LUC7 proteins define two major classes of 5′ splice sites in animals and plants

Abstract

Mutation or deletion of the U1 snRNP-associated factor LUC7L2 is associated with myeloid neoplasms, and knockout of LUC7L2 alters cellular metabolism. Here, we show that members of the LUC7 protein family differentially regulate two major classes of 5′ splice sites (5′SS) and broadly regulate mRNA splicing in both human cell lines and leukemias with LUC7L2 copy number variation. We describe distinctive 5′SS features of exons impacted by the three human LUC7 paralogs: LUC7L2 and LUC7L enhance splicing of “right-handed” 5′SS with stronger consensus matching on the intron side of the near invariant /GU, while LUC7L3 enhances splicing of “left-handed” 5′SS with stronger consensus matching upstream of the /GU. We validated our model of sequence-specific 5′SS regulation both by mutating splice sites and swapping domains between human LUC7 proteins. Evolutionary analysis indicates that the LUC7L2/LUC7L3 subfamilies evolved before the split between animals and plants. Analysis of Arabidopsis thaliana mutants confirmed that plant LUC7 orthologs possess similar specificity to their human counterparts, indicating that 5′SS regulation by LUC7 proteins is highly conserved.

Introduction

Eukaryotic protein-coding genes are often interrupted by non-coding introns, which are recognized and excised by the spliceosome. In the earliest steps of spliceosome assembly, U1 snRNP identifies potential intron boundaries in pre-mRNAs via RNA:RNA base pairing between the 5′ end of U1 snRNA and the 5′ splice site (5′SS), a conserved 9 nucleotide motif that demarcates the exon-intron boundary¹. While the 5′ end of U1 snRNA is nearly invariant, there is considerable variation in functional 5′SS motifs between eukaryotes^2,3 (Fig. 1A), indicating additional factors contribute to the recognition of pre-mRNA splicing substrates.

Fig. 1: Human LUC7 proteins impact 5′ splice sites with distinct composition.

A Sequence logos of annotated 5′SS in yeast (top) and humans (bottom). B Cryo-EM data of yeast Luc7p contacting U1 snRNA:5′SS duplex from Bai et al., 2018⁸. C Representative sequence logos of 5′SS for all cassette exons (left) and exons with significantly increased or decreased inclusion upon depletion of LUC7L2 or LUC7L3. D Schematic of 5′SS Balance score, which is defined as the number of consensus bases at positions +4/ + 5/ + 6 minus the number of consensus bases at –3/–2/–1 of a 5′SS. E Mean phyloP score for human 5′SS of each possible 5′SS Balance score. F Log-odds estimates of 5′SS differential inclusion versus skipping as a function of the 5′SS Balance score for each human LUC7 RNA-seq dataset. Point estimates are derived from 100 bootstrap samples, each containing 500 differentially included (FDR < 0.1, dPSI > 0) and 500 differentially skipped (FDR < 0.1, dPSI <0) events, with replacement. Error bars represent the 95% confidence intervals, calculated as the standard deviation of the 100 bootstrap estimates. Source data are provided as a Source Data file.

Recent work uncovered an unexpected connection between mRNA splicing and cellular metabolism, in which genetic ablation of U1 snRNP components in erythroleukemia cells promoted a shift towards increased use of oxidative phosphorylation (OXPHOS), promoting growth in galactose-containing media⁴. Of these components, knockout of LUC7L2, a non-constitutively bound U1 snRNP auxiliary factor, presented the strongest metabolic phenotype and altered the splicing of several hundred exons. The shift toward OXPHOS could be partially explained by changes in the splicing of key metabolic genes, including PFKM and SLC7A11.

Previous work on the LUC7 family suggests a role in 5′SS selection^5,6,7. Cryo-EM studies of the S. cerevisiae spliceosome showed that the yeast LUC7 is positioned adjacent to the U1 snRNA:5′SS duplex, interacting with the phosphodiester backbone of U1 snRNA during the initial stages of pre-spliceosome assembly^8,9,10. While yeast possess a single, constitutively U1-associated LUC7 protein, yLuc7p, mammals possess three: the closely related LUC7L and LUC7L2, and the more divergent LUC7L3^4,11. Immunoprecipitation of each human LUC7 protein recovered U1 snRNP proteins, confirming interaction with the snRNP. All three proteins crosslinked with splice sites in many pre-mRNAs and with bases at the 5′ end of U1 snRNA that base pair with the 5′SS^4,11. Despite their high conservation to yLuc7p, their association with U1 snRNP, and their broad impact on the transcriptome, human LUC7 proteins are absent from all published cryo-EM spliceosome structures.

Ablation of LUC7L2, but not of other LUC7 family members, promotes an oxidative phosphorylation metabolic state, supporting a functional distinction between these paralogs⁴. In iPSC-derived hematopoietic stem cells, low expression and/or loss of LUC7L2 via del(7q) impacts differentiation¹². Low expression and/or loss of LUC7L2 is also associated with the development of leukemia and other myeloid neoplasms¹³, suggesting that LUC7L2’s influence on splicing and cellular metabolism may contribute to leukemogenesis. However, the molecular functions of the LUC7 paralogs remain incompletely understood. Here, we investigated the distinct roles of LUC7s in pre-mRNA splicing. Unexpectedly, we found that different human LUC7 paralogs broadly impact the splicing of thousands of exons in a predictable, 5′SS-sequence-dependent manner. Experiments and analyses in a broad range of systems, from human cells to mutant plants, demonstrate that two subfamilies of LUC7 proteins regulate two newly defined classes of “left-handed” (LH) and “right-handed” (RH) 5′SS in opposing manners, helping to explain the distinct phenotypes of these paralogs.

Results

LUC7 family members impact distinct subclasses of 5′SS

In yeast pre-B complex, the second zinc-finger domain of yeast Luc7 (yLuc7p-ZnF2) forms salt bridges with the phosphodiester backbone of U1 snRNA while duplexed with the 5′SS⁸ (Fig. 1B). The three human LUC7 family members each possess two N-terminal zinc finger (ZnF) domains, but LUC7L3-ZnF2 is highly divergent compared to its paralogs (Figure S1A). Since human LUC7 paralogs also interact with human 5′SS^4,6,11, we reasoned that the divergence of LUC7L/LUC7L2 and LUC7L3 may have resulted in distinct sets of binding/regulatory targets. Therefore, we examined the 5′SS features of dysregulated exons in publicly available RNA-seq data from knockout (KO) or knockdown (KD) of these genes in human cells, including LUC7L2 KO in K562 erythroleukemia and HeLa cells⁴, and LUC7L, LUC7L2, and LUC7L3 KD in K562 cells¹¹. Analysis of 5′SS strength¹⁴ indicated that exons sensitive to depletion of LUC7 family members had either no significant difference in mean 5′SS strength relative to unchanged exons or had differences that were significant but very small in magnitude, ranging from 0 to 0.25 bits (Figure S1B). These observations suggest 5′SS strength contributes at most minimally to exon responsiveness to Luc7 perturbations.

