Diverse targets of SMN2-directed splicing-modulating small molecule therapeutics for spinal muscular atrophy

Abstract Designing an RNA-interacting molecule that displays high therapeutic efficacy while retaining specificity within a broad concentration range remains a challenging task. Risdiplam is an FDA-approved small molecule for the treatment of spinal muscular atrophy (SMA), the leading genetic cause of infant mortality. Branaplam is another small molecule which has undergone clinical trials. The therapeutic merit of both compounds is based on their ability to restore body-wide inclusion of Survival Motor Neuron 2 (SMN2) exon 7 upon oral administration. Here we compare the transcriptome-wide off-target effects of these compounds in SMA patient cells. We captured concentration-dependent compound-specific changes, including aberrant expression of genes associated with DNA replication, cell cycle, RNA metabolism, cell signaling and metabolic pathways. Both compounds triggered massive perturbations of splicing events, inducing off-target exon inclusion, exon skipping, intron retention, intron removal and alternative splice site usage. Our results of minigenes expressed in HeLa cells provide mechanistic insights into how these molecules targeted towards a single gene produce different off-target effects. We show the advantages of combined treatments with low doses of risdiplam and branaplam. Our findings are instructive for devising better dosing regimens as well as for developing the next generation of small molecule therapeutics aimed at splicing modulation.


INTRODUCTION
Spinal muscular atrophy (SMA), the leading genetic cause of infant mortality, results from low le v els of the Survival Motor Neur on (SMN) pr otein due to deletions or mutations of the SMN1 gene (1)(2)(3). SMN is a multifunctional protein involved in most aspects of RNA metabolism, including transcription, pre-mRNA splicing, mRNA trafficking, translation and stress granule formation ( 4 ). Consistently, most organs and tissues, such as the brain, gastrointestinal tract, heart, li v er, lung, kidney, muscles, pancreas, testis and ovary are intrinsically affected in SMA ( 5 ). SMN2 , a nearly identical copy of SMN1 present in humans, cannot compensate for the loss of SMN1 due to predominant skipping of exon 7 ( 6 , 7 ). Manipulation of splicing to r estor e SMN2 exon 7 inclusion provides a promising therapeutic avenue for SMA ( 8 ). Skipping of SMN2 exon 7 is driven primarily by a weak 3 splice site (3 ss) created due to a critical C-to-T mutation at the sixth position (C6U substitution in RNA) of the exon coupled with an inherently weak 5 ss ( 7 , 9-14 ). The 5 ss of SMN exon 7 is surrounded by multiple negati v e cis-elements, including the terminal stem-loop structure 2 (TSL2), intronic splicing silencer N1 (ISS-N1), a GC-rich sequence (GCRS) and an internal-stem formed by a long-distance interaction (ISTL1) (Figure 1 A) (14)(15)(16)(17)(18)(19)(20). Enhancing the accessibility of the 5 ss of exon 7 through either sequestration of ISS-N1 / GCRS by an antisense oligonucleotide (ASO) or disruption of TSL2 / ISTL1 has been shown to r estor e SMN2 exon 7 inclusion ( 21 ). Also, for ced r ecruitment of U1 small nuclear ribonucleoprotein (snRNP) at or downstream of the 5 ss of exon 7 has been found to promote exon 7 inclusion (22)(23)(24). Nusinersen (Spinraza ™), an ISS-N1-targeting ASO deli v ered through intrathecal administration, became the first Food and Drug Administration (FDA)-a pproved thera py for SMA in 2016 ( 25 , 26 ). Howe v er, an ASO-based therapy has limitations linked to poor body-wide deli v ery / distribution and potential off-target effects ( 27 , 28 ). Novel chemical modifications and careful size selection are needed to further advance ASO-based therapies ( 27 , 29 , 30 ). Zolgensma (also named onasemnogene abeparvovec-xioi), a gene replacement therapy with adenoassociated virus (AAV)-mediated SMN gene deli v ery, became the second approved treatment for SMA ( 31 ). Gene therapy has its own disadvantages; for example, uncontrolled expression of SMN may sequester and / or mislocalize cellular transcripts and proteins that interact with SMN (32)(33)(34). Consistently, a recent study found toxicity associated with the long-term AAV9-mediated SMN ov ere xpression ( 35 , 36 ).
Small molecules offer advantages of oral administration and excellent body-wide distribution ( 8 , 37 ). Risdiplam, an orally deli v erab le small molecule that r estor es SMN2 exon 7 inclusion, was recently approved for the treatment of SMA ( 38 , 39 ). Branaplam, another small molecule capable of restoring SMN2 exon 7 inclusion, has been considered for SMA therapy ( 40 ). Howe v er, clinical trials of branaplam wer e r ecentl y suspended due to ad verse effects. Additional small molecules that promote SMN2 exon 7 inclusion have been identified, although they have yet to enter clinical trials (41)(42)(43). Mechanisms by which small molecules stimulate SMN2 exon 7 inclusion appear to be complex and not yet fully understood. For instance, se v eral analogs of risdiplam were found to interact with GA-rich motifs of the single-stranded regions of both RN A and DN A (44)(45)(46). It has been proposed that binding of risdiplam to singlestranded regions of RNA helps recruit Far Upstream Element Binding Protein 1 (FUBP1) and its homolog KH-type Splicing Regulatory Protein (KHSRP) ( 45 ). Such recruitment may lead to sequestration of these proteins, affecting their targets. At least one analog of risdiplam has been shown to interact with a bulged adenosine residue present in the RN A:RN A duplex formed between the 5 ss of SMN2 exon 7 and U1 snRNA ( 47 ). Risdiplam-mediated stabilization of bulged adenosine has been proposed to r estor e the accessibility of the U1-C zinc finger for interactions with the minor groove of the U1:5 ss duplex ( 47 ). Branaplam (synonym: NVS-SM1) has also been shown to bind to the interface formed between the 5 ss of SMN2 exon 7 and U1 snRNP ( 48 ). It remains to be seen if such an interaction offers a high degree of specificity due to involvement of structural elements of SMN2 pre-mRNA. Employing highthroughput RNA sequencing (RNA-Seq), an early study analyzed off-target effects of 500 nM C3, a risdiplam derivati v e, in SMA patient fibrob lasts ( 37 ). While the authors confirmed significant change in expression of six genes, no validation was reported for any of the aberrantly spliced e v ents ( 37 ). Another RNA-Seq study analyzed the off-target effects of 100 nM branaplam in SMA patient fibroblasts and used quantitati v e polymerase chain reaction (qPCR) to validate 11 aberrant splicing e v ents ( 48 ). Howe v er, in most cases, qPCR results showing enhanced exon inclusion were not complemented with information about the corr esponding decr ease in exon skipping, making the findings somewhat difficult to interpret ( 48 ). A recent study compared the effects of 121 nM risdiplam (synonym used: compound 1), 605 nM risdiplam and 24 nM branaplam ( 38 ). Howe v er, the study was limited to examining only fiv e off-target splicing e v ents for STRN3 , FOXM1 , APLP2 , MADD and SLC25A17 by mapping of RNA reads, without independent validation of the aberrantly spliced products by qPCR or gel-based assays ( 38 ). Thus far, a systematic RNA-Seq analysis supported by complementary validations of different types of off-target effects associated with the small molecule therapeutics for SMA has not been performed.
Here we examine the transcriptome-wide effects of risdiplam and branaplam in SMA patient cells employing RNA-Seq. The study was aimed at capturing and comparing the broad spectrum of of f-target ef fects caused by these compounds. Employing qPCR and semi-quantitati v e PCR, we validated the findings of RNA-Seq at three different concentrations of each compound. The high concentrations of both compounds caused massi v e perturbations of the transcriptome within 24 h of treatment. The extent of alterations was significantly higher in the case of risdiplam than branaplam. Genes with aberrant expression were associated with important cellular processes including DNA replication, cell cycle, RNA metabolism, cell signaling and metabolic pathways. Among ad versel y affected splicing e v ents, we ca ptured m ultiple incidences of aberrant exon inclusion, exon skipping, intron retention, intron removal and alternati v e splice site usage. While low concentrations of the compounds partially r estor ed inclusion of SMN2 exon 7, the off-target effects wer e r educed in the case of risdiplam and were almost non-existent in the case of branaplam. The results of complementary experiments using minigenes expressed in HeLa cells suggested that the motifs within exons and their immediate downstream intronic sequences modulate the action of risdiplam and branaplam. Our findings also re v eal pre viously unreported novel exonic motifs associated with the stimulatory effect of risdiplam and branaplam on splicing of SMN2 exon 7. We show the advantages of combined treatments with low doses of the compounds in modulating splicing of SMN2 exon 7 while minimizing the potential off-target effects. Our findings are significant for devising the effecti v e dosing regimens as well as for de v eloping the next generation of small molecule therapeutics for the treatment of SMA.

Cell culture
Tissue culture media and supplies were purchased from Life Technologies (Gibco brand) unless indicated otherwise. GM03813 Type I SMA patient fibroblasts were obtained from Coriell Cell Repositories and were grown in minimal essential medium (MEM; Gibco, catalog #10370) supplemented with 2 mM GlutaMAX-1 and 15% fetal bovine serum (FBS) at 37 • C under 5% CO 2 . Human cervical adenocarcinoma (HeLa) cells were obtained from the American Type Culture Collection and were grown in Dulbecco's modified Eagle's medium (DMEM) supplemented with 10% FBS.

T r eatment with splicing-modulating compounds
Risdiplam and branaplam were purchased from Med-ChemExpr ess. To pr epar e stock solutions of risdiplam and branaplam, the compounds were resuspended in dimethylsulfoxide (DMSO) (Sigma) to a concentration of 5 mM. To fully dissolve the compounds, 1% hydrochloric acid was added dropwise to adjust the pH value. GM03813 cells were pre-plated 16 h before trea tment a t a density of 1.1 × 10 6 cells per 100 mm cell culture dish or 4 × 10 5 cells per 60 mm cell culture dish. The compounds were diluted in DMSO and then added to fresh cell culture medium so that the final concentration of DMSO was 0.1%. For treatments, cell culture medium was replaced with fresh medium containing the indicated concentration of DMSO alone, risdiplam or branaplam. Twenty-four hours later, cells were collected for RNA isolation followed by re v erse transcription-PCR (RT-PCR) or RNA-Seq library preparation.

Generation of minigenes and transfections
Hybrid minigenes were constructed using p SMN2 I6 ( 12 ). Briefly, the middle portion of the p SMN2 I6 minigene encompassing the last 121 nt of intron 6, the entire exon 7 and the first 165 nt of intron 7 of p SMN2 I6 was replaced with the exon of interest and its 100-200 nt long flanking intronic sequences (upstream and downstream). DNA fragments for cloning were generated by PCR using Phusion high fidelity DN A pol ymerase (Thermo Scientific) following the manufacturer's recommendations. Genomic DNA templa te was isola ted fr om SH-SY5Y neur oblastoma cells using the Qiagen DNeasy blood and tissue kit. First we generated the POMT2 hybrid minigene. In particular, the DNA fragment corresponding to POMT2 exon 11B and its flanking intron sequences (amplified from genomic DNA) was ligated by PCR with the fragment containing the last 279 nt of SMN2 intron 7 and the entire exon 8 (amplified from the pSMN2 I6 plasmid). As a r esult, a SalI r estriction site was also introduced at the junction of POMT2 intron 11B and SMN2 intron 7. The hybrid fragment was then treated with KpnI and NotI, gel purified and inserted into pSMN2 I6 digested with the same restriction enzymes. The remaining hybrid minigenes were generated by replacing POMT2 sequences with other sequences of interest using KpnI and SalI sites. Site-specific mutagenesis was achie v ed by twostep PCR using primers carrying specific base substitutions.
The MBNL1 minigene was constructed by amplifying an ∼2.5 kb genomic region from the beginning of exon 4 to the end of exon 6 of MBNL1. The resulting PCR amplification product was inserted into the pCI vector (Promega) using SalI and NotI sites. In order to facilitate the process of MBNL1 exon 5 mutagenesis, AgeI and SacII sites were introduced 228 nt upstream and 217 nt downstream of exon 5, respecti v ely, using a PCR-based approach. SMN2 minigenes with site-specific mutations in exon 7 and partially randomized exon 7 sequences were gener ated similar ly to as previously described ( 13 ). The identities of all minigene constructs were confirmed by Sanger sequencing.
Plasmids for transfections were purified with the QI-Apr ep spin minipr ep kit (Qiagen). Transfections wer e done using either HeLa cells in suspension (re v erse transfection) or pre-plated HeLa cells / GM03813 cells. Briefly, HeLa cells wer e pr e-pla ted a t a density of 1.1 × 10 5 cells per well of 24-well plates. GM03813 cells were pre-plated at a density of 1.8 × 10 5 per well of 6-well plates. Sixteen hours later, cells were transfected with 0.05 g of the minigene of interest and 0.45 g of empty pCI neo vector in the case of HeLa cells, and 0.1 g of the minigene of interest and 0.9 g of empty pCI neo vector in the case of GM03813 cells, using Lipofectamine 2000 (Invitrogen) as per the manufactur er's r ecommendations. Aliquots of 0.5 and 1.0 g of pCI neo alone were used as a transfection control for HeLa and GM03813 cells, respecti v ely. For re v erse transfection, HeLa cell suspensions containing 2.9 × 10 5 cells per transfection were combined with DNA-Lipofectamine 2000 complexes. These cells were then plated in 24-well plates. Six hours after transfection, cell culture medium was replaced with fresh medium containing DMSO alone, risdiplam or branaplam, as described above. After a further 24 h (30 h after minigene transfection), cells were collected for RNA isolation.

