Integrative multiomic analysis reveals co-ordinated alternative splicing in human bone marrow stromal stem cells Authors: Angelita Liang, Gene Hart-Smith, Moustapha Kassem, Abbas Jafari, Marc R. Wilkins Published: 2026 ABSTRACT Alternative splicing is essential for the production of diverse messenger RNAs from a single gene. However our understanding of alternative splicing and its regulation during processes such as osteoblast differentiation remains incomplete. Using differentiating hMSC-TERT4 cells as a model, we performed deep short- and long-read RNA sequencing, revealing a highly integrated multiphasic differentiation program. Analysis for splicing revealed extensive changes during lineage commitment, in genes associated with known transcriptional regulators including RUNX2, TEAD1, CTNNB1 and NFATC4. During late-stage matrix osteoblast differentiation (maturation), splicing changes occurred in genes encoding cytoskeletal and extracellular matrix proteins, and Golgi and vesicle trafficking proteins. Analysis of proteomic and phosphoproteomic profiles confirmed the presence of alternative splicing in proteins participating in these biological processes as well as in RNA splicing and autophagy, with splicing involving a variety of protein domains, regions and PTM sites. Together, the transcriptomic and protein splicing landscapes provide a comprehensive description of how mesenchymal stromal cells progressively acquire an osteoblastic phenotype.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-54912-3. Introduction RNA splicing involves the processing of nascent precursor mRNA molecules by the spliceosome. This generation of multiple RNAs from a single locus is a major form of transcriptional diversity. It has functional consequences in the transcriptome and proteome, modulating transcript localisation, transcript stability, protein function and protein-protein interactions1. Underscoring its importance, it is estimated that more than 90% of multi-exon human transcripts are alternatively spliced2. In addition, changes in alternative splicing patterns can drive lineage commitment and cell fate selection during differentiation3. Bone mass homeostasis is achieved by the coupling of bone formation by osteoblasts to bone resorption by osteoclasts during bone remodelling. Osteoblasts differentiate from bone marrow stromal stem cells (also known as mesenchymal stem cells, MSCs) – the integrity of this process is essential to maintain bodily reserves of MSCs and osteoblasts4. Osteoblast differentiation (and hence bone mass) can be improved by vitamin D3 intake, drugs such as selective estrogen receptor modulators (SERMs)5, or physical exercise that causes mechanical loading of bone6. However, dysregulated osteoblast differentiation can lead to decreases in bone mass and to osteoporosis; a major health care challenge especially in the aging population. Characterising the splicing landscape during osteoblast differentiation is necessary to establish a system-wide understanding of bone formation. Whilst the regulation of gene expression in osteogenic cultures has been well studied7–9, efforts are ongoing to characterise alternative splicing-mediated drivers of differentiation10. A study by Roosa et al. used an Affymetrix exon array to identify genes exhibiting altered splicing in response to mechanical loading of bone in rats11. Notably, these included extracellular matrix genes (collagen, fibronectin), cytoskeleton components (actinin, myosin, tropomyosin), and regulators of key signalling pathways involved in osteoblast differentiation such as TGF-β signalling (BMP6, BMP15) and those for Wnt signalling (SFRP4). Splice variants of known regulators of osteoblast differentiation such as Hepatocyte Growth Factor (HGF)12, POSTN (periostin)13 and RUNX2 have also been identified, the latter serving as a master transcription factor in osteoblasts. RNA-binding proteins and splicing factors that directly influence the specificity of the spliceosome are notable as they can each orchestrate splicing changes in a wide range of effector genes. Recently, a meta-analysis of publicly available gene expression and human RNA-binding protein data was conducted by Luo & Jiang14, highlighting 15 RBPs of interest. Additionally, splicing factors PTBP115, SNRPB16 and YBX117 have also been shown to orchestrate diverse splicing changes during osteoblast differentiation. Although alternative splicing can regulate osteoblast differentiation through post-transcriptional mechanisms, the translation of spliced mRNA can give rise to spliced peptides and proteins. Proteomic analysis of differentiating osteoblasts has been well described18–21, however the alternative splicing changes at the protein level remain understudied. To date, only a small set of genes have been confirmed to be alternatively spliced at the protein level, including ENC122, FOSB23, RUNX224, SP725 and TPM226. A more comprehensive understanding of how alternative splicing governs osteoblast differentiation from MSCs may offer new insights into mechanisms regulating bone formation, the pathogenesis of bone loss diseases, and new opportunities for therapeutic interventions. Here, we present an integrative analysis of time-series bulk short- and long-read RNA-Seq, proteomic and phosphoproteomic data for hMSC-TERT4 cells induced to differentiate to osteoblasts. Our work provides a comprehensive view of the stepwise osteogenic program, reinforcing the importance of alternative splicing in the bone formation program, and represents a substantial augmentation of known splice variants, especially at the protein level. Results Gene expression changes during osteoblast differentiation To model osteoblast differentiation of MSCs, we used hMSC-TERT4 cells – an immortalised human bone-marrow stromal cell line with well-established capacity for osteoblast differentiation27. We induced hMSC-TERT4 cells to differentiate as described previously8, with successful induction measured by increased alkaline phosphatase (ALP) activity and formation of mineralised matrix. To detect differential gene expression and alternative splicing events during osteoblast differentiation, we conducted deep short-read RNA-Seq, generating temporally resolved transcriptome profiles at time-points 0, 6, 12 h, as well as 1, 3-, 6-, 9- and 12-days post-induction of OB differentiation (Fig. 1A). This was done in biological triplicate, effectively yielding ~ 114 million short reads per timepoint. Fig. 1Osteoblast differentiation is marked by stage-specific gene expression signatures. A Overview of time-series transcriptomic and proteomic profiling. ONT: Oxford Nanopore. B Heatmap of enriched GO terms from differentially