Despite having similar strength, the 5′SS of exons impacted by depletion of LUC7L2 or LUC7L3 had strikingly different sequence compositions compared to unchanged exons (Fig. 1C). Specifically, exons more skipped following LUC7L2 KO had higher consensus matching at positions +3, +4, +5 and +6 on the intron side of the 5′SS motif, while those more included had increased consensus matching at positions –3, –2 and –1 in the exon. In contrast, KD of LUC7L3 yielded an opposite pattern, with exons more skipped following depletion having higher consensus matching at –3, –2 and –1, and those more included having higher consensus matching at positions +3, +4, +5 and +6. No differences were observed in the sequence logos at the upstream 5′SS or at the upstream or downstream 3′SS relative to affected exons (Figure S1C), supporting that the 5′SS sequence of the impacted cassette exon is the primary feature that predicts LUC7 regulation.

Analysis of human 5′SS logos (Fig. 1C) suggests that the extent of consensus matching on each side of the /GU is an important determinant of sensitivity to LUC7L2 versus LUC7L3 depletion, and previous studies^14,15,16,17 have noted the preferential co-occurrence of adjacent consensus bases in the intron or exon portion of the 5′SS motif. To quantitatively assess this feature, we devised a simple metric called the “5′SS Balance score” (or simply 5′SS Balance), which summarizes the distribution of consensus bases around the 5′SS. This score is calculated as the number of matches to consensus on the intron side at positions +4, +5, and +6, minus the number of consensus matches on the exon side at positions –3, –2, and –1 (Fig. 1D). At position –3, C or A are counted as consensus, and position +3 is ignored for the time being so that Balance scores are centered on 0. The 5′SS Balance ranges from –3 to +3, where LH 5′SS with more consensus matching on the exon side receive negative scores and RH 5′SS with more consensus matching on the intron side receive positive scores. “Balanced” exons with equal consensus matching on both sides of the /GU motif receive a score of 0.

Across all human exons, about 20% have Balance ≤ –2, while about 15% have Balance ≥ +2 (Figure S2A). The 5′SS Balance score is a property distinct from 5′SS strength (with a weak negative correlation observed, Figure S2B) and is generally conserved across vertebrates and between orthologous exons of human and mouse (Fig. 1E, Figure S2C). Together, these results imply that LH and RH 5′SS subclasses are abundant and stably maintained as subclasses of 5′SS.

5′SS Balance predicts sensitivity to LUC7L2 versus LUC7L3

To assess the extent to which 5′SS Balance could predict splicing changes in response to LUC7 depletion, we evaluated the log-odds of increased inclusion versus skipping across available LUC7 RNA-seq datasets. This analysis indicated that 5′SS Balance is strongly predictive, with LUC7L2 promoting the splicing of exons with RH 5′SS and repressing exons with LH 5′SS, and LUC7L3 promoting splicing of exons with LH 5′SS and repressing splicing of exons with RH 5′SS (Fig. 1F). Similar effects were observed for LUC7L2 KD in K562 cells and LUC7L2 KO in HeLa, while depletion of LUC7L gave results similar in direction to depletion of LUC7L2, but more modest in magnitude. To confirm these findings, we performed RNA-seq on HEK293T cells overexpressing each individual LUC7. Overexpression yielded the opposite effects on LH and RH exons as depletion for each paralog (Fig. 1F), supporting our conclusions that the splicing of LH 5′SS is promoted by LUC7L3, while RH 5′SS are promoted by LUC7L2 and LUC7L.

To systematically identify 5′SS motifs impacted by each LUC7 protein, we measured the impact of LUC7 proteins on the enrichment of each 5′SS 9mer (spanning positions –3 to +6) in exons differentially included or skipped. This measure, which we call “5′SS Enrichment”, uses a Dirichlet-multinomial model to approximate the log-odds of a given 5′SS sequence occurring in significantly included or skipped exons versus unchanged exons within an RNA-seq experiment. Positive enrichment values indicate over-representation of a 5′SS sequence in included exons and negative values indicate enrichment in skipped exons, while values near zero indicate the absence of bias for inclusion/skipping (or the presence in unchanged exons only).

As a positive control, we evaluated our approach on a recently published RNA-seq data where mutant U1 snRNA (g.3a>c) was overexpressed¹⁸ (Figure S3A). As expected, we recovered a very strong and significant enrichment of 5′SS with +6 G (complementary to U1 snRNA (g.3a>c) position 3) (Figure S3B). 5′SS Enrichment analysis of LUC7 perturbation experiments yielded many 5′SS motifs with strong enrichment in each direction (examples shown in Figure S3C). Hierarchical clustering of significant 5′SS Enrichment scores across all 8 LUC7 RNA-seq datasets yielded clear groups of 5′SS with many LH motifs promoted by LUC7L3 and repressed by LUC7L2 and LUC7L, and many RH motifs promoted by LUC7L2/LUC7L and repressed by LUC7L3 (Figure S3D).

Our analyses indicate that we can empirically define LH and RH 5′SS by their sensitivity to LUC7 proteins. Therefore, we defined a “LUC7 Score” as a ratio of position weight matrices (PWMs) for LUC7L2-promoted/LUC7L3-repressed RH motifs versus LUC7L3-promoted/LUC7L2-repressed LH motifs based on 5′SS Enrichment across seven LUC7 perturbations (Methods) (Fig. 2A). This score, which is strongly correlated with the 5′SS Balance (r = 0.86, p < 2.2e-16, Fig. 2D), predicted the direction of splicing change in a cross-validation test (Fig. 2B), performing slightly better than or like 5′SS Balance in every case. The LUC7 Score captures additional sequence features, including the base at +3 and the relative importance of different non-consensus bases at each position. For example, LUC7L2-promoted 5′SS favor –1U while LUC7L3-promoted 5′SS favor –1 A when not matching the consensus (G) at position –1.

Fig. 2: Enrichment of individual 5′ splice site motifs across LUC7 experiments defines two distinct classes.

A Volcano plot of 5′SS enrichment scores of individual 5′SS 9mers (each dot represents a 9-mer) from LUC7 meta-analysis (left). Sequence logos derived from significant 9mers (qvalue < 0.01) from LUC7 meta-analysis (right). B Receiving operator characteristic curve (ROC) of LUC7 score’s ability to discriminate held-out differentially included vs skipped splicing events (FDR < 0.1, no dPSI cutoff). C Distribution of LUC7 score and frequency for distinct 9mer 5′SS sequences (with /GT) in Gencode human genome protein-coding exons. D Heat map showing Pearson correlation between new measures – 5′SS Balance and LUC7 score – versus standard 5′SS measures for /GT donors in protein-coding exons. Source data are provided as a Source Data file.

The LUC7 Score groups human 5′SS into two similarly sized clusters (Fig. 2C). Comparison of our LUC7 Score to conventional 5′SS features shows that it is uncorrelated with splice site strength (r = 0.007, p = 0.65) and distinct from other previously studied features of 5′SS (Fig. 2D). The individual PWMs for the LH and RH motifs generated from LUC7 protein RNA-seq data are weakly, but significantly correlated, with the predicted minimum free energy of interaction with U5 and U6 snRNAs, respectively. Together, these findings indicate that the LUC7 Score quantitatively summarizes the sequence features of exons whose splicing depends on different LUC7 family members.

