We selected four genes for gradual modulation of their dosage in the human erythroid progenitor cell line K562 (Ulirsch et al., 2016): GFI1B, NFE2, MYB, and TET2. Two of the genes, GFI1B and NFE2, have been implicated in blood diseases and traits (Möröy et al., 2015; Polfus et al., 2016; Jutzi et al., 2019), and in our earlier work, we identified a broad transcriptional response to inhibition of GWAS-overlapping enhancers to these genes (Morris et al., 2023). MYB is a key transcription factor (Baker et al., 2014) and a downstream target of GFI1B (Morris et al., 2023). TET2 has a role in DNA demethylation and is unrelated to these transcriptional networks and is included in this study as a control with minimal expected trans effects. We refer to these four genes, targeted in cis for modulation of their regulation, as cis genes (Figure 1A).

To modulate the gene expression of the cis genes, we use K562 cells expressing CRISPRi (KRAB-dCas9-MECP2) and CRISPRa (dCas9-VPR) systems (see Methods). Both cell lines were hashed with DNA-conjugated antibodies against different surface proteins that allow pooled experiments. To obtain a wide range of dosage effects, we used four different single guide RNA (sgRNA) design strategies (Figure 1B): (1) targeting the transcription start site (TSS) as in the standard CRISPRi/CRISPRa approach, (2) tiling sgRNAs +/-1000 bp from the TSS in approximately 100 bp intervals, (3) targeting known cis-regulatory elements (CREs), and (4) using attenuated guides that target the TSS but contain mismatches to modulate their activity (Jost et al., 2020). We further included five non-targeting control (NTC) sgRNAs as negative controls.

The library of altogether 96 guides was transduced to a pool of K562-CRISPRi and K562-CRISPRa cells at low multiplicity of infection (MOI). After eight days, we performed ECCITE-seq (see Methods) to capture three modalities: cDNA, sgRNAs, and surface protein hashes (oligo-tagged antibodies with unique barcodes against ubiquitously expressed surface proteins). Instead of sequencing the full transcriptome, we used target hybridization to capture a smaller fraction of the cDNA and obtain more accurate expression readouts at a feasible cost. The subset of selected transcripts were picked from the transcriptional downstream regulatory networks of GFI1B and NFE2 identified previously (Morris et al., 2023), maintaining similar patterns of co-expression correlation across co-expression clusters (see Methods, Figure 1—figure supplement 1A). We targeted a total of 94 transcripts (Figure 1A), including the four cis genes, 86 genes that represent trans targets of GFI1B and/or NFE2 (Morris et al., 2023; Figure 1—figure supplement 1A), LXH3 that is not expressed in blood progenitors, GAPDH that is highly expressed and often considered an invariable housekeeping gene and the dCas9-VPR or KRAB-dCas9-MeCP2 transcripts.

We used the protein hashes and the dCas9 cDNA (indicating the presence or absence of the KRAB domain) to demultiplex and determine the cell line—CRISPRi or CRISPRa. Cells containing a single sgRNA were identified using a Gaussian mixture model (see Methods). Standard quality control procedures were applied to the scRNA-seq data (see Methods). To confirm that the targeted transcript capture approach worked as intended, we assessed concordance across capture lanes (Figure 1—figure supplement 1C). The final data set had 20,001 cells (10,647 CRISPRi and 9354 CRISPRa), with an average of 81 and 86 cells with a unique sgRNA for the CRISPRa and CRISPRi, respectively (Figure 1—figure supplement 1D).

