrupted drivers identified by differential expression (Figure 5F). Most notably, in set 100_DNAdam the significant downregulation of NDRG1 and PML coincided with depletion of oxidative phosphorylation and mitochondrial respiration, consistent with NDRG1-associated oxidative metabolism and the NDRG1-knockout advantage observed under transient oxygen limitation. Sets 101, 122, and 152 showed enrichment compatible with their resolved drivers (e.g., proliferation-associated programs in 101 and altered translational and biosynthetic programs in 122 and 152), whereas set 154_Biogenesis displayed no significantly disrupted members and only a generic biogenesis signature and is therefore additionally considered unresolved. These transcriptomes were derived from competitively enriched populations; therefore, the enrichment signatures partly reflect the proliferative state of the selected cells. Hence, they may be interpreted as hypothesis-generating corroboration of the assigned programs rather than as mechanistic proof. Overall, this combined modeling and transcriptomic approach resolved the majority of multiplex gene sets, linking observed growth advantages to molecular functions including stress signaling, transcriptional regulation, and DNA damage response. These findings highlight the utility of slope-based growth modeling for quantitative interpretation of arrayed co-culture CRISPR screening systems. Discussion This study establishes a high-throughput co-culture CRISPR array to validate genetic determinants of cellular fitness and growth in CHO cells using competitive population dynamics. Building on our prior genome-wide CRISPR screening in CHO cells,21 the primary objective was to develop a sensitive and high-throughput validation approach capable of resolving subtle growth/fitness effects that are poorly captured by conventional monoculture growth assays but are highly relevant for next-generation cell lines in industrial cell line development. Initial experiments targeting essential or growth-inhibitory genes confirmed that VCC-based metrics are insufficient to resolve fitness differences in pooled edited populations: fitness reductions were transient, detectable only during the second week of cultivation, and masked by rapid outgrowth of non-edited cells, consistent with reported measurement variability and biological compensation in mammalian suspension cultures.27,28,43 In contrast, a competitive co-culture integrates differences in cell proliferation or survival over many divisions, providing a cumulative fitness readout that is less affected by variability from passaging, editing efficiency, or handling.31,32 Notably, even a modest growth advantage (on the order of ∼1%) can compound over weeks to substantially alter culture composition, underscoring the need to detect such subtle effects.15 To implement this, the co-culture system pairs a Cas9-expressing CHO cell with a CRISPR-inactive stable GFP+ reference pool, enabling fitness quantification by GFP ratio measurements over time. Using a FACS-enriched stable polyclonal GFP-expressing cell pool mitigates clone-specific background effects and biases, a known issue in CHO due to genomic and epigenetic heterogeneity.3,11,12,44 This pooled design is consistent with prior findings that mixed cell populations improve reliability and signal-to-noise when measuring subtle effects and allow more accurate representation of the parental line.40 In this proof-of-concept, competitive co-culture enabled inference of fitness changes using a single-flow cytometric readout. Differences in cellular fitness were reflected as changes in population ratios, quantified by time-resolved GFP ratios, culminating in depletion of lethal KO variants after 16 days of cultivation, also verified by sequencing. Notably, these population-level shifts were readily resolved at endpoint, whereas IVCC measurements failed to detect corresponding differences, underscoring the limited sensitivity of absolute growth metrics for capturing cumulative fitness effects. Building on the establishment and validation of the co-culture CRISPR array, we next scaled the approach to systematic evaluation of all single-gene targets and multiplex gene sets derived from our previous genome-wide screen, observing expected behavior across all control conditions.21 Across seven independent batches spanning >100 days, the co-culture assay remained stable. However, a drift in GFP ratios was observed across controls, with a graded magnitude reflecting the extent of cellular perturbation: the strongest drift occurred in INT controls, which undergo double-strand break formation and DNA damage repair-associated cell-cycle arrest; an intermediate effect was observed in NT controls, which experience substrate delivery and processing without productive genome editing; and the weakest drift was detected in untreated controls subjected only to electroporation. This pattern is consistent with editing-associated stress responses and supports the use of INT controls as the most appropriate baseline for normalization.21,37 The use of ΔGFP and control-normalized ΔΔGFP metrics enabled batch- and end-time-resolved interpretation of short- and long-term growth trajectories. The co-culture assay correctly identified known lethal and growth-inhibitory gene knockout controls and, critically, revealed growth-enhancing knockouts that monoculture assays fail to detect. Across all batches, we identified 28 (plus the 10x_Ultra_Gene_Set) variant genetic changes that result in GFP clearance, i.e., accelerated growth or increased fitness. Moreover, the validated growth-enhancing variants were not randomly distributed but showed enrichment for a limited number of growth-regulatory programs rather than representing isolated single-gene effects. These programs prominently included chromatin remodeling and transcriptional regulation (e.g., SWI/SNF-associated factors such as BRD9), stress- and nutrient-responsive signaling (including the mTOR regulator DDIT4 and the stress-responsive factor NDRG1), and pathways integrating DNA damage, cell cycle control, and growth signaling (e.g., LATS2 and ZEB1). Collectively, these hits point to regulatory frameworks that couple environmental stress, transcriptional state, and proliferative control, rather than single metabolic bottlenecks, as primary constraints on CHO cell growth. The pathway-level assignments rest on the functional-genomic enrichment of knockouts and on transcriptomic correlations, and the present study is by design hypothesis-generating with respect to mechanism: it identifies and prioritizes fitness-relevant genes and programs but does not establish the causal molecular events through which an individual knockout confers a growth advantage. Dissecting these mechanisms, for example, through targeted pathway perturbation, time-resolved multi-omics, and isogenic clonal characterization, is an explicit objective of follow-up work. Recent omics investigations in CHO have also identified these same pathways as major determinants of cell-growth phenotypes. For instance, differential proteomic profiling of fast-growing versus slower CHO producer lines revealed elevated DNA replication, mitotic cell-cycle activity, and evasion of p53-mediated apoptosis in the high-growth clones.14 Similarly, a comprehensive analysis of the CHO cell surface proteome found enrichment of receptors and ligands driving PI3K-AKT/mTOR, MAPK (ERK/JNK/p38), JAK-STAT, and Rap1 signaling, pathways known to promote proliferation and survival, relative to slower-growing cell models.6 These pathway-level enrichments are consistent with prior transcriptomic and proteomic studies in CHO cells, which associate favorable growth phenotypes with coordinated regulation of cell cycle progression, stress adaptation, and global transcriptional state rather than single metabolic bottlenecks.5,6,13,14 Importantly, none of the growth-enhancing knockouts corresponded to genes essential for viability, consistent with large-scale CRISPR screens indicating that growth rate control and cell survival are governed by largely separate genetic programs.21,24,33,45 This distinction is particularly relevant for CHO