Randomization and in vivo selection of FOXM1 e x on 9
We introduced a 10 bp randomized region spanning from positions 70 to 79 of FOXM1 exon 9 in the context of the hybrid minigene. We employed a three-fragment PCR-based approach in which a 59mer oligonucleotide encompassing the 10 nt randomized region in the middle and flanking portions of FOXM1 exon 9 was used as template along with overlapping 5 and 3 fragments generated by PCR from the FOXM1 exon 9 hybrid minigene. The complete fragment was then digested with SalI and KpnI and ligated into ∼500 ng of minigene digested with the same enzymes. Individual clones were isolated and sequenced, and then used for transfection as described above. At 16 h after transfection, cells wer e tr eated with 0.1% DMSO, 250 nM risdiplam or 10 nM branaplam for 6 h and then RNA was collected for RNA isolation followed by RT-PCR.
For in vivo selection, 20 l of ligation reaction corresponding to ∼100 ng of plasmid was used directly for transfection in 6-well plates along with empty pCI vector to bring the final DNA concentration to 2 g per well, as performed previously ( 13 ). Sixteen hours after transfection, cells were treated with 0.1% DMSO, 250 nM risdiplam or 10 nM branaplam for 6 h and then RNA was collected for RNA isolation followed by RT-PCR. First, minigene-deri v ed transcripts were amplified by primers P1 and P2. Then, the Nucleic Acids Research, 2023, Vol. 51, No. 12 5951 band corresponding to inclusion of exon 9L was excised and purified by the 'crush and soak' method. The randomized region was then re-amplified using primers RR-F and RR-R to regenerate a double-stranded sequence corresponding to the original FOXM1-Random oligonucleotide (Supplementary Table S1). Amplification with flanking regions and ligation into the hybrid minigene were performed as described above. The entire process was repeated for a total of four cycles of selection.

RT-PCR and qPCR
Total RNA was isolated using TRIzol reagent (Life Technologies) following the manufacturer's instructions. Unless noted otherwise, RNA samples wer e tr eated with RQ1 RNase-free DNase (Promega) following the manufacturer's instructions and then re-purified using phenol:chloroform extraction and ethanol precipitation. cDNA was generated using SuperScript III re v erse transcriptase (Life Technologies) with 0.5-1.0 g of total RNA per 5 l reaction. The multi-exon skipping detection assay (MESDA) was performed similarly to as described previously ( 23 , 49-51 ). Of note, MESDA is a powerful technique to determine the relati v e abundance of multiple isoforms of SMN transcripts in a single experiment. Also, results of MESDA cannot be influenced by the amount of the template taken for PCR and / or the amount of sample loaded on the gel. cDNA templates for MESDA were generated using a gene-specific primer located within SMN exon 8. PCR for MESDA was carried out using primers located in exon 1 and the beginning of exon 8 (Supplementary Table S1). For semi-quantitative RT-PCR, cDNA was generated using oligo(dT) [12][13][14][15][16][17][18] (Life Technologies). The PCR step was carried out using Taq DN A pol ymerase (New England Biolabs) and cDNA template following the manufactur er's r ecommendations. PCR products were separated on nati v e polyacrylamide gels and visualized by ethidium bromide staining. Analysis and quantification of bands were performed using ImageJ software. Unless noted otherwise, PCR product identities were confirmed by sequencing as described previously ( 23 ). For qPCR, cDNA was pr epar ed using random primers (Promega) and 0.5 g of RNA per 5 l of re v erse transcriptase reaction. qPCR was performed using PowerUp SYBR green master mix (Life Technologies) on a QuantStudio 3 (Thermo Fisher) thermocycler according to the manufacturer's instructions. Relati v e e xpression was determined using the Ct method utilizing OAZ1 as the normalizing gene for gene expression assays or primers targeting constituti v ely spliced regions of the gene of interest for intron retention assays. All primers were obtained from Integrated DNA Technologies. Their sequences are listed in Supplementary  Table S1.

Libr ary gener ation and RNA-Seq
To confirm RNA integrity, TRIzol-isolated total RNA was characterized using an Agilent Bioanalyzer on an RNA nano chip (RIN ≥8). A 1 g aliquot of total RNA was then subjected to rRNA depletion using the NEBNext rRNA depletion kit v2 (Human / Mouse / Rat). Libraries were generated from rRNA-depleted RNA using the NEBNext Ultra II directional RNA library prep kit for Illumina. Libraries wer e bar coded for multiplexing using NEBNext Dual Index oligos for Illumina. The size distribution of libraries was determined using an Agilent Bioanalyzer DNA 1000 chip and quantified using a Qubit fluorimeter. Libraries were pooled together and sequenced on an Illumina Novaseq 60000 using an S23 flow cell following a 100 cycle, paired-end protocol.

RNA-Seq read mapping and bioinformatic analysis
Reads from RNA-Seq were mapped to the human r efer ence genome build GRCh38 using HISAT2 ( 52 ). For differential expr ession, mapped r eads wer e assigned to genes according to the Gencode v33 human transcriptome annotation ( 53 ) using the featureCounts script from the Subread software package ( 54 ). Differential expression was estimated using the DESeq2 R package ( 55 ). To identify differentially affected alternati v e splicing e v ents, mapped r eads wer e analyzed by rMATS ( 56 ). After initial identification, significant e v ents were subjected to the following filtering criteria: false discovery rate (FDR) < 0.05, ≥10 average junction reads supporting each isoform in at least one sample group, and a change in percentage spliced in (PSI) values of at least 0.1 for alternati v ely spliced exons or 0.05 for differential intron r etention / r emoval.
Significantly enriched Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways ( 57 ) were identified using WE-BGestalt (w e bgestalt.org). Differ entially expr essed genes were categorized by gene type, chromosomal location and number of transcripts per gene using Ensembl BioMart to filter gene lists. As a r efer ence list, we used all the genes with measurable expression (at least 10 reads per sample) in GM03813 cells. Transcription factors with enriched targets in differentially expressed genes were identified using ChEA3 ( 58 ). Splice site scor es wer e calculated using the tools hosted at http://rulai.cshl.edu/new alt exon db2/ HTML/score.html . Enriched sequence motifs were identified using MEME ( 59 ).

T r anscriptome-wide perturbations trigger ed by high concentrations of risdiplam and branaplam
In order to capture the transcriptome-wide perturbations caused by risdiplam and branaplam, we performed RNA-Seq on transcripts isolated from GM03813 Type I SMA patient fibroblasts treated with either compound for 24 h. Of note, these cells carry only the SMN2 gene. To select concentrations of interest, we first treated cells with four different doses of each compound (from 10 nM to 1 M for risdiplam and from 0.5 nM to 40 nM for branaplam) followed by examining SMN2 exon 7 splicing by semiquantitati v e PCR (Supplementary Figure S1A). We chose 1 M risdiplam (hereafter referred to as HiR) and 40 nM branaplam (hereafter referred to as HiB) as high concentrations based on the near total inclusion of SMN2 exon 7 (Figure 1 B). For low concentrations, we chose 50 nM risdiplam (hereafter referred to as LoR) and 2 nM branaplam (her eafter r eferr ed to as LoB). The selected high and low concentrations were in the range used in previously reported studies ( 42 , 44 , 48 ). Recently, a high concentration of an ISS-N1-targeting ASO has been shown to trigger off-target inclusion of SMN exon 6B that is deri v ed from an Alu element ( 28 , 60 ). Also, SMN exons 3, 5 and 7 undergo enhanced skipping under oxidati v e str ess conditions ( 49 , 51 ). The r esults of our RNA-Seq and MESDA showed that neither compound had an appreciable effect on individual splicing of SMN exons 3, 5 and 6B (Supplementary Figure S1B, C). As indicated in Figure 1 C, HiR altered expression of 10921 genes including 3670 genes with > 2-fold change. Of these, roughly half were upregulated and half were downr egulated (Figur e 1 C). Among highly expr essed genes, ther e was an apparent bias towards up-regulation compared with down-r egulation (Figur e 1 D). LoR alter ed (mostly upr egulated) the expression of 566 genes, with only one of them showing > 2-fold change (Figure 1  . For upregulated genes, the greatest overlap was between treatments with different doses of the same drug (Figure 1 E). Of the 367 genes affected by LoR, > 90% were also affected by HiR. All se v en genes upregulated by LoB were also upregulated by HiB. In contrast, only 479 genes out of 5626 and 1000 genes upregulated by HiR and HiB, respecti v ely, were common. Similarly, among downregulated genes, a greater overlap was found between LoR and HiR (153 genes, compared with 199 total LoR downregulated and 5295 HiR downregulated genes) and between LoB and HiB (7 genes, compared with 8 total LoB downregulated and 1187 total HiB downregulated genes) than between treatments with different drugs (Figure 1 E).
Genes upregulated and downregulated by HiR and HiB were located on all chromosomes (Supplementary Figure   S2). Of all the genes that ar e expr essed in GM03813 cells, ∼83% are protein coding, ∼13% are long non-coding RNA (lncRNA) genes and ∼4% ar e pseudogenes (Figur e 1 F). In the category of genes upregulated by HiR, protein-coding genes became enriched ( ∼91%), while in the downregula ted ca tegory, the rela ti v e proportion of affected genes showed only a slight change (Figure 1 F). In the case of HiB, protein-coding genes were over-represented in both up-and downregulated categories (Figure 1 F). Both HiR and HiB appeared to preferentially affect expression of genes that generate multiple alternati v e transcripts rather than genes with fewer alternative transcripts (Supplementary Figure S2B, E). We did not observe any significant overlap between enriched transcription factor target sites in HiR-and HiB-affected genes (Supplementary Figure  S2C, F). We performed pathway enrichment analysis on up-and downregulated genes. Among genes upregulated by HiR, we observed a significant enrichment of genes involved in RNA ma tura tion-and transla tion-rela ted processes such as ribosome , mRNA surveillance , microRNAs (miRN As) in cancer, RN A transport, ribosome bio genesis and spliceosome (Figure 1 G). In addition, we noted enrichment of genes involved in ubiquitin-mediated proteolysis and autophagy, suggesting an effect on protein homeostasis (Figure 1 G). Gi v en the low number of genes upregulated by LoR, the nature of affected pathways was distinct from those observed with HiR. Among pathways that involved genes downregulated by HiR were cell cycle, DNA replication, base excision repair, homologous recombination, general meta bolism (meta bolic pathways, carbon metabolism) and organelle function (lysosome, peroxisome) ( We performed motif enrichment analysis using 200 nt long promoter sequences extracted from the 200 most upr egulated and downr egulated transcripts by HiR and HiB (Supplementary Figure S3A-D). We observed enrichment of a G-rich motif with interspersed A residues in genes upregulated by HiR, which bears a strong resemblance to the target sequence of risdiplam identified in the body of SMN exon 7 ( 46 ) as well as the binding site of several zinc fingertype transcription factors (Supplementary Figure S3E). In genes downregulated by HiR as well as in genes up-and downregulated by HiB, we observed enrichment of a GCrich motif (Supplementary Figure S3B-D). Since the GCrich motif was present in all of three categories of genes, we assume this to be a common promoter element in genes expressed in fibroblasts. In addition, genes downregulated by HiR had enrichment of an A-rich motif with interspersed G residues, although it showed less consistent conservation than the motif found in HiR-upregulated genes (Supplementary Figure S3C).
We analyzed se v en types of alternati v e splicing e v ents, namely exon skipping (ESK), exon inclusion (EIN), incr eased intron r etention (IRT), impr oved intr on removal (IRM), alternati v e 5 ss usage (A5S), alternati v e 3 ss usage (A3S) and m utuall y e xclusi v e / mixed usage of neighboring exons (MXE) (Figure 1 H). HiR altered 2385 splicing e v ents, with the most common being ESK (33%) followed by EIN (18%), IRM (13%), MXE (13%), A5S (11%), A3S (8%) and IRT (4%) (Figure 1 I). LoR caused modest changes in splicing e v ents, with the most common being A3S and ESK (21% and 19%, respecti v ely), followed by EIN (15%), A5S (15%), MXE (12%), IRM (9%) and IRT (9%). Of the 835 alternati v e splicing e v ents triggered by HiB, 39% were EIN, followed by ESK at 24%. All other types of e v ents were < 10% each. LoB impacted 76 splicing e v ents that were mostly ske wed towar ds EIN, which constituted 51% of all e v ents. MXE (13%) and A3S (12%) were the next most common e v ents, and only a handful of e v ents were associated with other types of alternati v e splicing. The majority of affected splicing e v ents w ere distinct betw een HiR and HiB (Figure 1 J). The most significant overlaps were found in EIN, with 54 e v ents shared between HiR and HiB (13% of HiR-and 19% of HiB-triggered e v ents), 24 shared between HiB and LoB (8% of HiB-and 67% of LoB-riggered e v ents) and 16 wer e shar ed among HiR, HiB and LoB. We also observed a higher degree of overlap of IRM e v ents triggered by HiR and HiB (7% of HiR-triggered e v ents and 40% of all HiBtriggered e v ents). The remainder of the e v ents showed similar le v els of ov er lap, r anging from 5% to 12% for HiR and from 15% to 25% for HiB. Aside from EIN, no e v ents were prevalent enough in LoR or LoB to draw any definiti v e conclusions.
Of the 356 genes associated with EIN affected by HiR, the expression of 90 genes was downregulated, 95 genes remained unchanged and 171 genes wer e upr egulated (Figure 1 K). Using a hypergeometric test, we determined whether this corr elation r epr esented statistically significant under-or over enrichment (Figur e 1 K). Downregulated genes wer e under-r epr esented 1.3-fold compar ed with random chance, while upregulated genes were over-represented 1.38-fold. This suggests that EIN e v ents caused by HiR were strongly associated with increased transcript le v els.
Similarly, downr egulated genes wer e under-r epr esented in ESK-associated genes 1.25-fold, while upregulated genes wer e over-r epr esented 1.35-fold (Figur e 1 K). For IRT and IRM e v ents, only IRM showed a significant association with transcription up-regulation (1.4-fold enrichment, Figure 1 K). A3S, A5S and MXE were not significantly enriched among downregulated or upregulated genes. In the case of branaplam, a large number of genes with altered e xpression le v els were also coupled with aberrant splicing (Figure 1 K). In particular, transcripts undergoing EIN were enriched among both downregulated (2.85-fold) and upregulated (2.48-fold) genes, and IRT e v ents were strongly enriched (3.65-fold) among downregulated genes.