expressed genes. C Selected clusters for gene expression and their most significant GO terms (FDR < 0.01). Four stages of differentiation are highlighted. For full clustering results, see Table S1. D Timeline of important events during osteoblast differentiation. Osteoblast differentiation is marked by stage-specific gene expression signatures. A Overview of time-series transcriptomic and proteomic profiling. ONT: Oxford Nanopore. B Heatmap of enriched GO terms from differentially expressed genes. C Selected clusters for gene expression and their most significant GO terms (FDR < 0.01). Four stages of differentiation are highlighted. For full clustering results, see Table S1. D Timeline of important events during osteoblast differentiation. To verify the cell state transitions of differentiating MSCs, we produced one-against-the-rest comparisons for differentially expressed genes at each time-point (see Methods). Gene Ontology (GO) analysis (Fig. 1B) showed that genes expressed were enriched for mitotic processes until 1d before being downregulated by 3d, marking the onset of mitotic arrest. Reinforcing this, clustering of temporal gene expression profiles (Fig. 1C, clusters 4, 10, full details of clustering in Table S1) showed that genes responsible for cell cycle CCNA2/B1/B2/E1/E2/F, CDC25A/C, CDK1/2, CENPA/E/F/H/I/J/K/L/M/N/O/Q/U/W and E2F1/2/8 were collectively downregulated from 1-3d. Many other trends were observed concurrent with mitotic arrest at 3d. The GO terms describing “regulation of BMP signalling pathway” and “Wnt signalling pathway” showed peak significance of enrichment at 3d, suggesting these pathways are involved in cellular changes occurring concurrently (Fig. 1B). Genes enriched for “GTPase activator activity” and “actin binding” were also transiently upregulated at 3d (Fig. 1C, cluster 11). Commencing at 3d, genes enriched for “cell morphogenesis involved in differentiation” (including positive regulators of osteoblast differentiation FZD7, KLF2, MSX2 and RHOA/B and MSX2) were downregulated until 12d (Fig. 1C, cluster 3), suggesting they may only be required for initial stages of differentiation28. On the contrary, genes enriched for “PERK-mediated unfolded protein response” and “translation” reached a steady maximum at 3d (Fig. 1C, cluster 25) despite being initially downregulated upon induction of differentiation. The unfolded protein response is activated in osteoblasts to manage the high metabolic demand of extracellular matrix secretion29 and indeed, previous studies have shown that matrix secretion begins around 3d post-induction30,31. In contrast, “osteoblast differentiation”, “ossification”, “actin cytoskeleton organisation” and “extracellular matrix organisation” were significantly enriched up to 9d before being downregulated (Fig. 1B), indicating onset of late-stage matrix mineralisation30. Interestingly, expression consistently increased over the time course for genes enriched for “type I interferon signalling pathway” (including IFIT1/2/3, IFITM1, IRF7/9, OAS1/2/2L and STAT1) which exert both positive and negative regulatory effects on OB differentiation and matrix mineralisation32–36(Fig. 1C, cluster 24). Thus, the enrichment analysis suggests that MSCs underwent four distinct phases of OB differentiation over 12 days: mitotic arrest around 1-3d post-induction, continued lineage commitment and matrix secretion at 3d, and late-stage mineralisation between 9d and 12d (Fig. 1D). Classification and quantification of alternative splicing To detect alternative splicing events and their Percent Spliced-In (PSI) values from short-read RNA-Seq, we used two complementary bioinformatics tools – Junction Usage Model (JUM)37 and PSI-Sigma38. Both tools classify splicing events in terms of eight archetypal splicing modes39(Fig. 2A), with an additional category in JUM for “composite” events which are inseparable combinations of one or more of these modes. A total of 1305 (PSI-Sigma) or 983 (JUM) Variably Spliced Regions (VSRs) with at least one significant change between two time points were detected. There was a high level of agreement in genes called as differentially spliced by both JUM and PSI-Sigma, with most discrepancies between the splice events due to splice events detected in different genes (Figure S1A). We observed that JUM’s methodology was less sensitive for intron retention (IR) events, and it was prone to misclassifying alternative first/last exons as A5SS/A3SS events (Figures S1B-C). Hence, we only considered PSI-Sigma for exon extending events (ignoring 32 IR events from JUM). According to PSI-Sigma, intron retention events were the most prevalent differentially spliced events, followed by cassette exons (Fig. 2B); this was also true for constitutive splicing events (no significant change across the time-course). Fig. 2The consequences of alternative splicing during osteogenesis depend on the mode of splicing. A The modes of alternative splicing, classified as exon extending or exon skipping. B Number of events for each splice mode and quantity of events with predicted effects. See Methods for details of “NMD” and “PTC event” predictions. 32 intron retention events in Differential (JUM) have been omitted. C Number of Variably Spliced Regions (VSRs) in known Ensembl transcript regions. P values were calculated from 10,000 random shuffles along the transcriptome (see Methods). * P < 0.05. ** P < 0.01. D UpSet plot for the type of splice events in annotated Ensembl transcript regions (PSI-Sigma only). Frequencies are reported as unique intersections, hence each splice event is only counted in one intersection. A3SS: alternative 3’ splice site; A5SS: alternative 5’ splice site; IR: intron retention; MXS: mutually exclusive skipping. AFE: alternative first exon. ALE: alternative last exon. The consequences of alternative splicing during osteogenesis depend on the mode of splicing. A The modes of alternative splicing, classified as exon extending or exon skipping. B Number of events for each splice mode and quantity of events with predicted effects. See Methods for details of “NMD” and “PTC event” predictions. 