engineering, implying that cell proliferation could be accelerated by removing nonessential regulatory constraints or hyper-activating growth pathways without compromising core functions. In detail, many of the validated targets converged on a few shared regulatory axes controlling growth. For example, a chromatin remodeler like BRD9 influences global transcriptional accessibility, while nutrient and stress sensors such as DDIT4 modulate mTOR/p38 MAPK signaling, and BMP/TGF-β pathways adjust proliferation according to environmental conditions.46,47,48 In parallel, contact inhibition pathways (Hippo/YAP) link cell-cell interactions to proliferative control.49,50 Indeed, directly hyperactivating YAP in CHO cells was found to boost proliferation and even increase specific productivity by inducing growth-promoting effectors like c-MYC and mTORC1.51 Thus, the observation that many hits map to the same pathways suggests CHO cell growth is limited by a few regulatory programs rather than by isolated single-gene effects.51 Consistent with this idea, multiplex gene knockouts showed additive or partially synergistic fitness gains. The 10x_Ultra_Gene_Set, targeting multiple growth constraints simultaneously and derived from preliminary hits identified in batches 1–6, produced a markedly accelerated decline in GFP-positive cells across all batch 7 replicates, demonstrating that growth-promoting fitness effects can be combined and amplified in CHO cells. While these pooled perturbations remain genetically heterogeneous, extended co-culture naturally enriches for the most advantageous genotypes, enabling functional hit prioritization without immediate clonal isolation. To quantify fitness effects beyond endpoint ΔΔGFP values, a simple growth model was applied to the GFP ratio slopes to infer relative growth rate differences. Despite its simplifying assumptions, the inferred growth advantages aligned with independently measured growth rates, supporting the use of competitive co-culture as a quantitative proxy for growth performance. This modeling provides a basis for prioritizing candidates for follow-up engineering, while emphasizing that definitive validation will require isolating defined knockouts and testing them in production-relevant settings. Consistent with the slope-based inference, direct growth measurements confirmed that individual perturbations conferred modest but significant and consistent increases in growth rate, typically on the order of ∼1%–10% relative to controls. While such effects are unlikely to be immediately production-relevant in isolation, their magnitude is biologically meaningful and further multiplexing of 28x KO variant could enhance these effects further. This suggests that larger and potentially synergistic gains may require coordinated modulation of multiple functionally related genes or pathway-level interventions. Indeed, several multiplex gene sets exhibited stronger inferred and measured growth advantages than their individual components, motivating systematic deconvolution to distinguish additive enrichment from combinatorial effects and to identify rational gene combinations for subsequent engineering. To distinguish whether the fitness gains observed in multiplex gene sets arose from the joint action of multiple KOs or were attributable to individual perturbations within the set, we applied a stepwise deconvolution strategy integrating exclusion-based dissection with transcriptomic profiling; an approach conceptually aligned with established protocols for interpreting pooled multi-gene KO CRISPR screens.52,53 Growth advantages of gene sets were first compared with those of their constituent single-gene knockouts and in most cases, the phenotypes could be explained by the presence of individually advantageous perturbations, indicating enrichment of composite rather than synergistic effects, consistent with observations from large-scale genetic interaction mapping studies.52 For gene sets not fully resolved by this comparison, transcriptomic analysis at late screening stages enabled identification of functionally disrupted targets and revealed enrichment of stress response, transcriptional regulation, oxidative metabolism in mitochondria and DNA damage-associated pathways, mirroring strategies used in pooled perturbation-transcriptomics approaches to resolve complex genotypes.54,55 Although single-cell transcriptomics could in principle resolve co-causal perturbation effects, bulk RNA sequencing is appropriate here because functionally relevant variants are by design co-enriched under competitive enrichment, making their transcriptional consequences detectable at the population level. Unbiased gene set enrichment analysis of these transcriptomes (Figure S5) reinforced this picture, with several sets showing programs that converged on their disrupted drivers, most clearly the depletion of oxidative phosphorylation and mitochondrial respiration in set 100_DNAdam, consistent with down-regulation of NDRG1 and PML. Together, this approach enabled attribution of most multiplex phenotypes to defined genetic drivers while highlighting a subset likely arising from coordinated modulation of related regulatory programs. These results position singleplex and multiplex co-culture screening as a prioritization and enrichment tool for fitness-relevant perturbations, with the caveat that functional implementation of identified hits requires subsequent generation and characterization of stable monoclonal cell lines carrying the defined genetic modifications in isolation from competing populations. Extensive prior work in CHO cell engineering has shown that targeted modulation of apoptosis, metabolism, and stress-response pathways can substantially improve culture longevity and recombinant protein productivity under production-relevant conditions. Early studies demonstrated that deletion of pro-apoptotic regulators such as BAK and BAX yields apoptosis-resistant CHO cells with enhanced survival under bioreactor-relevant stresses,56,57,58 while more recent CRISPR-based approaches have refined these strategies to suppress caspase activation without compromising basal growth.59 In parallel, engineering of amino acid catabolism has repeatedly been shown to increase growth capacity, reduce toxic byproduct accumulation, and improve integrated viable cell density and productivity.16,60 Notably, many of these interventions act not by accelerating proliferation per se, but by improving cellular fitness through enhanced stress tolerance, metabolic efficiency, or survival at high cell density. Consistent with this, genome-wide and targeted CRISPR screens have identified genetic perturbations that rescue growth inhibition under process-relevant stress conditions, including histone deacetylase inhibitor treatment and nutrient-limited fed-batch cultivation.61,62 In this context, the growth- and fitness-enhancing perturbations identified here should be viewed as candidate fitness modifiers whose ultimate value must be evaluated in production-relevant settings. While the co-culture CRISPR array sensitively resolves subtle competitive advantages in suspension culture, translation to manufacturing benefit will require validation, individually and in rational combinations, in recombinant producer lines under fed-batch or perfusion conditions, where apoptosis resistance, metabolic robustness, and stress adaptation directly determine titer, product quality, and process stability.1,63 From an industrial perspective, growth rate remains a key determinant of development timelines, seed train efficiency, and process flexibility in CHO-based manufacturing.1,2,8 The results of the present study show that growth-limiting genetic programs in CHO cells can be systematically identified and functionally validated using a scalable suspension culture assay. At the same time, accelerating cell growth must be balanced against known trade-offs with specific productivity, metabolic load, and product quality, as slower proliferation has long been associated with higher cell-specific production in CHO systems (e.g., mild hypothermia or G1 arrest increases specific productivity at reduced growth rates),64 and genome-scale