LUC7 family members cross-regulate each other’s splicing via RH 5′SS

Many splicing regulatory factors (SRFs) negatively auto-regulate the expression of their own gene at the level of splicing and are often negatively cross-regulated by close paralogs through the same exons^19,20. This regulatory arrangement, which helps maintain a narrow range of activity for SRFs and SRF families, also occurs in the LUC7 family¹¹, with LUC7L2 repressing the expression of LUC7L by promoting the inclusion of an exon that results in a premature termination codon (PTC) in LUC7L⁴. Although such PTC-inducing exons are not annotated in the other LUC7 paralogs, we noted the strongly conserved alternative second exon (AE2) in LUC7L2 that replaces the coding sequence of a highly conserved N-terminal alpha helix that interacts with SM-ring component SmE in U1 snRNP⁹. Sequence alignment of the LUC7L2 AE2 and LUC7L PTC exons revealed that they are 82% identical at the nucleotide level, implying a common evolutionary origin; no comparable exon was found in LUC7L3.

Consistent with previous studies, we found that overexpression of either LUC7L or LUC7L2 promoted the inclusion of the LUC7L PTC exon in a minigene assay (Fig. 3A), suggesting that both paralogs repress LUC7L expression (Fig. 3B, C, left). We found that LUC7L2 AE2 is regulated in a similar manner to LUC7L PTC (Fig. 3B, C, right), supporting a pattern of negative auto- and cross-regulation analogous to that observed for PTBP1/PTBP2 and other closely related, often partially-redundant SRF genes. In contrast, LUC7L3 overexpression repressed the inclusion of the LUC7L and LUC7L2 negative-regulatory exons, promoting their canonical mRNA isoforms.

Fig. 3: Regulation of splicing by LUC7 proteins depends on 5′SS handedness.

A Schematic of the pSpliceExpress minigene construct used, in which an internal exon of interest with flanking splice sites is inserted into a minigene expressing rat insulin exons 2 and 3 along with the intervening intron. B Bar plots of the cassette exon’s mean percent spliced in (by RT-PCR with primers in flanking exons) across all experiments; bars are color-coded to match LUC7 paralog colors in (E). C Representative gel images from LUC7-related minigenes. Percent Spliced In values are shown at the bottom of each lane. D Mean PSI for the LUC7L2 AE2 minigene with mutagenized 5′SS (left), and representative RT-PCR gel images (right). E Proposed regulatory relationships between human LUC7 family members. F Bar plot of PSI for the SNRPC E2 minigene with wildtype RH 5′SS (left), mutagenized LH 5′SS (middle), and mutagenized RH 5′SS (right). G Bar plot of PSI for the XPA E3 minigene with wildtype LH 5′SS (left), mutagenized LH 5′SS (middle), and mutagenized RH 5′SS (right). For all minigene experiments reported (panels B, D, F and G), error bars represent the standard error of the mean. Brackets indicate significant comparisons. *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.00001, BH-corrected two-sided t-test. Source data are provided as a Source Data file.

Notably, the 5′SS sequence of the homologous LUC7L PTC exon and LUC7L2 AE2 are identical in sequence, of average strength and strongly RH in character (MaxEnt score = 8.7 bits, 5′SS Balance = +3 (RH), LUC7 Score = +7.6). To determine whether 5′SS properties contribute to the regulation of these exons, we swapped the RH 5′SS of LUC7L2 AE2 with an LH 5′SS of comparable strength (MaxEnt = 8.1 bits, Balance = –2, LUC7 Score = –8.4). We found that the autoregulatory LUC7L2 AE2 exon with this new 5′SS was spliced similarly to the wildtype (WT) under control conditions but that this change completely abolished splicing regulation by LUC7L2 (Fig. 3D) and was even sufficient to confer positive regulation via LUC7L3 (Figure S4A). Together, these observations indicate that the RH character of the AE2 5′SS is necessary for regulation by LUC7L2. Furthermore, it suggests that human LUC7 proteins auto and cross-regulate one another through the 5′SS sequence of regulatory exons. Promotion of LUC7L expression by LUC7L3 may help to maintain a specific ratio of LUC7L2 + LUC7L to LUC7L3 activity, perhaps to ensure that both RH and LH 5′SS are spliced efficiently (Fig. 3E).

To assess the generality of the 5′SS-dependence of LUC7 regulation, we conducted similar 5′SS mutagenesis experiments on minigenes containing the RH 5′SS of SNRPC exon 2, which encodes the N-terminal ZnF domain that interacts with U1:5′SS duplex opposite to Luc7, and on the LH 5′SS of XPA constitutive exon 3. As above, changing the SNRPC exon 2 WT 5′SS (MaxEnt = 8.4 bits, Balance = +3, LUC7 Score = +13.7) to a LH 5′SS (MaxEnt = 7.6 bits, Balance = –3, LUC7 Score = –6.3) abolishes all positive and negative regulation by LUC7 proteins (Fig. 3F). However, mutation to a different RH 5′SS (MaxEnt = 8.3 bits, 5′SS Balance = +3, LUC7 Score = +9.7) preserved regulation similar to the WT sequence. This demonstrates that RH 5′SS are necessary for positive and negative regulation by LUC7L/LUC7L2 and LUC7L3, respectively. Similarly, switching the LH 5′SS of XPA exon 3 (MaxEnt = 8.1 bits, Balance = –2, LUC7 Score = –8.4) to a different LH 5′SS (MaxEnt = 9.5 bits, Balance = –2, LUC7 Score = –3.3) mimics the regulation of the WT LH 5′SS. However, splice site switching a LH 5′SS to a RH 5′SS (MaxEnt = 7.4 bits, Balance = +3, LUC7 Score = +10.6) abolished regulation by LUC7L/LUC7L2 and yielded repression by LUC7L3 (Fig. 3G).

These results indicate that RH 5′SS are necessary, but not sufficient, to confer positive regulation by LUC7L or LUC7L2, but may be sufficient to confer repression by LUC7L3 (Fig. 3G). Altogether, our mutagenesis experiments suggest that the “handedness” (LH or RH character) of the 5′SS sequence is a key determinant of regulation by LUC7 family members.

Specificity for RH 5′SS is seen by CLIP and conferred by LUC7 N-terminal domains

We hypothesized that positive regulation of RH 5′SS by LUC7L2 occurs through a direct mechanism involving stabilization of U1 snRNP’s interactions with RH 5′SS by LUC7L2. To assess LUC7L2 RNA binding transcriptome-wide, we analyzed published eCLIP data⁴. Consistent with previous analyses, LUC7L2 was found to crosslink predominantly to the 5′SS and a region ~25 nt upstream of the 3′SS (Fig. 4A, top). To assess whether LUC7L2 crosslinks more to RH 5′SS, we binned exons by their LUC7 scores and then examined the Pearson correlation between the mean LUC7 score of the bin and eCLIP crosslinking enrichment in windows around the 3′SS and 5′SS of cassette exons. This analysis, performed on 4 eCLIP experiments in two separate human cell lines, revealed increased LUC7L2 crosslinking around the 5′SS of cassette exons with more RH 5′SS (Fig. 4A, bottom), supporting our hypothesis. Increased crosslinking was also observed to a region ~25 nt upstream of the 3′SS near the expected location of U2 snRNP binding; this binding might result from cross-exon interactions of U1 and U2 snRNPs related to exon definition.

