Heterogenous nuclear ribonucleoprotein D-like controls endothelial cell functions

2.1 HNRNPDL regulates endothelial cell functions

In order to investigate whether HNRNPDL affects physiological functions of the endothelium, we analyzed proliferation, migration and sprouting after knockdown of HNRNPDL in HUVECs (Figure 1 and Supplementary Figure S1A). HUVECs were transfected with an siRNA targeting HNRNPDL (si_DL_A) or a non-silencing control siRNA (si_ctrl). We achieved a reduction in HNRNPDL protein levels to ∼10 % (Supplementary Figures S1A and S1B). Crystal violet assays were performed to assess changes in proliferation. Knockdown of HNRNPDL decreased HUVEC numbers to 60 % after 48 h and 51 % after 72 h in comparison to the control (Figure 1A). To analyze HUVEC migration, we performed transwell-migration assays. HNRNPDL knockdown dramatically reduced the migration of HUVECs to 16 % (Figure 1B and C). Further, to test how HNRNPDL affects cellular processes important for angiogenesis, we analyzed spheroid sprouting capacity after knockdown of HNRNPDL in 3D cultures. HNRNPDL knockdown reduced spheroid sprouting of HUVECs to 23 % (Figure 1D and E), underlining the importance of HNRNPDL for efficient angiogenesis.

Figure 1:

HNRNPDL affects endothelial cell proliferation, migration and sprouting. (A) Cell viability of HUVECs 48 h and 72 h after knockdown of HNRNPDL. n = 3. (B) Quantification of migrating HUVECs after knockdown of HNRNPDL. n = 4. (C) Representative pictures used to quantify the numbers of migrated HUVECs in (B). (D) Quantification of the cumulative sprout length per spheroid in HUVECs after knockdown of HNRNPDL, with or without VEGF-A stimulation (0.05 mg/ml). n = 4. (E) Representative pictures of spheroids and sprouts used to quantify the cumulative sprout length per spheroid in (D). All bar graphs are reported as mean values ± standard deviation. Statistical analysis was done using student’s t-test, two-tailed, paired. (**) p value < 0.01, (*) p value < 0.05, n.s. = not significant.

VEGF-A is a pro-angiogenic factor that stimulates endothelial cell sprouting in blood vessels and promotes spheroid sprouting of HUVECs in cell culture (Yancopoulos et al. 2000). Notably, while VEGF-A stimulation increased sprouting of control cells, it was not able to restore sprouting after HNRNPDL knockdown (Figure 1D and E). Together, the results clearly demonstrate the importance of HNRNPDL for general physiological endothelial cell functions i.e., proliferation, migration and sprouting; processes affecting the formation of new blood vessels.

The observed reduction in HUVEC proliferation after knockdown of HNRNPDL is consistent with its oncogenic function, where HNRNPDL has been shown to promote proliferation in diverse cancer cell types (Ji et al. 2020; Wu et al. 2008; Zhang et al. 2018; Zhou et al. 2014). Thus, the positive effect of HNRNPDL on proliferation seems to be cell type-independent.

In general, RBPs have been suggested to modulate angiogenesis by controlling the expression of cytokines, growth factors and cell cycle regulators. Recently, interleukin enhancer binding factor 3 (ILF-3) and quaking (QKI) were shown to promote angiogenesis by mRNA stabilization. In human coronary artery endothelial cells, ILF-3 induction promoted proliferation, migration and tube formation by stabilizing pro-angiogenic transcripts, such as the mRNA encoding the cytokine IL-8 (CXCL8) (Vrakas et al. 2019). QKI knockdown reduced cell cycle progression and spheroid sprouting in HUVECs, by reducing cell cycle-associated genes. In particular, QKI stabilizes the mRNA encoding CCND1, which promotes G1/S phase transition (Azam et al. 2019). Whether HNRNPDL induces the severe changes in the phenotype of HUVECs by targeting such mRNAs is not known.