modeling demonstrates that growth rate and specific productivity represent competing cellular objectives that cannot be simultaneously maximized.65,66 Moreover, comparisons of CHO clones and culture conditions show that enhancements in specific productivity often coincide with reduced growth, indicating allocation of cellular resources away from proliferation toward biosynthesis (e.g., antibody production).67,68 Common process interventions such as low-temperature cultivation or sodium butyrate addition deliberately decelerate CHO cell growth to prolong culture viability and boost specific protein productivity.69 In such growth-inhibited cultures, cells typically accumulate in G0/G1 arrest and activate stress-response pathways, resulting in improved per-cell output.70 This exemplifies the inverse relationship between rapid cell proliferation and optimal recombinant protein production. Nonetheless, certain genetic engineering and expression system strategies can alleviate this trade-off. For instance, concurrent overexpression of MYC and XBP1s has been shown to improve CHO cell growth rate and product titer simultaneously,4,71,72 and inducible transgene expression systems have been developed that allow recombinant protein production to be triggered at defined stages of culture, enabling separation of growth and production phases.73 Moreover, hybrid expression systems combining inducible and constitutive promoters enable selection of fast-growing clones under uninduced conditions during cell line development, while full productivity is subsequently triggered during the production phase, effectively decoupling the growth and productivity requirements across culture stages.60 Thus, the co-culture CRISPR array serves as an upstream screening platform to pinpoint promising growth-enhancing (or fitness-reducing) gene perturbations for later evaluation in production cell lines under process-relevant conditions.51,72 In conclusion, the competitive co-culture CRISPR platform provides a sensitive and practical framework for fitness hypothesis testing and growth engineering in CHO cells. By resolving subtle, pathway-level growth constraints undetectable in conventional assays, this approach bridges genome-scale functional genomics with industrial cell line development. Integrating competitive dynamics with multiplex perturbations and quantitative modeling provides a rational basis for engineering next-generation CHO host cells with improved growth performance while maintaining recombinant protein productivity and product quality. Limitations of the study This work was designed as a screening and prioritization study to establish a scalable co-culture fitness assay. Several limitations result from that scope. First, sgRNAs were selected based on prior screen ranks rather than being re-validated individually in the arrayed format; therefore, as with all CRISPR screens, false negatives cannot be categorically excluded, if guides underperform under the chosen delivery conditions. Relatedly, knockout efficiency varies across guides and target loci, owing to differences in intrinsic guide activity, chromatin accessibility, and the number of alleles requiring biallelic disruption, so the fraction of fully functional knockouts differs between genes. A strongly fitness-relevant gene edited at low efficiency may therefore generate a weaker competitive signal than a moderately relevant gene edited at high efficiency, biasing apparent effect sizes and potentially producing false negatives. The competitive format partially buffers this effect, because even a small fully edited subpopulation is progressively enriched or depleted over extended co-culture and the KO-score trajectory reports the functionally edited fraction directly; nonetheless, absolute ΔΔGFP magnitudes should be interpreted as efficiency-dependent lower bounds on the intrinsic fitness impact rather than efficiency-independent measures of it, and definitive ranking of hits will require knockout-efficiency-controlled validation with confirmed biallelic editing. Off-target effects were minimized via established guide design rules; using two sgRNAs per gene reduces the likelihood of shared off-target-driven phenotypes, but all hits intended for engineering will require orthogonal validation with independent guides and genotype confirmation. Second, the assay captures integrated competitive growth and fitness, which inherently reflects contributions from multiple underlying biological processes like reduced spontaneous death, improved stress tolerance, or improved recovery after electroporation and DNA repair. Additional phenotyping will be needed to assign mechanisms to individual hits. Third, screening was performed in 96-well plates at small volumes. Quantitative scaling requires confirmation, as typical for all high-throughput CHO screening platforms. This method is therefore best viewed as a small-scale hypothesis testing tool, with scale-up confirmation required for downstream application. Fourth, multiplex gene sets introduce genotype heterogeneity as high-order editing yields mixtures of incomplete sub-combinations from which the most beneficial sub-combinations are automatically enriched through long-term cultivation, but the assay does not directly identify the exact edit combinations responsible without follow-up reconstruction of defined multi-knockout genotypes. Transcriptomic profiling helped prioritize candidates within unresolved sets, but imperfect transcript knockdown and late-culture adaptation can limit causal interpretation. Fifth, the study used proprietary Sartorius Stedim Cellca, GmbH, CHO host-cell systems. The CHO host cell line may be accessible under a research or commercial licensing agreement, and the final co-culture pools may be available under a material transfer agreement. These access restrictions may limit independent reproduction in the identical cellular background. Finally, growth and fitness were measured in a non-producer context. Before industrial deployment, growth-enhancing edits must be evaluated for potential trade-offs in recombinant protein productivity, product quality attributes, and long-term stability in relevant producer cell lines and bioprocess settings. Resource availability Lead contact Requests for further information and resources should be directed to, and will be fulfilled by, the lead contact, Jannis Peter Marzluf (jannis.marzluf@sartorius.com). Materials availability This study did not generate new unique reagents. The parental and engineered CHO host cell lines used in this study are proprietary to Sartorius Stedim Cellca, GmbH. Access to the CHO host cell line may be considered under a research or commercial licensing agreement. The final co-culture pools generated and used in this study may be made available from the lead contact under a material transfer agreement (MTA), subject to material availability, institutional approval, applicable legal and commercial restrictions, and execution of appropriate confidentiality and use agreements. Restricted access to these proprietary CHO cell-line systems is acknowledged as a limitation of the study. Data and code availability •Raw RNA-seq data generated in this study have been deposited in the NCBI Sequence Read Archive under BioProject accession PRJNA1489724 and are publicly available as of the date of publication. The accession number is listed in the key resources table. Processed RNA-seq data supporting the transcriptomic analyses are provided in the supplemental tables, including the raw gene count matrix (Table S5), GeTMM-normalized expression matrix (Table S6), DESeq2-derived results (Table S7), and GSEA outputs for the profiled multiplex gene sets compared with INT controls (Table S8). The CHO host cell line used in this study is proprietary to Sartorius Stedim Cellca GmbH and may be accessible under a research or commercial licensing agreement. The final co-culture pools generated and used in this study may be available from the lead contact under a material transfer agreement, subject to material availability, applicable legal and commercial restrictions, and execution of appropriate confidentiality and use agreements.