Validation of expression of representative genes dysregulated by risdiplam and branaplam
We validated the results of RNA-Seq by qPCR using 14 representati v e genes that were aberrantly expressed in the presence of HiR and / or HiB. For validation purposes, we used risdiplam at 1 M, 250 nM and 50 nM as high (HiR), intermediate (InR) and low (LoR) concentrations, respecti v ely. For branaplam, we used 40 nM, 10 nM and 2 nM as high (HiB), intermediate (InB) and low (LoB) concentrations, respecti v ely (Figure 2 A). Considering both compounds were dissolved in DMSO, for all comparisons we used DMSO control as base le v el e xpr ession. We included untr eated control to monitor if DMSO itself had an effect on the expression of genes in question. We began with the examination of expr ession of thr ee genes, YTHDF2 , GGNBP2 and SRSF3 , which, according to RNA-Seq, wer e upr egulated by HiR but not HiB. YTHDF2 codes for an N -6-methyladenosine (m 6 A) 'reader' protein that binds m 6 A-modified RNAs and regulates their stability ( 61 ). Interestingly, lung adenocarcinoma cells have been found to display ele vated e xpression of YTHDF2 ( 62 ). The GGNBP2 gene codes for a protein that interacts with gametogenetin during spermatogenesis; the protein is also associated with amyotrophic lateral sclerosis (ALS), ovarian cancer, breast cancer and chromosome 17Q12 deletion syndrome (63)(64)(65)(66)(67). SRSF3 codes for SRp20 that is highly expressed in cancers ( 68 ). SRSF3 regulates alternati v e splicing and RNA stability; it also mediates m 6 Adependent mRNA splicing and nuclear export along with YTHDC1 ( 69 , 70 ). Our qPCR validation confirmed ∼1.6and > 4-fold increases in YTHDF2 expression by InR and HiR, respecti v el y, w hereas LoR as well as branaplam produced no effect (Figure 2 B). GGNBP2 le v els wer e upr egulated by ∼1.9and ∼4.3-fold by InR and HiR, respecti v ely (Figure 2 B). GGNBP2 le v els were not significantly affected by LoR or any of the treatments with branaplam (Figure  2 B). We observed up-regulation of SRSF3 by risdiplam in a concentration-dependent manner, with HiR showing the most prominent change of ∼6.3-fold (Figure 2 B). A negligible but statistically significant increase of SRSF3 le v els was also observed with HiB.
We next examined the expression of CDC37L1 , CACNB1 and ZFP82 that, according to our RN A-Seq anal ysis, were upregulated by both compounds. CDC37L1 codes for a tumor suppressor that acts by negati v ely regulating cy clindependent kinase 6 ( 71 ). CACNB1 codes for the beta subunit of a voltage-dependent calcium channel and its d ysregula tion has been linked to pathological conditions, including malignant hyperthermia and headache ( 72 ). ZFP82 codes for a zinc finger transcription factor that displays tumor suppressor properties ( 73 ). Our qPCR results validated strong up-regulation ( ∼3-fold) of CDC37L1 by HiR, while InR and HiB induced an ∼1.5-fold increase in its expression ( We validated the expression of two genes, namely LIG1 and SORT1 , that were downregulated by risdiplam but not branaplam as per our RNA-Seq analysis (Figure 2 E). LIG1 codes for a protein involved in joining of Okazaki fragments during DNA replication, and its mutation is associated with delayed growth, cancer and possibly immunodeficiency ( 76 ). SORT1 codes for Sortilin, a protein that interacts with apolipoprotein B100 (apoB100) in the Golgi and regulates plasma low-density lipoprotein le v els ( 77 ). Dysregulation of SORT1 is associated with coronary artery diseases ( 78 ). Our qPCR results confirmed down-regulation of LIG1 by risdiplam in a concentration-dependent manner, whereas the effect of branaplam on LIG1 expression was minuscule (Figure 2 E). SORT1 was significantly downregulated by HiR and InR, w hile brana plam treatments triggered only slight reductions in the expression of this gene (Figure 2 E). We tested the effect of the compounds on the expression of PDXDC1 and EGR1 shown to be downregulated by both HiR and HiB as per our RNA-Seq results (Figure 2 F). PDXDC1 codes for a protein that participates in phospholipid metabolism and is associated with eleva ted circula ting lipid le v els based on GWAS (genomewide association studies) ( 78 ). EGR1 codes for a transcription factor that controls cell growth and survival by targeting multiple tumor suppressors ( 79 ). Validation by qPCR confirmed that PDXDC1 e xpression le v els were strongly reduced by both compounds in a concentration-dependent manner, with ∼5-fold reduction by HiR and ∼10-fold reduction by HiB (Figure 2 F). EGR1 was strongly downregulated by InR and HiR as well as by HiB (Figure 2 F). The effect of the low concentration of the compounds ap-peared to be some what ske wed as DMSO itself slightly upregulated EGR1 compared with untreated cells (Figure 2 F). Finally, we verified the expression of DOCK11 and PDS5B shown to be downregulated by branaplam by our RN A-Seq anal ysis (Figure 2 G). DOCK11 codes for a protein that is a member of the DOCK family of guanine nucleotide exchange factors. The protein regulates the actin cytoskeleton and cell motility. DOCK11 d ysregula tion has been linked to immunodeficiency disorders ( 80 ). PDS5B codes for a protein r equir ed for pairing of sister chromatids during cell division, and down-regulation of PDS5B has been associated with breast cancer ( 81 ). Our qPCR results valida ted tha t branaplam downregula ted DOCK11 in a concentration-dependent manner, with the strongest effect ( ∼3-fold decrease) caused by HiB (Figure 2 G). PDS5B was downregulated by ∼20% with InB and by ∼80% with HiB ( Figure 2 G).
To gain mechanistic insights, we examined the effect of HiR and HiB on expression of the representati v e genes at different time points using qPCR. While a quick response to a compound may support a direct and / or high affinity engagement with a target, the delayed response could be due to indirect and / or low affinity engagement with a target. We chose 3 h treatment as the earliest time point as both HiR and HiB induced a substantial increase in inclusion of SMN2 exon 7 at this time (  Figure S3). Many of these genes harbor stretches of purine-rich motifs within their TSS regions. Howe v er, we could not find a definiti v e correlation between the effect of the compounds on transcription and a specific TSS motif(s)

Analysis and validation of e x on inclusion events triggered by risdiplam and branaplam
In order to better understand the mechanisms behind the of f-target splicing modula tion by risdiplam and branaplam, we compared the conservation of sequences at the 5 ss of Nucleic Acids Research, 2023, Vol. 51, No. 12 5957 each exon whose inclusion was triggered by HiR and / or HiB (Supplementary Figure S4). Our analysis indicated sequence conservation from two bases upstream to six bases downstream of each exon / intron junction (Supplementary Figure S4A). The first six intronic bases matched the consensus 5 ss, GUAAGU. After the invariant GU dinucleotide, the next highest conservation was observed for G at the fifth intronic position, followed by As at the third and fourth intronic positions. U at the sixth intronic position had the lowest conservation. We also looked at the last two exonic positions (positions -2 and -1) ( Supplementary Figure S4). In human exons, A at the -2 position and G at the -1 position are represented at > 55% and ≥75% instances, respecti v el y. Exons, inclusion of w hich was promoted by HiR, matched this consensus rather w ell; how e v er, HiB-only sensiti v e e xons had a noticeable enrichment of G at their -2 position and A at their -1 position (Supplementary Figure  S4A). This enrichment was e v en stronger in the exons affected by both compounds.
We analyzed the splicing-rele vant conte xt for 30 exons (EIN e v ents) strongly affected by HiR onl y, HiB onl y or both (Supplementary Figure S4B). These exons were present in genes of varying sizes, from 18.3 to 702 kb. Of the candidate exons we selected, the HiR targets tended to be located within larger genes, but this difference was not statistically significant. Like wise, e xon positions within each gene varied widely, from the second exon to the 29th. Sizes of the affected exons and their flanking introns also showed no clear pr efer ence, with affected exons being as short as 11 bp and as long as 1193 bp, and flanking introns as short as 40 bp and as long as 61.9 kb (Supplementary Figure S4B). We also compared the strengths of the 3 ss and 5 ss of the exons in question. The mean scores of the constituti v e 3 ss and 5 ss are 7.9 and 8.1, respecti v ely. HiR targets had mean 3 ss and 5 ss scores of 3.9 and 7.0, respecti v ely. Most targets had either a weak 3 ss or a weak 5 ss, but rarely both (Supplementary Figure S4B). In contrast, HiB targets had the mean 3 ss score of 6.6 and the mean 5 ss score of 7.0. Interestingly, for exons whose inclusion was promoted by both drugs, the mean 3 ss score was 7.9, the same as for constituti v e e xons, while the mean 5 ss score was 6.1 (Supplementary Figure  S4B). This suggests that the functional overlap of the compounds may be centered around correction of a weak 5 ss. Notably, 8 of the 10 top risdiplam-only target exons had a G at the last position, whereas in the top branaplam-only targets, 3 had a G, 1 had a U and 6 had an A residue. Strikingly, of the top 10 targets of both compounds, all 10 had an A at the last position of the exon, suggesting that this position is important for the effect of both drugs ( Supplementary Figur e S4B). Ther e was also a striking conservation of an A residue at the third intronic position, which was shared by most targets of branaplam, but not of risdiplam.
Employing semi-quantitati v e RT-PCR followed by analysis of the amplified products in a gel-based assay, we performed independent validations of nine EIN e v ents captured by RNA-Seq. We first analyzed EIN e v ents for three coding e xons: MBNL1 e xon 5, TT C28 e xon 20 and EXOC1 exon 12 shown to be impacted by HiR but not by HiB (Figure 3 B, lower panel). The MBNL1 protein plays an important role in RNA metabolism and, inclusion of MBNL1 exon 5, which codes for a nuclear location signal, could be critical for general splicing regulation ( 82 ). Sequestration of MBNL1 in the nucleus by repeat-containing transcripts is associated with myotonic dystrophy ( 83 ). Both the 3 ss and 5 ss of MBNL1 exon 5 are quite weak, unlike most targets of risdiplam ( Figure 3 A; Supplementary Figure S4B). Like most targets of risdiplam alone, the last position of MBNL1 exon 5 is a G residue. The proteins coded by TTC28 and EXOC1 regulate the mitotic cell cycle and exocytosis, respecti v ely. Aberrant e xpression of TT C28 and EXOC1 is associated with melanoma and platelet secretion defects, respecti v ely ( 84 , 85 ). Both TTC28 exon 20 and EXOC1 exon 12 have a weak 3 ss and strong 5 ss (Figure 3 A). The last position of TTC28 exon 20 is a G, while the last position of EXOC1 exon 12 is an A. Our PCR results confirmed the stimula tory ef fect of InR and HiR on inclusion of MBNL1 e xon 5, TT C28 e xon 20 and EXOC1 exon 12 (Figure 3 B, lower panel). The effect of risdiplam was particularly robust on splicing of shorter exons, for example TTC28 exon 20 and EXOC1 exon 12. Of note, the fact that risdiplam promoted inclusion of MBNL1 exon 5 and TTC28 exon 20 despite the absence of an A nucleotide at the last position of these exons suggests a different mechanism of risdiplam action from the one proposed for modulation of SMN2 exon 7 splicing. As expected, none of the branaplam concentrations used had any significant effects on splicing of MBNL1 exon 5 and TTC28 exon 20 (Figure 3 B).
We next validated the splicing of coding exons of POMT2 , STRN3 and FOXM1 shown to be affected by both compounds according to our RNA-Seq analysis (Figure 3 C). POMT2 , STRN3 and FOXM1 code for Omannosyltr ansfer ase, a signaling protein and a transcription acti vator, respecti v ely. Mutations and / or aberrant splicing of all three genes are associated with pathogenic conditions, including muscular dystrophy and cancer (86)(87)(88). All three exons in question have a nearly identical 5 ss, with an A at the last exonic position and a canonical GUAAGU intronic sequence (Figure 3 A). Their 3 ss varies in strength: POMT2 exon 11B and FOXM1 exon 9 have a relati v ely strong 3 ss, while that of STRN3 exon 8 is weaker. FOXM1 exon 9 contains a second, internal 5 ss (9S) that is also predicted to respond to risdiplam and branaplam treatment. This splice site shares many sequence characteristics with the others, namely an A at the last exonic position and a GUAAGU intronic sequence. Supporting the findings of RNA-Seq, the results of PCR showed an increase in inclusion of POMT2 exon 11B in the presence of risdiplam and branaplam in a dose-dependent manner (Figure 3 C). Of note, POMT2 exon 11B is a previously unannotated exon, whose identity we confirmed by sequencing. While analysis of RNA-Seq re v eals stimulation of STRN3 exon 8 inclusion by HiR and HiB, no information could be deduced about co-skipping e v ents involving STRN3 exons 8 and 9 (  can be included in two isoforms, short (9S) and long (9L), due to the alternati v e 5 ss usage. Branaplam caused inclusion of both isoforms in a dose-dependent manner, while risdiplam primarily induced usage of the downstream 5 ss to include the 9L isoform. In the case of HiB, we also captured inclusion of an unannotated downstream exon. We confirmed its identity by sequencing and termed it 9B (Figure 3 C). Our findings provide new information about coskipping e v ents compared with earlier studies that analyzed the of f-target ef fects of risdiplam and branaplam on splicing of exons 8 and 9 of STRN3 and exon 9 of FOXM1 by RNA-Seq or qPCR ( 38 , 48 ). Of note, a gel-based method remains the only reliable technique to accurately quantify co-skipping of two or more exons in a single lane.
We examined three coding exons that showed enhanced inclusion by HiB but not by HiR. These were KDM6A exon 28, ANXA11 exon 16B and SH3YL1 exon 11 (Figure 3 D, upper panel). Proteins encoded by KDM6A , ANXA11 and SH3YL1 regula te demethyla tion, calcium-dependent phospholipid binding and phospha tid ylinositol biosynthesis, respecti v el y. Patho genic conditions associated with aberrant expression of KDM6A , ANXA11 and SH3YL1 include pediatric cancer, ALS and diabetic nephropathy, respecti v ely (89)(90)(91). All three exons had a somewhat weaker than average 5 ss, while the 3 ss of KDM6A exon 28 and that of SH3YL1 exon 11 were quite strong. In contrast, ANXA11 exon 16B had a weak 3 ss (Figure 3 A). Both KDM6A exon 28 and ANXA11 exon 16B ended with A residues, while the last position of SH3YL1 was a G. Our PCR results confirmed stimulation of KDM6A exon 28 inclusion by InB and HiB (Figure 3 D, lower panel). HiR also stimulated inclusion, but to a much lesser extent. In addition, InB and HiB triggered a significant retention of KDM6A intron 28. According to the RNA-Seq analysis, HiB had a complex effect on processing of the ANXA11 exons 16L and 16B (

