TriTower-m6Am

N6,2'-O-dimethyladenosine (m6Am) sits at exactly one place in a transcript: the first transcribed nucleotide, and only when that nucleotide is an adenosine. That positional constraint makes the prediction problem unusual. In the DTC-m6Am benchmark every sample is a 41-nucleotide window centered on an adenosine, positives and negatives alike, so nothing about the candidate base separates a methylated site from an unmethylated one. All of the discriminative signal lives in the 20 flanking nucleotides, and every predictor is really a choice about how to encode that flank — k-mer statistics, hand-crafted physicochemical features, or a plain one-hot code.

TriTower-m6Am, from Kaifeng Xiong and Jianhua Jia at Jingdezhen Ceramic University, turns that choice into the experiment. Three independent towers read the same window through three representations. A semantic tower pools per-position embeddings from a frozen RNA-FM backbone; a sequential tower runs a bidirectional LSTM over a raw one-hot code and keeps the hidden state at the central position; a structural tower builds a 41-node graph from the RNAfold-predicted secondary structure and processes it with a relational graph convolutional network. The three probabilities are fused by AUC-weighted soft voting.

Keeping the towers separate through training is what makes the model auditable: each can be ablated cleanly, each fusion weight is a number a user can read off, and because the graph convolution gives every edge type its own weight matrix rather than shared attention, the ablation that validates the design also reports which structural relationship a prediction rests on.

#Key Features

• Three complementary views of one window: Pairwise CKA similarity between the tower representations runs 0.55–0.67 — correlated, but far from redundant.
• Position-aware BellPooling: The semantic tower collapses 41 positions into one vector using learned weights initialized from a bell-shaped prior centered on the candidate site.
• Typed structural edges: The graph carries backbone (i, i+1), base-pairing, and neighborhood (i, i+2) edges as three distinct relations, so their contributions stay separable rather than averaged into one structural score.
• Readable fusion weights: Voting weights come out near-uniform — RNA-FM 31.6%, BiLSTM 34.5%, RGCN 33.9% — giving a per-prediction contribution trace without a post-hoc explainer.
• Small trainable footprint: With the backbone frozen, the three heads total 329,666 parameters and score a sample in 6.18 ms at 42 MB of peak GPU memory.

#Technical Details

RNA-FM (rna_fm_t12, pretrained on 23.7 million ncRNA sequences) is used strictly as a precomputed embedding extractor, producing a 41 × 640 matrix per window that a convolutional head compresses to 48 dimensions. The BiLSTM tower uses two layers of 64 units per direction; the structural tower stacks two RGCN layers of width 64 over 21-dimensional node features. Training uses the DTC-m6Am benchmark of Huang et al. — 40,700 windows at roughly 10:1 negative-to-positive ratio, with a balanced 640-window independent test set — under 5-fold stratified cross-validation with class-weighted, label-smoothed BCE. On the held-out set the ensemble reaches AUC 0.776, MCC 0.440, sensitivity 0.888, specificity 0.522 and F1 0.750, against DTC-m6Am's 0.765 / 0.411 / 0.800 / 0.530 / 0.740 re-evaluated from its official checkpoint. Individual towers reach AUC 0.768 (RGCN), 0.762 (RNA-FM) and 0.756 (BiLSTM), and parameter-matched "wide" single towers of 318K–549K parameters top out at 0.767 — the paper's strongest evidence that the gain is a representation effect rather than a capacity effect.

#Applications

The model targets the screening step ahead of antibody-based mapping, ranking candidate cap-proximal adenosines so that miCLIP or m6Am-seq effort goes to the likeliest sites first and recovering roughly nine additional true sites per hundred relative to the prior best method. Prevalence sets the practical ceiling: evaluated against BCA-motif negatives drawn from the GENCODE v44 transcriptome at a 1:100 imbalance, AUC rises to 0.951 while average precision falls to 0.315, and even at specificity 0.989 that leaves 342 false positives among 32,000 negatives. Transcriptome-wide use therefore requires pre-filtering on cap-proximal annotation.

#Impact

The ablations are the part most likely to outlast the benchmark numbers. Backbone adjacency alone recovers AUC 0.757 of the full graph's 0.768, while base-pairing edges shift AUC by 0.002 — a hierarchy consistent with m6Am being installed co-transcriptionally, before the nascent transcript has folded, and one that a shared-attention graph network would have averaged away. The limits are stated by the authors: specificity of 0.522 means roughly half of non-m6Am adenosines on the balanced test set are still misread, training and evaluation are human-only, the graph depends on RNAfold predictions that are about 73% accurate per base pair, and the positives in every split trace back to the same batch of miCLIP experiments. Code is public under the MIT license with the benchmark data included; the pretrained checkpoints the repository advertises have not been posted, and the training script is the only entry point, with no inference path for new sequences.