•The custom code used for the GFP co-culture time-course analysis, including endpoint fitness metrics (ΔGFP and ΔΔGFP), statistical hit calling, and peak-slope growth modeling, is provided in Data S1. The corresponding algorithms, formulas, software packages, parameters, and thresholds are described in the STAR Methods.•This paper does not report proprietary algorithms for RNA-seq processing or analysis. RNA-seq read processing, quantification, normalization, differential-expression analysis, and enrichment analysis were performed using standard open-source tools and explicitly described parameters, as detailed in the STAR Methods. The internal Sartorius workflow repository is not publicly released because it contains company-internal infrastructure, file organization, sample identifiers, and reference-resource handling; however, the deposited raw sequencing data, processed count tables, differential-expression results, enrichment outputs, software versions, parameters, and formulas provided in this paper enable independent reanalysis of the reported RNA-seq results. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request. •Raw RNA-seq data generated in this study have been deposited in the NCBI Sequence Read Archive under BioProject accession PRJNA1489724 and are publicly available as of the date of publication. The accession number is listed in the key resources table. Processed RNA-seq data supporting the transcriptomic analyses are provided in the supplemental tables, including the raw gene count matrix (Table S5), GeTMM-normalized expression matrix (Table S6), DESeq2-derived results (Table S7), and GSEA outputs for the profiled multiplex gene sets compared with INT controls (Table S8). The CHO host cell line used in this study is proprietary to Sartorius Stedim Cellca GmbH and may be accessible under a research or commercial licensing agreement. The final co-culture pools generated and used in this study may be available from the lead contact under a material transfer agreement, subject to material availability, applicable legal and commercial restrictions, and execution of appropriate confidentiality and use agreements. Raw RNA-seq data generated in this study have been deposited in the NCBI Sequence Read Archive under BioProject accession PRJNA1489724 and are publicly available as of the date of publication. The accession number is listed in the key resources table. Processed RNA-seq data supporting the transcriptomic analyses are provided in the supplemental tables, including the raw gene count matrix (Table S5), GeTMM-normalized expression matrix (Table S6), DESeq2-derived results (Table S7), and GSEA outputs for the profiled multiplex gene sets compared with INT controls (Table S8). The CHO host cell line used in this study is proprietary to Sartorius Stedim Cellca GmbH and may be accessible under a research or commercial licensing agreement. The final co-culture pools generated and used in this study may be available from the lead contact under a material transfer agreement, subject to material availability, applicable legal and commercial restrictions, and execution of appropriate confidentiality and use agreements. •The custom code used for the GFP co-culture time-course analysis, including endpoint fitness metrics (ΔGFP and ΔΔGFP), statistical hit calling, and peak-slope growth modeling, is provided in Data S1. The corresponding algorithms, formulas, software packages, parameters, and thresholds are described in the STAR Methods. The custom code used for the GFP co-culture time-course analysis, including endpoint fitness metrics (ΔGFP and ΔΔGFP), statistical hit calling, and peak-slope growth modeling, is provided in Data S1. The corresponding algorithms, formulas, software packages, parameters, and thresholds are described in the STAR Methods. •This paper does not report proprietary algorithms for RNA-seq processing or analysis. RNA-seq read processing, quantification, normalization, differential-expression analysis, and enrichment analysis were performed using standard open-source tools and explicitly described parameters, as detailed in the STAR Methods. The internal Sartorius workflow repository is not publicly released because it contains company-internal infrastructure, file organization, sample identifiers, and reference-resource handling; however, the deposited raw sequencing data, processed count tables, differential-expression results, enrichment outputs, software versions, parameters, and formulas provided in this paper enable independent reanalysis of the reported RNA-seq results. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request. This paper does not report proprietary algorithms for RNA-seq processing or analysis. RNA-seq read processing, quantification, normalization, differential-expression analysis, and enrichment analysis were performed using standard open-source tools and explicitly described parameters, as detailed in the STAR Methods. The internal Sartorius workflow repository is not publicly released because it contains company-internal infrastructure, file organization, sample identifiers, and reference-resource handling; however, the deposited raw sequencing data, processed count tables, differential-expression results, enrichment outputs, software versions, parameters, and formulas provided in this paper enable independent reanalysis of the reported RNA-seq results. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request. Acknowledgments We thank the colleagues in the Corporate Research division at Sartorius Stedim Cellca, GmbH, for their valuable input, continued feedback, and assistance with cell culture operations throughout this project. This study was funded by Sartorius Stedim Cellca, GmbH. Author contributions J.P.M., conceptualization, investigation, writing – original draft, methodology, validation, visualization, formal analysis, data curation, and software; L.H., writing – review and editing, investigation, methodology, and validation; D.H., writing – review and editing and software; A.-C.L., writing – review and editing, project administration, and supervision; C.Z., writing – review and editing, conceptualization, supervision, resources, and funding acquisition. Declaration of interests This work is part of pending European patent applications (EP25181372.1 and EP25167486.7) filed by the authors and Sartorius Stedim Cellca, GmbH, related to co-culture-based high-throughput growth rate engineering and modified CHO cells with a genetically engineered faster growth rate. J.P.M., C.Z., and A.-C.L. are inventors on these applications. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work, the authors used an AI-based language model API (ChatGPT, OpenAI) for drafting and grammar refinement. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERChemicals, peptides, and recombinant proteinsL-glutamine (200 mM)Sartorius Stedim Cellca GmbHCat#03-020-1BTrueCut™ Cas9 Protein v2 (SpCas9)Thermo Fisher ScientificCat#A36499QuickExtract™ DNA Extraction SolutionLucigenCat#QE090504Cell® SmartCHO Stock and Adaptation MediumSartoriusCat#CFP3FB1106PuromycinGibco™, Thermo Fisher ScientificCat#A1113802Trypan blue 0.4%Gibco™, Thermo Fisher ScientificCat#15250061DPBSGibco™, Thermo Fisher ScientificCat#14190094Critical commercial assaysNEON™ Transfection System 100 μL KitThermo Fisher ScientificCat#MPK10096NEON™ Transfection System 10 μL KitThermo Fisher ScientificCat#MPK1096Experimental models: cell linesCHO-DG44 (DHFR) −/−Sartorius proprietaryN/ACHO mock cell lineSartorius proprietaryN/ACHOCas9+ cell lineSartorius proprietaryN/ADeposited dataRaw sequencing dataNCBI SRA BioProject: PRJNA1489724N/AOligonucleotidesTrueGuide™ Synthetic sgRNAsThermo Fisher ScientificSee Table S7Recombinant DNApRSGEGP-U6-sg-EF1-TagGFP2-2A-PuroCellectaCat#SVCRU6EP-LpMockSartorius proprietaryN/ASoftware and algorithmsGeneious Prime (v11.0.11)BiomattersGeneious | Bioinformatics Software for Sequence Data AnalysisGraphPad Prism (v10.6.1)DotmaticsHome - GraphPadInference of CRISPR Edits (ICE) (v3.0)SynthegoEditCo - CRISPR Performance AnalysisBD FACSDiva™ softwareBD BiosciencesBD FACSDiva™ Software | BD BiosciencesPython (v3.13.2)Python Software Foundation.74Python