Fig. 4: LUC7 specificity for subclasses of 5′SS depends on the structured N-terminal region.

A eCLIP enrichment of LUC7L2 around windows of constitutive and cassette exon splice sites, with crosslinks aggregated into 10 nt bins (top). Pearson correlation between LUC7 score and LUC7L2 eCLIP enrichment is shown, with significant correlations (p < 0.05) identified by Pearson’s correlation test. B Protein domain structure of human LUC7 proteins and experimentally investigated chimeric proteins. C Change (“delta”) in percent spliced in (qRT-PCR) from transfected minigenes containing different internal exons of introns following overexpression of LUC7 WT or chimeric cDNA (shown at right). Mean is plotted with error bars representing standard error of the mean, and bars are color-coded by the Balance score of the internal exon’s 5′SS. D Heat map of Pearson correlations of delta percent spliced values for each LUC7 WT or chimeric cDNA overexpressed in (C). Source data are provided as a Source Data file.

A multiple sequence alignment of human LUC7 paralogs with yLUC7p suggested that yLuc7p is more similar overall to LUC7L and LUC7L2, particularly in their ZnF2 domains (Figure S1A), while LUC7L3 diverges in ZnF2, potentially contributing to functional differences relative to LUC7L/LUC7L2. Given that human LUC7L and LUC7L2 are more similar in sequence and possess similar target specificity in RNA-seq and minigene assays, we hypothesized that the divergent structured domains dictate LUC7 paralog specificity for distinct 5′SS subclasses. To test this hypothesis, we generated chimeric proteins by fusing the structured N-terminal regions of LUC7L2 or LUC7L3 with the C-terminal intrinsically disordered region (IDR) of the other paralog (Fig. 4B) and assessed the impact on exons with different types of 5′SS with a small library of minigene constructs with different RH and LH 5′SS exons of interest. Consistent with the pattern established above, 24-hour overexpression of wildtype (WT) LUC7L or LUC7L2 invariably promoted the inclusion of exons with RH 5′SS and repressed inclusion of exons with LH 5′SS, while overexpression of WT LUC7L3 repressed inclusion of RH 5′SS and promoted inclusion of exons with LH 5′SS (Fig. 4C). Western blot analysis indicated minimal impact on endogenous LUC7 protein levels after 24 h of overexpression, reinforcing that cross-regulation does not substantially affect our findings (Figure S4B). Considering the chimeric proteins, we found that the LUC7L2 N-terminus fused to LUC7L3 C-terminus (LUC7L2::LUC7L3) yielded similar changes in splicing as LUC7L2 (or LUC7L) overexpression for both LH and RH 5′SS, though somewhat weaker in magnitude (Fig. 4C). On the other hand, the reverse chimera, LUC7L3::LUC7L2, repressed RH 5′SS inclusion, often more potently than WT LUC7L3, and activated one of the two LH 5′SS studied. This data (Fig. 4D) supports that the structured N-terminal domains of LUC7L2 and LUC7L3 are adequate for establishing the overall repression/activation patterns of LH versus RH 5′SS. Additionally, our results indicate that the C-terminal regions of these paralogs perform similar, but not identical, functions in splicing.

AMLs with low expression of LUC7L2 have distinct splicing and expression patterns

Loss of one copy of LUC7L2 occurs with monosomy 7 or loss of the long arm of chromosome 7 (-7/del(q)), which is one of the most frequent cytogenetic abnormalities found in both myelodysplastic disorders (MDS) and myeloid tumors, including acute myeloid leukemias (AMLs)²¹. In a differentiation phenotype-rescue screen, LUC7L2, along with EZH2, HIPK2, and ATP6V0E2, was identified as haploinsufficient genes responsible for mediating the hematopoietic defect associated with -7/del(q)¹². Furthermore, the LUC7L2 locus is recurrently mutated in patients with MDS and AML, and low expression of LUC7L2 is identified as a common characteristic among patients who harbor mutations in other MDS-associated splicing factors²².

These links between LUC7L2 and AML pathogenesis motivated us to focus on AMLs with copy number variation (CNVs) for LUC7L2 and to hypothesize that AMLs with reduced LUC7L2 expression (LUC7L2^Low) would have reduced inclusion of RH versus LH 5′SS compared to other AMLs. To test this hypothesis, we identified 14 or 8.8% of AML samples that harbor deletions spanning the LUC7L2 locus. As expected, AMLs with fewer copies of LUC7L2 express roughly 2-fold lower LUC7L2 expression on average (Fig. 5A). While we did not detect differences in the expression of LUC7L or LUC7L3 mRNA in these tumors, we observed that both the LUC7L and LUC7L2 autoregulatory exons were less included in AML samples with low expression of LUC7L2 (not shown), as predicted by our LUC7 regulatory model (Fig. 3E).

Fig. 5: Transcriptomic analysis of Acute Myeloid Leukemias with low expression of LUC7L2.

A Normalized gene expression values for human LUC7 family in LAML samples, colored by whether samples possess a LUC7L2 CNV loss (CNV log₂ value < –0.5). A two-sided t-test was used to test differences in LUC7 paralog expression between LUC7L2^+/+ (n = 159) and LUC7L2 CNV loss (n = 14). B Overlap of LUC7L2 CNV loss samples and LUC7L2^Low expression samples. C Mean dPSI per 5′ splice site sequence for differentially spliced exons when comparing LUC7L2^Ctrl versus LUC7L2^Low expression samples. A linear regression line (black) with a 95% confidence interval is shown, and the Pearson correlation and Pearson correlation test are displayed. D Gene set enrichment analysis of differentially expressed genes comparing LUC7L2^Low versus LUC7L2^Ctrl expression samples. E Copy-number variation analysis for LUC7L2 and LUC7L3 loci across all TCGA cancer types. ***p < 0.001, Bonferroni-corrected two-sided Wilcoxon-rank sum test. For panels A and E, boxplots display the data distribution within each group. The box represents the interquartile range (IQR), spanning from the 25th percentile (Q1) to the 75th percentile (Q3), with the horizontal line inside the box indicating the median (50th percentile). The whiskers extend to the most extreme data points within 1.5 times the IQR from Q1 and Q3. Source data are provided as a Source Data file.

Differential splicing analysis comparing 16 LUC7L2^Low AMLs versus 157 LUC7L2^Ctrl samples recapitulated our earlier results from LUC7L2-depleted human cell lines, in which differentially skipped and included exons possess right- and left-handed 5′SS sequences, respectively (Fig. 5C, Figure S4C). Conducting the same splicing analysis on shuffled AML groups completely abolished the observed relationship between 5′SS handedness and splicing regulation (Figure S4D). Together, these analyses show that LUC7L2^Low AMLs inefficiently splice RH 5′SS relative to LH 5′SS, supporting a role for reduced LUC7L2 levels in shaping the transcriptomes of these tumors.