Next, we calculated the expression fold change for each of the four cis genes targeted by each sgRNA in the two cell lines (CRISPRi/a), comparing each group of cells with its respective NTC sgRNA group (see Methods). We first confirmed that the sgRNAs targeting the transcription start site (TSS) up- and down-regulated their targets (Figure 1C, Figure 1—figure supplement 1F). When looking at all sgRNAs at once, across the four genes, we observed a 2.3-fold range (Figure 1E), with a minimum 72% reduction and maximum 174% increased expression (log2(FC) values from –1.83–0.80). However, the range varied between the genes, with GFI1B covering the widest range of gene expression changes (gene expression ranging between 0.28–1.42 fold), while MYB expression could not be pushed higher than 1.13 fold (Figure 1E). The direction of the effects were consistent with the cell lines of origin, where 98.88% of the significant perturbations (Wilcoxon rank test at 10% FDR, n=89) were correctly predicted based on the direction of the target gene fold change. The predicted on- and off-target properties of the guides (Doench et al., 2014; Doench et al., 2016; McKenna and Shendure, 2018) did not correlate with the fold changes in the cis genes (Figure 1—figure supplement 2A), suggesting that the observed effects represent true cis-regulatory changes. The fold changes were also robust to the number of cells containing a particular sgRNA (Figure 1—figure supplement 2B, top).

We verified that the fold change estimation was not biased depending on the expression level of the target gene at the single-cell level, which can vary due to drop-out effects or binary on/off effects of the KRAB-based CRISPRi system (Noviello et al., 2023). By splitting cells with the same sgRNA based on the normalized expression of the cis gene (0 vs. >0 normalized UMIs, Figure 1—figure supplement 3A), we observed highly concordant transcriptome gene expression effects between the two groups (Figure 1—figure supplement 3B). This indicates that the dosage changes per guide were not primarily driven by the changing frequency of binary on/off effects, and the use of pseudo-bulk fold changes provides a robust estimation of cis gene fold changes. These patterns are further supported by the cells forming a gradient rather than distinct clusters on a UMAP (Figure 1—figure supplement 3C).

The fold change patterns differed between sgRNA designs (Figure 1D, left). As expected, sgRNAs targeting the TSS showed strong perturbations in gene expression. However, sgRNAs tiled +/-1 kb from the TSS provided a broader and more gradual range of up- and downregulation across the target genes, sometimes surpassing the effects of TSS-targeting sgRNAs. Attenuated sgRNAs with mismatch mutations resulted in a range of gene silencing effects in the CRISPRi line, as expected based on their original design (Jost et al., 2020). However, these attenuated sgRNAs did not exhibit such a dynamic range in the CRISPRa modality, although a significant correlation existed between the silencing or activating effect size and the distance of the mismatch from the protospacer adjacent motif (PAM) when considering all data points together (Figure 1—figure supplement 2C). The sgRNAs targeting distal cis-regulatory elements (CREs) showed both inhibiting and activating effects, even though both the CRISPRi and CRISPRa constructs were initially designed to inhibit or activate transcription from the promoter and initial gene body region. Nonetheless, the number of known CREs per gene is typically limited. Given its simplicity and the ability to achieve both up- and downregulation of the target gene, we consider the tiling sgRNA approach, with a simple design that only requires annotation of the TSS, as a useful method for gradually modulating gene dosage with CRISPRi/a systems.

Having designed guides targeting both distal and local neighbouring regulatory regions of the four transcription factors (TFs) and ensuring minimal bias in fold changes due to sgRNA’s biochemical properties, we investigated the cis features that determine the strength of dosage perturbation. We observed substantial differences in the effects of the same guide on the CRISPRi and CRISPRa backgrounds, with no significant correlation between cis gene fold changes (Figure 2A). However, in both modalities, the strongest effects on gene expression were observed when the guides were close to the transcription start site (TSS) (Figure 2B, excluding NTC and attenuated sgRNAs), although the peaks of strongest activation or repression differed between the modalities. In the CRISPRi modality, the maximum effect was located within the gene body at +238 bp from the TSS (Figure 2B, bottom), consistent with previous studies that used essentiality as a proxy for expression (Sanson et al., 2018). However, in the CRISPRa modality, the maximum average fold changes occurred closer to the TSS at around –99 bp (Figure 2B, bottom), as also shown for CD45 (Legut et al., 2020).

