m6A-FORM

Almost every m6A site predictor works the same way: take a candidate adenosine, cut a fixed-length window centered on it, and classify that window on its own. Only a small fraction of transcriptomic adenosines are ever methylated, so most of that inference lands on positions that were never plausible candidates — and scoring each window independently treats several methylated adenosines inside one enriched region as unrelated events rather than a local cluster.

m6A-FORM changes the unit of analysis. MeRIP-seq, the workhorse of transcriptome-wide m6A profiling, returns methylation-enriched peak regions rather than single bases, and m6A-FORM treats those peaks as the biological prior: it pretrains on peak-derived RNA sequence, then recasts base-resolution site calling as token-level sequence labeling over a whole peak, so every candidate adenosine in a region is predicted jointly in one pass. The same encoder is then reused for two questions site callers do not address — which m6A reader, writer, or eraser binds a site, and whether a YTHDF2-bound site is associated with transcript decay.

The model comes from Ting-He Zhang, Sumin Jo, Shou-Jiang Gao, and Yufei Huang at the University of Pittsburgh and the UPMC Hillman Cancer Center Cancer Virology Program. It is a separate model from M6AFormer, a CNN-Transformer site classifier from a different group, despite the near-identical name; the two share neither authors, architecture, nor released artifacts. m6A-FORM and m6AFORM are the same model.

#Key Features

• Peak-level pretraining corpus: Masked language modeling over 24,909,934 mRNA sequence windows derived from 22,548,379 MeRIP-seq peaks across 143 human studies, concentrating the encoder on methylation-enriched contexts rather than generic transcript sequence.
• Sequence labeling instead of windowed classification: m6A-FORM-sites labels each 3-mer token by its central nucleotide, so one forward pass over a peak calls every adenosine in it and captures dependencies among neighboring candidates.
• Span-aware k-mer masking: Overlapping 3-mers share two of three bases, so masking a token alone is trivially reversible; masking its immediate neighbors too forces the model to use broader context.
• One encoder, three regulatory tasks: m6A-FORM-sites, m6A-FORM-RWEBind, and m6A-FORM-decay are fine-tuned from the same pretrained weights, spanning site discovery, regulator binding, and RNA fate.
• Attribution-guided interpretation: Integrated Gradients over the fine-tuned heads recovers regulator-specific RBP neighborhoods, tying predictions back to interpretable sequence context.

#Technical Details

m6A-FORM is a BERT-like transformer encoder following the DNABERT recipe: RNA is tokenized into overlapping 3-mers at stride 1, with a maximum input of 512 tokens, and pretrained with masked language modeling (80/10/10 mask/random/keep, no next-sentence prediction). Pretraining ran up to 120,000 AdamW steps on NVIDIA L40S and A100 40GB GPUs, with the masking probability rising from 0.15 to 0.25 over the final 20,000. Site calling is trained against 131,320 base-resolution sites supported by at least two independent technologies across m6A-Atlas v2.0 and GLORI, over 528,452 peaks from five human cell lines split 70/10/20 at the peak level. m6A-FORM-sites reaches a mean PR-AUC of 0.635 and ROC-AUC of 0.988 — at least 0.14 PR-AUC above DeepSRAMP — while running roughly 10.7-fold faster than adenosine-centered inference, and an ablation trained from scratch on the identical setup loses most of that PR-AUC gain. m6A-FORM-RWEBind covers 104,244 sequences across 19 regulators (13 readers, 5 writers, 1 eraser) from POSTAR3 CLIP-seq, improving median PR-AUC by 0.09 over iDeepS and RNAProt with significant gains on 15 of the 19. m6A-FORM-decay, trained on YTHDF2 PAR-CLIP peaks paired with knockdown half-life changes, reaches PR-AUC 0.167 against 0.150 for m6ABERT-decay.

#Applications

The intended use is functional annotation of the human m6A epitranscriptome from MeRIP-seq data a lab already has: convert peaks to single-base calls, ask which reader or writer is likely to engage each site, and flag the sites whose reader engagement points toward decay. The authors run this at scale on 67 MeRIP-seq samples spanning 24 human tissues, producing 317,986 non-redundant sites and a recurrence analysis that isolates 19,631 tissue-conserved sites — enriched near stop codons, more heavily methylated, more often predicted to bind YTHDF2, and embedded in denser RBP neighborhoods than sporadically detected sites. It is a hypothesis-generation tool for RNA biologists and cancer researchers prioritizing methylation events for wet-lab follow-up.

#Impact

m6A-FORM is an argument that domain-matched pretraining beats generic sequence pretraining for epitranscriptomics: corpus, input representation, and fine-tuning formulation are all shaped by how MeRIP-seq data actually arrive, and the from-scratch ablation shows the pretraining is doing real work. Its dependence on antibody-based enrichment is also its main limitation, since peak calling inherits MeRIP-seq's biases in sequencing depth, antibody specificity, and tissue coverage, and the two-technology support requirement pushes genuine low-signal sites into the negative set. Everything reported is human, sequence-only, and from a preprint that has not completed peer review. The code repository and the tissue-landscape data compendium are public, but neither the pretrained encoder nor the fine-tuned task heads are distributed as downloadable checkpoints.