Abstract
RNA G-quadruplexes (rG4s) are key regulatory elements in gene expression, yet the effects of genetic variants on rG4 formation remain underexplored. Here, we introduce G4mer, an RNA language model that predicts rG4 formation, classifies rG4 subtypes, and evaluates the effects of genetic variants across the transcriptome. G4mer significantly improves accuracy over existing methods and uncovers subtype-specific differences in mutational sensitivity and evolutionary constraint, highlighting sequence length and flanking motifs as important rG4 features. Applying G4mer to 5'5′<math> <msup><mrow><mn>5</mn></mrow> <mrow><mo>'</mo></mrow> </msup> </math> untranslated region (UTR) variations, we identify variants in breast cancer-associated genes that alter rG4 formation and validate their impact on structure and gene expression. These results demonstrate the potential of integrating computational models with experimental approaches to study rG4 function, especially in diseases where non-coding variants are often overlooked. To support broader applications, G4mer is available as both a web tool and a downloadable model.