Release Python 3.13.2 | Python.orgpandas (v2.2.2)McKinney et al.75pandas · PyPINumPy (v1.26.4)Harris et al.76numpy · PyPISciPy (v1.13.1)Virtanen et al.77scipy · PyPIMatplotlib (v3.9.2)Hunter, J.D.78matplotlib · PyPIseaborn (v0.13.2)Waskom et al.79v0.8.1 (September 2017) — seaborn 0.13.2 documentationSnakemake (v7.8.3)Mölder et al.80Snakemake — Snakemake 7.8.3 documentationFastQC (v0.12.1)Andrews81GitHub - s-andrews/FastQC: A quality control analysis tool for high throughput sequencing data · GitHubTrimmomatic (v0.39)Bolger et al.82Home - TrimmomaticKallisto (v0.48.0)Bray et al.83Releases · pachterlab/kallistoR (v4.4.2)R Core Team84Index of/bin/windows/base/old/4.2.0Bioconductor (v3.20)Huber et al.85https://doi.org/10.1038/nmeth.3252tximport (v1.34.0)Soneson et al.86Bioconductor - tximportedgeR (v4.4.2)Chen et al.87https://doi.org/10.1093/nar/gkaf018DESeq2 (v 1.46.0)Love et al.88N/Agprofiler2 (v0.2.3)Kolberg and Raudvere89https://doi.org/10.32614/CRAN.package.gprofiler2WikiPathwaysAgrawak et al.90https://doi.org/10.1093/nar/gkad960GO.db (v3.20.0)Bioconductor91Bioconductor - GO.dborg.Hs.e.g.dbBioconductor92Bioconductor - org.Hs.eg.dbAnnotationHub (v3.14.0)Bioconductor93Bioconductor - AnnotationHubOtherNEON™ Transfection SystemThermo Fisher ScientificInvitrogen Neon Transfection System Je 1 | Buy Online | Invitrogen™ | Fisher ScientificCASY® Cell Counter & AnalyzerOMNI Life ScienceCell Counter Casy | Your Home for Cell ResearchFluent® automated liquid handlerTecan Group Ltd.Fluent®Automated Workstation - TecanCellavista™ imaging systemSynentecCELLAVISTA 4BD FACSAria™ cell sorterBD BiosciencesBD FACSAria™ III | High Sensitivity Flow Cytometer4D-Nucleofector™ SystemLonzaElectroporation and Nucleofector® Technology | LonzaiQue® Screener v3SartoriusiQue® 3 High-Throughput Screening (HTS) Cytometry Platform | Sartorius REAGENT or RESOURCESOURCEIDENTIFIERChemicals, peptides, and recombinant proteinsL-glutamine (200 mM)Sartorius Stedim Cellca GmbHCat#03-020-1BTrueCut™ Cas9 Protein v2 (SpCas9)Thermo Fisher ScientificCat#A36499QuickExtract™ DNA Extraction SolutionLucigenCat#QE090504Cell® SmartCHO Stock and Adaptation MediumSartoriusCat#CFP3FB1106PuromycinGibco™, Thermo Fisher ScientificCat#A1113802Trypan blue 0.4%Gibco™, Thermo Fisher ScientificCat#15250061DPBSGibco™, Thermo Fisher ScientificCat#14190094Critical commercial assaysNEON™ Transfection System 100 μL KitThermo Fisher ScientificCat#MPK10096NEON™ Transfection System 10 μL KitThermo Fisher ScientificCat#MPK1096Experimental models: cell linesCHO-DG44 (DHFR) −/−Sartorius proprietaryN/ACHO mock cell lineSartorius proprietaryN/ACHOCas9+ cell lineSartorius proprietaryN/ADeposited dataRaw sequencing dataNCBI SRA BioProject: PRJNA1489724N/AOligonucleotidesTrueGuide™ Synthetic sgRNAsThermo Fisher ScientificSee Table S7Recombinant DNApRSGEGP-U6-sg-EF1-TagGFP2-2A-PuroCellectaCat#SVCRU6EP-LpMockSartorius proprietaryN/ASoftware and algorithmsGeneious Prime (v11.0.11)BiomattersGeneious | Bioinformatics Software for Sequence Data AnalysisGraphPad Prism (v10.6.1)DotmaticsHome - GraphPadInference of CRISPR Edits (ICE) (v3.0)SynthegoEditCo - CRISPR Performance AnalysisBD FACSDiva™ softwareBD BiosciencesBD FACSDiva™ Software | BD BiosciencesPython (v3.13.2)Python Software Foundation.74Python Release Python 3.13.2 | Python.orgpandas (v2.2.2)McKinney et al.75pandas · PyPINumPy (v1.26.4)Harris et al.76numpy · PyPISciPy (v1.13.1)Virtanen et al.77scipy · PyPIMatplotlib (v3.9.2)Hunter, J.D.78matplotlib · PyPIseaborn (v0.13.2)Waskom et al.79v0.8.1 (September 2017) — seaborn 0.13.2 documentationSnakemake (v7.8.3)Mölder et al.80Snakemake — Snakemake 7.8.3 documentationFastQC (v0.12.1)Andrews81GitHub - s-andrews/FastQC: A quality control analysis tool for high throughput sequencing data · GitHubTrimmomatic (v0.39)Bolger et al.82Home - TrimmomaticKallisto (v0.48.0)Bray et al.83Releases · pachterlab/kallistoR (v4.4.2)R Core Team84Index of/bin/windows/base/old/4.2.0Bioconductor (v3.20)Huber et al.85https://doi.org/10.1038/nmeth.3252tximport (v1.34.0)Soneson et al.86Bioconductor - tximportedgeR (v4.4.2)Chen et al.87https://doi.org/10.1093/nar/gkaf018DESeq2 (v 1.46.0)Love et al.88N/Agprofiler2 (v0.2.3)Kolberg and Raudvere89https://doi.org/10.32614/CRAN.package.gprofiler2WikiPathwaysAgrawak et al.90https://doi.org/10.1093/nar/gkad960GO.db (v3.20.0)Bioconductor91Bioconductor - GO.dborg.Hs.e.g.dbBioconductor92Bioconductor - org.Hs.eg.dbAnnotationHub (v3.14.0)Bioconductor93Bioconductor - AnnotationHubOtherNEON™ Transfection SystemThermo Fisher ScientificInvitrogen Neon Transfection System Je 1 | Buy Online | Invitrogen™ | Fisher ScientificCASY® Cell Counter & AnalyzerOMNI Life ScienceCell Counter Casy | Your Home for Cell ResearchFluent® automated liquid handlerTecan Group Ltd.Fluent®Automated Workstation - TecanCellavista™ imaging systemSynentecCELLAVISTA 4BD FACSAria™ cell sorterBD BiosciencesBD FACSAria™ III | High Sensitivity Flow Cytometer4D-Nucleofector™ SystemLonzaElectroporation and Nucleofector® Technology | LonzaiQue® Screener v3SartoriusiQue® 3 High-Throughput Screening (HTS) Cytometry Platform | Sartorius Experimental model and study cell line details CHO DG44 cells were originally obtained from Lawrence Chasin (Columbia University, New York, USA) and had been previously adapted in-house for serum-free suspension growth in chemically defined media. A mock control cell line had been generated and validated prior to the start of the present study, as described previously,21 and served as the parental background for all experiments reported here. Additionally, also the CHOCas9+ Cell Line was generated in a previous study.21 Cells were maintained in serum- and antibiotic-free chemically defined 4Cell® SmartCHO Stock and Adaptation Medium (Sartorius, Ulm, Germany) supplemented with 6 mM L-glutamine under suspension conditions and cultivated in shake flasks at 36.8 °C, 7.5% CO2, and 80% relative humidity unless stated otherwise. Cultures were subcultivated every two to four days to an inoculation density of 1E5 cells/mL or 2E5 cells/mL. Cells were routinely cultured in 96-well plates (300 μL) using automated liquid handling (Fluent®, Tecan, Männedorf, Switzerland) with custom scripts or manually expanded in shake flasks (25 mL) for cell banking and limited experimentation. Cell density and viability were monitored using Cellavista™ imaging with trypan blue exclusion (Synentec, Elmshorn, Germany) or a CASY® Cell Counter & Analyzer (OMNI Life Science, Bremen, Germany). All parental cultures were routinely tested and confirmed negative for mycoplasma contamination. Method details Generation of GFP+ controls GFP-positive control cell pools were generated by transfecting DHFR competent CHO mock cells with a commercial GFP expression vector (pRSEGP-U6-sg-EF1-TagGFP2-2A-Puro), which enables constitutive GFP expression and puromycin resistance. The construct was used exclusively as a fluorescent marker and selection cassette for competitive co-culture assays. Transfections were performed using the NEON™ Transfection System with the NEON™ Transfection System 100 μL Kit (Thermo Fisher Scientific, Waltham, MA, USA). Briefly, 1E6 cells were electroporated with 10 μg plasmid DNA using settings of 1400 V, two pulses, and a pulse width of 20 ms. Following electroporation, cells were resuspended at a density of 3E5 cells/mL in 4Cell® SmartCHO Stock and Adaptation Medium (Sartorius, Ulm, Germany) supplemented with 6 mM L-glutamine. Two days post-transfection, puromycin was added at a final concentration of 10 μg/mL to select for GFP-expressing cells. Selection was maintained for 14 days, after which puromycin was withdrawn and cells were allowed to recover under standard culture conditions. Recovered GFP-positive cell pools were subsequently expanded and used for downstream evaluation. To mitigate expression instability and reduce heterogeneity in GFP intensity, the generated GFP-pools were subsequently pooled and subjected to fluorescence-activated cell sorting (FACS). Sorting was performed on a BD FACSAria™ system using BD FACSDiva™ software (BD Biosciences, San Jose, CA, USA). Cells were gated based on GFP fluorescence intensity and sorted into four distinct populations using