LeafCutter2

Summary

LeafCutter2 keeps LeafCutter's annotation-free intron clustering but adds a classifier that labels each splice junction productive or unproductive (premature-termination-codon-bearing, hence degraded by nonsense-mediated decay), using only start and stop codon positions as annotation input.[1] This turns a purely descriptive splicing quantification into a functionally interpreted one: an unproductive splicing event predicts reduced protein output of its host gene, which is a mechanistic hypothesis a plain intron-usage ratio cannot make.

Source Status

Preprint: this page is compiled from the April 2025 bioRxiv preprint, which had not been peer reviewed at time of ingestion. confidence: medium accordingly.

Method

The junction classifier runs a breadth-first search over exon paths joined by observed splice junctions, computing whether each path yields a valid open reading frame terminating at the annotated stop codon. Junctions that can participate in a complete protein-coding path are annotated protein-coding; those that never can remain unproductive.[1] Dynamic programming avoids re-walking redundant paths.

One workflow change from LeafCutter matters: because unproductive transcripts are degraded by NMD and therefore depleted in polyA RNA-seq, LeafCutter2 adds rarely used junctions back into clusters after the standard filtering step and recomputes relative usages — otherwise the events of interest would be filtered away as noise.[1]

Four published PTC-position rules affecting NMD efficiency are additionally implemented; all four predicted a junction's enrichment in nascent versus steady-state RNA-seq. For example, PTCs in long exons (≥407 nt) versus short exons (<100 nt) differed by log2 fold-change 0.29 in that enrichment, indicating PTCs in long exons trigger NMD less efficiently.[1]

Validation

• Across HeLa SMG6/SMG7 double-knockdown and LCL nascent-versus-steady-state RNA-seq, LeafCutter2 extracted ~1 million unique junctions, classifying 269,331 (~28%) as unproductive, 33% as coding, and 39% as too close to a UTR to call.[1]
• Junctions classified unproductive were overwhelmingly enriched in NMD-factor knockdown and in nascent RNA-seq, exactly as predicted if they are genuine NMD substrates.[1]
• Against GENCODE v46, 97% of LeafCutter2-"productive" junctions lie within protein-coding transcripts, and unproductive ones fall almost exclusively in transcripts annotated nonsense_mediated_decay (48%), protein_coding_CDS_not_defined (33%) or retained_intron (7%).[1]
• Performance matches using GENCODE's own NMD annotation directly, while increasing the number of annotated unproductive junctions 12.1-fold — the practical argument for the method.[1]
• Usage stratification is informative: 71% of the most-used junctions are productive, but only 14% of the rarest-used are.[1]

Genetic analysis: u-sQTLs

LeafCutter2 distinguishes p-sQTLs (variants affecting clusters producing only productive isoforms) from u-sQTLs (variants affecting mixed clusters that can yield unproductive isoforms). Across 49 GTEx tissues:[1]

	Unique clusters	Median per tissue	Genes affected
p-sQTL	32,334	3,082	9,387
u-sQTL	16,816	1,800	5,107

Fewer u-sQTLs are found, which the authors attribute to depletion of unproductive junction reads in polyA RNA-seq.[1]

The mechanistic signature is clean and is the paper's central genetic result:

• Alleles increasing unproductive splicing have strong negative effects on host gene expression, significant in the majority of the 49 tissues. p-sQTL effects show no such correlation in any tissue.[1]
• u-sQTLs show stronger eQTL enrichment than p-sQTLs across all 49 tissues, and stronger pQTL enrichment in brain.[1]
• No enrichment with H3K9ac histone acetylation QTLs — consistent with NMD being post-transcriptional, and a useful negative control.[1]
• 225 of 2,897 GWAS loci (7.8%) across 20 complex traits colocalise with at least one u-sQTL in at least one tissue. Example: rs7212573, where the minor allele increases unproductive splicing of ASB16, decreases ASB16 expression, and increases bipolar disorder risk.[1]

Alzheimer's disease application

Applied to ROSMAP brain RNA-seq (DLPFC n = 806, anterior cingulate n = 603, posterior cingulate n = 449) and integrated with FunGen-xQTL eQTL/pQTL/haQTL data, PTWAS identified 17 AD risk genes driven by unproductive splicing; three — PICALM, TSPAN14, *CASS4 — also showed strong SuSiE-COLOC colocalisation between u-sQTL and AD GWAS plus colocalisation with at least one eQTL context.[1] TSPAN14 effects were consistent across all three brain regions and most pronounced in excitatory neurons; CASS4* showed microglia-specific expression with a distinct secondary eQTL signal.[1]

Note on strength of evidence: mediation analysis confirmed splicing significantly affects TSPAN14 expression, but the proportion of expression variance explained was modest, and CASS4 showed no significant mediation. The authors present these as regulatory hypotheses supported by colocalisation, not as established causal chains.[1]

Cross-species finding

Applied across human and six non-human species, LeafCutter2 found a conserved developmental-stage-specific increase in unproductive splicing during testis maturation in every species examined.[1] This is relevant to LeafCutter's earlier observation that testis is the major outlier for unannotated splicing — the two findings are consistent with permissive splicing in spermatogenesis.

See Also

• LeafCutter — the parent method; LeafCutter2 reuses its clustering unchanged.
• RNA-seq-Derived Molecular Phenotypes — where u-sQTLs sit among the phenotype definitions available from RNA-seq.
• Statistical Colocalization — the SuSiE-COLOC and ColocBoost analyses used here.
• AlphaGenome — predicts splice junctions but not their productive/unproductive status; see the gap noted on the phenotypes page.

Citations

[1] Buen Abad Najar, C. F., Feng, R., Dai, C., Fair, B., Hauck, Q., Li, J., Cao, X., Dey, K. K., De Jager, P., Bennett, D., ADSP FGC, Liu, X., Wang, G. & Li, Y. I. (2025). Genetic and functional analysis of unproductive splicing using LeafCutter2. bioRxiv (preprint, posted 8 April 2025). Supports: the BFS junction-classification algorithm and rarely-used-junction re-addition; NMD-rule findings; all validation figures against knockdown/nascent RNA-seq and GENCODE; p-sQTL/u-sQTL counts and the expression-correlation and eQTL/pQTL/haQTL enrichment results; the 7.8% GWAS colocalisation figure and ASB16 example; the ROSMAP/AD analysis and PICALM/TSPAN14/CASS4 results; the cross-species testis finding. Location: Results ("Overview of LeafCutter2"; "Robust detection of NMD splice junctions"; "Analysis of unproductive splicing in non-human species"; "Unproductive splicing mediates the effect of genetic variants on gene expression across human tissues"; "Genetic basis of unproductive splicing reveals novel regulatory mechanisms in Alzheimer's disease"); Figs. 1–5; Code availability. Source paper: 2025.04.06.646893v1.full.pdf