Analysis and validation of e x on skipping events triggered by risdiplam and branaplam
We examined the sequence context of the splice sites for 30 affected exons whose skipping was increased by HiR and / or HiB (Supplementary Figure S5). As compared with the average 147 nt size of an internal exon, the median sizes of exons undergoing increased skipping triggered by HiR only, HiB only or both were relati v ely short, at 55, 104 and 80 nt, respecti v ely. Howe v er, intronic sequences flanking these skipped exons did not show any size pr efer ence. In general, skipped exons tended to have a slightly weaker than average 3 ss. The average 3 ss str engths wer e 6.7, 6.7 and 6.3 for exons affected by HiR only, HiB only and both treatments, respecti v ely. We noted an interesting trend for the 5 ss of exons undergoing a compound-induced increase in skipping. Although the 5 ss of exons affected by HiR and HiB treatment only were relati v ely strong on average (average scores of 8.0 and 9.1, respecti v ely), e xons whose skipping was increased by both HiR and HiB tended to have a very weak 5 ss, with an average score of 4.2 (Supplementary Figure  S5). Unlike exons with increased inclusion after treatment with these compounds, there was no preference for an A residue at the last exonic position.
We employed PCR to validate r epr esentati v e ESK e v ents for coding exons of three genes: DST , TEAD1 and THOC5 . Their splicing was impacted by risdiplam but not by branaplam. Proteins coded by DST , TEAD1 and THOC5 regulate microtubule organization, transcription and RNA metabolism, respecti v ely. Aberrant e xpression and / or splicing of DST (also known as BPAG1 ), TEAD1 and THOC5 are associated with neurological disorders and cancer (92)(93)(94). DST exon 92 has a relatively weak 3 ss and 5 ss, while TEAD1 exon 5B and THOC5 exon 10 have a weak 3 ss, but a strong 5 ss. Like most candidates undergoing ESK e v ents, all three have a G at the last exonic position. Supporting the findings of RNA-Seq, PCR results showed enhanced skipping of DST exon 92, TEAD1 exon 5B and THOC5 exon 10 in the presence of InR and HiR (Figure 4 B, lower panel). Although not initiall y a pparent from RNA-Seq analysis, the results of PCR also re v ealed a small but noticeable stimula tory ef fect of HiB on splicing of THOC5 exon 10 (Figure 4 B). We confirmed ESK events of three coding exons triggered by both HiR and HiB: ODF2L exon 14, MAST2 exon 8 and KIF23 exon 8 (Figure 4 C, upper panel). Proteins coded by ODF2L , MAST2 and KIF23 regula te cilium assembly, peptid yl-serine phosphoryla tion and transport of organelles during cell division, respecti v el y. Patho genic conditions associated with aberrant expression of ODF2L , MAST2 and KIF23 include ciliopathy, li v er cancer and chronic myeloid leukemia, respecti v ely (95)(96)(97). Both ODF2L exon 14 and KIF23 exon 8 have a weak 3 ss and 5 ss, while MAST2 has a strong 3 ss and a slightly weaker than average 5 ss. The last exonic position was G in the case of ODF2L and MAST2 , while KIF23 exon 8 had a U residue. The PCR results showed that InR and HiR produced a nearly equal effect on splicing of ODF2L exon 14, suggesting a lower risdiplam concentration threshold for maximum inhibitory effect that was limited to ∼50% skipping of ODF2L exon 14 (

Analysis and validation of alternative splice site usage triggered by risdiplam and branaplam
We examined the sequence context of the top 15 A5S e v ents affected by HiR and / or HiB (Supplementary Figure S6A). For A5S e v ents affected by risdiplam only, all of the 5 ss with increased usage had lower scores than the 5 ss with decreased usage, suggesting that risdiplam may be acting by directly strengthening a weak 5 ss. We only observed a Nucleic Acids Research, 2023, Vol. 51, No. 12 5961 single 5 ss with an A residue at the last position of the exon, and its usage was decreased by risdiplam treatment. There was less consistency among sequences when 5 ss usage was influenced by both compounds. In particular, in three cases, the two alternati v e splice sites were of similar str ength, wher eas in the other two the splice site with increased usage after treatment was the stronger of the two. Two of the 5 ss with increased usage had an A residue at the last exonic position and three had the canonical G residue. Among the top fiv e A5S e v ents affected by branaplam only, there was no clear trend in regard to the relati v e strength of the alternati v e 5 ss, but all fiv e 5 ss whose usage was promoted by HiB had A residues at the last exonic position.
We selected r epr esentati v e coding e xons of fiv e genes, namely NCOR2 , NFATC4 , CLEC16A , CNOT1 and PAXBP1 , to validate by PCR the A5S e v ents identified by RNA-Seq ( Figure 5 A-D). NCOR2 , NFATC4 and PAXBP1 code for transcription-associated factors and their d ysregula tions are linked to Crohn's disease, carcinogenesis and myopathic hypotonia, respecti v ely (101)(102)(103). CLEC16A and CNOT1 code for autophagy-and RNA turnover-related factors, and their dysr egulations ar e associated with Parkinson's disease and impaired neurological de v elopment, respecti v ely ( 104 , 105 ). NCOR2 e xon 47 has two competing 5 ss that are 138 nt away from each other. Both splice sites are weak, and neither harbors an A residue upstream of the GU (Figure 5 A). Consistent with analysis of RNA-Seq, the results of PCR showed increased usage of the downstream 5 ss upon treatments with InR and HiR (Figure 5 B, lower panel). As expected, the results of PCR did not capture any noticeable effect of branaplam on the 5 ss usage of NCOR2 exon 47. In the case of NFATC4 , two competing 5 ss of exon 10 are separated from each other by 324 nt. Usage of the downstream 5 ss, which is strong, is reduced in favor of usage of the upstream weak 5 ss ( Figure  5 A, B). In addition to alternati v e usage of the 5 ss of exon 10, analysis of RNA-Seq re v ealed retention of intron 10 (Figure 5 B, upper panel). This suggests that risdiplam may function by weakening the downstream 5 ss, rather than str engthening the upstr eam 5 ss. Supporting the findings of RNA-Seq, PCR results confirmed increased usage of the upstream 5 ss caused by risdiplam in a dose-dependent manner (Figure 5 B, lower panel). Although not re v ealed by RNA-Seq, the results of PCR also showed an increase in usage of the downstream 5 ss of exon 10 of NFATC4 in the presence of branaplam (Figure 5 B, lower panel). Here again, selection of the 5 ss in response to treatments was not determined by the identity of the nucleotide at the last exonic position, as it is a G residue in both 5 ss. PCR did not re v eal any appreciab le change in the retention of intron 10 of NFATC4 by any treatment, although these results should be interpreted with caution as longer introncontaining transcripts are likely to be poorly amplified. We analyzed the sequence context of the top 15 A3S e v ents that were affected by HiR and / or HiB according to our RNA-Seq analysis (Supplementary Figure S6B). We could not discern any common sequence feature as a determinant of A3S e v ents, suggesting that the mode of compound action may be complex and / or indirect. We selected r epr esentati v e coding e xons of six genes, namely RPL22L1 , DOCK7 , ELMO2 , SREK1 , ELF2 and SART3 , to validate A3S e v ents identified by RNA-Seq (Figure 5 E). RPL22L1 , SREK1 and SART3 code for RNA-binding proteins, and their aberrant expression is linked to carcinogenesis (106)(107)(108). Dysregulation of ELF2 , a transcription factor-coding gene, is also associated with carcinogenesis ( 109 ). Aberrant expression of DOCK7 that codes for a guanine nucleotide exchange factor is associated with epileptic encephalopathy ( 110 ). ELMO2 codes for a phagocytosis-associated protein, and homozygous mutation in ELMO2 has been linked to Ramon syndrome ( 111 ). RPL22L1 exon 3 has two alternati v e 3 ss separated from each other by 67 nt. The upstream 3 ss, usage of which is increased in risdiplam treatment, is stronger than average, while the canonical downstream 3 ss is quite weak (  dose-dependent manner, with the strongest effect produced by HiB (Figure 5 G, lower panel). ELF2 exon 7 has two alternati v e 3 ss separated from each other by 36 nt ( Figure  5 H, upper panel). PCR validation confirmed branaplaminduced usage of the upstream strong 3 ss over the downstream weaker 3 ss in a dose-dependent manner (Figure 5 H, upper panel). SART3 exon 5 has two alternative 3 ss that are 54 nt apart (Figure 5 H, lower panel). Both are somewhat weaker than average. Results of PCR validated HiBinduced inclusion of the longer form of exon 5 of SART3 , confirming that usage of the upstream 3 ss was stimulated by HiB (Figure 5 H, lower panel).