2.2 HNRNPDL target mRNAs are involved in cell cycle and motility regulation

In order to identify how HNRNPDL affects endothelial cell function, we performed transcriptomic analysis after HNRNPDL knockdown in HUVECs. To identify the earliest time-point after siRNA transfection when a robust reduction of HNRNPDL was reached, a time-course experiment was performed. This was achieved after 36 h (Figure 2A). Thus, this time point was chosen for RNA sequencing to increase the chance to observe changes mainly in primary targets and minimize secondary effects. Total RNA was extracted from HUVECs transfected with one siRNA targeting HNRNPDL (si_DL_A) or a non-silencing control siRNA (si_ctrl), polyA-selected and subjected to deep sequencing (n = 2). Approximately 35 million paired-end reads were obtained for each sample. We detected expression of 13,381 genes with expression being defined as mean transcripts per million [TPM] >1 in either control or HNRNPDL knockdown cells. From these, 3878 genes (29 %) changed significantly in their abundance (adjusted p value < 0.05, Figure 2B and Supplementary Table S1), including 866 >1.5-fold up-regulated and 988 >1.5-fold down-regulated genes.

Figure 2:

HNRNPDL regulates target mRNAs involved in cell cycle regulation and cytoskeleton remodeling. (A) Western blot of HNRNPDL and β-actin after siRNA-mediated knockdown for 24, 36 or 48 h. HNRNPDL knockdown for 36 h was chosen as condition for RNA sequencing. n = 2. (B) MA plot showing differentially expressed genes in HNRNPDL knockdown experiments (n = 2). (C) KEGG analysis of the DEGs. (D, E) RT-qPCR quantification of CCND1 and TPM4 after HNRNPDL knockdown in HUVECs with si_DL_A (D, n = 4) and si_DL_B (E, n = 3). All bar graphs are reported as mean values ± standard deviation. Statistical analysis was done using student’s t-test, two-tailed, paired. (**) p value < 0.01, (*) p value < 0.05.

Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis of the differentially expressed genes was used to identify significantly changed pathways in response to HNRNPDL knockdown (Figure 2C and Supplementary Table S2). With an adjusted p-value of 3.37 × 10⁻⁸, “cell cycle” was most significantly affected. Interestingly, the mRNA of the major cell cycle regulator cyclin D1 (CCND1) controlling G1/S phase transition, was the most highly reduced mRNA amongst cell cycle genes (Figure 2C, red dot). CCND1 controls cell cycle progression by interaction with cycline-dependent kinases 4 or 6. In cancer, CCND1 is often upregulated and leads to excessive cell proliferation (reviewed in Tchakarska and Sola 2020). Thus, reduction of CCND1 mRNA is in concordance with the observed reduction in proliferation after HNRNPDL knockdown. We used RT-qPCR to verify and quantify the reduced mRNA levels of CCND1 in response to HNRNPDL knockdown. Two different siRNAs (si_DL_A and si_DL_B) were transfected to ensure that the observed changes are no off-target effects of the siRNA (si_DL_A) used for RNA-seq. The knockdown of HNRNPDL with si_DL_A as well as si_DL_B was confirmed via Western Blot (Supplementary Figure S1A). HNRNPDL protein levels were reduced to ∼10 % for si_DL_A and to ∼13 % for si_DL_B (Supplementary Figure S1B). RT-qPCR verified the expected reduction in CCND1 mRNA abundance after knockdown of HNRNPDL with both siRNAs (Figure 2D and E).

Further, we confirmed induction of the most abundantly expressed tropomyosin, TPM4, at mRNA as well as protein level by RT-qPCR and Western Blot, respectively (Figure 2D and E, Supplementary Figure S2). Tropomyosins affect cell migration by modulating the dynamics of actin cytoskeleton, which forms along the leading edge of a migrating cell (Khaitlina 2015; Schaks et al. 2019). Interestingly, TPM4 is discussed as biomarker in multiple cancer types, with contrasting results regarding its influence on cell migration and disease prognosis, which might depend on the identity of TPM4 protein isoforms (Jeong et al. 2017; Luo et al. 2022; Sheng and Chen 2020; Yang et al. 2018; Zhao et al. 2019). TPM4 induction in HUVECs is concordant with reduced cell migration after HNRNPDL knockdown, as reported for colon, cervical and breast cancer (Jeong et al. 2017; Luo et al. 2022; Yang et al. 2018).