increasing GFP intensity thresholds. For each gate, approximately 200,000 cells were collected per well into 6-well plates containing pre-warmed culture medium. Sorted populations were expanded separately and cell banked to be used in downstream co-culture fitness assays. Genome editing and knockout validation To confirm CRISPR competence, the Cas9-expressing cell lines were tested for genome-editing by transfecting synthetic sgRNA molecules targeting specific genomic loci. All gene targets and sgRNA sequences for testing are shown in Table S1 sgRNA sequences were designed using Geneious Prime (v11.0.11; Biomatters, Auckland, New Zealand), targeting early exons common to all transcript variants to induce frameshift mutations and disrupt protein function. Top sgRNAs were selected based on predicted on-target efficiency36 and off-target specificity,35 ensuring that potential off-target sites resided within non-coding regions. Off-target analysis was performed against the fully sequenced and annotated host genome. The used sgRNAs were screened in previous internal studies conducted at Sartorius Stedim Cellca GmbH.21,40,94 TrueGuide™ Synthetic sgRNAs (Thermo Fisher Scientific, Waltham, MA, USA) were delivered at 7.5 or 20.0 pmol into 2E5 CRISPR-competent CHOCas9+ cells. As controls for editing efficiency, the same sgRNAs were complexed with TrueCut™ Cas9 Protein v2 (Thermo Fisher Scientific, Waltham, MA, USA) at a 1:1 molar ratio (7.5 pmol each) to form ribonucleoprotein (RNP) complexes. RNPs were electroporated into 2E5 cells using the NEON™ Transfection System and the 10 μL Kit (Thermo Fisher Scientific, Waltham, MA, USA), applying optimized standard conditions: 1700 V, 20 ms pulse width and 1 pulse. To verify the functionality of generated knockouts, sequencing was performed. Genomic DNA was extracted from cell pools using QuickExtract™ DNA Extraction Solution (Lucigen, Middleton, WI, USA). PCR was performed to amplify regions flanking the sgRNA cut site (∼200–300 bp upstream and downstream), and amplicons were purified and subjected to Sanger sequencing (Microsynth Seqlab GmbH, Göttingen, Germany). Sequencing data were analyzed using the Inference of CRISPR Edits (ICE) tool (v3.0; https://ice.synthego.com, Synthego, Menlo Park, USA) following the developer’s guidelines. ICE analysis provided both the InDel percentage, representing the overall editing efficiency, and the KO score, which reflects the proportion of cells harboring frameshift InDels or fragment deletions likely to result in functional gene KO. Candidate genes, sgRNA selection and array composition Candidate genes and functional gene sets were selected from a previously published genome-wide CRISPR knockout screen in CHO cells,21 which identified candidate genetic perturbations associated with enhanced cellular growth. In that study, six sgRNAs per gene were screened, enabling gene-level fitness ranking. From this dataset, 235 growth-associated genes were found and grouped into 155 functionally annotated gene sets based on shared molecular pathways and biological processes.21 For arrayed validation in this study, all targets were drawn from this ranked hit list. For single-gene validation, two sgRNAs per gene were selected based on the highest log2 fold-change contributions observed in the original screen data. All sgRNA sequences used for the CRISPR array are shown in Table S4. Multiplex gene sets were assembled by selecting the top-ranked sgRNA per constituent gene, following the gene-set definitions established in the original screen.21 Control conditions included intron-targeting sgRNAs (INT) to induce CRISPR-associated DNA damage without disrupting coding sequences, non-targeting sgRNAs (NT) to control for sgRNA delivery burden, and wild-type electroporation controls lacking sgRNA substrate. All variants were screened in an arrayed 96-well format across seven experimental batches, each containing internal controls to enable batch- and day-matched normalization. Co-culture assay design and execution Competitive co-cultures were established by mixing a CRISPR-competent CHOCas9+ population with a CRISPR-deficient GFP-positive reference population, enabling fitness quantification as changes in the GFP-positive fraction over time. Unless stated otherwise, co-cultures were initialized at an approximately 1:1 ratio (GFP+:GFP−), such that only the GFP-negative subpopulation underwent genome editing following sgRNA delivery. sgRNA substrates (single-gene or multiplex sets) were delivered by electroporation at 20 pmol for singleplex approaches and 50 pmol total sgRNAs for multiplex approaches using a 4D-Nucleofector™ System (Lonza, Basel, Switzerland) with pulse code DG-208 applied sequentially twice per well. Following electroporation, cultures were allowed to recover and were maintained by regular passaging every 3–4 days. For arrayed screening, co-cultures were distributed into 96-well plates at 300 μL per well using automated liquid handling (Fluent®, Tecan Group Ltd., Männedorf, Switzerland). At each passage, cultures were thoroughly resuspended to minimize settling bias, cell density was determined, and cultures were re-seeded to a defined inoculation density (typically 1E5 to 2E5 cells/mL) using automated liquid handling. Cell cultures were homogenized using a 96-channel pipetting arm immediately prior to aspiration for sampling or passaging using an automated four-point well resuspension routine with two aspiration–dispense cycles per point in each well to ensure uniform cell suspension and minimize settling effects. Following resuspension, liquid transfer was performed using the 96-channel pipetting arm for sampling. Aspirated volumes were dispensed into DPBS (Gibco™, Thermo Fisher Scientific, Waltham, MA, USA), supplemented with trypan blue for viability staining, and viable cell concentration (VCC) was determined using a Cellavista (Synentec, Elmshorn, Germany) instrument. For passaging, cells were resuspended prior to splitting, and an 8-channel pipetting arm was used to allow flexible volume dispensing per well. Defined split ratios were implemented through automated volume calculations on a per-well basis, including automated preparation and distribution of fresh media to achieve the desired final culture volumes and cell densities. All aspiration, dispensing, and volume calculations were executed under scripted liquid handling control to ensure consistency and reproducibility across plates and batches. All liquid handling steps, including cell seeding, media exchange, passaging, and re-seeding, were executed in a controlled and precise manner to minimize operator-dependent variability. To limit temperature- and CO2-induced perturbations, plate exposure time outside the incubator was strictly minimized. Consequently, plates were processed in discrete batches, each comprising two to four plates, allowing individual handling and rapid return to incubation. All automated liquid handling procedures were validated for method accuracy, precision, and reproducibility prior to experimental use (data not shown). System performance qualification and routine calibration were performed in accordance with the manufacturer’s specifications to ensure consistent volumetric fidelity and positional accuracy across runs. Cultivation continued for up to 116 days for 3,000+ parallel cultures depending on the number of experimental batches run. Sampling for flow cytometry was performed at every second passage where feasible or at predefined time points, depending on resource availability. Cultures were terminated when further expansion was no longer feasible under the established passaging schedule or when sufficient GFP time-course data had been acquired. Experimental batches were processed as independent runs, each comprising up to four 96-well plates and including internal controls for normalization and quality control. These controls comprised intron-targeting (INT), non-targeting (NT), wild-type electroporation controls (WT), and GFP-only controls to monitor transfection efficiency and flow cytometry gating performance. Batch start dates were defined as the day of sgRNA delivery, and all subsequent measurements were aligned