Enhancer, tiling, and TSS sgRNAs targeted regions of the genome with different compositions of histone marks, annotated by ENCODE, in K562 cells (Luo et al., 2020; Figure 2C), which allowed us to investigate the impact of chromatin state on the strength of cis gene dosage modulation. The magnitude of cis gene fold changes varied significantly depending on the presence of specific marks or peaks, which again differed between the two modalities (Figure 2D). In the CRISPRa cell line, the strongest effects were observed when guides were located in regions with open chromatin marks, such as DNase or ATAC peaks. In contrast, the strongest repression by CRISPRi occurred in genomic regions with the presence of H3K27ac, H3K4me3, and H3K9ac marks. These differences may be explained by the distinct mechanisms of action of the activator and repressor domains. MeCP2 and KRAB repressor domains recruit additional repressors that silence gene expression through chromatin remodelling activities, such as histone deacetylation (Lupo et al., 2013). On the other hand, the VPR activation fusion domain may only require Cas9 to scan the open chromatin and recruit RNA polymerase and additional transcription factors to activate transcription. Overall, while a few sgRNAs have a strong effect in both CRISPRi and CRISPRa cell lines, a single guide library containing guides optimized for both modalities enables a range of gradual dosage regulation. However, larger data sets are needed for more careful modelling of the ideal dosage modulation designs and to understand how both cis-regulatory features, feedback loops, and other mechanisms contribute to the outcomes.

We then turned our attention to the remaining 91 genes captured by our custom panel and determined the relative expression fold change of each trans gene, compared to NTC in each unique guide perturbation (see Methods). Principal component analysis (PCA) performed on all pseudo-bulk fold changes demonstrated the removal of batch effects from the cell lines and revealed a clear direction of the cis gene dosage effect in the first three principal components (Figure 3—figure supplement 1B). This finding suggests that dosage modulation is the primary determinant of trans effects. The PCA indicated that the dosage modulation of GFI1B and MYB is reflected in opposite directions in PC1 and PC2, while the trans responses of NFE2 are captured by PC3.

Using a false discovery rate (FDR) cutoff of 0.05, all 91 trans genes except for the neural-specific TF LHX3 (negative control) exhibited a significant change in expression upon perturbation of any of the TFs. The observed trans-effects were well correlated with perturbations of these genes in other data sets (Figure 3—figure supplement 1C and D). Among all measured fold changes, the most extreme negative effect sizes were observed in cis genes, with the top 10 being predominantly reductions in GFI1B expression. This indicates that cis downregulation tended to surpass the endogenous expression limits. In contrast, the largest increases in gene expression were observed through trans mechanisms, where KLK1 and TUBB1 reached the largest expression values when GFI1B was strongly upregulated, or SPI1 and DAPK1 when GFI1B was strongly downregulated. These findings suggest that the CRISPRa approach did not reach a biological ceiling of overexpression.

Inspecting trans responses as a function of cis gene modulation, we observed that the number of expressed genes and the mean absolute expression changes of trans genes exhibited gene-specific correlations with cis-gene dosage (Figure 3A, Figure 3—figure supplement 1E). Perturbations in GFI1B led to the most pronounced trans responses, with positive dosage changes resulting in larger effect sizes compared to decreasing TF gene expression, where the effect plateaued. NFE2 exhibited similar patterns but with a smaller magnitude. In the case of MYB, trans responses were observed when decreasing the expression of this TF, but the effects of upregulation are largely unknown as we were unable to increase MYB expression beyond 0.35. As expected, given the unrelatedness of TET2 to the trans network, dosage modulation of this gene had minimal trans effects with the least pronounced trend when compared to TET2 dosage, so we excluded it from subsequent analyses.

Upon clustering the changes in expression of trans genes based on the cis gene dosage change linked to each sgRNA, we identified distinct clusters exhibiting different dosage-response patterns (Figure 3B for GFI1B, Figure 3—figure supplements 2–5A for all cis genes). Further examination of the gene expression fold changes for each individual trans gene in relation to the TF fold changes revealed a diverse range of response patterns (Figure 3C, Figure 3—figure supplements 2–5B for all cis genes). These responses exhibited both linear and nonlinear forms, including some instances of non-monotonic gene expression responses for certain trans genes within the GFI1B trans network (e.g. GATA2 in Figure 3C, Figure 3—figure supplement 6E).