32 intron retention events in Differential (JUM) have been omitted. C Number of Variably Spliced Regions (VSRs) in known Ensembl transcript regions. P values were calculated from 10,000 random shuffles along the transcriptome (see Methods). * P < 0.05. ** P < 0.01. D UpSet plot for the type of splice events in annotated Ensembl transcript regions (PSI-Sigma only). Frequencies are reported as unique intersections, hence each splice event is only counted in one intersection. A3SS: alternative 3’ splice site; A5SS: alternative 5’ splice site; IR: intron retention; MXS: mutually exclusive skipping. AFE: alternative first exon. ALE: alternative last exon. Oxford nanopore long-read RNA sequencing validates splicing events To validate splice events detected via short reads, we conducted Oxford Nanopore (ONT) sequencing of polyA RNA at timepoints 0 and 6 h, as well as 1-, 3-, 9- and 12-days post induction (see Fig. 1A). A total of 117,316,560 long reads were generated with a median length of 1,099 bases. Small (1-3nt) inaccuracies in raw ONT reads were corrected using StringTie240 with the short-read RNA-Seq data as a reference. The ONT data recapitulated most differential splicing events detected by short-read sequencing (80.7% for PSI-Sigma, 98.7% for JUM); this included those we noted above to be in Variably Spliced Regions (Figure S1A). The ONT-validated splice events were thus used for further analyses. When considering assembled ONT transcripts, 146,084 unique multiexon transcripts were present in the data, of which 4,872 were identical to an Ensembl transcript (all exon start and end boundaries matching), and 35,208 were almost identical to an Ensembl transcript, differing only in the start and end co-ordinates (i.e. transcriptional start site or polyadenylation site). The remaining 72.6% did not directly match to Ensembl transcripts. At exon level, of the 1,305,038 unique ONT exons, 242,895 ONT exons had an equivalent in Ensembl, 165,054 had one common boundary with an Ensembl exon and 21,152 had two common boundaries but each boundary with mutually exclusive sets of Ensembl exons. Transcript fate depends on the mode of alternative splicing We sought to predict the locations in annotated transcripts where splicing occurred and their functional consequence at the transcriptional level. By referring to Ensembl annotations, we found that alternative splicing most frequently occurred in protein-coding sequences of transcripts, followed by 5’ and 3’ UTR regions (Fig. 2C). A permutation test (see Methods) showed the number of VSRs affecting 5’ UTRs, start codons and stop codons were significantly (P < 0.01) higher than what would be expected by chance, suggesting precise targeting of these functional regions by alternative splicing machinery. There was an association between the types of splice events and their location in functional regions of transcripts (Fig. 2D). IR events in protein-coding transcripts occurred more frequently in the 3’ UTR than the 5’ UTR and protein-coding regions. A5SS events were mostly in the 5’ UTR, and A3SS events tended to be in the 3’ UTR and coding regions. It is known that human UTRs are extensively modulated by intron retention41. Indeed, our data extends this observation, showing that exon extending events (Fig. 2A) tend to occur in UTR regions whereas exon skipping events (Fig. 2A) tend to occur in protein-coding sequences. The modulation of UTRs by splicing events may be of consequence for regulatory elements in mRNAs, given that UTRs can contain miRNA motifs, micro-ORFs and regions of ribosome binding42,43. In-frame termination codons associated with transcript degradation can be introduced by alternative splicing. To investigate the extent to which this occurs during OB differentiation, we used our PSI-Sigma and JUM splice regions to overwrite Ensembl exons to predict the introduction of in-frame premature termination codons or frameshifts (see Methods). For differential splicing events, exon extending events A3SS, A5SS and IR were more likely to introduce premature termination codons compared to exon skipping events MES, MXS and cassette events (15–36% vs. 7.6–9.4% respectively) (Fig. 2B, “PTC event”). This difference was smaller for constitutive splicing events (4.6–29% for exon extending events vs. 6.1–14% for exon skipping events). It has been reported that many splice events are not frame-preserving44. In our data, a similar proportion of exon skipping events altered the reading frame when compared to exon extending events (16–30% vs. 16–32%) (Fig. 2B, “alters frame”). Together, this means that exon-extending events during differentiation are likely to modulate nonsense-mediated decay mechanisms independently of other indirect mechanisms such as downstream frameshifts. We investigated the frequency at which premature termination codons are produced by upstream splicing events that cause a frameshift. We defined frameshifts as a change from the reference reading frame due to alternative splicing. Using our ONT long read data and existing start and end codon annotations in Ensembl (see Methods), we performed in silico translation to estimate the NMD potential of ONT transcripts. Transcripts were considered to be NMD if they had a premature termination codon in all three frames; this is a conservative approach but reduces false positives. We predicted the total nonsense transcripts to be 113 (8.2%) PSI-Sigma, 72 (7.4%) JUM differential splice events and 411 (7.2%) PSI-Sigma, 406 (5.3%) JUM constitutive splice events. These rates are similar to, although lower than, the ~ 10% reported for NMD elsewhere (reviewed by Kurosaki et al.45) whereby the lower rates likely reflect false negatives from our conservative detection approach. Despite exon skipping events altering frame as frequently as exon extending events, we noted that intron retention was the predominant mode of splicing associated with the predicted nonsense transcripts (55 differential, 113 for constitutive PSI-Sigma events; many IR events are contained in the composite category in JUM) (Fig. 2B, “NMD”). Introns associated with frameshift were not of different length to those which were not associated with frameshift (P = 0.14, Wilcoxon two-sided test). While our conservative detection approach could possibly bias the types of splice events found associated with NMD, our intron retention results were consistent with previous studies which showed it to be a predominant mode of modulating total gene expression through the production of non-protein coding transcripts39. Our predictions suggest that stop codons produced by splicing occur mainly through intron retention in differentiating MSCs, and that sequence changes due to intron retention can be either direct or indirect. We investigated the correlation of time-series splice event PSI profiles and expression profiles of their parent genes. We noted a complex relationship, evidenced by the Pearson correlation coefficients (r values) (Figure S1D). The r values across the dataset follow a bimodal distribution; this is a statistically significant departure from normality which would arise if there were no relationship between splicing and expression. However, even for frame-altering exons, those with premature termination