post hoc to the corresponding batch start date, as described in the quantification and statistical analysis section. Flow cytometry acquisition and GFP quantification GFP-positive fractions were quantified by high-throughput flow cytometry as the primary readout for relative fitness/growth in the competitive co-culture assay. Measurements were performed using an iQue Screener v3 system (Sartorius, Ann Arbor, MI, USA), which operates with fixed-time acquisition and does not use event-based stopping rules. At each sampling time point, cultures were resuspended using automated plate shaking to ensure homogeneous cell suspension, and samples were acquired directly from the culture plates immediately after subcultivation, without intermediate washing or staining. Data acquisition was performed with a fixed acquisition time of 2 s per well, ensuring a consistent sample volume across all conditions. GFP was excited at 488 nm and detected using a standard FITC/GFP detection channel. Instrument control and acquisition were performed using manufacturer-provided software (Sartorius, Ann Arbor, MI, USA). Events were gated using a standardized, hierarchical strategy applied uniformly across all experiments (Figure S2). First, cellular events were identified based on forward- and side-scatter properties (FSC-H vs. SSC-H) to exclude debris and non-cellular events (Figure S2A). Second, singlets were selected from the cell population using FSC-A versus FSC-H pulse geometry to exclude doublets and aggregates (Figure S2B). Third, the singlet population was refined using FSC-H versus SSC-H to define the viable cell population and remove remaining outliers (Figure S2C). Within the resulting viable singlet gate, GFP-positive events were identified using GFP-negative parental controls to define the GFP positivity threshold. Representative GFP intensity distributions for the three co-culture pools (CC1, CC2, and CC3) are shown in Figures S2D–S2F. In each case, GFP-positive and GFP-negative subpopulations were clearly resolved, enabling accurate quantification of GFP-positive fractions over time. The GFP-positive fraction was reported as the percentage of GFP-positive events among gated viable singlet cells. GFP gating thresholds were periodically verified using GFP-negative and GFP-positive reference controls to ensure stability across long-term experiments. No adaptive or dynamic gating was applied, all gates were defined using fixed criteria derived from controls and applied consistently during downstream analysis. Because the iQue Screener v3 uses fixed-time rather than fixed-event acquisition, absolute event counts per well varied naturally as a function of cell density, viability. To characterize acquisition variability, total singlet counts and GFP-positive event counts were compiled and analyzed globally across all experimental days, samples, and batches (Figure S6). Wells yielding zero recorded events were excluded from downstream analyses. No additional filtering or exclusion based on low but nonzero event counts was applied, and all remaining measurements were retained irrespective of absolute count magnitude. Wells exhibiting clear technical acquisition failures (e.g., incomplete sample uptake or aberrant scatter distributions) were excluded prior to analysis such exclusions were rare and did not affect downstream statistical analyses due to sufficient biological replication. Cell growth measurements and derived metrics Viable cell concentration (VCC) and cell viability were measured at each passage using Cellavista™ automated imaging with trypan blue exclusion (Synentec, Elmshorn, Germany) after automated sampling with TECAN (Fluent®, Tecan Group Ltd., Männedorf, Switzerland) or a CASY® Cell Counter & Analyzer (OMNI Life Science, Bremen, Germany), according to the manufacturers’ instructions. Growth curves were generated by plotting VCC as a function of aligned culture day to enable comparison across independent experimental runs. Where indicated, VCC values were normalized to the corresponding wild-type or control condition measured within the same experimental batch and at the same culture day. Integrated viable cell concentration (IVCC) was used as a cumulative measure of cell growth over defined culture intervals. IVCC values were computed from time-resolved VCC measurements using numerical integration as described in formula 1, Table 1. For long-term cultures, passaging frequency and split ratios were selected to maintain cell densities within the operational range of the culture system and to minimize confounding effects arising from nutrient depletion. VCC-derived growth metrics were used as readouts to monitor culture behavior and confirm expected phenotypes but were not employed as primary fitness inference parameters. In screening experiments, primary growth/fitness inference was based on GFP-ratio dynamics in competitive co-culture. Quantification and statistical analysis Statistical analyses and most data visualizations were performed using GraphPad Prism (v10.6.1; GraphPad Software, USA) unless stated otherwise. Details on sample sizes (n), replicate definitions, summary statistics (mean ± SD), and the statistical tests applied are provided in the corresponding figure legends. No randomization or blinding was performed, as experiments were conducted in vitro and involved predefined experimental groups. No data points were excluded unless explicitly stated. Statistical Analysis of Co-Culture fitness assays Flow cytometry–derived GFP-positive fractions were used as a relative measure of population fitness in competitive co-culture assay. Raw GFP measurements were aggregated across time points and aligned to batch-specific start dates to enable comparison across independent experimental runs. All data processing and statistical analyses were performed using Python v3.13.2,74 unless otherwise stated, with tabular data handling implemented in pandas v2.2.275 and numerical operations performed using NumPy v1.26.4.76 For each sample, GFP trajectories were normalized to a batch-specific reference value (Day 0), reflecting the initial co-culture composition. Endpoint fitness effects were quantified as ΔGFP, defined as the difference between the final observed GFP fraction and the corresponding Day 0 reference for a given sample (Formula 2; Table 2). The evaluation time point for each sample (“Last_Day”) was defined as the latest aligned culture day with a non-missing GFP measurement, ensuring that endpoint comparisons were made at matched culture durations to batch and day correct controls. To account for batch effects and time-dependent drift, ΔΔGFP values were calculated by normalizing each sample’s ΔGFP to the mean ΔGFP of matched control populations measured in the same batch and at the same evaluation day (Formula 3; Table 2). Intron-targeting controls (INT) were used as the primary normalization baseline unless stated otherwise. For each control group, ΔΔGFP distributions were computed using batch- and day-matched ΔGFP values, ensuring that normalization reflected contemporaneous control behavior rather than global averages. Statistical testing was performed at the gene or gene-set level using replicate-level ΔΔGFP values. For each gene (or multiplex gene set) within each batch, ΔΔGFP distributions were compared against the corresponding INT control distribution from the same batch and evaluation day using one-sided Student’s t-tests (alternative hypothesis: mean ΔΔGFP <0), as implemented in SciPy v1.13.1.77 Tests were only performed when at least two biological replicates were available for both the target and control groups on a given day and target gene. When a gene was evaluated across multiple batches in rare cases, stratum-level p-values were combined using Fisher’s method SciPy v1.13.177 to obtain a single gene-level significance estimate. Multiple-testing correction across genes was performed using the Benjamini–Hochberg false discovery rate (FDR) method, implemented in custom Python code using NumPy v1.26.4.76 Hits were defined as genes or gene sets with an adjusted p-value <0.05 and a mean endpoint ΔΔGFP below a 2x SD effect-size threshold derived from the variability of