2.3 Dual regulation by HNRNPD and HNRNPDL paralogs

Paralogous RBPs often affect similar mRNA targets (Rahman et al. 2013; Vuong et al. 2016). Further, paralogs tend to control each other’s expression by cross-regulation. Likewise, HNRNPD and DL control their own as well as each other via alternative splicing of poison exons in their 3′-UTRs (Figure 3A, indicated in red). Cross-regulation of HNRNPD expression by HNRNPDL is exerted by controlled inclusion of exon 9 into the HNRNPD mRNA. Inclusion of exon 9 subsequently triggers degradation of the mRNA by NMD (Kemmerer et al. 2018). Knockdown of HNRNPDL is expected to reduce exon 9 inclusion into the HNRNPD mRNA and thus, to increase HNRNPD protein levels, due to the reduced production of NMD-targeted mRNA isoforms.

Figure 3:

Dual regulation of CCND1 and TPM4 by HNRNPD and DL. (A) Exon cluster of HNRNPD and HNRNPDL. (B) RT-qPCR quantification of HNRNPDL, HNRNPD, CCND1, and TPM4 after double knockdown of HNRNPDL and HNRNPD in HUVECs all bar graphs are reported as mean values ± standard deviation. Statistical analysis was done using student’s t-test, two-tailed, paired. (**) p value < 0.01, n.s. = not significant. n = 4–5. (C and D) 12-mer motifs derived by GraphProt (Maticzka et al. 2014) analysis from RBNS data for HNRNPD (C) and HNRNPDL (D). (E and F) Correlation of HNRNPDL, HNRNPD, ZFP36, and ELAVL1.

To analyze if the observed changes in target mRNA levels could be due to altered HNRNPD protein levels, we analyzed exon 9 inclusion levels and the levels of all four HNRNPD protein isoforms (Figure 3A and Supplementary Figure S3). Indeed, knockdown of HNRNPDL led to reduced inclusion of HNRNPD exon 9 in HUVECs (Supplementary Figure S3A). However, Western blot analysis showed that HNRNPD protein levels were not changed after HNRNPDL knockdown (Supplementary Figure S3B). This finding is most likely explained by the extremely low basal expression of exon 9-including mRNA isoforms in HUVECs. Thus, a further reduction in exon 9-containing isoforms does not significantly increase HNRNPD protein levels. In sum, although the cross-regulation of HNRNPD expression by HNRNPDL is functional in HUVECs, HNRNPD is not the driver of the observed changes in CCND1 and TPM4 mRNA abundances.

To examine if HNRNPD protein reduction can affect HNRNPDL target mRNA abundances in HUVECs, we investigated the impact of simultaneous HNRNPD and DL knockdown on the levels of CCND1 and TPM4 mRNAs. HUVECs were transfected with identical amounts of siRNAs: control (si_ctrl), control and HNRNPDL (si_ctrl+si_DL_A), and HNRNPDL and HNRNPD (si_DL_A+si_D). HNRNPD levels increased with HNNRPDL knockdown in line with the reduced exon 9 inclusion and strongly decreased in the double knockdown (Figure 3B). This change in mRNA abundance of HNRNPD is also reflected in HNRNPD protein levels (Supplementary Figure S4).

HNRNPD has been shown to be a negative regulator of CCND1 in several cell types including cervical carcinoma (HeLa), human osteosarcoma (U2OS), colon carcinoma (HCT116) and adenocarcinoma (LoVo) as well as primary normal human skin fibroblasts HFSN1 (Al-Khalaf et al. 2011; Lal et al. 2004; Tian et al. 2020). This suggests that simultaneous HNRNPD knockdown could potentially counteract the reduction in CCND1 mRNA observed after HNRNPDL knockdown. However, CCND1 mRNA was not affected by the additional HNRNPD reduction in HUVECs (Figure 3B). Thus, CCND1 might either be not targeted by HNRNPD in HUVECs or is more dependent on HNRNPDL status. Notably, regulation of CCND1 levels by HNRNPD were shown to be complex, with varying results found in different cellular systems and at different HNRNPD levels. One reason for this might be that the CCND1 mRNA is targeted by a multitude of stabilizing and destabilizing RBPs as well as lncRNAs and miRNAs, the sum of which will ultimately determine CCND1 mRNA levels (Al-Khalaf et al. 2011; Azam et al. 2019; Cao et al. 2017; Chen et al. 2019, 2008; Ghosh and Adhya 2018; Lal et al. 2004; Liu et al. 2021; Yoneda et al. 2020).