To assess potential functional consequences of low LUC7L2 expression in AML, we performed gene set enrichment analysis (Fig. 5D). These analyses revealed that LUC7L2^Low AMLs have increased expression of genes associated with oxidative phosphorylation relative to other AMLs, suggesting that perturbation of LUC7L2 expression in AMLs may mimic the metabolic state previously observed in K562 LUC7L2^KO cells⁴. Other pathways associated with cell cycle and DNA damage repair, such as E2F target genes, were also generally increased. Altogether, these results signify  that 5′SS subclasses are differentially regulated by LUC7 proteins in AML and that low expression of LUC7L2 may promote a transcriptomic and/or metabolic state conducive to AML pathogenesis²³.

To explore whether variations of LUC7L2 and LUC7L3 may be involved in other cancers, we examined the CNV of these genes across the entire TCGA cohort (CNV data for LUC7L were not available). This analysis revealed that several tumor types, including glioblastomas (GBM), kidney chromophobes (KICH), and cutaneous melanomas (SKCM) exhibit increased copy number of LUC7L2, with KICH also showing decreased copy number of LUC7L3, suggesting dysregulation of U1 snRNP and potential skewing of 5′SS selection toward RH 5′SS in these cancers (Fig. 5E). Additionally, kidney renal papillary cell carcinoma (KIRP) had increased copy number of LUC7L3. Some of these observations reflect well-established chromosomal aberrations commonly found in specific tumor subtypes, e.g., trisomy 7, which is common in GBM and SKCM associated with EGFR amplification, would also yield an additional copy of LUC7L2 in these tumors. Changes in LUC7 expression due to CNVs may generally contribute to the observed splicing variation found in many cancer subtypes and might contribute to metabolic changes as well.

Evolutionarily related Luc7 proteins possess conserved specificities for 5′SS subclasses

To explore the origins of the divergence in function between the LUC7L3 and LUC7L2/LUC7L, we searched for homologs of these proteins and performed multiple sequence alignments. We identified at least one annotated Luc7 family member in 7 out of 9 major eukaryotic supergroups (Figure S5A), indicating that this gene family is very ancient. We also performed multiple sequence alignment of Luc7 proteins in 33 plants, animals, and fungi, spanning a deep split in eukaryotic phylogeny²⁴. The resulting dendrogram supports that there are two major subfamilies of Luc7 proteins, represented by LUC7L2 and LUC7L3, and that both subfamilies were present before the split between plants and animals/fungi (Fig. 6A). Human LUC7L2 and LUC7L, which have similar 5′SS specificity via our minigene and transcriptome analyses, duplicated more recently in the last common ancestor (LCA) of jawed vertebrates approximately 425 million years ago (Fig. 6B). Intriguingly, while animals and plants retained both LUC7L2-like and LUC7L3-like proteins, none of the initially analyzed fungi possess a member of the LUC7L3 subfamily (but see below).

Fig. 6: Related LUC7 proteins possess conserved specificities for 5′SS subclasses.

A Maximum likelihood phylogenetic tree built from multiple sequence alignment of Luc7 proteins from 33 animal, fungal and plant species, adding Trichomonas vaginalis Luc7 protein as an outgroup. Two main clusters are shaded to indicate Luc7 subfamilies; individual proteins represented by symbols indicating clade of origin. B Representation of presence/absence and likely duplication/loss events for Luc7 subfamilies overlaid on the eukaryotic phylogenetic tree. C Correlation matrix of 5′SS models learned from dinucleotide features of 5′SS differentially spliced in human and plant Luc7 RNA-seq experiments. D Sequence logos of top or bottom 10% unique 5′SS sequences identified by dinucleotide 5′SS models. E Residue conservation estimated from multiple sequence alignment of LUC7L2-type proteins from 6A), projected onto the yLUC7p (from Bai et al., 2018) cryo-EM protein structure.

The deep conservation of Luc7 family members across animals and plants provided an opportunity to compare the molecular functions of Luc7 proteins across eukaryotic supergroups. Arabidopsis thaliana has three Luc7 genes: two members of the LUC7L2-like subfamily (LUC7A and LUC7B) and one member of the LUC7L3-like subfamily (LUC7RL), which are important for development and stress resistance²⁵. To assess whether evolutionarily related LUC7 proteins possess analogous 5′SS specificities, we performed RNA-seq on every possible combination of Arabidopsis thaliana luc7 single, double and triple mutants, and carried out differential splicing analysis.

Consistent with previous work, analysis of cassette exons revealed that luc7 single mutants have relatively modest splicing phenotypes: luc7a and luc7b showed a trend toward higher LUC7 score in skipped exons (like LUC7L2), while luc7rl showed the opposite trend (like LUC7L3), but only luc7a reached statistical significance (Figure S5B-C). Strikingly, a similar analysis of the luc7a luc7b double mutant showed a very strong bias for RH 5′SS among skipped exons, phenocopying depletion of human LUC7L2. This observation remained consistent for luc7a luc7b double mutants with two different alleles of luc7a (Figure S5D), and the skipping of three exons with RH 5′SS was validated in these mutants using RT-PCR (Figure S5E). Similarly, luc7rl mutants trended toward negative LUC7 scores among skipped exons, reaching significance in one case (Figure S5D). No bias for LH or RH 5′SS in differentially skipped or included exons was observed in the luc7 triple mutant.

To assess the overall similarity of the 5′SS dysregulated in human and plant LUC7 datasets, we devised a model that identifies 5′SS nucleotide pairs, consecutive or non-consecutive, as predictive features of 5′SS dysregulated within an RNA-seq data set. By scoring all human 5′SS with each model and evaluating the correlation of these scores, we found that the LUC7 RNA-seq datasets formed two main clusters that mirrored their phylogenetic relationship. One cluster included LUC7L3, luc7rl, and luc7rl-containing double mutants and a second cluster encompassed LUC7L2, LUC7L, and all A. thaliana mutants featuring luc7a and/or luc7b (Fig. 6C). Furthermore, the top 10% of 5′SS sequences most sensitive to LUC7 subfamily depletion were very similar between humans and plants, as predicted by our dinucleotide model. Depletion of LUC7L2 subfamily members consistently results in more skipping of RH 5′SS with +5 G and –1U (Fig. 6D). Notably, the amino acid identity of the LUC7L2 subfamily is highly conserved in ZnF2 across plants and animals, suggesting that ZnF2 in LUC7L2-type proteins facilitates the splicing of exons with 5′SS that have +5 G in both humans and plants (Fig. 6E). In contrast, depletion of LUC7L3 subfamily members results in more skipping of LH 5′SS with both –1 G and +5H (Fig. 6D). These observations indicate that the two subfamilies of LUC7 proteins in plants have distinct activities on 5′SS subclasses, paralleling our observations in humans.

While the 5′SS subclasses impacted by human and plant LUC7 proteins are very similar overall, we do observe some subtler species-specific differences. For example, Arabidopsis LUC7RL promotes LH 5′SS with –1 G, rather than LH 5′SS with a –2 A/–1 G pair promoted by human LUC7L3 (Fig. 6D). Overall, our observations support that orthologous human and A. thaliana LUC7 proteins have largely retained their ancestral specificities for specific 5′SS subclasses over 1.5 billion years of evolution.