codons and NMD-associated exons, not all splice events exhibited a negative correlation with expression (only 33.7%, 26.6% and 29.2% of respective events with r < − 0.5). This suggests that factors such as RNA kinetics and stability are likely to influence observed PSI and aggregate expression levels as recorded by RNA-Seq. Osteoblast differentiation is associated with co-ordinated changes in alternative splicing for major regulatory pathways To further understand the relationship between alternative splicing and gene expression, we compared their transcriptome-wide changes during OB differentiation (Fig. 3A). A total of 972 (PSI-Sigma) or 906 (JUM) differentially spliced genes, being those with at least one significantly changing splice event, were detectable between any two time points of differentiation (Fig. 3B) with 750 (PSI-Sigma) or 611 (JUM) differentially spliced genes detectable at all timepoints of differentiation. Splicing-specific GO terms included “RNA splicing”, “transcriptional coregulator activity”, “histone modification” and “Golgi to plasma membrane protein transport” (Fig. 3A; Table S2). Gene expression-specific terms included “ERK1 and ERK2 cascade”, “response to glucocorticoid”, “Wnt signaling pathway” and “canonical glycolysis”. Genes changing both in splicing and gene expression showed other GO terms. Indicative of synergistic splicing and expression changes, these included “mitotic cell cycle”, “response to unfolded protein” and processes related to cytoskeletal/ECM organisation. These terms also reflected the partial overlap between genes which were differentially spliced and those which were differentially expressed (Fig. 3B). When we compared genes which were differentially spliced and differentially expressed to those that were differentially spliced but not differentially expressed, the relative distribution of A3SS, A5SS and IR events were not significantly different (χ² = 9.91, df = 5, P = 0.078) suggesting that our detection of differential splicing is unlikely to be influenced by unspliced pre-mRNA39. Our data shows that changes in the transcriptome during osteogenesis cannot be fully explained by gene expression alone, as some biological processes are exclusively regulated by or in conjunction with splicing. PCA plots showed that both gene expression and alternative splicing signatures could resolve cells according to differentiation time point (Fig. 3C). In addition, the rate of significant changes in splicing events over the time course was similar to that of gene expression (Fig. 3D). Notably, the first major increase in both splicing and expression occurred during the 1d-3d interval, coinciding with the onset of matrix mineralisation and mitotic arrest (Fig. 1D). This suggests that expression and splicing are temporally co-ordinated. Fig. 3Coordinated changes in alternative splicing affect molecular drivers of differentiation. A Comparison of GO terms enriched from differentially expressed/spliced genes. B Number of genes differentially expressed/spliced. C Principal component analysis showing variance in expression or splicing data. D Cumulative changes in expression or splicing over time. Expressed as a proportion of total changes over the time-course. E Co-clustering of time-series PSI profiles for JUM and PSI-Sigma splice events. Gene ontology for each cluster shown as heatmap. NA in scale: no genes in the GO term were represented in cluster. For full clustering results, see Table S3. F Schematics depicting the splice events of notable significance. Black arrows indicate the exon/junction of interest. Genomic co-ordinates are in hg38. Grey track: Oxford Nanopore (ONT) long read; red track: PSI-Sigma splice event; purple track: JUM splice event; blue track: Ensembl transcripts. *Junction present but incorrectly classified by tool. Coordinated changes in alternative splicing affect molecular drivers of differentiation. A Comparison of GO terms enriched from differentially expressed/spliced genes. B Number of genes differentially expressed/spliced. C Principal component analysis showing variance in expression or splicing data. D Cumulative changes in expression or splicing over time. Expressed as a proportion of total changes over the time-course. E Co-clustering of time-series PSI profiles for JUM and PSI-Sigma splice events. Gene ontology for each cluster shown as heatmap. NA in scale: no genes in the GO term were represented in cluster. For full clustering results, see Table S3. F Schematics depicting the splice events of notable significance. Black arrows indicate the exon/junction of interest. Genomic co-ordinates are in hg38. Grey track: Oxford Nanopore (ONT) long read; red track: PSI-Sigma splice event; purple track: JUM splice event; blue track: Ensembl transcripts. *Junction present but incorrectly classified by tool. To understand the functional focus of alternative splicing during OB differentiation, we examined differentially spliced genes at each timepoint. We co-clustered the PSI profiles of differential splicing events from PSI-Sigma and JUM, yielding a total of 15 clusters with P < 0.05 (see Methods). Clusters 1, 2, 3, 4 and 15 displayed significant (P ≤ 0.05) enrichments for “transcription coregulator activity” (MF) and “transcription factor binding” (MF) (Figs. 3E; Table S3), suggesting splicing modulates OB differentiation through transcriptional complexes. Indeed, these clusters contained (master) transcription factors known to regulate OB differentiation. Cluster 1 (Fig. 3E) contained DDX5 [IR] (pictured, Fig. 3F), NFATC4 [A5SS] (pictured, Fig. 3F), RERE [composite] and TEAD1 [cassette]. Cluster 2 (Fig. 3E) contained RUNX2 [cassette] (pictured, Fig. 3F), ATF2 [cassette], CTNNB1 [IR], HDAC7 [cassette] and TCF3 [MES]. The function of some of these splice events have been well characterised. For example, the intron retention event in the 3’ UTR of CTNNB1 confers lower transcript stability46, whereas spliced RUNX2 proteins localise in the nucleus to exert transcription factor activity47. MSCs are known to express the RUNX2 isoform which skips exon 6, whereas more differentiated osteoblastic lineage cells tend to exclusively express the full-length transcript of RUNX2 (corresponding to Ensembl transcript ID ENST00000359524)10. Our data corroborates the observed isoform switch from RUNX2 ΔE6 to RUNX2. Of additional importance was the fact that clusters 2, 3 and 4 had clear inflection points during 1-3d. Aside from transcription factors, these clusters were also significantly enriched for “cell cycle” and “protein kinase activity” (Fig. 3E). This coincides with observations of gene expression changes as described (Fig. 1B). Together, this evidence highlights