To accurately characterise the dosage response, we employed both linear and nonlinear modelling approaches (Figure 3D), which allowed us to quantitatively assess the extent of nonlinear responses by comparing the goodness of fit of these models using the Akaike Information Criterion (AIC). For the nonlinear model, we utilized a sigmoid function with four free parameters (Figure 3D, right). These parameters represented the slope at the inflection point (b, indicating the rate of increase or decrease in expression), the minimum and maximum asymptotes (c and d, representing the lower and upper limits of fold change), and the value of cis gene expression at which the inflection point occurs (a). To prevent overfitting, we implemented a 10-fold cross-validation scheme, which yielded reliable predictions on the left-out data (Pearson r=0.71–0.88 for all trans genes in the GFI1B, MYB, and NFE2 networks, Figure 3—figure supplement 6C). Additionally, the predicted parameter a was centred around zero, as expected, since the input data represents relative fold changes (Figure 3—figure supplement 7). Since a sigmoid function cannot capture non-monotonic responses, we employed a loess regression as an alternative approach for the few genes that exhibited non-monotonic responses (see Methods, Figure 3—figure supplement 6D and E). For the vast majority of genes, the sigmoid (or loess) fit was remarkably good, partially due to the low level of noise in the targeted scRNA-seq data.

We compared the performance of the linear vs. nonlinear models with the ∆AIC (AIClinear−AICr\begin{document}$AIC_{linear} -\, AIC_{r}$\end{document}), where a positive ∆AIC means that the sigmoid model captures the variance better in the dosage response than in the linear model. This showed that most GFI1B-dependent dosage expression responses are better fit by the sigmoid model (median ∆AIC = 18.7, with 70.4% of all trans genes with a significant response having ∆AIC >2, Figure 3E). The responses to dosage modulation of MYB and NFE2 were also better captured by the nonlinearities, but to a lesser extent (0.14 and 3.4 median ∆AIC, with 20.8% and 40.7% of all trans genes dosage responses having ∆AIC >2 for MYB and NFE2, respectively, Figure 3—figure supplement 6A). The broader range of GFI1B expression modulation, providing more data to detect nonlinear trends, likely contributes to this difference. When ignoring those genes classified as unresponsive (genes that their expression did not change upon the TF modulator, see Methods), even more responses of the remaining trans genes were better explained by a sigmoidal model with 83.6%, 26.3%, and 63.2% of these having a ∆AIC >2, for GFI1B, MYB, and NFE2 respectively. A similar trend holds even when limiting the models to be fitted to those data points that correspond to a hypothetical one copy loss or gain of the cis gene (Figure 3—figure supplement 6B), where the median ∆AIC of responsive genes are 7.05, 0.05, and 3.6 for GFI1B, MYB, and NFE2 trans responses. Overall, this shows that trans responses to TF dosage are dominated by nonlinear behaviours even when the TF dosage changes are not extreme but within biologically plausible ranges.

Utilising a model that effectively captures the variance in our data provided the ability to predict unmeasured TF dosage points and facilitated a direct comparison of trans effects across different cis genes. Employing the sigmoid model (and loess for those with non-monotonic responses), we estimated the continuous expression of trans genes on a uniform fold-change scale across the spectrum of GFI1B, MYB, and NFE2 expression changes (Figure 4A). This estimation was carried out within the empirically observed range of all three cis genes, spanning from log2(FC) –1.83–0.51. Subsequent hierarchical clustering of trans gene responses revealed six major clusters of distinct response patterns. For the majority of trans genes, the response to GFI1B and MYB was opposite, with only two small clusters displaying exceptions. Notably, GFI1B generally induced the most substantial response, while NFE2 triggered the smallest range of trans gene response.

Next, we collected diverse annotations for the trans genes to explore the connections between their regulatory properties, disease associations, and selective constraints concerning their response to TF dosage (Figure 4B and C). To quantify these relationships, we assessed significant differences in belonging to these qualitative annotations using the Wilcoxon rank test (Figure 4D) and correlated parameters from the sigmoid model with quantitative gene metrics (Figure 4E). We hypothesized that genes with annotated selective constraint, numerous regulatory elements, and central positions in regulatory networks would exhibit greater robustness to TF changes. Indeed, housekeeping genes demonstrated a considerably smaller dosage response range (Figure 4F). Moreover, genes classified as unresponsive were enriched in the housekeeping category (odds ratio = 2.14, Fisher test p-value = 0.024). The link between selective constraint and response properties is most apparent in the MYB trans network. Specifically, the probability of haploinsufficiency (pHaplo) shows a significant negative correlation with the dynamic range of transcriptional responses (Figure 4G): genes under stronger constraint (higher pHaplo) display smaller dynamic ranges, indicating that dosage-sensitive genes are more tightly buffered against changes in MYB levels. This pattern was not reproduced in the other trans networks (Figure 4E).