In contrast, not much is known about the posttranscriptional regulation of TPM4. PTB (polypyrimidine tract binding protein) destabilizes the TPM4 mRNA when bound by the lncRNA SFTA1P in cervical cancer cell lines (Luo et al. 2022). ARE-mediated decay of TPM4 has not been investigated so far. Interestingly, TPM4 mRNA induction was significantly increased by double knockdown of HNRNPD and DL compared to the knockdown of HNRNPDL alone (Figure 3B). Thus, the two paralogs can act additively in mRNA reduction. This contrasting behavior of HNRNPD and DL towards different targets is reminiscent of other paralogs such as HNRNPL and LL as well as PTB and nPTB, which can act redundant, additive or even competitive on shared targets (Rahman et al. 2013; Vuong et al. 2016). Here, the outcome in target regulation can be the result of subtle differences in RNA-binding preferences between the paralogs.

To evaluate this further, we analyzed publicly available RNA Bind-n-Seq (RBNS) data (Van Nostrand et al. 2020) to compare the RNA-binding preferences of HNRNPD and HNRNPDL. Using GraphProt (Maticzka et al. 2014), models were trained using default parameters on samples of 100,000 training sequences and tested with a 10-fold cross-validation. The obtained binding motif for HNRNPD showed the expected very strong preference for adenine and uridine residues as typically found in AREs (Figure 3C). The binding preference is highly similar to the one from HNRNPDL (Figure 3D). This is expected, as the two paralogs share highly similar RNA-binding domains (>74 % sequence homology), which is indicative for similar motif preferences (Ray et al. 2013). Direct comparison of the top ten RBNS hexamers identified for HNRNPD and HRNPDL indicates an elevated preference of HNRNPDL for longer A-stretches in this ARE motifs (Supplementary Table S3).

Next, we visualized the top ten RBNS hexamers in the 3′-UTRs of the CCND1 and TPM4 mRNAs as potential binding sites for HNRNPD and DL (Supplementary Figures S5 and S6). A 9 nt long HNRNPDL binding site resulting from overlapping hexamer motifs is present in a 470 nucleotide long AU-rich region of the 3′-UTR of CCND1, that was previously described to increase CCND1 mRNA stability (Deshpande et al. 2009). Thus, HNRNPDL likely is able to interact with this regulatory region, subsequently modulating mRNA abundance. Interestingly, several high scoring HNRNPD hexamers are also present, in this AU-rich region, in line with regulation of CCND1 mRNA levels by HNRNPD observed in other cells types.

The 1656 nt long TPM4 3′-UTR (hg38_knownGene_ENST00000643579.2 Supplementary Figure S6), contains fifteen hexamer motifs of HNRNPDL and two overlapping ones of HNRNPD. These overlapping hexamers can also be recognized by HNRNPDL (Supplementary Figure S6). Thus, the TPM4 3′-UTR contains recognition motifs of both HNRNPD and DL allowing for the additive regulation observed in the knockdown experiments. In sum, both 3′-UTRs contain several of the AU-rich hexamers identified as recognition motifs of HNRNPD and DL by RBNS, allowing for dual regulation.