Loss of the LUC7L3 subfamily is associated with altered 5′SS composition

Our observations that LUC7L3-subfamily genes promote LH 5′SS and LUC7L2-subfamily genes promote RH 5′SS in both A. thaliana and humans implies that the specificity of these two subfamilies evolved prior to the divergence of animals/fungi from plants. Since none of the fungal species studied possess a LUC7L3 ortholog, it appears likely that this subfamily was lost from the fungal lineage soon after its divergence from animals, such that most modern fungal genomes have evolved largely in the absence of a LUC7L3 protein. To explore how this gene loss may impact the 5′SS composition within a species, we compiled annotated /GT 5′SS from the genomes of 33 eukaryotes, representing diverse plants, animals and fungi, and summarized the positional information content in each lineage (Fig. 7A). This analysis revealed distinct patterns of information content between 5′SS of plants, animals and fungi, with greater similarity between organisms within each clade. Consistent with previous studies, all analyzed fungi have strongly biased composition on the intron side of the 5′SS, with little or no bias at any exonic position²⁶. Meanwhile, all animals, except for 3 Protostomes (insects and nematodes) studied, have moderate biases at both exonic and intronic positions²⁷. Plant 5′SS have a slightly stronger bias on the exon side than animals but are otherwise more like animals than fungi²⁸.

Fig. 7: LUC7 proteins co-evolve with 5′SS motifs.

A Mean position-specific information content (calculated as in Irimia et al., 2019³⁸) of GT-type 5′SS motifs color-coded by animals (n = 13), fungi (n = 9), and plants (n = 11). Error bars represent standard error of the mean. B Density plot of the distribution of 5′SS subclass frequencies each organism, vertical dashed line reflects mean LUC7 score of the clade. C Frequency of dinucleotide features that define classical LH and RH 5′SS subclasses in 5 representative eukaryotes. Source data are provided as Source Data file.

Assessing 5′SS subclasses, we observed that plants and animals (other than Protostomes) have roughly equal frequencies of LH versus RH 5′SS. In contrast, the RH class is far more prevalent in modern fungi (Fig. 7B) and two intron-rich fungi were 4–15 times more likely to have a classical –1H/ + 5 G RH 5′SS than a –1 G/ + 5H LH 5′SS. Notably, we found that a member of a very early diverging fungal subphylum, Mucoromycota, has a LUC7L3-type gene and possesses a higher proportion of LH 5′SS than other fungi (Fig. 7C). Together, these data suggest long-term coevolution between LUC7 subfamilies and 5′SS subclasses, with depletion of LH 5′SS and concomitant loss of the LUC7L3-subfamily occurring early in fungal evolution.

Discussion

In both animal and plant lineages, our findings support a model where LUC7L2/LUC7L and LUC7L3 play significant roles in splicing regulation by tuning the recognition of distinct subclasses of 5′SS. Biochemical assays demonstrate that LUC7 proteins associate with U1 snRNP and crosslink near 5′SS, while structural evidence suggests that the highly conserved regions of the LUC7 proteins directly interact with the U1:5′SS duplex⁸. Our minigene experiments validate that the influence LUC7 proteins have on pre-mRNA splicing is dependent on specific nucleotide features of 5′SS, which are succinctly summarized by the 5′SS Balance score. Our domain-swapping experiments reveal that LUC7 structured regions are largely sufficient to confer specificity for 5′SS subclasses. The most straightforward model is a U1 stabilization model in which the LUC7L/LUC7L2 structured regions stabilize the interaction of U1 with RH 5′SS, and the LUC7L3 structured region stabilizes U1 interactions with LH 5′SS, thus influencing 5′SS choice. However, a model in which LUC7L/LUC7L2 and LUC7L3 preferentially destabilize interactions with LH and RH 5′SS, respectively, is equally consistent with our data.

Additional evidence consistent with our model comes from recent work in S. cerevisiae where humanized LUC7L/LUC7L2-ZnF2 rescued cell viability and promoted more flexible 5′SS recognition, which contrasts with the inviability of yeast that were complemented with humanized LUC7L3-ZnF2²⁹. These findings are consistent with our own experimental and evolutionary observations, in which the functionally similar LUC7L and LUC7L2 promote splicing of exons with fungal-like RH 5′SS, while the human LUC7L3 ZnF2 promotes recognition of 5′SS unlike those seen in the budding yeast genome.

While we favor a preferential U1:5′SS stabilization model, the manner in which LUC7 proteins mechanistically influence the pre-mRNA splicing reaction is not definitively known. The absence of LUC7 proteins in human spliceosome structures likely reflects their transient interactions as auxiliary splicing factors. Alternatively, our findings may imply that in vitro assembly of spliceosomes on single pre-mRNA substrates does not fully capture the diversity of human spliceosomal complexes. Nonetheless, using specific LH or RH pre-mRNA sequences in cryo-EM experiments may stabilize these intermediate splicing complexes and allow for the visualization of LUC7 proteins or other factors that impact the selection of 5′SS subclasses. Importantly, the LH and RH 5′SS subclasses described here have sequence features that would provide more stable interactions with the U5 and U6 snRNAs, respectively^14,17,30, so regulation of 5′SS subclasses at later stages of splicing is also possible.

Despite the deep conservation of splicing mechanisms involving the core spliceosome and snRNAs, splice site motifs vary substantially between distantly related species, likely reflecting clade-specific preferences for splicing substrates. For example, our results indicate that the absence of LUC7L3 genes may contribute to the preference for RH 5′SS in most fungi. When comparing RH 5′SS between species, we found that animal RH 5′SS—but not LH 5′SS or RH 5′SS from plants or fungi—have a clear preference for U1-complementarity at +7A and +8U bases (Fig. 1E, Figure S5F), suggesting that additional clade-specific preferences for splicing substrates are likely to exist. These variations may help explain the contrasting experimental results seen between humans and fungi regarding hyperstabilized U1 snRNP:5′SS interactions^31,32. Furthermore, the absence of U1-associated proteins like LUC7L3 from S. cerevisiae, C. neoformans, and S. pombe suggests that these common model organisms of splicing may not recapitulate all aspects of mammalian 5′SS choice.

The essentiality, large target space, and reasonably well-defined rules of pre-mRNA splicing make it a compelling target for the therapeutic modulation of gene expression. Our findings linking the activity of LUC7 proteins with specific 5′SS sequence features may have implications for the future advancement of small molecule regulators of splicing. For example, our proposed mechanism in which LUC7 proteins modulate splicing via recognition of U1:5′SS RNA duplex structures implies that LUC7 proteins will likely influence the specificity of small molecules that stabilize U1 snRNP:5′SS interactions. Furthermore, the synthetic lethal relationship between LUC7L and LUC7L2 suggests that splicing therapeutics specifically targeting the recognition of RH 5′SS will be more effective in AML patients with monosomy 7 or LUC7L2 mutations.

Methods

Cloning

Human LUC7L, LUC7L2 and LUC7L3 ORFs were amplified from human cDNA and cloned into pcDNA3.1(+)IRES GFP (Addgene #: 51406). Domain swap constructs were synthesized as gblocks from IDT and cloned into pcDNA3.1(+)IRES GFP. Exons and flanking intronic regions used for pSpliceExpress minigenes were PCR amplified from human male genomic DNA using primers with attB overhangs and subsequently recombined into pSpliceExpress using BP Clonase II (Thermo Fisher, cat no. 11789020).

