The excision of introns from pre-mRNA is an essential step in mRNA processing. We developed LeafCutter to study sample and population variation in intron splicing. LeafCutter identifies variable splicing events from short-read RNA-seq data and finds events of high complexity. Our approach obviates the need for transcript annotations and circumvents the challenges in estimating relative isoform or exon usage in complex splicing events. LeafCutter can be used both to detect differential splicing between sample groups and to map splicing quantitative trait loci (sQTLs). Compared with contemporary methods, our approach identified 1.4–2.1 times more sQTLs, many of which helped us ascribe molecular effects to disease-associated variants. Transcriptome-wide associations between LeafCutter intron quantifications and 40 complex traits increased the number of associated disease genes at a 5% false discovery rate by an average of 2.1-fold compared with that detected through the use of gene expression levels alone. LeafCutter is fast, scalable, easy to use, and available online.
We thank X. Lan and other members of the Pritchard lab for helpful discussions and comments. This work was supported by a CEHG fellowship (Y.I.L.), the Howard Hughes Medical Institute (J.K.P.), and the US National Institutes of Health (NIH grants HG007036, HG008140, and HG009431 to J.K.P., and MH107666 to H.K.I.).
Integrated Supplementary Information
Bar plots showing the number of alternatively used junctions annotated from our GTEx analyses that were found in Intropolis6. phenopredict8 was used to predict the tissue type corresponding to the SRA samples analyzed in Intropolis. For each set of junctions, the proportion of junctions that were found (at least 1 read) in any SRA sample (Any), or found in samples which were predicted to be from testis (Testis) are highlighted. The predicted tissues with the highest number of supported junctions are colored in purple. Eighty-six percent of all novel alternatively used testis junctions from our LeafCutter analysis could be found in testis samples within SRA (not including GTEx).
Junctions in GTEx tissues. (a) Distribution of the number of different GTEx tissues in which junctions predicted to be absent, or present in three commonly-used annotation databases, could be detected. (b) Relative junction usage in multiple GTEx organs of annotated and unannotated junctions identiffed in four GTEx organs. (c) Distribution of LeafCutter clusters from GTEx samples in terms of their splicing types. Clusters with only annotated junctions and clusters with unannotated junctions were further separated.
Comparison between beta-binomial and Dirichlet-multinomial models for differential splicing analyses, performed on 10 male brain vs. heart samples from GTEx. Two approaches for combining per-intron p-values into cluster level introns are compared: Bonferroni correction and Fisher's combined test. Bonferroni is very conservative, as expected. Fisher's combined test has considerably lower power than the multinomial approaches. However, only v2 of the Dirichlet-multinomial (which uses a per intron concentration/overdispersion parameter) is well calibrated under permutations.
Cumulative distributions of differential splicing test P values (1-posterior for MAJIQ) for the comparison of all YRI versus CEU LCLs (red). The distribution of test P values for the permuted comparisons are also shown (black). *Cuffinks2 reports 19 signiffcantly differentially spliced genes in the 3 vs. 3 comparison, but none in the other comparisons.
Receiver operating characteristic (ROC) curves of LeafCutter, Cuffinks2, rMATS and MAJIQ for evaluation of differential splicing of genes with transcripts simulated to have varying levels of differential expression. Top panel shows ROC curves when excluding genes that were not tested by each respective methods. The bottom plot includes genes that were not tested in the calculation of true positive rate.
LeafCutter is effective even with as few as 8 samples. Here we performed differential splicing analysis of 4 male brain vs 4 male muscle samples, and compared to results using 220 samples. a) p-values under permutations are well-calibrated. b-c) p-values and effect sizes are highly correlated between the two sample size datasets. d) Signiffcant disparity in effect sizes between the two sample sizes is primarily driven by an intron being unique to a tissue when N = 8.