To get a more detailed insight into the correlation of HNRNPD and DL in relation to other ARE-binding proteins, we first examined publicly available single cell transcriptome data from the Genotype-Tissue Expression (GTEx) Project (GTEx Consortium 2020) to identify a tissue with highly expressed HNRNPDL in healthy cells. Single cell tissue gene expression for HNRNPDL (ENSG00000152795.17) in vascular endothelial cells (Supplementary Figure S7A) is highest in breast mammary tissue. Thus, we used breast tissue data to analyze the relation between the expression of HNRNPDL, HNRNPD and two other ARE-binding RBPs, ZFP36 (zinc finger protein 36, also TTP) and ELAVL1 (ELAV like RNA-binding protein 1, also HuR) (Figure 3E, Supplementary S7B and S7C). As expected, the expression of HNRNPD and DL is highly positively correlated (Supplementary Figures S7C and S7E), which likely reflects the cross-regulation by AS-NMD. Further, HNRNPDL levels are also positively correlated with ZFP36 or ELAVL1, albeit two-fold lower than with HNRNPD. Further, we compared expression levels in breast cancer tissue to see if RBP networks are disturbed under pathological conditions (https://www.cancer.gov/tcga) (Figure 3F and Supplementary Figures S7D and S7E). In general, the positive correlation between the different RBPs decreases in tumor tissue, suggestive of pathological alterations in RBP levels (Figure 3E and F).

Together, the data suggest that AREs and ARE-binding RBPs, such as HNRNPDL, affect angiogenic processes by modulating mRNA abundances. The comparison of healthy and tumor tissue demonstrates the changing RBP landscape, which in turn will affect mRNA levels in a cell. However, how differential target regulation is achieved and differs depending on the cell type is still unclear and warrants further investigation.

4.1 Cell culture and transfection

HUVECs (Lonza, CC-2519) were cultured in T75 flasks in EBM medium (Lonza) supplemented with 10 % FCS (Thermo Fisher Scientific) and EGM single quots [hEGF, Hydrocortison, GA-1000, bovine brain extract (Lonza)]. HUVECs were split once a week and only passages 2 and 3 were used for experiments. For siRNA transfection, 3.5 × 10⁵ cells were seeded in 60 mm cell culture dishes. 24 h later, transfection was performed using Lipofectamine RNAiMAX according to the manufacturer’s protocol with varying amounts of siRNA (60 pmol siRNA for single knockdown, 120 pmol for double knockdown) in OptiMEM medium. After 4 h the transfection mixes were removed, and fresh medium was added.

4.2 RNA isolation

Total RNA from HUVECs was isolated using the miRNeasy Mini kit (Qiagen), including the optional on-column DNA digest with the RNase-Free DNase Set (Qiagen). After isolation, 500 ng RNA were quality checked on a 1 % agarose gel.

4.3 RNA sequencing

For the HNRNPDL knockdown experiments (n = 2), total RNA from HUVECs transfected either with a non-silencing control siRNA (si_ctrl) or one siRNA targeting HNRNPDL (si_DL_A) was used to generate polyA-enriched strand-specific cDNA libraries. Libraries were sequenced on a HiSeq 2000 (Illumina), obtaining 76 nt paired-end reads. Approximately 35 million reads were obtained for each library.

4.4 RNA-seq data analyses

Reads were mapped to the human genome (GRCh38/hg38 assembly) using the splice-aware aligner STAR version 2.5.4b (Dobin et al. 2013). The htseq-count script from the HTSeq python package version 0.6.1p1 (Anders et al. 2015) was used to count reads within genes annotated in GENCODE version 32. Differential expression analysis between normal and HNRNPDL knockdown conditions was performed using the R/Bioconductor package DESeq2 (Love et al. 2014) using the method apeglm for effect size shrinkage (Zhu et al. 2019). Here, differentially expressed genes were defined based on an adjusted p value < 0.05 and a shrunken absolute fold change >1.5. This identified 1854 significantly differentially regulated genes, including 866 upregulated and 988 downregulated genes. For all analyses, we considered only expressed genes with a mean expression level of >1 transcript per million (TPM) in at least one of the two conditions and at least 10.23 mean normalized read counts (baseMean).

4.5 KEGG analysis

The KEGG enrichment analysis was performed with clusterProfiler (v. 3.16.0) (Yu et al. 2012) for the genes that were significantly upregulated (n = 1924) or downregulated (n = 1954). 64 KEGG pathways were significantly enriched (adjusted p value < 0.05, Supplementary Table S2). Selected enriched pathways are shown in Figure 2C.