splicing as a complementary mode of regulation for transcriptional drivers of OB differentiation. Matrix mineralisation involves the trafficking of matrix vesicles to the membrane along microtubules in a Golgi-lysosome-dependent process48,49. Consistent with our identification of the onset of matrix mineralisation after 9d, splice events in cluster 5 and 6, which remained constant until a sudden change from 9-12d, were enriched for genes involved in “post-Golgi vesicle-mediated transport (BP)”, “focal adhesion (CC)” and “microtubule cytoskeleton (CC)” (Fig. 3E). The genes associated these enrichments included regulators of post-Golgi transport KLHL20 (pictured, Fig. 3F), ANK2 (pictured, Fig. 3F), KIF13A, MACF1 and SEC16A. Concurrent splicing changes in the Golgi transport and cytoskeletal machinery could help prepare osteoblasts for the metabolic demand and logistics of matrix vesicle trafficking. Splicing factors are modular subunits of the spliceosome which dictate the specificity and kinetics of splicing. We found that splicing factor genes themselves underwent extensive differential splicing and were significantly enriched (P < 0.05) in clusters 1, 2, 4, 5, 6 and 7. Splicing events in cluster 7 were associated with the greatest number of splicing factors; these included HNRNPA2B1 (pictured, Fig. 3F), PRPF3 and SRSF11. The alternative splicing of splicing factors is indicative of complex regulation of the splicing landscape by an upstream splicing factor network, which is known to occur but remains incompletely understood50. Extracellular matrix proteins often act as autocrine and paracrine signalling factors for bone cells in addition to providing a structural matrix for mineralization51. We found that all clusters except for 12 and 15 were significantly enriched (P ≤ 0.05) for “extracellular matrix structural constituent” (MF). This included many genes essential for bone function, including, FN1 and POSTN and collagen paralogs (COL1A1/1A2/5A1/6A3/8A1/16A1)52, displaying highly complex splicing patterns. These patterns of splicing may render the secreted matrix appropriate for bone. In comparison to gene expression, splicing provides a detailed, independent account of the behaviour of differentiating osteoblasts. Our data shows that alternative splicing and gene expression changes are synchronised and multiphasic during osteoblast differentiation. Confirmation of functional alternative splicing changes in the proteome To examine the consequences of alternative splicing in the proteome, and to further validate splice events seen in transcripts, we employed proteome data from undifferentiated hMSC-TERT4 cells as well as phosphoproteome data from undifferentiated (MSC) to 1d53,54. New proteomic analyses were done for cells at day 12 post-induction (see Fig. 1A). To detect peptides confirming the translation of alternatively spliced regions, we searched our MS/MS data against custom in-house peptide databases based on junctions and exons detected by JUM and PSI-Sigma. Varying a method previously described55, we also matched splice events to the ONT data to obtain exon connectivity to aid the in silico translation process (Figure S2, see Methods). Using the MaxQuant search engine56, we filtered for PEP < 0.01 and score > 30 at the peptide level yielding 1,485 splicing-associated peptides, of which 812 and 673 unambiguously implicated the translation of JUM junctions and PSI-Sigma exons respectively (Fig. 4A; see Methods). The full table of validated spliced peptides is in Table S4. An example where MS/MS data supports the existence of splice events is given in Fig. 4B, where the RNA and proteome evidence for the splicing of ATF2 is co-visualised. The corresponding assignment of MS2 fragmentation peaks is also shown in Fig. 4C. Further examples of distinct protein isoforms, which are consistent with known mRNA splicing events documented in Ensembl, are given in Figure S3. For peptides unambiguously assigned to a differentially spliced junction or exon (see Methods), and thus those that change during differentiation, they were most commonly seen in cassette and mutually exclusive exons (for peptides associated with JUM junctions, many mutually exclusive exons were grouped under the “composite” category) (Fig. 4D). In contrast, A3SS, A5SS and IR events were less frequent. Interestingly, we did find two examples where peptides confirmed IR events and where translation frame was preserved (HM13 - intron 20:31547607–31548676, PAGR1–16:29819602–29820877). Overall, our findings were consistent with observations at the transcript level, where exon-extending events are less likely to be translated. Fig. 4A subset of alternative splicing changes alter the proteome and phosphoproteome. A Number of unique splice junctions/exons validated by proteome evidence. B Co-visualisation of RNA, proteome and phosphosite evidence for a spliced peptide of a cassette exon (black arrow) in ATF2. Reference transcripts (Ensembl) shown in blue. C Fragment ion spectrum supporting splice peptide in panel (B). D Number of splice peptides validated per splice mode. E GO terms for alternatively spliced proteins. F Predicted alterations in protein domains annotated in Ensembl BioMart. Seg: Low complexity region; MobiDB-lite: Disordered region; ncoils: Coiled-coiled domain; TMHMM: Transmembrane helix; SignalP: Signal peptide. G Top 5 predicted alterations in interpro domains. H Predicted alterations in PTM sites. Left panel: phosphorylation sites identified in this study. Right panel: annotated PTM sites in dbPTM.* P < 0.05 A subset of alternative splicing changes alter the proteome and phosphoproteome. A Number of unique splice junctions/exons validated by proteome evidence. B Co-visualisation of RNA, proteome and phosphosite evidence for a spliced peptide of a cassette exon (black arrow) in ATF2. Reference transcripts (Ensembl) shown in blue. C Fragment ion spectrum supporting splice peptide in panel (B). D Number of splice peptides validated per splice mode. E GO terms for alternatively spliced proteins. F Predicted alterations in protein domains annotated in Ensembl BioMart. Seg: Low complexity region; MobiDB-lite: Disordered region; ncoils: Coiled-coiled domain; TMHMM: Transmembrane helix; SignalP: Signal peptide. G Top 5 predicted alterations in interpro domains. H Predicted alterations in PTM sites. Left panel: phosphorylation sites identified in this study. Right panel: annotated PTM sites in dbPTM.