The relationship between the response of trans genes and intrinsic gene properties differed between GFI1B, MYB, and NFE2 trans network responses. We also performed a similar analysis comparing the sigmoid parameters to network properties using the approach outlined by Minaeva et al., 2025 and obtained inconsistent results between TF regulons (Figure 4—figure supplement 1A and B). This suggests that the link between commonly annotated gene properties and the gene responses are complex and highly context-specific, as in our data from a single cell line, they differed between the upstream regulators that were manipulated. Thus, much more data is needed before transcriptional responses can be predicted from gene properties, and conversely, to understand the cellular mechanisms that lead to the annotated gene properties.

Moving beyond the characterization of mechanisms of transcriptome regulation, a key question is how gradual dosage variation links to downstream cellular phenotypes, and whether these responses exhibit analogous nonlinear patterns. To address this question, we correlated our findings with the expression profiles of various cell types in order to study the myeloid differentiation process, a phenotype well-characterized for our K562 model that has been used as a reliable system for investigating erythroid differentiation within myeloid lineages (Wang et al., 2013) and blood tumours (Salvadores et al., 2020). Specifically, leveraging single-cell expression data for bone marrow cell types from the Human Cell Atlas and Human Biomolecular Atlas Project (Granja et al., 2019), we filtered the expression data to the targeted genes in our study. After aggregating data across donors and normalising expression across cell types (Figure 5—figure supplement 1A), we compared the expression patterns resulting from each unique transcription factor dosage modulation in relation to each unique cell type expression state. The ensuing correlation can then be construed as a ‘phenotype,’ signifying the similarity between the transcriptional state induced by the TF increase or decrease and the transcriptional state of a specific blood lineage cell type.

Such analyses recapitulate known biology, with GFI1B upregulation (Möröy et al., 2015) and MYB downregulation (Jayapal et al., 2010) being crucial factors promoting erythrocyte maturation (Figure 5A). The downregulation of NFE2, instead, was negatively related to platelet differentiation. Analysing the correlations as inferred phenotypes suggests potential nonlinear relationships (Figure 5—figure supplement 1B), but these trends should be considered hypotheses that require experimental validation. In summary, this points to cellular phenotypes resulting from gradual TF dosage modulation.

Many of the analyzed trans genes are associated with physiological traits and diseases (Figure 4). Understanding the nonlinear trends in the expression of these genes is of particular interest, as it helps comprehend how genes with physiological impacts may be buffered against upstream regulatory changes, and how their dosage changes as a response to upstream regulators contrasts with genetic variants that contribute to diseases and traits. Additionally, knowing the underlying dosage-to-phenotype curve of a gene can be crucial if this is considered a biomarker for identifying or treating disease. To investigate this, we analyzed whether OMIM genes for rare diseases and Mendelian traits or GWAS genes for different blood cell traits (Figure 5B) that are part of the trans networks of genes affected by GFI1B, MYB, or NFE2 perturbation are enriched for nonlinear dosage responses. As seen in Figure 4, the trans response properties of each gene are highly specific to the regulators and thus analyzed in parallel for each cis gene network. An enrichment for nonlinear responses was observed for MYB trans network genes associated with disease and for blood traits related to white blood cells and reticulocytes. These enrichments are particularly interesting given that most trans genes that were sensitive to MYB dosage modulation did not respond with a nonlinear trend (Figure 3—figure supplement 6A).

