G4mer

A canonical RNA G-quadruplex is easy to write down as a pattern: four runs of at least three guanines separated by loops of one to seven nucleotides. Rule-based scorers such as cGcC and G4Hunter work from that pattern, as do the structural classes annotating rG4-seq data. The trouble is that experimentally detected rG4s routinely break the rule — a G-run interrupted by a bulge, a loop far longer than seven nucleotides, a stack of only two quartets — and whether a guanine-rich stretch actually folds depends on sequence the pattern never inspects at all, including the flanking bases on either side of the motif.

G4mer replaces the pattern with a learned representation. It was developed by the BioCiphers group at the University of Pennsylvania, with structural and reporter validation by collaborators at the University of North Carolina at Chapel Hill, and published in Nature Communications in November 2025 after a 2024 preprint. The model is a fine-tune of mRNAbert, a BERT encoder the same group pretrained with masked language modeling over the whole human transcriptome. Because that encoder already carries a general representation of mRNA sequence, the rG4 heads can be fit on roughly five thousand labeled windows and still transfer to sequence lengths and assay protocols they never saw.

The second half of the model's remit is variant effect. Scoring a reference window and its mutated counterpart yields a signed change in rG4 probability, applied by the same checkpoint — no refitting per cohort — to gnomAD, ClinVar UTR variants and Penn Medicine BioBank exomes.

#Key Features

• Flanking sequence enters the prediction: rG4 windows are extended into transcript context out to 70 nt, and integrated-gradient attribution shows guanines outside the motif contributing positively to the call — an effect pattern-matching scorers cannot express.
• Eight-way subtype classification: a second fine-tuned head sorts a predicted rG4 into canonical, long-loop, bulged, two-quartet, G≥40%, potential G-quadruplex, potential G-triplex, or unknown, which is what makes subtype-specific mutational sensitivity measurable.
• Variant scoring by paired inference: the difference between reference and alternate scores ranks a variant as rG4-breaking or rG4-stabilizing, turning a structure predictor into a transcriptome-wide annotation track for noncoding variation.
• Stable across sequence length: performance holds from 14-nt oligos to kilobase transcripts, where the convolutional baselines degrade sharply as inputs grow.
• Interactive web tool: precomputed predictions per gene and transcript, plus gnomAD variant annotations, are browsable and downloadable without a local install.

#Technical Details

mRNAbert is a compact BERT — 6 transformer layers with 6 attention heads each — trained by masked language modeling on GENCODE GRCh38 mature transcript sequences tokenized into overlapping 6-mers, capped at 510 tokens plus [CLS]/[SEP]. Pretraining ran 200,000 steps on four NVIDIA Tesla P100 GPUs over three weeks. Fine-tuning used 5,454 unique rG4 sequences from the rG4-seeker reprocessing of HeLa rG4-seq data, mapped back to GRCh38.v29 transcripts to recover flanks and capped at 70 nt; the released checkpoints are roughly 46M parameters each. On ten-fold cross-validation of subtype classification G4mer reaches 0.74 accuracy, 0.91 ROC-AUC and 0.68 PR-AUC against 0.54, 0.80 and 0.44 for the CNN-based rG4detector. On 795 held-out G4RNA database sequences spanning 24 experimental protocols and 14–1368 nt, it holds PR-AUC near 0.85 at every length band, reaching 0.82 on both circular dichroism sequences (median 30.5 nt) and in cellulo luciferase sequences (median 210 nt). Estimated against rG4-negative regions of highly expressed HeLa transcripts, the false discovery rate is 1.86%.

#Applications

The immediate use is triaging noncoding variation. UTR variants are numerous, poorly annotated, and frequently returned as variants of uncertain significance; an rG4-breaking score gives a mechanistic hypothesis to rank them by. The paper works this end to end: a PheWAS over Penn Medicine BioBank genotypes surfaced an rG4-breaking variant in the EPN3 5' UTR associated with breast cancer (p = 7.4e-7), alongside a variant in the mismatch-repair gene MSH6. Both predictions were then tested at the bench — circular dichroism confirmed that the EPN3 variant disrupts an existing rG4 while the MSH6 variant induces one, and dual-luciferase assays showed the corresponding changes in downstream protein output.

#Impact

G4mer's contribution is showing that a pretrained RNA encoder generalizes across rG4 assay chemistry, the practical obstacle in this field: rG4-seq, SHAPE, DMS probing and circular dichroism disagree on what counts as a folded quadruplex, and models fit to one protocol have not carried over. Coupling that to population-scale variant catalogues connects RNA secondary structure to human genetics evidence directly. The scope is human only — pretraining, fine-tuning and every downstream analysis use human sequence and HeLa-derived labels. The analysis covers single-nucleotide variants alone, and the experimental validation, being reporter- and spectroscopy-based, reaches only a handful of loci. Weights, code and data are released under a University of Pennsylvania academic license permitting non-commercial use only.