Analysis and validation of differential intron retention triggered by risdiplam and branaplam
We examined the characteristics of introns with IRM e v ents triggered by risdiplam and / or branaplam. There were no clear trends among the splice sites of IRM introns regarding their sequences or strengths (Supplementary Figure S7A). Somew hat surprisingl y, onl y one IRM intron identified by RNA-Seq is flanked by a 5 ss with an A at the last exonic position. The majority of affected introns were relati v ely short ( < 1 kb), although this is not unexpected since most longer retained introns are likely to be degraded by RNA surveillance mechanisms. Using qPCR, we validated the IRM e v ents of six r epr esentati v e introns identified by RNA-Seq (Figure 6 A-D). In order to avoid ambiguities of expression, we normalized each intron retention value using a constituti v ely spliced region of the same gene. We also performed standard qPCR normalization against a housekeeping gene and compared changes in total transcript levels with changes in intron-retained transcripts, along with untreated control and intermediate concentrations of the compounds (Supplementary Figure S7).
Genes used for validation of IRM e v ents were RBM5 , SPOCD1 , PNISR , NUDT22 , HNRNPM and NFATC4 . RBM5 , PNISR (also known as SFRS18 ) and HNRNPM code for RNA-binding proteins, and d ysregula tions of these genes are associated with carcinogenesis, psoriasis and cogniti v e deficits, respecti v ely (112)(113)(114). SPOCD1 and NFATC4 code for a transcription factor and a transcription-associated factor, respecti v ely. Aberrant e xpression of these genes is linked to glioma and oncogenesis ( 102 , 115 ). NUDT22 codes for a protein with DP-sugar diphosphatase activity and metal ion binding activity; howe v er, the pathological significance of NUDT22 has not yet been established ( 116 ). The results of qPCR confirmed a significant reduction in retention of RBM5 intron 6 by HiR (Figure 6 B). Of note, this is coupled with an increase in RBM5 total transcript le v els (Supplementary Figure S7D). RBM5 exon 6 is longer than most targets of IRM e v ents at ∼2.5 kb, and its 5 ss is quite weak (Figure 6 A; Supplementary Figure S7A). Validating the findings of RNA-Seq, the results of qPCR showed concentration-dependent reduction in SPOCD1 intron 12 retention triggered by risdiplam (Figure 6 B). The 5 ss of SPOCD1 intron 12 is of average strength, while its 3 ss is str ong. PNISR intr on 10 has a somewhat weak 5 ss and a very weak 3 ss. Typical of other IRM targets, PNISR intron 10 is > 1 kb in length (Figure 6 A). qPCR results confirmed a decrease in PNISR in-tron 10 retention by HiR and HiB (Figure 6 C). InB also reduced PNISR intron 10 retention, but to a lesser degree. Interestingly, LoR and InR produced an increase in PNISR intron 10 retention. The opposite effects between high and low concentrations of risdiplam could be due to secondary ef fects tha t might come into play at high risdiplam concentration. NUDT22 intron 1 does not have any exceptional features, except for its small size (Figure 6 A). In line with the findings of RNA-Seq, the results of qPCR showed reduced retention of NUDT22 intron 1 by InR, HiR, InB and HiB (Figure 6 C). The results of qPCR also validated HiB-induced reduction in HNRNPM intron 6 retention.
Howe v er, we observ ed the opposite effect on retention of this intron in the presence of InR, HiR and LoB ( Figure  6 D). These differences could be attributed to secondary effects caused by a high concentration of branaplam. Supporting the findings of RNA-Seq, qPCR validations confirmed the dose-dependent reduction in the retention of NFATC4 intron 10 by branaplam (Figure 6 D). In contrast, qPCR results showed an increase in NFATC4 intron 10 retention by risdiplam in a dose-dependent manner ( Figure  6 D). Of note, NFATC4 intron 10 uses two alternati v e 5 ss, usage of which is regulated in contrasting manners by risdiplam and branaplam ( Figure 5 ). Since both intron retention and 5 ss usage are regulated in opposite manners by the two compounds, we hypothesize that the two e v ents are linked.
Similar to IRM, IRT e v ents caused by risdiplam and / or brana plam involved mostl y short introns, with onl y two exceptions of introns being longer than 1 kb (Supplementary Figure S7B). In addition, introns whose retention was increased by branaplam tended to have a somewhat weak 5 ss, but other than that we observed no clear trends regarding sequence composition or splice site strengths for the affected introns. There was little overlap between IRT e v ents triggered by risdiplam and branaplam treatments (Figure 1 J); those e v ents that we identified were not strongly affected compared with those that were triggered by risdiplam only or branaplam only (Supplementary Figure  S7B). Using qPCR, we validated the IRT e v ents for OGT , C8orf33 , DGKA and SMTN genes that code for glycosyltr ansfer ase, an unknown protein, diacylglycerol kinase and a structural protein, respecti v ely (Figure 6 E). Expression of OGT , C8orf33 , DGKA and SMTN has been correlated with cardiovascular disease, hepatocellular carcinoma, acute myeloid leukemia and hypertension, respecti v ely (117)(118)(119)(120). OGT intron 4 stands out among other IRT e v ents due to its relati v ely long size and high le v el of baseline r etention (Figur e 6 E, F). It has a strong 5 ss, but its 3 ss is rather weak. The results of qPCR confirmed the increase in OGT intron 4 retention by risdiplam at all concentrations, with the effect being most pronounced at HiR (Figure 6 F, lower panel). LoB and InB also caused a small but significant increase in OGT intron 4 retention. Howe v er, HiB showed no significant effect on the splicing of OGT intron 4. The concentr ation-dependent contr adictory effects could be due to multiple factors including secondary effects that are likely to become more prominent at the high concentr ation of br anaplam. C8orf33 intron 1 is quite short at 88 nt, with a slightly weaker than average 5 ss and a strong 3 ss (Figure 6 E). Consistent with the results of RNA-Seq,  Figure S7H).

Mechanism of risdiplam and branaplam action captured by minigenes
To uncover if the effects of the small molecules were directly linked to motifs located within the affected exons and their flanking intronic sequences, we employed hybrid exon trapping cassettes using an SMN2 minigene (p SMN2 I6) as the 'backbone'. We generated 11 hybrid minigenes by replacing SMN2 exon 7 and its flanking intronic sequences with an exon of interest and its flanking intronic sequences (Supplementary Figure S8). These minigenes harbored EXOC1 exon 12, TTC28 ex on 20, TMEM50B ex on 5B, SH3YL1 exon 11, POMT2 exon 11B, STRN3 exon 8, FOXM1 exon 9, SLC25A17 exon 3, MADD exon 21, ARHGAP12 exon 17 and ATG5 exon 3. Of note, the offtarget effects of risdiplam and / or branaplam on splicing of STRN3 exon 8, FOXM1 exon 9, SLC25A17 exon 3, MADD exon 21 and ATG5 exon 3 have been also validated independently ( 38 , 48 ). Splicing of exonic sequences examined in this study provided di v erse structural conte xt (Supplementary Figures S9-S12). To determine a potential overlapping mechanism of splicing modulation by risdiplam and brana plam, we created m utant minigenes in w hich the nucleotide at the last exonic position was substituted with a C residue. For a few minigenes, we also tested the effect of other mutations, exonic and intronic, on the ability of the small molecules to affect splicing. To ensure efficient transfection, we performed all hybrid minigene-based experiments in HeLa cells. Of note, the effects of HiR and HiB on splicing of the endogenous genes in HeLa cells were largely similar to those observed in GM03813 fibroblasts (Supplementary Figure S8). Due to the facts that minigene e xpression was dri v en by a strong constituti v e promoter, and the minigene context was heterologous, we did observe some differences in the le v el of e xon inclusion / skipping between the minigenes and the endogenous genes. Importantly, the minigenes responded to the treatments with the small molecules, allowing interrogations into the mechanism of action of the compounds.
We observed HiR-induced enhanced inclusion of EXOC1 exon 12 and TTC28 exon 20 in tr anscripts gener ated from the wild-type minigenes (Figure 7 A). Supporting the findings for SMN2 exon 7, an A-to-C substitution at the last position of EXOC1 exon 12 reduced the ability of risdiplam to promote inclusion of the exon, although this substitution itself had an inhibitory effect on splicing (Figure 7 A). In the case of the TTC28 minigene, risdiplam promoted wild-type exon 20 inclusion despite the nucleotide at the last position of this exon being G (Figure 7 A). Changing G to C abolished exon 20 inclusion, with and without risdiplam. Interestingly, this G-to-C substitution also activated a downstream cryptic 5 ss (Cr), usage of which was somewhat promoted by branaplam (Figure 7 A). We observed a HiB-linked increase in inclusion of TMEM50B exon 5B and SH3YL1 exon 11 in transcripts generated from the minigenes (Figure 7 B). In the case of TMEM50B exon 5B, a U-to-C substitution at the last exonic position resulted in increased exon inclusion. Importantly, the ability of branaplam to promote inclusion of the mutated TMEM50B exon 5B was retained (Figure 7 B). These results supported that the presence of a specific nucleotide at the last exonic position is not a critical factor for the branaplam effect on TMEM50B exon 5B splicing. Similar r esults wer e obtained for SH3YL1 exon 11. Inter estingly, a G-to-C substitution at the last position of SH3YL1 exon 11 caused predominant skipping of exon 11, yet HiB was able to effecti v ely r estor e inclusion of the mutated ex on; ho we v er, one of the two cryptic 5 ss was used in this case (Figure 7 B). The SH3YL1 sequence contains two AGU motifs near the 5 ss of exon 11 which could serve as targets for branaplam. Mutation of the A residue within both motifs to C (mutant M2) fully r estor ed SH3YL1 exon 11 (Figure  7 B). While not useful for assessing the mechanism of the branaplam effect on splicing, SH3YL1 mutant M2 underscored the role of residues outside of the 5 ss in determining the outcome of exon 11 splicing. It is possible that the inter action of br anaplam with regions away from the 5 ss may have a similar effect to substitutions used in SH3YL1 mutant M2.
Using minigenes, we were able to capture the stimulatory effect of both HiR and HiB on splicing of POMT2 exon 11B, STRN3 exon 8, FOXM1 exon 9, SLC25A17 exon 3 and MADD exon 21 (Figure 7 C). In most instances, branaplam produced a stronger stimulatory effect than risdiplam, suggesting differences in the mechanisms of their action. Yet, ther e wer e many similarities between risdiplam and branaplam with respect to how exonic and intronic mutations impacted the ability of these small molecules to influence splicing. In the case of POMT2 , an A-to-C substitution at the last position of exon 11B resulted in complete loss of the ability of HiR and HiB to promote exon 11B inclusion (Figure 7 C). Interestingly, another single A-to-C substitution introduced at the 10th position of POMT2 intron 11B (mutant M2) actually increased the ability of both compounds to promote exon 11B inclusion (Figure  7 C). These results underscored the critical role of a few intronic nucleotides immediately downstream of the 5 ss in the effect of both compounds on splicing. We noted that   Figure S4). Mutation of this third position to G promoted skipping of POMT2 exon 11B and suppressed the ability of the compounds to promote exon 11B inclusion, as did a mutant harboring all three mutations (Figure 7 C).
In the case of STRN3 exon 8, an A-to-C substitution at the last exonic position suppressed the ability of HiR and HiB to increase inclusion of this exon. An A-to-G mutation at the third position of intron 8 of STRN3 triggered near total skipping of STRN3 exon 8 and se v erely impacted the ability of the compounds to promote exon inclusion (Figur e 7 C). Inter estingl y, this intronic m utation had different consequences for the risdiplam and branaplam treatments, as only HiB but not HiR produced a small but noticeable increase in STRN3 exon 8 inclusion (Figure 7 C, mutant M2). We observed somewhat complex results for the impacts of HiR and HiB on splicing of FOXM1 exon 9 that has an additional 5 ss within exon 9; usage of this splice site generates a shorter exon 9 isoform (exon 9S). While HiR promoted inclusion of both short and long isoforms of exon 9 by suppressing the skipping of exon 9, HiB pr efer entially promoted inclusion of exon 9S (Figure  7 C). An A-to-C substitution at the last position of the long exon 9 reduced skipping of exon 9 in favor of the inclusion of the long isoform of exon 9 (exon 9L). Interestingly, HiR promoted inclusion of both 9S and 9L in the mutated exon 9. In contrast, HiB pr efer entially promoted inclusion of FOXM1 exon 9S in the presence of the substitution (Figure 7 C). These results provided a clear example in which the role of the last exonic position turned out to be more critical for the action of branaplam than of risdiplam. HiR and HiB promoted inclusion of multiple isoforms of SLC25A17 exon 3. An A-to-C substitution at the last position of SLC25A17 exon 3 substantially reduced, but did not completel y abro gate, the stim ula tory ef fect of HiR and HiB on splicing of this exon. Similar r esults wer e obtained in the case of an A-to-C mutation at the last position of MADD exon 21.
We analyzed two skipping e v ents induced by HiB only: splicing of ARHGAP12 exon 17 and ATG5 exon 3. A Gto-C substitution at the last position of ARHGAP12 exon 17 triggered almost total skipping of this exon (Figure 7 D). Ther efor e, w e w er e unable to captur e the inhibitory effect of HiB on this mutant. We noted the presence of a 5 sslike sequence early in exon 17 of ARHGAP12 that has an A r esidue upstr eam of the GU dinucleotide. We hypothesized that this could serve as a target for branaplam to induce exon skipping. Introducing an A-to-C substitution at this position did not have an impact on splicing in the absence of branaplam. Yet, this exonic mutation fully suppressed the inhibitory effect of HiB on splicing of ARHGAP12 exon 7 (Figure 7 D). In the case of ATG5 exon 3, a G-to-C substitution at the last exonic position of exon had a limited effect on exon 3 splicing. Howe v er, this mutation enhanced the inhibitory effect of HiB on its splicing (Figure 7 D). An additional A-to-C substitution in exon 3 of ATG5 at a 5 sslike sequence similar to the one in ARHGAP12 exon 17 did not have an impact on splicing in the control condition. Yet, this exonic mutation fully suppressed the inhibitory effect of HiR on splicing of ATG5 exon 3 (Figure 7 D).
Inclusion of MBNL1 exon 5 is promoted by risdiplam but not branaplam in both GM03813 fibroblasts and HeLa cells (Figure 3 ; Supplementary Figure S8). Similar to SMN2 e xon 7, MBNL1 e xon 5 is 54 nt long. Howe v er, unlike SMN2 exon 7 that has an A residue at the last exonic position, MBNL1 exon 5 has a G residue at this position. Hence, MBNL1 exon 5 could provide a useful model for uncovering the role of critical exonic positions / motifs in risdiplaminduced splicing changes without the domina ting ef fect of the last exonic position in the context of a 54 nt long exon. The initial MBNL1 hybrid minigene showed an aberrant pattern of splicing. Hence, we generated an MBNL1 minigene encompassing the entire MBNL1 genomic sequence from exon 4 through exon 6. In HeLa cells, the splicing pattern of this newly created minigene showed nearly 100% inclusion of exon 5, making the minigene useless for studying the stimulatory effect of HiR. On the other hand, in GM03813 fibroblasts, MBNL1 exon 5 skipping in transcripts expressed from the minigene constituted 30-40%. Ther efor e, to perform our MBNL1 minigene-related studies, we used GM03813 cells. As expected, HiR but not HiB increased inclusion of MBNL1 exon 5 in transcripts generated from the minigene (Figure 7 E, lanes 1-3). A G-to-C substitution at the last position of MBNL1 exon 5 resulted in complete skipping of this exon and fully suppressed the effect of risdiplam (Figure 7 E, mutant M7). A recent work has shown a critical role for GA-rich sequences in risdiplam-mediated inclusion of SMN2 exon 7 ( 46 ). To specifically examine the effect of exonic A and G residues on HiR-induced inclusion of MBNL1 exon 5, we made six additional minigene mutants. Risdiplam retained some ability to stimulate exon 5 inclusion in tr anscripts gener ated from all six of these mutants, but to a varying degree . Mutants M1 and M2 incorporated four G-to-A and A-to-G mutations, respecti v ely, in the 5 portion of exon 5, and still retained a strong ability to respond to risdiplam treatment (Figure 7 E). Risdiplam affected exon 5 splicing the least in mutants M4 and M5. M4 contained three G-to-A substitutions from position 24 to 29, while M5 contained four Ato-G substitutions in the region from position 38 to 43 (Figure 7 E). Our results indicate that both A and G residue(s) within the middle and 3 portion of MBNL1 exon 5 might be r equir ed for risdiplam-induced inclusion of this exon. Howe v er, mutant M6 that carried four G-to-A substitutions in this exon still retained a significant ability to respond to risdiplam (Figure 7 E). These substitutions led to complete elimination of all G residues in the region between position 30 and position 53 of exon 5.