* P < 0.05 Alternative splicing modulates specific protein domains and PTM sites We investigated the function of proteins that involve alternative splicing. GO terms for alternatively spliced proteins were consistent with those observed at the RNA level (Fig. 4E). In particular, “focal adhesion”, “actin cytoskeleton” and “Golgi vesicle transport” indicate remodelling of the cytoskeleton, actin-binding proteins and Golgi proteins by alternative splicing, which in turn shape matrix secretion and mineralisation. On the other hand, “autophagosome membrane” which includes LAMP2, ATG16L1 and GABARAP, suggests that alternative splicing plays a role in regulating autophagy during osteoblast differentiation57. Interestingly, “extracellular matrix structural constituent” contained POSTN, COL1A1, COL1A2, COL6A3, FN1, as well as ABI3BP and TNC, the latter two which are positive regulators of osteolineage commitment. For ABI3BP, its intrinsically disordered region at positions 391–818 was partially spliced out, which is likely to affect protein-protein interactions. ABI3BP interacts with integrin β1 to activate mitotic arrest and lineage commitment58. Hence, alternative splicing of ABI3BP may fine-tune downstream signalling as required during the differentiation process. For TNC (tenascin C), the EGF-like and Fibronectin type III domains were spliced out, which could alter its interactions with other bone matrix proteins59,60. Since tenascin C is expressed in matrix vesicles, its alternative splicing may allow for the structural and functional specialisation of secreted bone matrix proteins. We further investigated the regions of proteins and their domains, for which we had proteomic evidence of alternative splicing. In general, except for transmembrane regions (TMHMM) and secretion signals (SignalP), all annotated regions of proteins were significantly enriched within differential VSRs (Fig. 4F). This is consistent with previous reports that protein domains and intrinsically disordered regions are common regulatory targets of alternative splicing61. In terms of specific protein domains, fibronectin, collagen and RNA splicing factors were most frequently altered (Fig. 4G). In particular, the von Willebrand factor type A domain, which facilitates protein-protein interactions, was truncated by splicing in COL6A3 (100% of domain) and CPNE7 (~ 22% of domain), highlighting a possible role of alternative splicing in the remodelling of protein-protein interactions. We investigated the degree to which phosphorylation sites, experimentally validated in this study, were modulated by splicing. The number of sites affected by splicing constituted 73% (JUM) or 76% (PSI-Sigma) of the number predicted in dbPTM, which was statistically overrepresented (P < 0.01) (Fig. 4H). Although to a lesser extent, the predicted acetylation, ubiquitination, methylation and malonylation sites from dbPTM were also statistically significant in their association with spliced regions of proteins. Together this suggests that splicing can precisely modulate the availability of PTM sites, thus forming an important regulatory process. Overall, protein changes from alternative splicing during OB differentiation are associated with matrix secretion, mineralisation, ECM maturity and RNA splicing. The regions that are dynamically spliced have a tendency to occur in protein domains, intrinsically disordered regions and in association with PTM sites. This is supported by previous observations1,62. The selective truncation of these critical regions could alter protein interactions and enzyme activity. Further, the removal of phosphorylation sites has the potential to rewire kinase signalling networks, which are instrumental in osteolineage commitment63. Discussion Through an integrative analysis of the transcriptome as well as the proteome and phosphoproteome, we have dissected the changes in alternative splicing during the differentiation of bone marrow-derived MSCs into osteoblasts. The cell state transitions involved in OB differentiation have been investigated in recent studies which conducted high-throughput profiling of miRNA and mRNA expression, as well as single-cell RNA-Sequencing7,30,64–67. In addition to recapitulating the pre-osteoblast states previously reported, our study shows patterns in splicing which are independently informative of the phenotypes of these states and would be missed if only gene expression were considered. At 1d-3d post-induction, commitment to the OB lineage occurs. Whilst this manifests in gene expression as mitotic arrest and induction of the unfolded protein response, splicing patterns revealed widespread alterations in key transcriptional regulators of differentiation. At 9d-12d post-induction, the onset of matrix mineralisation occurs. Gene expression shows changes in cytoskeleton and extracellular matrix genes, but splicing shows changes in cell adhesion molecules and Golgi/microtubule-mediated trafficking. These changes are likely to remodel osteoblast mechanosensation, clustering and exocytosis so that mineralisation is possible. Indeed, the splicing factors PTBP1 and ESRP1 are involved in the response of osteoblasts to extracellular matrix stiffness15. Thus, integrating gene expression and splicing affords a more complete view of the differentiation process of osteoblasts. The functional consequences of splicing events are diverse and are linked to the geometry of splicing as well as the functional regions in which splicing occurs. Exon skipping events (cassette, MES, MXS) mostly affect coding regions, and some of these were validated in our proteome and phosphoproteome data. This is consistent with observations that cassette exons are the most common alternative splicing events in the proteome1. On the other hand, exon extending events (A3SS, A5SS and IR) are predicted to modulate nonsense-mediated decay; further studies are required to determine if cellular NMD machinery engages with the RNA molecules containing these splice events. It is unequivocal that through co- and post-transcriptional regulation and protein sequence alterations, alternative splicing affects proteins crucial for matrix secretion and mineralisation, lineage commitment and kinase signalling (via the removal of PTM sites). Some proteins with modulated phosphorylation sites were found to be highly spliced at the RNA level. These include KLC1, MACF1 and TPM1, splicing factors such as HNRNPH1, SRRM2, TRA2A and TRA2B, as well as microtubule components DST, KLC1, MACF1, SEPTIN9, TUBA1A and TUBA1B. Despite this, we observed almost all genes only had one protein sequence isoform in the proteome data. This has been reported in a previous study which used the LC-MS dataset for uninduced MSCs employed in this study53,54, where 1.004 protein isoforms per gene were recorded. In our study, many alternatively spliced peptides may not have been detected because they are non-protein coding68, or