BioGraphX-RNA

Where an RNA transcript ends up inside a cell — nucleus, ribosome, endoplasmic reticulum, exosome — is a primary determinant of what it does, and mislocalization is implicated in cancer and neurodegeneration. Computational predictors of RNA subcellular localization have converged on end-to-end deep sequence models that classify accurately but reveal little about why a transcript was assigned to a compartment, and they typically ignore the base-pairing and stacking interactions that shape a transcript's folded form.

BioGraphX-RNA inverts that emphasis. Developed by Abubakar Saeed and Waseem Abbas at Government College University Faisalabad and posted to bioRxiv in February 2026, it translates a primary nucleotide sequence into an explicit multiscale interaction graph using biophysical rules — no 3D structure and no thermodynamic folding software required — and reads out 149 interpretable descriptors from that graph. Those hand-engineered features are combined with frozen embeddings from RiNALMo through a gated fusion layer, so that every prediction carries a quantified split between the sequence-derived and structure-derived evidence behind it.

The model is the RNA counterpart of the authors' earlier protein-oriented BioGraphX encoding. Its stated goal is not to top a leaderboard but to show that explicit biophysical constraints can match a much larger black-box predictor while staying auditable: the trained head carries only 2.05 million parameters on top of the frozen language model.

#Key Features

• Rule-based interaction graphs: Sequences are encoded as nucleotide networks over five interaction types — canonical Watson–Crick pairs, wobble pairs, non-canonical pairs, base stacking, and phosphate backbone connectivity — under linear-distance constraints and a minimum loop size of four nucleotides.
• 149-dimensional physicochemical descriptor: The encoding yields 65 graph-topology features, 15 hybrid cooperative-interaction features, 27 compartment-specific knowledge-guided profiles, 17 frustration (conformational-conflict) features, and 25 global biophysical statistics.
• Gated fusion with a frozen language model: A sigmoid gate weights the graph branch against 1,280-dimensional RiNALMo embeddings per molecule, rather than concatenating them blindly.
• Per-molecule modality attribution: Gate values quantify reliance on structure versus sequence. miRNA sits near equilibrium at 49.1% physics weight, while mRNA leans on sequence at 39.8% physics weight, consistent across compartments.
• SHAP-based feature interpretation: Periodic GC content (GC autocorrelation at lag 1) ranks as a top positive feature for nuclear and ribosomal localization, and structural accessibility correlates with exosome targeting.

#Technical Details

The network is a dual-branch multilayer perceptron. A 1,280-dimensional RiNALMo branch and a 149-dimensional physics branch each pass through Linear–BatchNorm–ReLU–Dropout (rate 0.484) into a shared 512-dimensional space; the concatenated 1,024-dimensional vector is modulated element-wise by a sigmoid gate and classified by a 1,024→256→9 head producing multi-label scores over nine compartments. RiNALMo runs in inference mode only, with sliding windows of 2,046 tokens at stride 1,536 and mean-pooled CLS representations for long transcripts.

Training used the human portion of the DeepLocRNA dataset, spanning nine compartments (Exosome 22,335 sequences, Nucleus 13,352, Cytoplasm 10,026, Ribosome 5,226, Membrane 3,356, Cytosol 2,587, ER 1,977, Microvesicle 1,958, Mitochondrion 33), with mRNAs truncated to their first and last 2,500 nucleotides, five-fold cross-validation, focal loss, AdamW, and per-class thresholds tuned by Matthews correlation. CD-HIT-EST filtering at 80% identity confirmed cross-fold homology had negligible effect (maximum AUROC change 0.0100). Against DeepLocRNA, macro AUROC is 0.7575 ± 0.0054 on mRNA (0.7493), 0.9228 ± 0.0137 on miRNA (0.8681), and 0.5600 ± 0.0191 on lncRNA (0.5786). The graph features alone reach 0.9396 ± 0.0045 on miRNA, exceeding both RiNALMo alone and an RNAfold partition-function graph baseline (0.9139 ± 0.0138).

#Applications

The primary use is annotating localization for uncharacterized human transcripts and prioritizing candidates for imaging or fractionation follow-up, with the gate values and SHAP attributions turning each call into a testable hypothesis about which biophysical signal drives it. Because the encoder is a standalone stage that emits a fixed 149-feature table per sequence, it can also be reused as a compact structural descriptor in other RNA models, and the low compute footprint makes transcriptome-scale scoring feasible without GPU-heavy infrastructure.

#Impact

BioGraphX-RNA is an early-stage academic preprint, not yet peer reviewed, and its contribution is methodological: it shows that interpretable biophysical graph features can substitute for learned structural representations on well-structured RNAs. The limitations are reported plainly by the authors. lncRNA localization remains hard, with the frozen language model alone giving the best ranking performance and macro AUROC below DeepLocRNA's. In a strict blind cross-species test, models trained on human data transfer poorly to mouse — macro AUROC 0.4889 for mRNA, 0.5195 for miRNA, and 0.5434 for lncRNA, with only nuclear retention rising meaningfully above chance — so the biophysical features capture human-specific compositional bias rather than conserved localization grammar. The public repository ships the encoding pipeline, training script, inference script, and encoded feature tables; the per-fold checkpoints the inference script expects are not included in it.