Context-specific role of cis-elements in pr efer ential splice site selection by small molecules
We chose FOXM1 exon 9 as a model to evaluate the role of cis-elements away from the 5 ss in splicing modulation by small molecules. FOXM1 exon 9 contains two competing 5 ss separated from each other by 80 nt. Usage of the upstr eam and downstr eam 5 ss r esults in inclusion of the short (9S) and long (9L) isoforms of exon 9, respectively. Both risdiplam and brana plam substantiall y reduced FOXM1 exon 9 skipping by promoting the usage of both 5 ss, with risdiplam pr efer entially promoting usage of the downstream 5 ss and branaplam that of the upstream 5 ss (Figure 3 C).
We generated a pool of FOXM1 minigenes in which a 10 nt long region from positions 70 to 79 of exon 9L was randomized. This exonic region is located roughly halfway between the 5 ss of exons 9S and 9L. Ther efor e, this r egion would be positioned within the intronic sequence if the upstr eam 5 ss wer e to be used for inclusion of exon 9S. This allows us to sim ultaneousl y probe the role of both exonic and intronic cis-elements in the effects of the small molecules on the usage of a specific 5 ss. We randomly selected 20 clones from the pool of minigenes. Our selected clones had a minimum and maximum of four and nine substitutions within the 10 nt str etch, r especti v ely (Figure 8 A). The wild-type sequence of the 10 nt region we randomized contains fiv e cytosine, four adenosine and one uridine residue. Hence, our approach also evaluated the effect of the loss of the potential small molecule-responsi v e AC-rich motif(s) in the 5 ss selection. In order to capture subtle differences between the mutant minigenes, we utilized intermediate concentrations of both drugs (InR, 250 nM; InB, 10 nM) and a shorter treatment time (6 h). We examined the effect of risdiplam and branaplam on three splicing e v ents, namely skipping of exon 9 and inclusion of exons 9S and 9L (Figure 8 B). Gi v en the large number of substitutions generated by randomization, we expected some mutations to create cryptic splice sites. Hence, our experimental design also offered an opportunity to ask how small molecules might impact the usage of two competing 5 ss in the presence of a cryptic splice site in between.
We first measured the effect of InR and InB on transcripts generated from the wild-type FOXM1 minigene. Both InR and InB pr efer entially promoted inclusion of exon 9S, although a small stimulatory effect on inclusion of exon 9L was also captured (Figure 8 B). Supporting the compoundspecific differences, the stimulatory effect on inclusion of exon 9L was somewhat stronger in the case of InR than InB. These results are consistent with those observed with 24 h treatment with HiB and HiR (Figure 7 C). We next analyzed the splicing pattern of 20 randomly selected mutant minigenes. Supporting the critical role of motifs away from the 5 ss, most mutants showed splicing pattern distinct from those observed with the wild-type minigene (Figure 8 B). We identified eight mutants (U3, U8, U11, U15, U18, U22, U23 and U24) that displayed predominant ( > 80%) skipping of exon 9 and barely detectable levels of exon 9L inclusion. Some of these mutants also sho wed lo w but detectable levels of inclusion of exon 9S. While most of these mutants had lost CC-and / or AC-rich motifs, the nature of substitutions varied widely (Figure 8 A). Three additional mutants (U6, U12 and U13) showed > 60% skipping of exon 9 and differ ent degr ees of inclusion of exons 9S and 9L. Here again, the substitutions varied significantly. Mutant U4 that harbored six substitutions showed comparable levels of exon 9 skipping and exon 9S inclusion, accompanied by near total loss of the usage of the downstream 5 ss. Mutant U9 carried eight substitutions and showed a splicing pattern similar to that of the wild-type minigene. Interestingly, exon 9 substitutions in U9 created an 8 nt long pyrimidine-rich motif tha t incorpora ted two wild-type residues immedia tely downstream of the randomized region. Mutant U21 had six substitutions and showed ∼50% exon 9 skipping and ∼27% and ∼23% inclusion of exons 9L and 9S, respecti v el y. This m utant retained the CCA motif from the wildtype sequence as well as acquiring a CCCC motif. In four minigene mutants (U1, U14, U16 and U25), exon 9L was pr efer entially included, accompanied by decreased skipping of exon 9 and usage of the upstream 5 ss. Of note, these FOXM1 minigene mutants had all acquired one or more GA-rich motifs that were absent in the wild-type sequence (Figure 8 A). We also identified two mutants (U2 and U10) in which a cryptic 5 ss (Cr1) immediately upstream of the area of randomization was activated. In both instances, the usage of Cr1 substantially reduced skipping of exon 9. In the case of mutant U2, activation of Cr1 also decreased the usage of both the upstream and downstream 5 ss. In the case of the U10 mutant, activation of Cr1 enhanced usage of the downstream 5 ss, while usage of the upstream 5 ss was slightly reduced. We did not observe a single minigene mutant in which inclusion of exon 9S was the major splicing e v ent.
We compared the effect of InR and InB on splicing of the above mutants. With the exception of U3 and U4, mutants with enhanced skipping of exon 9 showed increased inclusion of exon 9S upon treatment with InR and InB (Figure  8 B). These results suggested that our chosen region of inv estigation was de void of cis-elements associated with the small molecule-induced inclusion of exon 9S. In the case of the U3 mutant minigene, only risdiplam but not branaplam promoted inclusion of exon 9S, underscoring the difference in composition and / or positioning of motifs that respond differently to risdiplam and branaplam (Figure 8 B). Such motifs may fall outside of the 10 nt long sequence that we randomized but may still be responsi v e to trans -factors that ar e r ecruited b y cis-elements or modified b y structures generated as a result of randomization. The results of U4 minigene splicing were rather surprising as both InR and InB reduced the usage of the upstream 5 ss but promoted the usage of the downstream 5 ss and enhanced the skipping of exon 9 (Figure 8 B). Once again, these r esults underscor ed the subtle differences in the mechanism of action of branaplam and risdiplam in selection of a specific 5 ss. The presence of a strong cryptic 5 ss (Cr1) in mutants U2 and U10 completely suppressed the ability of risdiplam and branaplam to promote usage of the downstream 5 ss, but not the upstream one. This is probably due to sequestration of small molecule-responsi v e cis-elements by U1 snRNP recruited at the cryptic 5 ss. It is also possible that the small moleculeresponsi v e cis-elements recruit low affinity binding factors that are easily displaceable by U1 snRNP recruited at Cr1.
One of the major findings of our FOXM1 mutant minigene screening was the disproportionate usage of the two 5 ss in the presence of small molecules. For example, a small molecule-dependent increase in the usage of the downstream 5 ss was always accompanied by an increase in the usage of the upstream 5 ss. Howe v er, an increase in the usage of the upstream 5 ss did not always result in an increase in the usage of the downstream 5 ss. This could be due to the loss of one of the small molecule-responsi v e motifs that favor the usage of the downstream 5 ss. It is also possible that factors recruited by cis-elements created by substitutions directly or indirectly inhibit the usage of the downstream 5 ss. To identify small molecule-responsi v e motifs that e xclusi v ely fav or the usage of the do wnstream 5 ss, we performed an in vivo selection experiment. This experiment also allowed us to test whether small molecule-responsi v e intronic motifs that favor usage of the upstream 5 ss could be fully eliminated. We transfected HeLa cells with the pool of FOXM1 exon 9 minigenes containing the 10 randomized bases (see above). We then treated cells with InR or InB for 6 h and collected total RNA. Following RT-PCR of minigene-deri v ed transcripts, we isolated the PCR fragment corresponding to the exon 9L-included product. We then used two-stage PCR to regenerate the full-length FOXM1 minigene pool, which was used for the next round of selection (Figure 8 C). We performed four rounds of selection for exon 9L inclusion in total (Figure 8 D; Supplementary Figure S13). Over the first three rounds, we observed a steady increase in the proportion of the exon 9L band in all thr ee tr eatments (CTR, InR and InB) (Supplementary Figure S13). By round 4, exon 9L inclusion had stopped increasing, indica ting tha t we had reached the limits of selection. We then analyzed 48 randomly chosen FOXM1 minigenes from pool 4 selected in the presence of InR and InB by sequencing and determining their splicing pattern (Supplementary Tables S2 and S3). Confirming that the selection was nearly complete, we observ ed se v eral duplicate sequences in the selected minigene clones. The majority of clones had extremely high inclusion of exon 9L even without the treatment with either small molecule. This suggested that the selection primarily favored small moleculeindependent motifs that strengthened the downstream 5 ss. Consistently, motifs selected in the presence of risdiplam and branaplam were similar in compositions, mostly centered around [CU]GGACC, with some slight differences depending on the parameters chosen for motif identification (Supplementary Table S4). Yet we were able to isolate many minigenes that retained significant skipping of exon 9. We used these minigenes to evaluate the effect of the small molecules on the usage of both the downstream and upstr eam 5 ss. Inter estingly, most of these minigenes showed an increase in the usage of the upstream 5 ss in the presence of InR and InB (Figure 8 D). These results once again supported that small molecule-responsi v e motif(s) involved in the usage of the upstream 5 ss fall outside of the randomized region. We had hoped to select at least one clone with the capability to suppress skipping of exon 9 in addition to specifically promoting usage of the downstream 5 ss.
Howe v er, our selection process appeared to be primarily dri v en by sequence motifs that promoted inclusion of exon 9L e v en in the absence of risdiplam and branaplam. Our results suggest that the selection for a particular FOXM1 exon 9 splicing event was driven by cis-elements in the small molecule-independent mode. Future experiments will determine if small molecules are capable of suppressing the usage of a competing 5 ss while also pre v enting the skipping of an exon.