4.6 RBNS data analysis

We used GraphProt (Maticzka et al. 2014), which was developed with CLIP-seq and RNAcompete data in mind, to find motif enrichments from RBNS data for HNRNPD and HNRNPDL, respectively (ENCODE project; experiment IDs: ENCSR175OMA, ENCSR055HDN (Van Nostrand et al. 2020)). We used the negative controls (ENCSR141NLX, ENCSR394HHV, ENCSR599EMY (Van Nostrand et al. 2020)) to train a classification model and used randomly 100,000 training sequences for the training phase. We validated the models with a 10-fold cross-validation.

4.7 RT-qPCR analyses

RT-qPCR quantification was performed as in Weigand et al. (2012). Oligonucleotide sequences are listed in Supplementary Table S4. All PCR products were verified by sequencing.

4.8 Western blotting

HUVECs were lysed in lysis buffer [137 mM NaCl, 10 % glycerol, 20 mM Tris-HCl pH 8.0, 2 mM EDTA pH 8.0, 1 % Igepal, 5 μl protease inhibitor cocktail (Sigma-Aldrich, St. Louis, USA)] for 20 min on ice. The samples were centrifuged (15 min at 17,000 g, 4 °C) and cell debris was removed. The protein content of the samples was determined in three technical replicates according to the Bradford method. 10 μg protein were loaded onto 10 % SDS polyacrylamide gels. Total lane protein (loading control) was visualized using 2,2,2-trichloroethanol (Ladner et al. 2004). Subsequently, proteins were blotted onto PVDF membranes (Bio-Rad, Hercules, USA) and membranes were blocked using 2 % skim milk. Primary antibodies are listed in Supplementary Table S5 Horseradish peroxidase-conjugated anti-mouse or anti-rabbit IgG (Jackson ImmunoResearch, West Grove, USA) were used as secondary antibodies. Blots were developed with the ECL system (Bio-Rad) or Amersham ECL select (GE Healthcare, Chicago, USA) for weaker signals. Images were detected using the ChemiDoc Imaging System (Bio-Rad). The Western Blots were quantified using Image Lab v 6.1.0 build 7 (Bio-Rad).

4.9 Cell viability

HUVECs were transfected as described above. 48 or 72 h after transfection, HUVECs were fixed with 0.5 % formaldehyde in PBS and stained with 0.5 % crystal violet in PBS. After three washing steps with PBS the cells were incubated with 33 % acetic acid. Samples were transferred to a 96-well plate. Absorption at 620 nm was measured in a TECAN infinite M 200 Pro plate reader (Tecan, Männedorf, Switzerland). Absorption was normalized to untreated HUVEC control cells.

4.10 HUVEC migration assay

24 h after transfection with si_ctrl or si_DL_A, HUVECs were serum-starved [EBM medium without FCS (Lonza Group, Basel, Switzerland)]. Again 24 h later, HUVECs were washed with 1× PBS and detached using Trypsin-EDTA [0.05 % trypsin, 0.02 % EDTA in 1× PBS (Thermo Fisher Scientific, Waltham, USA)]. HUVECs were resuspended in starvation medium (EBM medium without FCS). 5 × 10⁴ cells in 100 μl starvation medium were seeded per insert in a 24-well transwell chamber [Thincert cell culture inserts, pore diameter 8 μm, translucent PET membrane (Greiner Bio-One, Kremsmünster, Austria)]. Transwell inserts were placed into a lower well containing 700 μl EBM medium with supplements or starvation medium as a negative control. After 5 h (37 °C, 5 % CO₂), cells were removed from the inner side of the insert with a cotton swab. Migrated cells were washed with 1× PBS and fixed with methanol. Migrated cells were stained using crystal violet (0.5 % in 1× PBS) and the membranes were mounted using Pertex (Histolab, Gothenburg, Norway). No cells were detected in the negative control. Migrated cells in six fields per well from random sites of the transwell insert membrane were counted using Fiji software (version 1.51f).