INT controls (−18.3%). Summary statistics are reported as mean ± standard deviation unless stated otherwise. Data visualization, including volcano plots, ranked effect-size plots, density distributions, and time-course summaries, was performed using Matplotlib v3.9.278; kernel density estimates were generated where applicable using seaborn v0.13.2.79 For time-resolved comparisons of individual gene knockouts against a single neutral reference condition (Figures 1D, 1E, 3D, and 3F), statistical significance was assessed using one-way ANOVA with Dunnett’s multiple-comparisons post-hoc test in GraphPad Prism (v10.6.1), comparing each target group’s mean value at a given time point to the corresponding WT or FUT8 (Figures 3D and 3F) reference group. A significance threshold of p < 0.05 was applied throughout. For comparisons of measured growth rates across multiple knockout groups and gene sets against INT-controls after extended cultivation (Figures 5C and 5D), statistical significance was assessed using Brown-Forsythe and Welch ANOVA in GraphPad Prism (v10.6.1), followed by Benjamini–Hochberg FDR correction across comparisons. Groups were classified as significantly increased, significantly decreased, or not significant at an adjusted p < 0.05. Final figure assembly and complementary statistical analyses were performed using GraphPad Prism (v10.6.1). Quantification of growth dynamics using peak slope–based modeling To quantify dynamic growth effects beyond endpoint fitness metrics, a slope-based growth model was applied to time-resolved co-culture GFP trajectories. This approach captures the strongest transient decline in GFP-positive fraction and translates it into an inferred change in relative growth rate. Time-course GFP measurements (% viable GFP-positive cells) were analyzed per gene by first computing the mean GFP trajectory across all biological replicates. Data were processed in Python v3.13.274 using pandas v2.2.275 and NumPy v1.26.76 Time-series restructuring and visualization were performed using Matplotlib v3.9.2,78 and linear regression was implemented using SciPy v1.13.1.77 For each gene, the mean GFP trajectory (GFP% vs. aligned culture day) was scanned using sliding contiguous windows containing at least three and at most eight time points. For each window, a linear model of formula 4 in Table 3 was fitted, where m represents the slope (% GFP per day). The peak negative slope was defined as the most negative slope observed across all eligible windows. In cases of near-identical slopes (within a tolerance of 10−6), preference was given to windows centered closest to the median time point of the experiment to avoid edge-driven artifacts. The selected window is reported by its start day, end day, and number of contributing points, ensuring interpretability and traceability of the inferred slope. To account for global co-culture drift and batch-specific effects, the same peak-slope window identified for each target was applied to the INT-control. For each batch in which the target gene was measured, an INT-control mean GFP trajectory was computed and a control slope was estimated over the identical time window. The final control-corrected peak slope metric was defined as in formula 5. Negative values of Δs indicate a decline in the GFP-positive fraction relative to INT-control dynamics. To translate slope effects into interpretable growth parameters, Δs values (% GFP per day) were mapped onto an inferred relative growth-rate model. All targets were initialized at a standardized starting GFP fraction (f) of 0.5. The endpoint GFP fraction after one day was estimated as in formula 6, where values were constrained to (0 and 1) to ensure numerical stability and biological plausibility. Growth rate changes were then computed using a logit-based transformation according to formula 7. This change was added to a fixed reference growth rate for INT-controls μINT = 0.9241 days−1 (Doubling time of 18 h) to obtain an inferred target growth rate according to formula 8. Inferred doubling times are calculated according to formula 9. For selected gene sets (e.g., top 15 hits based on ΔΔGFP or Δs), annotated time-course plots were generated showing mean GFP trajectories, fitted peak-slope segments, and corresponding target and control slopes (Figure S4). This slope-based quantification complements endpoint fitness metrics by capturing transient growth advantages and providing a mechanistic link between competitive dynamics and inferred cellular growth rates. Transcriptomic profiling of unresolved gene sets For transcriptomic profiling of unresolved multiplex gene sets, co-culture samples were harvested directly from validation experiments. For each condition, 1E6 cells were collected to standardize RNA input and support comparable library complexity across samples. Total RNA was extracted from pooled bulk co-culture samples, and poly(A)-enriched cDNA libraries were prepared using the Eurofins Genomics InVIEW Transcriptome service according to the provider’s standard protocols. Libraries were sequenced externally by Eurofins Genomics Europe Sequencing GmbH (Konstanz, Germany) as 150-bp paired-end reads on an Illumina NovaSeq 6000 system, with >30 million reads per sample. RNA-seq processing was performed using a Snakemake v7.8.380-based workflow composed of standard open-source tools. Raw reads were quality-checked with FastQC v0.12.181 and adapter- and quality-trimmed with Trimmomatic v0.39.82 Trimmed paired-end reads were pseudoaligned and quantified with kallisto v0.48.083 against a proprietary Sartorius Stedim Cellca GmbH Cricetulus griseus transcriptome. The reference transcriptome used internal Sartorius gene identifiers linked to NCBI RefSeq/LOC gene identifiers and public ortholog annotations. Downstream analyses were performed in R v4.4.284 with Bioconductor v3.20.85 Transcript-level kallisto estimates were summarized to gene-level counts using tximport v1.34.0.86 Genes with total counts <10 across all samples were removed before downstream analysis. For the expression matrix used in the manuscript, gene-length-corrected trimmed mean of M-values values were calculated with edgeR v4.4.287 by applying TMM normalization factors to gene-level counts and converting to RPKM using the per-gene median effective transcript length estimated by kallisto across samples. The GeTMM-normalized expression matrix was used for descriptive expression comparisons and visualization without additional log transformation or pseudocount addition. Differential expression analysis was conducted at the level of gene sets to support interpretation of multiplex knockout phenotypes. Formal differential-expression testing was performed with DESeq2 v1.46.088 using gene-level counts, results are provided in Table S7. In parallel, for the gene-wise expression heatmaps, sample groupings and gene-set annotations were used and processed in Python v3.13.274 using pandas v2.2.275 and NumPy v1.26.476 for data handling and numerical operations. For each gene within a gene set, log2 fold changes were calculated by comparing the mean GeTMM expression of the target group to the intron-targeting control group. Statistical significance was assessed using Welch’s two-sample t test (unequal variance) as implemented in SciPy v1.13.1,77 and results are indicated by significance tier in the corresponding differential-expression heatmaps (∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001). No multiple-testing correction was applied, as analyses were performed in a targeted, hypothesis-driven manner to support interpretation of multiplex phenotypes rather than genome-wide discovery. GSEA differential expression results were visualized as DESeq2 derived gene-wise log2 fold-change heatmaps generated using Matplotlib v3.9.2.78 Gene Set Enrichment Analysis (GSEA) was performed with gprofiler2 v0.2.389 using GO, WikiPathways,90 and transcription-factor annotation sources. GO.db v3.20.091 and org.Hs.e.g.db92 retrieved through AnnotationHub v3.14.093 were used for annotation. Processed raw-count and GeTMM matrices, sample metadata, and differential-expression outputs are provided with the RNA-seq deposition and supplemental data.