due to limitations in the LC-MS methodology employed to generate our proteome data. Indeed, a recent study by Sinitcyn et al.69 conducted proteome profiling for six different human cell lines, employing six different digestion enzymes, high pH reversed-phase fractionation, followed by HCD and ETD-based tandem mass spectrometry. In that study, 64% of frame-preserving splice events of highly expressed genes were validated in the proteome, and even then, some constitutive splice events were missing from proteins which otherwise had the highest coverage in the dataset. Thus, challenges remain to comprehensively characterise the downstream consequences of removing protein sequences due to splicing, and by extension, the domains and PTM sites associated with these sequences. Experimental model and subject details hMSC-TERT4 culture and osteoinduction hMSC-TERT4 cells were plated at a density of 6,250 cells/cm2 in 6-well plates in standard growth medium. At 70% to 80% confluence, the medium was replaced with osteogenic medium consisting of minimum essential medium (MEM) + 10% fetal calf serum (FCS) and 1% penicillin/streptomycin supplemented with 10 mM β-glycerophosphate (Calbiochem, La Jolla, CA, USA), 50 μg/mL L-ascorbic acid (Wako Chemicals GmbH, Neuss, Germany), 10 nM dexamethasone (Sigma Chemical Co.,St. Louis, MO, USA), and 10 nM calcitriol (1,25-dihydroxyvitamin D3; kindly provided by Leo Pharma, Ballerup, Denmark). The medium was replaced every 3 days. The control cells were grown in standard growth medium. Method details Illumina short-read bulk RNA-Seq Details of the library preparation, sequencing, and data processing of RNA-seq have been described elsewhere53. Briefly, cells were harvested at eight time-points post-osteoblast differentiation induction (0, 6, 12, 24, 72, 44, 216, and 288 h). Three independent biological replicates were taken at each time-point for a total of 24 samples. Total RNA was isolated using TRIzol (Invitrogen) and high purity RNA was obtained from samples using a GenElute mammalian total RNA miniprep kit (Sigma-Aldrich; according to the manufacturer’s instructions). For each sample, 0.5 µg RNA was validated for purity using a 1% agarose gel (Sigma-Aldrich) in Tris–acetate–EDTA (TAE; Invitrogen). For mRNA-Seq sample preparation, the Illumina TruSeq RNA Sample Prep Kit (version 2) was used according to the manufacturer’s instructions, with 2 µg of total RNA for each of the three biological replicates as input. Samples were sequenced on a HiSeq 2000 with TruSeq v3 SBS reagents to generate 100 bp paired-end reads. Oxford Nanopore long-read bulk RNA-Seq Cells were harvested at six time-points post-osteoblast differentiation induction (0, 6, 24, 72, 216 and 288h). One biological replicate was taken at each timepoint for a total of 6 samples. Total RNA was isolated using TRIzol (Invitrogen) and high purity RNA was obtained from samples using a GenElute mammalian total RNA miniprep kit (Sigma-Aldrich; according to the manufacturer’s instructions). Approximately 2 μg of purified RNA was resuspended in ~20 μl TE buffer (10 mM Tris pH 8.0, 1 mM EDTA pH 8.0), and 600-700 ng total RNA was validated for purity using the TapeStation (Agilent), with samples achieving RNA Integrity Numbers (RINs) ranging from 9.0 to 9.7. RNA was multiplexed using the PCR-cDNA barcoding kit (SQK-PCB111.24) for long-read sequencing on one Oxford Nanopore PromethION flow cell. This generated 117,316,560 reads with a Q-score of at least 9.00, and a median read length of 1,099 bp. Phosphoproteome measurement by LC-MS The protocol to generate phosphoproteome profiles for MSCs using tandem LC-MS has been previously reported54. Proteome measurement by LC-MS The protocol to generate proteome profiles for uninduced MSCs using tandem LC-MS has been previously reported53,70 and an identical protocol was employed in this study for the proteomic analysis of 12d OBs, except proteolytic peptide samples from 12d OBs were subjected to three rounds of LC-MS/MS using MS/MS exclusion lists, whereas uninduced MSCs were subject to five rounds. Illumina short-read RNA-Seq data processing The quality of RNA-seq reads were confirmed with FastQC (version 0.11.9). Reads were aligned to the reference human genome (GRCh38 build) using STAR (version 2.7.2b) in transcriptome alignment mode for gene expression analysis or two-pass mode for splicing analysis71. This yielded 38,031,243 reads per sample, effectively 114,093,730 reads per timepoint when summed across all replicates. For gene expression analysis, transcriptome-aligned BAM files were quantified using RSEM. The expected counts from RSEM were normalized using RUV in residual mode (RUVr) and differential expression was detected using EdgeR. For splicing analysis, BAM files from two-pass mapping were processed according using the JUM or PSI-Sigma pipelines. For JUM, the time-series pipeline was run with the following parameters: --JuncThreshold 5; --fileNum_threshold 3; --IRthreshold 5; --Readlength 100. For PSI-Sigma, pairwise comparisons were done between each timepoint for a total of 28 comparisons, using the following parameters: --type 1; --fmode 3; --nread 1; --irmode 0. For DEXSeq re-analysis of PSI-Sigma data, we used the DEXSeqDataSet function as input – exons associated with the same “Event Region” (referred to here as Variable Splice Regions) were given the same groupID, a matrix of counts reflecting the numerator of the new PSI index were used as countData and a matrix of counts reflecting (denominator – numerator) of the new PSI index were used as input for the alternativeCountData option. This allowed a fair comparison between the results from JUM and PSI-Sigma. Differentially spliced exons or junctions were defined as having Benjamini FDR < 0.01 and PSI change > 15% between any two time-points. Constitutively spliced exons or junctions were defined as having Benjamini FDR > 0.01 and PSI change < 15% between all time-points. For transcript reconstruction, BAM files from each RNA-Seq sample were processed using Strawberry72 using the “--no-quant” option without specifying a reference GTF. GTF files for each sample were combined into a single GTF using StringTie merge with default settings. Principal component analysis of RNA-seq data was done using the “prcomp” function in R. False discovery rates (FDR) were obtained by adjusting raw P-values using the Benjamini-Hochberg correction. All gene ontology analyses were conducted using the “GOHyperGAll” systemPipeR package in R. For clustering of scaled gene expression profiles, a rectangular 5 × 5 map was made, yielding a total of 25 clusters using the “kohonen” package in R with default parameters. For clustering of combined