Exonic motifs associated with the SMN2 e x on 7 splicing modulation by risdiplam and branaplam
The reported mechanisms of risdiplam stimulation of SMN2 exon 7 inclusion are based on studies with risdiplam analogs C5 and C2. While C5 contacts 54A, an A residue at the last exonic position ( 47 ), C2 interacts with an GArich motif (GAA GGAA GG) located in the middle of the exon ( 45 , 46 ) (Figure 9 ). It has been proposed that 54A and the GA-rich motif serve as the primary and secondary binding sites of risdiplam ( 46 ). Currently, it is not known if additional exonic motifs may contribute to risdiplaminduced inclusion of SMN2 exon 7. The proposed mechanism of branaplam action on SMN2 exon 7 inclusion is based on studies with one of the brana plam analo gs, NVS-SM2, which is shown to stabilize the interaction of U1 snRNP with the 5 ss of SMN2 exon 7 ( 48 ). Currently, it is not known whether the above-mentioned GA-rich motif in the middle of SMN2 exon 7 may also be involved in the branaplam-linked increase in exon 7 inclusion. We hypothesized that additional exonic nucleotides / motifs may play a role in the risdiplam-and branaplam-elicited effect on exon 7 splicing. In search of such motifs, we utilized a library of SMN minigene mutants with partially randomized exon 7 sequences. We used only those mutants that showed predominant skipping of exon 7 (Supplementary Figure S14). In many cases, both compounds failed to stimulate exon 7 inclusion despite the presence of 54A, as well as absence of the inhibitory C6U mutation associated with skipping of exon 7 (Supplementary Figure S14). For example, neither risdiplam nor branaplam was able to promote exon 7 inclusion in the 00-49NS minigene, in which exon 7 contained 54A and the intact GA-rich motif, and did not carry the inhibitory C6U mutation (Supplementary Figure S14). These results implicate the potential involvement of additional exonic sequences and / or factors associated with them in stimulation of SMN exon 7 inclusion by the compounds.
Using in vivo selection of the entire e xon, we pre viously reported thr ee r egions that contribute to SMN exon 7 splicing r egulation. These ar e extended inhibitory tract (Exinct), the conserved tract and 3 -cluster ( 13 ). Exinct and 3 -cluster are negati v e regulators, while the conserved tract is a positi v e regulator of exon 7 splicing ( 13 ). The above-mentioned GArich motif falls within the conserved tract ( Figure 9 ). We tested the effect of HiR and HiB on splicing of SMN2 exon 7 that carried single, double and quadruple substitutions known to promote exon 7 skipping. As previously reported, a single G-to-A substitution at the first exonic position (mutant M1) caused nearly total skipping of SMN2 exon 7 (Figure 8 A, lane 4) ( 13 ). Yet, both compounds were able to rescue inclusion of the exon with this mutation (Figure 8 A, lanes [5][6]. Similarly, the inhibitory effect of U-to-G substitutions at the third and fourth positions of exon 7 on its splicing was overcome by HiR and HiB treatment ( Figure  9 ). As expected, substitutions within the conserved tract increased exon 7 skipping. Most of these mutations also negati v ely impacted the ability of the compounds to promote exon 7 inclusion (Figure 9 ). Mutant M5 is of particular interest, since it carried a double nucleotide substitution located outside of two proposed sites of risdiplam / branaplam action (AC-to-UU substitutions at positions 36 and 37), and yet these substitutions se v erely impacted the ability of both compounds to promote exon 7 inclusion (Figure 9 A, B). Based on these results and a pr evious r eport, we propose that GG residues at the 25th and 26th positions and AC residues at the 36th and 37th positions are critical for small molecule-induced inclusion of SMN exon 7. Previous reports support the role of risdiplam-induced recruitment of FUBP1 and / or KHSRP at the GA-rich motif within exon 7 ( 45 ). It is likely that FUBP1 and / or KHSRP also contact AC residues at the 36th and 37th positions of exon 7. Alternati v ely, it is possible that an additional factor that contacts the AC motif is recruited by FUBP1 and / or KHSRP. Such a mechanism may be m utuall y inclusi v e to the small molecule-mediated recruitment of U1 snRNP at the 5 ss of exon 7 (Figure 9 C).

Effect of combined treatment of risdiplam and branaplam
We examined the effect of the combined treatment of risdiplam and branaplam at two concentrations and at two time points using GM03813 fibroblasts. At 50 nM risdiplam and 2 nM branaplam, we observed a reduction in SMN2 exon 7 skipping from 44% to 30% and 20% at 6 h of treatment, respecti v ely (Figure 10 A). There was an ∼1.6-fold increase in SMN2 exon 7 inclusion when concentrations of compounds were doubled. Similar results were obtained when treatment times were increased from 6 h to 24 h. When 50 or 100 nM risdiplam was combined with 2 or 4 nM branaplam, we observed a dose-dependent increase in the stimula tory ef fect on inclusion of SMN2 exon 7 at both 6 and 24 h of treatment. Howe v er, unlike the outcome of the individual treatments, some of the combined treatments elicited a noticeably stronger response at 24 h in comparison with 6 h (Figure 10 A, right panel). Based on the calculations of the combined effects of the treatments ( 121 ), all 24 h treatments showed additi v e and / or synergistic effects. The strongest synergistic effect was observed at 24 h with 100 nM risdiplam and 4 nM brana plam. Interestingl y, for a gi v en concentration of risdiplam, increasing the concentr ation of br anaplam was more beneficial, showing a greater effect on SMN2 exon 7 inclusion than vice versa. For example, at 50 nM risdiplam, we observed a remarkable ∼1.90fold increase in exon 7 inclusion at 6 and 24 h when the br anaplam concentr ation was doubled. In contr ast, at 2 nM brana plam, we observed onl y a 1.20-and 1.37-fold increase in exon 7 inclusion at 6 and 24 h, respecti v el y, w hen the risdiplam concentration was doubled. We saw a similar trend when the effect of doubling br anaplam concentr ations in the presence of 100 nM risdiplam was compared with that of doubling risdiplam concentration in the presence of 4 nM branaplam. A better outcome of increasing concentrations of branaplam rather than risdiplam in combined treatments could be due to multiple factors including diminishing potential negati v e of f-target ef fects of risdiplam by keeping it low combined with indirect positi v e effects of branaplam on SMN2 exon 7 splicing.
To gain insights into the effects of combined treatments on off-target splicing e v ents, we analyzed the splicing pattern of r epr esentati v e e x ons sho wn to be impacted by risdiplam and / or branaplam, including POMT2 exon 11, MBNL1 exon 5, SH3YL1 ex on 11, ODF2L ex ons 14 and 15, DST exon 92 and ARHGAP12 exon 17 (Figure 10 B-G). Our PCR results confirmed the stimulatory effect of all concentrations of the compounds used except for LoR in the case of POMT2 exon 11B (Figure 3 C). Consistently, at 6 h of treatment, 50 and 100 nM risdiplam showed a negligible effect on splicing of POMT2 exon 11, whereas 2 and 4 nM brana plam noticeabl y stim ulated inclusion of this exon (Figure 10 B). At 6 h, we did not observe any additi v e effect of the combined treatment of compounds. Howe v er, effects became additi v e or slightl y synergistic w hen the treatment time was increased from 6 h to 24 h (Figure 10 B). Inclusion of MBNL1 exon 5 and SH3YL1 exon 11 was stimulated by risdiplam alone and branaplam alone, respecti v ely ( Figure  3 ). Splicing of neither exon was strongly affected at low concentrations of the compounds. Consistently, we observed only mild effects of the combined treatments on splicing of MBNL1 exon 5 and SH3YL1 exon 11 (Figure 10 C, D). Co-skipping of ODF2L exons 14 and 15 was promoted by both compounds, although the effect was less significant at LoR and LoB (Figure 4 C). Hence, only the combined treatment using high concentrations of the compounds produced an appreciable effect on co-skipping of ODF2L exons 14 and 15 at 24 h (Figure 10 E). Considering that skipping of DST exon 92 is induced by risdiplam but not by branaplam (Figure 4 B), we noted a small effect of risdiplam but no effect of branaplam or an added effect of both compounds on DST exon 92 at 6 or 24 h (Figure 10 F). Skipping of ARHGAP12 exon 17 is promoted by higher concentrations of branaplam (Figure 4 D). Howe v er, at the concentr ations of br anaplam and risdiplam used for the combined treatments, we did not observe any significant changes in splicing of this exon (Figure 10 G). We also examined the effect of the combined treatments on transcription of fiv e genes impacted by risdiplam and / or brana plam, namel y YTHDF2 , ZFP82 , LIG1 , EGR1 and DOCK11 (Supplementary Figure S15). Except for YTHDF2 that showed an ∼1.8fold increase in expression le v els in response to 100 nM risdiplam, we did not observe any changes at 6 h treatments with individual compounds. We did not capture any noticeable change in expression levels of these genes using the combined treatments at the 6 h time point. At the 24 h time point, treatments with 100 nM risdiplam caused ∼1.4and ∼1.5-fold increases in expression of YTHDF2 and DOCK11 , respecti v ely. At the same time, e xpression of DOCK11 was significantly downregulated by both concentrations of branaplam. In addition, we captured a very small but significant decrease in expression of LIG1 at 24 h of treatment with 50 nM risdiplam. Howe v er, this decrease was not observed in the case of the combination treatments. Overall, our results supported the benefits of the combined treatments using lower concentrations of the compounds on SMN2 exon 7 splicing with seemingly minimized off-target effects.