Cell Line Authentication

HEK293T cell line authentication was performed at ATCC using STR profiling and referenced to ATCC’s internal database.

Cell Culture & Chemical Transfection

HEK293T RMCE cell lines were cultured in Advanced DMEM supplemented with 5% FBS, 25 mM HEPES and Glutamax and tested negative for mycoplasma. For each experiment, cells were plated 24 h in advance in 24-well plates. The following day, cells were transfected with 500 ng of 95:5 w/v of a cDNA overexpression vector and minigene reporter respectively using Lipofectamine LTX (Thermo Fisher, cat. no. 15338100).

Minigene Reporter Assay

RNA was extracted 24 h after transfection using Qiagen RNeasy Mini kit (cat. no. 74104) according to manufacturer’s instructions with the optional on-column DNAse digestion (cat. No 79254). RNA was eluted in nuclease free water and quantified using Nanodrop. For each RNA sample, we used 125 ng of RNA input into a 12.5 µL LunaScript Multiplex One Step Master Mix for RT-PCR (cat. no. E1555S). After PCR cycling, we mixed PCR samples with NEB 6X loading dye and loaded 5 µL of PCR products on a 3% agarose gel infused with ethidium bromide, which was run for 40 min at 150 V. Images were acquired using Azure Biosystems c600 with UV imager.

Percent Spliced In Calculation

Agarose gel images were manually quantified using chromatograms in ImageJ. Percent Spliced In values were calculated by taking the signal intensity of the larger band and dividing it by the sum of the signal intensity of the included product and the skipped product.

LUC7 Overexpression and RNA-seq

Flag-tagged human LUC7L, LUC7L2 and LUC7L3 ORFs were transfected into HEK293 RMCE cells as described above. RNA was extracted 24 h after transfection using Qiagen RNeasy Mini kit (cat. no. 74104) according to manufacturer’s instructions with the optional on-column DNAse digestion (cat. no. 79254). RNA was eluted in nuclease-free water and quantified using Nanodrop. Illumina-compatible libraries were prepared by MIT BioMicroCenter using NEB II Ultra Directional RNA with poly(A) selection and sequenced on NovaSeq 6000 with 2 × 150 bp reads.

Analysis of LUC7 paralog RNAseq data

We downloaded RNAseq data for LUC7L2 KO in HeLa and K562 erythroleukemia cells (GSE157917) and LUC7L, LUC7L2, LUC7L3 KD in K562 cells (E-MTAB-9709). For splicing analyses, we used rMATS 4.1.1³³ with default settings.

In any given RNA-seq experiment, only a subset of exons likely change in splicing, with most other exons having Δψ near zero. The Δψ estimates reported by rMATS, however, are direct summaries of the read counts and are affected by sampling variation in the read counts which may artificially inflate changes for exons with low read counts. Commonly, shrinkage estimates are used in differential expression analyses to account for this issue. Shrinkage considers the set of all effect sizes to constrain the noise in estimates from low read count events. Doing so, however, requires parameter estimates and associated uncertainty, which are not directly reported by rMATS.

For each exon, we reconstruct the effect size in log-odds scale δ from the rMATS read counts and approximate a standard deviation σ describing the uncertainty in δ using the rMATS p-value. We pass these effect sizes and standard deviations to ashr using the ‘normal’ option, which assume that the proportions of up- and down-regulated exons are equal (Stephens, 2017). Using these shrinkage estimates, we reconstruct an estimate of Δψ^∗ (see Supplemental Methods).

5′SS Enrichment Calculation

We used a Dirichlet-multinomial model to calculate the log-odds of whether a given 5′SS was more likely to be involved in a significantly included event vs significantly skipped event. For each dataset, we excluded all events with fewer than 10 junction-count reads on average across all samples. First, we combined the counts of 5′SS from significant included and skipped events (FDR < 0.1) and their respective background sets, which consisted of an equal number of unregulated 5′SS that were matched for both PSI and expression level, into a count matrix. To avoid division by zero, we added a pseudocount of 1 to every observed 5′SS sequence and accounted for class imbalance by dividing each column by a weight that reflected the fraction of included or skipped events. For example, if there were 4,000 significantly included events and 1000 significantly skipped events, the significantly included counts and their respective background set were divided by 0.8 and skipped events and their respective background were divided by 0.2.

Then, we used this count matrix as the alpha parameters for Dirichlet-multinomial model, and simulated drawing from the posterior distribution 2500 times. For each draw we calculated the log-odds of a given 5′SS being enriched in the included versus skipped set. The posterior distribution of log odds generated from the Dirichlet-multinomial model was used to calculate the posterior mean and the posterior standard deviation, which were both passed to ashr for shrinkage using the uniform option. From the output, we plotted the PosteriorMean estimates in 5′SS Enrichment plots, which can be directly interpreted as the log-odds of a given sequence occurring in the differentially included exon set over the differentially skipped exon set. For hierarchical clustering, we used scaled 5′SS enrichment scores to cluster 5′SS sequences by their activity across each LUC7 paralog RNA-seq dataset using Euclidean distance and ward.D2 linkage.

5′SS Enrichment meta-analysis

To calculate a single statistic describing the likelihood of a given 5′SS to be LH (LUC7L2-repressed/LUC7L3-promoted) versus RH (LUC7L2-promoted/LUC7L3-repressed) using all the LUC7 RNA-seq datasets (except for LUC7L KD), we performed a modified 5′SS enrichment analysis. For each dataset, we applied the same filtering criteria as above and used a significance threshold of FDR < 0.1 to group exons in differentially included, differentially skipped or background sets. We then set aside a random 12.5% fraction of differentially included or skipped events as the held-out test set (Fig. 2B) to evaluate our LUC7 score. Background sets of unregulated exons for differentially included or skipped exon sets were matched for PSI and expression and were threefold larger in size. For each of these 4 groups, we sampled 5000 exons with replacement. This approach was repeated for every RNA-seq dataset (except for LUC7L knockdown from¹¹). The events were then aggregated into a single table such that each RNA-seq data was equally represented in the final 5′SS enrichment analysis.

To account for opposing effects on LH and RH 5′SS subclasses for different experiments, the direction of differentially included and skipped events (and their respective background sets) from LUC7L3 KD, LUC7L OE and LUC7L2 OE analyses were flipped, such that included events represented RH 5′SS and skipped events represented LH 5′SS. Then we performed a 5′SS enrichment analysis as described above using a Dirichlet-multinomial model and simulated drawing from the posterior 10,000 times. As above, the distribution of log odds generated from the Dirichlet-multinomial model was used to calculate the posterior mean and the posterior standard deviation, which were both passed to ashr for shrinkage using the uniform option. From this output, we plotted the PosteriorMean estimates in 5′SS enrichment plots and the associated qvalue. LH 5′SS (LUC7L2-repressed/LUC7L3-promoted) were defined as 5′SS with negative meta-5′SS enrichment scores with qvalue < 0.01 and and RH 5′SS (LUC7L2-promoted/LUC7L3-repressed) were defined as 5′SS with positive meta-5′SS enrichment scores with qvalue < 0.01.