PSI-Sigma and JUM PSI profiles, a fuzzy ensemble method was used. To do this, 30 rectangular Kohonen self-organising maps were made, with dimensions 2 × 3, 3 × 2, 2 × 4, 4 × 2, 2 × 5, 5 × 2, 3 × 4, 2 × 6, 2 × 7, 7 × 2, 4 × 4, 8 × 2, 2 × 9, 9 × 2, 5 × 4, 2 × 10, 2 × 11, 11 × 2, 6 × 4, 3 × 8, 2 × 13, 13 × 2, 7 × 4, 2 × 14, 5 × 6, 10 × 3, 8 × 4, 16 × 2, 2 × 17 and 17 × 2. These dimensions were chosen to be uniformly distributed and centred around 20, the approximate number of clusters along the diagonal in the consensus matrix when sweeping across dimensions greater than 20. Clustering was repeated with five different random number seeds. For each item, “neighbours” were defined to be items which co-clustered in more than 50% of the 150 maps generated. Each unique pair of items were used to create a network, with the number of pairwise neighbours in common representing network edge weightings. Fuzzy clusters were detected using ClusterONE in Cytoscape, with minimum cluster size: 25, node penalty: 2, haircut threshold: 0.1, merging method: “multi-pass”, similarity: “match coefficient”, overlap threshold: 0.8, seeding method: “from every node”. Detected clusters were filtered for P < 0.05. Prediction of transcript consequences due to alternative splicing To predict premature termination codons, we used the in-house script “NMD_flagger.R”. In this script, three-frame translation of all assembled transcripts was performed using the “SeqinR” package in R. Transcripts with at least 4 exons and stop codons in all three frames occurring less than 51 nt from the last exon junction (premature termination codons) were considered NMD candidates. To predict PTC events and frame-shifts resulting from the inclusion of PSI-Sigma exons, we used the in-house script “poison_exon_finder.R”. In this script, parent PSI-Sigma VSRs were matched to Ensembl transcripts. These transcripts were modified computationally by appending alternative exons from PSI-Sigma: PSI-Sigma alternative exons were inserted into the VSRs of these matched Ensembl transcripts, fully displacing any Ensembl exons which overlapped the alternative exons. PSI-Sigma alternative exons were considered to be PTC events if they introduced a stop codon upstream of the annotated stop codon for the original Ensembl transcript. To reduce the likelihood of false positives, only transcripts with annotated stop codons were considered. The inclusion of alternative exons was deemed to cause a downstream frame-shift if the nucleotide length difference between alternative exons and any displaced reference exons was not an integer multiple of 3. Significance P-values indicating VSR enrichment within Ensembl CDS, 5’ UTR, 3’ UTR, start codons and stop codons were calculated using the regioneR package in R. Custom peptide database generation Alternative junctions detected by JUM were processed using the in-house script “three_frame_translate_junctions_ORF.based.R”. Alternative exons detected by PSI-Sigma were processed using the script “three_frame_translate_exons_ORF.based.R”. In these scripts, junctions and exons were matched to transcripts in the hg38 v98 Ensembl GTF as well as the GTF generated from the ONT long read data. Junctions were matched with a tolerance of ±2 nt on both sides; exons were matched with a ±0 nt tolerance. For each matched transcript, the cDNA sequence was extracted according to the GRCh38 assembly. For junctions, this sequence covered a 50 nt upstream and 50 nt downstream window for a total of 100 nt; for exons, we considered the full sequence of the exon. In silico three-frame translation was performed using the SeqinR package in R. To reduce the likelihood of including untranslated sequences in our database, we discarded amino acids occurring after a stop codon without any intervening start codon (methionine). For junction peptides, we required the peptide to cross the splice junction. Virtual peptides shorter than 7 AA were removed. All remaining unique virtual peptides were used to create a FASTA file. LC-MS/MS data processing Raw files were searched against our custom peptide databases or human-only sequences in the SwissProt canonical and isoforms databases (November 2019 release) with the included contaminants database in MaxQuant 1.6.17.0 56. For splice peptide identification, peptide hits with FDR < 0.01 were considered. For protein identification, additional filtering was done for FDR < 0.01 at the protein level, yielding 3,101 distinct proteins. Of the total 36,570 phosphosites identified, after filtering for localization probability in addition to PEP and score filtering, 23,633 sites with localisation probability ≥ 0.75 remained. To evaluate the validity of splice-associated peptide identifications, we compared identifications from the custom database search against identifications from SwissProt73and contaminants database search. Peptides identified in both searches were considered to be valid. If there was a discrepancy with the identifications between the two searches, the identification with the greater score was preserved. In order for splice-associated peptide identifications to be considered as evidence to support the translation of JUM junctions, the peptides were required to span at least two amino acids on both sides of the splice junction. On the other hand, peptide identifications associated with PSI-Sigma exons were required to be unambiguously associated with the protein-coding region contained in the alternative exon i.e. the peptide must be specific to one transcript and must not match to any protein-coding region shared by multiple transcripts in the Ensembl reference. To predict whether protein regions were modulated by alternative splicing, we considered VSRs with at least one event observed in the proteome. For each VSR, junctions (JUM) or exons (PSI-Sigma) were matched to annotated Ensembl transcripts and alternatively spliced protein coding (CDS) regions were inferred from the cDNA sequence. For genome visualisation of identified peptides and phosphopeptides, the “peptides.txt” file from MaxQuant was processed using the in-house script “MQtobed.R” with score cutoff of 30, FDR cutoff of 0.01 and localisation cutoff of 0.75 (for phosphopeptides only). The Ensembl GRCh38 v98 GTF as well the file “table_ENSP_to_uniprot_entry_mapping_38.98.txt” was used as additional input. A note on nomenclature and terminology In accordance with the approach that PSI-Sigma uses to detect splicing, we define “Variable Splice Regions” (VSRs) as regions where alternative splicing occurs. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3 Supplementary Material 4 Supplementary Material 4 Supplementary Material 5 Supplementary Material 5