DISCUSSION
Here we examine the transcriptome-wide off-target effects of risdiplam and branaplam, two splicing-modulating small molecules designed for the treatment of SMA. We analyzed the transcriptome of SMA patient fibroblasts treated with 50 nM (LoR) and 1000 nM (HiR) risdiplam. For br anaplam, we used concentr ations of 2 nM (LoB) and 40 nM (HiB). The highest concentrations employed in this study fall within the concentration ranges of those reported in in vitro and in vivo studies ( 37-40 , 44-48 ). Expression levels of 10921 and 2187 genes were significantly altered by HiR and HiB, respecti v ely. LoR and LoB affected expression of 566 and 15 genes, respecti v ely. Both compounds impacted genes located on all chromosomes, while genes producing between 2 and 15 alternati v ely spliced transcripts were affected the most. The unexpected high perturbations of the transcriptome caused by HiR could be attributed to its adverse effects on processes upstream of protein synthesis such as DN A replication, DN A repair, spliceosome function, mRN A transport, mRN A surveillance, ribosome biogenesis and ribosome function. In contrast, top pathways impacted by HiB were primarily signaling cascades. The highly expressed genes were more upregulated than downregulated by HiR, whereas a nearly equal number of the highly expressed genes were upregulated and downregulated by HiB. A recent report showed a tight interaction of a risdiplam analog with GA-rich motifs in single-stranded regions of DNA ( 46 ). It is possible that risdiplam interacts with GA-rich motifs in promoter regions in a similar fashion, increasing DNA accessibility for positi v e or negati v e transcription factors recruited in a context-specific manner. Incidentally, the promoter region of risdiplam-upregulated genes harbors GA-rich motifs that bear a resemblance to that of the target genes of se v eral zinc finger proteins (Supplementary Figure S3). In particular, analysis of RNA-Seq data supported the association of ZNF148 in regulation of genes impacted by risdiplam (Supplementary Figure S2C). Both compounds caused abnormal expression of lncRNAs.
Gi v en potential interactions of small molecules with unique structures formed by lncRNAs and the role of lncRNAs in transcription ( 122 ), it is possible that some of the transcriptional changes are brought about by the perturbed status of lncRN As. rRN As are kno wn tar gets of small molecules due to their complex secondary and high-order structures ( 123 ). A potential rRNA structure stabilized by interactions with a small molecule may lead to the deregulation of ribosome assemb ly. Hence, it is concei vab le that risdiplam-induced up-regulation of genes associated with ribosome biogenesis is triggered by a feedback mechanism due to improper ribosomal assembly.
For validation purposes, we used intermediate concentrations in addition to the lowest and the highest concentrations of the compounds. Our validations confirmed the aberrant expression of 14 genes randomly selected from RNA-Seq data. In most instances, the intermediate concentrations of the compounds also produced the expected changes in gene expression that were captured at the highest concentrations of the compounds. Each affected gene we valida ted is associa ted with important function(s) and / or linked to pathological condition(s). Our time course experiment captured noticeable changes in the expression of many genes at or before 6 h of the treatment with HiR and HiB. These early affected genes, including CACNB1 , ZFP82 , PDXDC1 , EGR1 , YTHDF2 , GGNBP2 , DOCK11 and PDS5B , could be considered as the primary targets of HiR and / HiB (Figure 2 I-N). Aberrantly expressed primary targets at the early time points may unleash cascading secondary effects resulting in the perturbed le v els of transcripts we captured at 24 h of treatments. Gi v en the role of YTHDF2 in RNA decay ( 61 ), it is likely that many of the downregulated transcripts observed at 24 h could be due to their degradation by YTHDF2 in the case of the treatment with HiR. Transcripts downregulated by YTHDF2 may include transcription and / or splicing factors with potential to produce e v en lar ger do wnstream effects. Although not examined in this study, important protein factors could also be the primary targets of the drugs. Perturbed protein structures, aberr ant post-tr ansla tional modifica tions and mislocalizations of proteins have potential to alter the relati v e abundance of transcripts that we analyzed at 24 h treatment with HiR and HiB.
Both compounds impacted se v en types of mechanistically distinct splicing e v ents: EIN, ESK, IRT, IRM, A5S, A3S and MXE. We captured a total of 2385 and 835 aberrant splicing e v ents caused by HiR and HiB, respecti v ely. Howe v er, these numbers should be interpreted with caution as some of the alternati v ely spliced transcripts harboring pr ematur e stop codons could be degraded by nonsense-mediated decay and / or other surveillance mechanisms ( 124 ). ESK and EIN were the most frequent e v ents triggered by HiR and HiB, respecti v el y. There were onl y 54 EIN and 44 ESK e v ents shared between HiR and HiB, underscoring the differences in their mechanism of action on splicing regulation. LoR and LoB triggered aberrant splicing affecting 58 and 76 e v ents, respecti v ely. We observ ed limited overlaps between splicing e v ents affected by HiR and LoR, suggesting that additional targets were engaged at high risdiplam concentrations. In contrast, a larger overlap in splicing e v ents affected by HiB and LoB suggested tha t the HiB concentra tion was not high enough to engage widespread targets and cause massi v e perturbations in splicing. Both transcription and splicing affect each other through complex and poorly understood mechanisms in which the structure of chr omatin, pr omoter elements and cis-elements within nascent RNA attached to RN A pol ymerase II determine the rate of transcription and the outcome of splicing ( 125 ). Our analysis re v ealed a potential connectivity between transcript levels and the compoundspecific splicing e v ents, including EIN, ESK, IRM and IRT. The impacted splicing e v ents were observed in both up-and downregulated genes, suggesting di v erse mechanisms were at play in transcription-coupled splicing regulation. Notab le EIN e v ents linked to increased gene e xpression were the risdiplam-specific inclusion of MBNL1 exon 5, branaplam-specific inclusion of KDM6A exon 28 and POMT2 exon 11B triggered by both compounds (Figure  3 C; Supplementary Tables S5 and S6). Of note, although HiB promoted inclusion of POMT2 exon 11B more efficiently than HiR (Figure 3 C), the greater change in expression of this gene was observed at HiR, suggesting that the correlation between splicing and transcription may not be entirely straightforward (Supplementary Tables S5 and S6).  Table S6). Enhanced skipping of TEAD1 exon 5B and THOC5 exon 10 observed under HiR treatment was associated with increased gene expression (Figure 4 B; Supplementary Table S7). Both of the HiR-triggered IRM e v ents that involved RBM5 intron 6 and SPOCD1 intron 12 coincided with an increase in total transcript le v els ( Figure  6 B; Supplementary Figure S7D; Supplementary Table S8). Although IRT was generally associated with lower gene expr ession in HiB tr eatment, SMTN intron 17 r etention was associated with an increase in transcript le v els (Supplementary Figure S7H; Supplementary Table S9).
We validated the aberrant splicing e v ents caused by risdiplam and branaplam using gel-based assays that accuratel y ca ptured the exon-skipped and exon-included transcripts in the same lane. For quantification of the intronretained transcripts, we used a qPCR-based assay. We performed our valida tions a t three concentrations. In most instances, w e w er e able to captur e ef fects a t the intermedia te concentrations of the compounds. These validations confirmed the effect of risdiplam and / or branaplam on six categories of splicing e v ents, namely EIN, ESK, A5S , A3S , IRM and IRT. For EIN e v ents, we e v aluated the conserv ation of the 8 nt long motif encompassing the last two nucleotides of the exon and the first six nucleotides of the downstream intron. We did not find a major deviation from the consensus 5 ss (GUAAGU), except for the sixth intronic position where the U residue was the least conserved. In the case of HiR-induced EIN, we observed the usual r epr esentation of > 55% A and > 75% G residues at the -2 and -1 e xonic positions, respecti v ely. These results support that an A residue at the last exonic position is not a general determinant for an EIN e v ent affected by risdiplam ( 44 ). Consistent with previous studies, G and A residues at positions -2 and -1, respecti v ely, were preferred in the case of EIN induced by HiB ( 48 ). Unlike EIN e v ents that affected exons of all sizes, ESK e v ents trigger ed by HiR and / or HiB pr eferentially influenced small e xons. Accor ding to our analysis, ESK e v ents generally occurred in exons with a weak 3 ss, although a weak 5 ss was also a factor in cases where ESK was triggered by both compounds. There were no preferred consensus motifs associated with A5S , A3S , IRM and IRT e v ents caused by the two compounds.
The impact of risdiplam and branaplam on splicing could be due to direct engagement with the target and / or through indirect effects via alteration of concentra tions / localiza tions of splicing factors. An effect of a compound on splicing at the early time points of treatment would support a direct engagement with the target. Indeed, we captured se v eral splicing e v ents as early as at 3 h of treatment with HiR and HiB. These are inclusion of SMN2 exon 7, POMT2 exon 11 and SH3YL1 exon 11 as well as skipping of ODF2L exon 14, DST exon 92 and ARHGAP12 exon 17 (Supplementary Figure S16). To further examine a direct interaction of the compounds with their targets, we employed minigenes that largely recapitulated the effects of the compounds on splicing of their corresponding endogenous genes (Figure 7 ; Supplementary Figure S8). Employing minigenes, we confirmed splicing of several e xons, including TT C28 e xon 20, TMEM50B e xon 5B, SH3YL1 exon 11, ARHGAP12 exon 17, ATG5 exon 3 and MBNL1 exon 5, that did not encompass an A residue at the last exonic position, yet responded to the treatment with risdiplam / branaplam. In the TMEM50B exon 5B mutant that carried a U-to-C substitution at the last exonic posi-tion, the ability of branaplam to promote inclusion of this e xon was preserv ed (Figure 7 B). An A-to-C mutation at the last exonic position of FOXM1 exon 9L did not abrogate the stimula tory ef fect of risdiplam on exon 9L inclusion (Figure 7 C). Howe v er, abrogation of e xonic AC-rich motifs inhibited the ability of the small molecules to stimulate FOXM1 exon 9L inclusion (Figure 8 B). These results supported the role of novel small molecule-responsive cis-elements upstream of the 5 ss in alternati v e 5 ss selection. In the case of MBNL1 exon 5, a G-to-C mutation at the last exonic position abolished the ability of risdiplam to stimulate exon 5 inclusion (Figure 7 E). On the other hand, a G-to-C mutation at the last exonic position of ATG5 exon 3 enhanced the ability of branaplam to stimulate exon 3 skipping (Figure  7 D). These results support that both compounds modulate splicing in a context-specific manner where the last exonic position together with other cis-elements may play an important role. We should also keep in mind that RNA structure plays a pivotal role in splicing regulation ( 126 , 127 ). Secondary and high order RNA structur es ar e proposed to serve as potential targets of small molecule therapeutics ( 122 ).
High concentrations of risdiplam and branaplam have been recently shown to interact with small motifs in the bulged region of the secondary structures of RNAs ( 128 ). The predicted secondary structures of exons / introns affected by the compounds show similar motifs within the single-stranded regions (Supplementary Figures S9-S12). Hence, it is likely that these motifs provide interaction pockets for risdiplam and branaplam. Splicing is regulated by a combinatorial control of cis-elements and trans -acting factors that define the context of an exon (129)(130)(131). Our screening of randomized sequences of SMN exon 7 showed the inability of the compounds to promote exon inclusion when it is se v erely inhibited by mutations within the exons (Supplementary Figure S14). Howe v er, not all inhibitory mutations suppressed the stimula tory ef fects of the compounds. In particular, muta tions tha t weakened the 3 ss of SMN exon 7 had the least effect on the ability of the compounds to stimulate exon 7 inclusion. We hypothesize that exonic mutations that interfere with recognition of the 5 ss of exon 7 would suppress the ability of the compounds to promote inclusion of this e xon. Such e xonic mutations may not be restricted to the GA-rich motifs proposed to interact with risdiplam. Supporting this argument, we uncovered the role of a novel AC-containing exonic motif in small molecule-induced promotion of SMN exon 7 inclusion (Figure 9 ). Similarly, we captured the role of AC-rich exonic motif(s) in the small molecule-induced enhanced usage of the downstream 5 ss in the case of FOXM1 exon 9 (Figure 8 ). It is likely that the small molecules dir ectly or indir ectly help recruit factors that recognize ACcontaining exonic motifs. Such recruitment may be enabled by a small molecule-assisted interaction of FUBP1 and / or KHSRP with the upstream GA-rich motif as proposed in the case of risdiplam ( 45 ). Interestingly, a GA-rich motif is also present upstream of the AC-rich motif within FOXM1 exon 9 (Supplementary Figure S11). It is possible that the small molecule-mediated network of RNA-protein and pr otein-pr otein interactions on the exonic sequences may facilitate recruitment of U1 snRNP to the 5 ss, which itself is shown to independently interact with the small molecule and U1 snRNA (Figure 9 C). Based on the observation that the same mutation(s) influenced the effects of risdiplam and branaplam on splicing of SMN exon 7 and FOXM1 exon 9 (Figures 8 and 9 ), it is likely that these compounds share common mechanisms of splicing regulation in these instances.
SMA is a model disease in w hich m ultiple a pproaches could be employed to modulate SMN2 exon 7 splicing as a potential therapy ( 8 ). Howe v er, therapeutic benefits of a splicing-modulating compound are not guaranteed without proper evaluation of its off-target effects. The long-term consequence of off-target effects could be detrimental if repeated high doses are administered during the course of the protracted tr eatment. A r ecent study in a mouse model of se v ere SMA showed better efficacy of risdiplam when combined with an ISS-N1-targeting ASO ( 132 ). Howe v er, high efficacy of splicing-modulating drugs observed in in vivo studies in mouse models cannot be extrapolated to humans gi v en the di v ergences of intronic sequences that harbor the vast r epertoir e of human-specific regulatory elements. With the uni v ersal availability of high-throughput sequencing at an affor dab le cost, it is now possible to examine the genome-wide off-target effects at the very outset of a drug de v elopment process. A systematic e valuation of of f-target ef fects is critically important not only for understanding the mechanism of drug action but also for de v eloping the next generation of highly specific drugs. Additionally, a thorough evaluation of the concentration-dependent effects of a drug in question on the transcriptome would help determine the dosing and frequency of drug administra tion. W hile the recent approval of the splicing modulator risdiplam for the treatment of SMA is welcome news, many concerns remain due to uncomfortable side effects including rash, fe v er, diarrhea, joint pain (arthralgia), ulcers of the mouth area and urinary tract infections ( 39 ). It is likely that some of these side effects could be attributed at least in part to the off-target effects we report here. The relati v ely fewer off-target effects associated with branaplam could be due to the high efficacy of this compound at low nanomolar concentrations. Yet, we captured considerable perturbations in gene expression and pre-mRNA splicing, especially aberrant exon inclusion, even at 40 nM branaplam. Interestingl y, brana plam is also being considered for the treatment of Huntington's disease ( 133 ). Hence, findings reported here have relevance for diseases other than SMA. The contrasting account of the off-target effects of two splicing-modulating small molecules provides a general insight into how critically important the concentration of a compound used for therapeutic intervention could be. On an optimistic note, both compounds showed limited offtarget effects at low concentr ations. Br anaplam in particular remained highly ef ficacious a t a concentration as low as 2 nM while changing the expression of only 15 genes. The results of our combination treatments showed strong stimula tory ef fect on SMN2 exon 7 splicing at low doses of the compounds, with potentially minimized off-target disturbances. These findings augur well for de v eloping the ne xt generation of small molecule therapeutics with e v en better efficacies and fewer off-target effects.

DA T A A V AILABILITY
All RNA-Seq data are available publicly at the NCBI Sequence Read Archi v e, accession number SRP334251.

SUPPLEMENT ARY DA T A
Supplementary Data are available at NAR Online.