Nucleotide-level evolutionary constraint and population variant tolerance at human splice junctions scale continuously with transcript architecture representation

Abstract The spliceosome must accurately recognize splice donor and acceptor sites to ensure correct pre-mRNA processing. While canonical splice sites are evolutionarily constrained, the extent to which this constraint scales with a junction's representation across a gene's transcript architectures remains incompletely understood. Here, I present a systematic analysis of 485,994 unique splice junctions (derived from 3,903,107 transcript-junction observations) across 19,129 human protein-coding genes. Using a continuous "architecture fraction" metric representing the proportion of distinct annotated transcript architectures containing a given junction I demonstrate that evolutionary conservation, measured by phyloP scores across 100 vertebrate species, scales positively with architectural representation. This association is strongest at the immediate splice boundaries (positions − 1 to + 2) and persists after controlling for gene-level confounders using within-gene centered regression models. Position-specific analyses across 80 nucleotide positions (− 20 to − 1 and + 1 to + 20) revealed that conservation in exonic flanks exhibits a pronounced 3-bp periodicity, reflecting dual constraints of splice-site recognition and reading-frame maintenance. Extending this framework to population genetics, I show that junctions with lower architecture fractions harbor a 2.1-fold higher density of population variants (1000 Genomes Phase 3) and that local variant burden significantly predicts exon skipping (p = 3.14 × 10⁻²). Furthermore, 69.5% of low-architecture splicing events preserve the reading frame. These findings establish that the evolutionary and population-genetic constraint acting on human splice junctions is highly heterogeneous and directly proportional to their functional representation across the transcriptome.