Position Weight Matrix and LUC7 score calculation

Sequence logos were created from LH and RH 5′SS sequences identified from meta-5′SS enrichment analyses. Individual PWM were created by calculating the observed frequency of a nucleotide at a given position, assuming a uniform nucleotide distribution. Pseudocounts of 0.1 were used to avoid division by zero. The LUC7 score was calculated by taking the ratio of the LUC7L2-promoted/LUC7L3-repressed RH PWM over the LUC7L3-promoted/LUC7L2-repressed LH PWM.

Free energy predictions of U1, U5, U6 and 5′SS sequences

We used ViennaRNA 2.5³⁴ to model free energy predictions between all 5′SS and the 5′end of U1 snRNA (ATACTTACCUG), U6 snRNA (ATACAGAGA) and U5 snRNA loop 1 (GCCUUUUAC) using RNAcofold with default parameters.

LUC7L2 eCLIP analysis

Publicly available LUC7L2 eCLIP data was downloaded from European Read Archive (PRJNA663333). 10 nucleotide UMIs were extracted from reads and appended to read name. Then Illumina adaptors were removed from reads using cut-adapt with the following settings (-a AGATCGGAAGAGCACACGTCTGAACTCCAGTCA \ --minimum-length 18 \ --quality-cutoff 6 \ --match-read-wildcards \ -e 0.1). This was performed twice to ensure the complete removal of adaptor sequences. Using STAR 2.7.3a, we created a genomic index using GRCh38.primary_assembly.genome.fa and gencode.v38.primary_assembly.annotation.gtf with --sjdbOverhang 65. Trimmed eCLIP reads were then aligned to this genome using STAR with default settings and resulting bam files were then deduplicated with umi-tools. Crosslink counting was restricted to cassette exons observed in LUC7L2 KO datasets around 250 nt windows surrounding four splice sites (upstream 5′SS, cassette exon 3′SS, cassette exon 5′SS and downstream 3′SS) and crosslinks in 10 nt bins were summed together.

For both cell lines, 2 pulldowns and 2 size-matched (SM) inputs were considered. Crosslink counts in each library were first normalized by library size. Cassette exons were then stratified into 10 bins using their 5′SS LUC7 score (1 = most LH, 10 = most RH) and the number of crosslinks for each LUC7 score bin were summed together. Position-based eCLIP enrichment was calculated as the ratio of normalized eCLIP crosslinks to normalized SM-input crosslinks at each 10 nt bin. eCLIP enrichment within each 10 nt bin was then correlated with the mean LUC7 score of binned cassette exons.

Phylogenetic and cross-kingdom splice site analysis

For our phylogenetic analysis, we manually chose 13 animals, 9 fungi and 11 plants from Esembl that represented diverse species in each eukaryotic kingdom. For each of these species, we used BLAST from UniProt to identify putative Luc7 paralogs and performed multiple sequence alignment using MAFFT³⁵. We then created a maximum likelihood tree with RAxML³⁶ and annotated the final protein tree using ggtree³⁷. For each of the species used in our phylogenetic analysis, we downloaded their genomes and associated gene annotation files from Ensembl and extracted their splice site sequences using bedtools 2.29.2. For each species, we filtered for only /GT splice sites and calculated the information content per position of the 5′SS.

Arabidopsis thaliana lines and growth conditions

All A. thaliana mutant lines used in this study are in Columbia (Col-0) background and were previously described²⁵. Seeds of wild type (Col-0), luc7 single mutants (luc7a-1, luc7a-2, luc7b-1 and luc7rl-1), luc7 double mutants (luc7a-1 luc7b-1, luc7a-2 luc7b-1, luc7a-1 luc7rl-1 and luc7b-1 luc7rl-1) and a luc7 triple mutant (luc7a-2 luc7b-1 luc7rl-1) were surface sterilized with chlorine-gas and then grown on half-strength Murashige Skoog (MS) plates containing 0.8% phytoagar in continuous light at 22 °C for 10 days. Seedlings were collected and flash-frozen in liquid nitrogen. Total RNA was isolated using RNeasy® Plant Mini Kit (Qiagen, cat. nos. 74904) according to the manufacturer’s instructions. mRNA stranded library preparation and sequencing (PE150) was done by Novogene (Cambridge, United Kingdom) using an Illumina Novaseq6000 system. To validate the alternative splicing found in RNA-seq, 1 µg of RNA was treated with DNase I (Thermo Fisher Scientific). cDNA synthesis was carried out using the RevertAid First Strand cDNA Synthesis kit (Thermo Fisher Scientific) with 100 µM oligo-dT. RT-PCR was then performed using Taq DNA Polymerase and the products were analyzed on a 2% agarose gel.

Patient stratification for TCGA analysis and differential splicing analysis

TCGA patient GDC-PANCAN.masked_cnv and normalized RNA-seq expression values were downloaded from (https://xenabrowser.net/datapages/?cohort=GDC%20Pan-Cancer%20(PANCAN)&removeHub=https%3A%2F%2Fxena.treehouse.gi.ucsc.edu%3A443). To assess per-sample CNV for LUC7 loci, we filtered for genomic segments spanning LUC7L2 (chr7:139341360-139422599) and LUC7L3 (chr17:50719544-50756213) for each patient sample. AML samples with average log2 CNV < –0.5 were considered to have LUC7L2 loss. To distinguish LUC7L2 low-expressing AMLs from AML samples expressing control levels of LUC7L2, we converted the normalized gene expression for LUC7L2 to Z scores and selected 16 patients with Z score < –1.5 were selected as LUC7L2 low expressing AMLs. Of the 151 LAML with RNA-seq data, 14 LUC7L2 low samples were available through GDC Data Portal. Differential splicing analyses using rMATs was performed to compare these LUC7L2 low expression samples with the remainder of the LAML cohort. As a control, we shuffled the patient labels and carried out the same differential splicing analysis. Statistical tests for Fig. 5E were conducted using a two-sided Wilcoxon rank sum test comparing the CNV distributions for LUC7L2 and LUC7L3 for up to 100 individuals sampled for each TCGA cancer-type and 100 samples selected randomly from all other cancers.

GSEA of LUC7L2^Low vs LUC7L2^Ctrl AML samples

Gene set enrichment was performed in the Xena browser web portal. The same 16 samples identified as LUC7L2 low-expressing AMLs were selected and compared to the remainder of the cohort with available data.

Dinucleotide 5′SS model and scoring

To identify and score dinucleotide features contributing to 5′SS inclusion or skipping, we calculated the frequency of adjacent and non-adjacent dinucleotide pairs in the 5′SS of differentially spliced exons, normalized by the background frequency from a set of unregulated exons. The log-odds ratio of a dinucleotide feature’s occurrence in differentially included versus skipped exons was calculated, and the variance was estimated by sampling the unregulated set 100 times. The resulting mean and standard deviation were shrunk using the ashr package, and only significant dinucleotides (q-value < 0.01) were retained. For each RNA-seq dataset, the posterior log-odds of dinucleotides in the 9mer were summed to score the 5′SS.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.