Despite nonlinear responses not being significantly enriched among disease genes across all trans networks, the responses of the same trans gene can show very different dosage responses depending on the upstream regulator being tuned. In Figure 5C, we highlight several disease-associated genes (linked to one or more disease phenotypes Amberger et al., 2015). FOXP1, a haploinsufficient and potentially triplosensitive transcription factor implicated in intellectual disability, exhibited a strong and dose-dependent response, particularly to varying levels of GFI1B. A similar pattern was observed for NFIA, another haploinsufficient gene involved in developmental disorders. However, it is difficult to interpret their expression response in K562 cells when their most apparent phenotypic effects likely derive from other cell types. RHB is the Rhesus blood type gene, where a common deletion of the gene causes the Rh blood type in homozygous individuals, with a strong nonlinear response to GFI1B levels. A particularly interesting gene is TUBB1, part of β-tubulin, that causes autosomal dominant macrothrombocytopenia or abnormally large platelets. Here, K562 cells are a reasonable model system, being closely related precursors to megakaryocytes that produce platelets. Interestingly, GFI1B loss also causes a macrothrombocytopenia phenotype in mice (Beauchemin et al., 2017), and in our data, TUBB1 expression decreases quickly as a function of decreased GFI1B expression but then plateaus at a level that corresponds to loss of one copy of TUBB1. This raises the hypothesis that low GFI1B levels may cause macrothrombocytopenia at least partially via reducing TUBB1 expression.

To construct the vector harbouring the CRISPRa system (pJDE003), the CRISPRi (KRAB-dCas9-MeCP2) gene fusion of pGC02 was replaced with dCas9-VPR cassette, which was PCR amplified (Q5 High-Fidelity 2X Master Mix, NEB M0492L) from the plasmid pCC_05 (Morris et al., 2023) with primers oJDE005 and oJDE006 following instructions from manufacturer. pGC02 was digested with XbaI-FD and BamHI-FD (Thermo Fisher FD0685 and FD0054) and sequentially dephosphorylated with FastAP (Thermo Fisher EF0651) following the manufacturer’s recommendations. The digested pGC02 vector and the PCR insert with the CRISPRa system (previously treated with a 15 min DpnI enzyme incubation, Thermo Fisher FD1704) were assembled by Gibson assembly using a 2:1 insert:vector ratio with Gibson Assembly Master Mix (NEB E2611S). Assemblies were transformed into NEB 5-alpha E. coli competent cells and single colonies were picked and sequence validated by Sanger sequencing. Frozen stock of the correct construct cells were regrown for plasmid Maxiprep extraction (QIAGEN 12362) for subsequent virus production.

Lentivirus was produced by polyethylenimine linear MW 25000 (Polysciences 23966) transfection of HEK293FT cells with the transfer plasmid containing a Cas9-VPR effector, packaging plasmid psPAX2 (Addgene 12260) and envelope plasmid pMD2.G (Addgene 12259). After 72 hr post-transfection, cell media containing lentiviral particles was harvested and filtered through 0.45 μm filter Steriflip-HV (Millipore SE1M003M00). One volume of Lentivirus Precipitation Solution (Alstem VC100) was added to the collected lentivirus, mixed and stored overnight at 4°C. The mix was centrifuged for 30 min at 1,500g, and the pellet of lentiviral particles were resuspended in 1/10th of the original volume of DMEM media. Lentivirus vials were frozen at -80°C and later thawed for transduction.

To construct the monoclonal K562 cell line with the CRISPRa system, the dCas9-VPR lentivirus was transduced into one million K562 cells using 100 µl of 10X concentrated lentivirus in a total volume of 1 ml (high MOI). After 24 hr, the media was replaced with fresh IMDM, and 48 hr after transduction, blasticidin (A.G. Scientific B-1247) was added to a final concentration of 10 µg/µl for 16 days. Monoclonal cell lines were sorted by FACS (Sony Cell Sorter SH800) into a 96-well plate. The presence of dCAS9 protein in several growing clones was confirmed by western blot (Primary antibody: Purified anti-CRISPR CAS9 antibody; BioLegend 844302, Secondary antibody: LI-COR 925-32212) and protein levels were normalized to GAPDH (Primary antibody: GAPDH (14C10) Rabbit; Cell Signalling Technology 2118S. Secondary antibody: LI-COR 925-68073).