4.11 HUVEC spheroid sprouting assay

24 h after transfection, HUVECs were detached using Trypsin-EDTA (0.05 % trypsin, 0.02 % EDTA in 1× PBS). 500 HUVECs were seeded per 50 μl drop as hanging drop culture in EBM medium supplemented with 20 % EBM-methylcellulose medium [1.2 % (w/v) methylcellulose (Carl Roth, Karlsruhe, Germany) in EBM medium without supplements]. Approximately 100 drops were created for each condition (si_ctrl or si_DL_A). 24 h later, spheroids were transferred to a collagen matrix. Collagen (Sigma-Aldrich) was mixed with 10 % 10× M199 medium (Sigma-Aldrich). The collagen mix was prepared for all samples, so that the pH is the same for every condition. Spheroids were harvested in 1× PBS and centrifuged for 3 min with 500 g. The supernatant was removed. 2.6 ml methylcellulose-FCS (20 % FCS, 80 % EBM-methylcellulose medium) was added to the spheroids. The samples (si_ctrl or si_DL_A) were split between two reaction tubes. Spheroids in one of the two reaction tubes were stimulated with 0.05 mg/ml VEGF-A protein (VEGF₁₆₅, Peprotech). The collagen medium mix was then neutralized using 0.2 M NaOH. 1.3 ml collagen mix was added to each spheroid sample. Samples were transferred immediately into the inner wells of a 24-well plate (4 wells per sample) without producing air bubbles and incubated for 24 h at 37 °C, 5 % CO₂. Outer wells of the 24-well plate were filled with 1× PBS beforehand to prevent draining of the spheroids. 24 h later, spheroids and their sprouts were photographed using a Zeiss (Oberkochen, Germany) Axiovert 200 microscope (10× objective). The cumulative sprout length in µm of 10 spheroids per condition and biological replicate was measured using Fiji software (version 1.51f).

4.12 Statistical analyses

All bar graphs are reported as mean values ± standard deviation. Statistical analysis was done using student’s t-test, two-tailed, paired.

4.13 Bioinformatic analysis of HNRNPDL expression and correlation

We used public RNA expression data from breast tissue (GTEx, (GTEx Consortium 2020)) and breast cancer tissue (data generated by the TCGA Research Network: https://www.cancer.gov/tcga) to compare physiological and pathological expression of HNRNPDL to HNRNPD, ZPF36, and ELAVL1.

Data was obtained using the recount3 v 1.8.0 (Wilks et al. 2021) package from Bioconductor using the create_rse function and then the compute_read_counts function. The breast tissue data was selected with the keywords file_source: “gtex”, project_type: “data_sources” and project: “BREAST”. The breast tumor data was selected with project: “BRCA”. From the breast cancer samples only samples from female donors were used, only the sampleType: “Primary Tumor” was used and only samples with one sample per donor were used. If more than one tumor sample exists from the same donor, all are removed.

Both the breast tissue data and the breast cancer data were processed following the following steps to obtain comparable RPKM measurements: (i) Genes with low expression were filtered with the filterByExpr function from edgeR Bioconductor package v 3.40.2 (McCarthy et al. 2012) with default filter settings. (ii) For breast tissue data, batch effect correction was performed on the batches provided by GTEx (smnabtcht). For batch correction the ComBat_seq algorithm from the sva Bioconductor package v 3.46.0 (Leek et al. 2023) was used. (iii) TMM values were calculated from the read count data by dividing the read counts by a TMM normalization factor, which was calculated with calcNormFactors from edgeR. (iv) Then RPKM values were calculated with edgeR rpkm function setting log = TRUE.

Expression of HNRNPDL to HNRNPD, ZPF36, and ELAVL1 is shown in box-and-whisker plots. Correlations are calculated as Pearson correlations. Dotplots are made with the corrplot package (v 0.92). Pearson’s R is given as color scale and the areas of circles show the absolute value of corresponding correlation coefficients. For correlations with a p-value > 0.001 no dot is shown.

4.14 Accession numbers

The RNA-seq data of the HNRNPDL knockdown experiments have been deposited with the Gene Expression Omnibus database (GEO; www.ncbi.nlm.nih.gov/geo/) and are available under the accession number GSE169378.