To select the final monoclonal CRISPRa cell line, the three clones with the highest protein expression in the western blot were subjected to functional validation to test for activation activity. Lentiviral guides designed from Legut et al., 2020 targeting CD4 (Anti-CD4 Mouse Monoclonal Antibody (FITC), BioLegend, 300505), which is lowly expressed in K562, CD19 (Anti-CD19 Mouse Monoclonal Antibody (APC), BioLegend, 302211) with null expression, and CD45 (PE anti-human CD45, BioLegend, 368509) with intermediate expression, were independently transduced into all three monoclonals, and after puromycin selection, the expression of this markers was screened by FACS at day 4 and at day 10 or 11 after transduction. The clone with the strongest and most consistent activity was selected. K562 cells were obtained as fresh stocks from ATCC for this work and not authenticated further. Cells were regularly passaged and tested for the presence of mycoplasma contamination with a Mycoplasma PCR Lookout assay.

The four selected dosage genes were GFI1B, NFE2, MYB, and TET2. GFI1B, and NFE2 were chosen due to their reported trans effects following the inhibition of their cis-CREs (Morris et al., 2023). MYB was selected for being downstream of the GFI1B network, and TET2 was selected as an unrelated gene to those transcriptional networks. Both MYB and TET2 qualified as oncogene and tumour suppressor functions in K562 (https://depmap.org/portal/), making them ideal choices to determine the impact of growth effects on the experiment.

The additional 88 genes captured by targeted sequencing were selected based on the significant trans effects of GFI1B and NFE2 inhibition from Morris et al., 2023 including two control genes, GAPDH and LHX3 genes, one highly constantly expressed ‘housekeeping gene’ and the other with no reported expression in K562 cell lines, respectively. The remaining 86 genes were selected based on Morris et al. to include (1) 29 genes that overlapped between the NFE2 and GFI1B network, (2) 47 trans genes for GFI1B, 10 trans genes for NFE2. The number of trans genes selected from each unique network was proportionally chosen, given the size of each trans network, and oversampling TFs and TF targets as defined in Garcia-Alonso et al., 2019, as well as maintaining the proportional co-expression cluster structure as defined in Morris et al., 2023; Figure 1—figure supplement 1A and B. Additional filters in the selection included a minimum expression of 0.1 mean UMI/cell and the lack of alternative 5' splice isoforms and a unique Ensembl ID.

The 10X Probe Full Custom Panel design tool was used to design the targeted gene expression probe library. A total of 687 probes (~15%) were discarded because they covered transcript regions with a median coverage per base of <3 reads/bp (for medium and highly expressed genes) or <1 reads/bp (for lowly expressed genes). All probes for LHX3 (0 median reads/bp) were retained. In addition, 93 probes covering the entire transcript sequence of the dCas9-VPR and KRAB-dCas9-MeCP2 transcript were included, resulting in a final total of 4405 probes. The xGen Custom Hybridization Capture Panel of biotinylated oligos was ordered and synthesized at IDT.

The sgRNA library contained a total of 96 guides (51 tiling, 8 TSS, 20 attenuated, 12 enhancer, and 5 non-targeting controls). All guides were designed to not contain the U6 terminator sequence, repeats of five or more consecutive G, C, or As, as well as not falling in the genomic region where K562 cell line has alternative alleles compared to the human genome reference (Hg38). All guides were scored with FlashFry McKenna and Shendure, 2018 to obtain off-target and on-target activity scores that allowed the selection of the best scoring guides. Tiling guides were designed to target different regions of the promoter, TSS and beginning of the gene body of each dosage gene, spanning a total average distance of 1400 bp (TSS in the centre), each being on average distant from one another of 110 bp. The sequences of the two TSS guides were obtained from Replogle et al., 2022. The sgRNAs targeting enhancers were picked from previously reported work that showed a CRISPR-based evidence of enhancer activity (GFI1B Morris et al., 2023), (NFE2 Xie et al., 2019), (MYB Li et al., 2021). The five attenuated guides for each gene were manually designed following the rules described in Jost et al., 2020 to span a range of activities, including a single point mutation on the best scoring guide that targeted the TSS.

Overhangs with homology regions to the pGC03 plasmid (18 bp downstream and 22 bp upstream) were added to the sgRNA sequence to be able to directly clone the ssDNA oligos into the plasmid. The 96 sgRNAs were ordered in IDT as single-stranded DNA oligos (total 60 bp) in a 96-well plate to 100 pmol scale. The oligos were pooled at equimolar concentration and diluted to a final concentration of 0.2 µM. The library was cloned into the BsmBI-digested plasmid pGC03 using 10 reactions of the NEBuilder HiFi DNA Assembly kit following the manufacturer’s instructions. All the reactions were pooled and the DNA precipitated using isopropanol, GlycoBlue (Thermo Scientific AM9515) and 50 mM NaCl for 15 min at RT. Following two washes with ethanol 70%, the assembly was resuspended with 15 µl of 0.2X TE.

To transform the library into E. coli, 1 µl of the assembly was mixed with 25 µL of Endura cells under manufacturer’s electroporation conditions, then plated onto 245 × 245 mm square LB 100 µg/ml Carbenicillin plates. The plates were grown ON, and >2.5e5 transformants were obtained, ensuring the complexity of the library was maintained at >1000 cells per unique sgRNA. All colonies were collected and subjected to maxiprep using the Maxi Fast-Ion Plasmid Kit, Endotoxin Free kit (IBI Scientific IB47123). The representation of the library was assessed through MiSeq shallow sequencing (Illumina).

The lentiviral library was produced by transfecting ~80 million HEK293FT cells with a transfer plasmid containing the 96 sgRNA library, along with the packaging plasmid psPAX2 and envelope plasmid pMD2.G, using polyethylenimine linear MW 25000. The supernatant media was replaced with fresh D10 10% BSA 6 hr after transfection, and the virus was collected and filtered through 0.45 µm filters after 48 hr. The lentiviral library was then concentrated 2X using the Precipitation Lentiviral Solution, aliquoted, and stored at -80°C for subsequent transduction.

Both CRISPRi and CRISPRa K562 cell lines were independently transduced with different titers of the lentiviral sgRNA library at a low MOI (one sgRNA per cell). Twenty hours post-transduction, the cell media was replaced with fresh IMDM 10% Serum Plus II Blasticidin 5 µg/mL, and 4 hr later, Puromycin (Invivogen ant-pr-1) was added at a final concentration of 2 µg/mL to select for cells with sgRNA integration. The transduction batch with an infection rate of ~10% was selected, and cells were sorted to near purity using FACS to remove dead cells. Cells were maintained at >90% survival and a maximum confluency of 700,000 cells/mL. On day 8 post-transduction, the cells were collected and prepared for cell hashing.

Cell hashing was performed as previously described using four hashtag-derived oligonucleotides (HTOs) in a hyperconjugation protocol (Stoeckius et al., 2017). Each transduced cell line was split into four batches of 500,000 cells, resulting in a total of 8 different hashes. After incubation and washes, all 8 hashed batches were pooled together and run in two reaction lanes of the 10X Chromium Next GEM Single Cell 5’ Reagent Kit v2 (single indexing, PN-1000265, and PN-1000190). The manufacturer's protocol was followed with modifications stipulated in the ECCITE-seq protocol (Mimitou et al., 2019). For each GEM reaction, 42,000 cells from the hash pool were used to obtain approximately 21,000 total cells, including ‘multiplets’ (multiple cells per droplet counts). Gene expression (cDNA), hashtags (HTOs), and guide RNA (Guide-derived oligos, GDOs) libraries were constructed following the 10x Genomics and ECCITE-seq protocols (https://cite-seq.com/eccite-seq/) with minimal modifications. Specifically, the antibody pool protein tag library steps were ignored, and a custom-designed probe library was used to enrich the cDNA for the genes of interest in the 10X Targeted Gene Expression protocol (10X PN-1000248). The resulting libraries were sequenced using an Illumina Nextseq 500/550 Mid-Output v2.5 Kit (150 cycles). The targeted enrichment of the dCas9 transcripts was performed separately using an independent probe library and was sequenced together with additional HTO libraries using the Illumina Miseq Reagent Kit v3 (150 cycles).