SpongeScope v1

Predicts whether a long non-coding RNA and a microRNA interact. 697,857 parameters; runs on CPU in about 5 ms per pair once the lncRNA's embedding is cached.

The number that matters

Evaluated cold-start: test lncRNAs share no MMseqs2 sequence cluster (80% identity) with any training lncRNA, and 10% of miRNA seed families are held out entirely.

group AUROC AUPR positive rate ECE
cold lncRNA 0.841 0.303 0.096 0.009
seed-controlled 0.753 0.312 0.163 0.016
cold lncRNA + miRNA 0.853 0.297 0.087 0.021

precision@10 per lncRNA is 0.274 against a 0.096 base rate β€” test ten candidates, get about 2.7 real ones instead of one.

What this model does not establish

A precision improvement is not statistically established. Against a gradient-boosted baseline on the same features, AUROC gains are significant on 3 of 6 paired comparisons and positive on all 6; no AUPR gain is significant. With 149 test lncRNAs the AUPR confidence interval is Β±0.03-0.05, so differences below ~0.05 cannot be resolved on this benchmark at all. The honest claim is that it ranks better, not that it is more precise.

Most of the signal is in the seed-site scanner, not the network. A gradient-boosted model over the same candidate sites reaches 0.824 AUROC. The neural model's contribution is the cross-attention between site context and the miRNA, worth about 0.025 AUPR; removing it drops performance to baseline.

Two negative results, reported rather than buried

Tissue context adds nothing. Eleven expression features (co-expression across organs, co-occurrence, breadth, abundance) were built and tested. Every univariate AUROC fell between 0.487 and 0.508 β€” chance. A gradient-boosted model given the features directly showed no significant gain. Two structural reasons: the labels come from CLIP experiments in cell lines, so organ-level co-expression cannot explain detection; and the negative sampling already draws from real, abundant miRNAs, so co-expression is equally high on both sides by construction.

Site-level supervision made it worse. Labelling each candidate site by overlap with a real CLIP peak (122,141 site labels) and adding an auxiliary head: all six paired comparisons negative, one significantly so. The auxiliary loss was learned β€” site loss fell to 0.150 β€” so the model can locate sites; knowing which one is real simply does not improve the pair call. The scanner already contains the true site for 94.9% of positive pairs, so localization was never the bottleneck.

Training data and its biases

ENCORI CLIP-supported lncRNA-miRNA pairs (15,962 positives over 1,313 lncRNAs and 638 miRNAs), GENCODE 50 transcripts, miRBase mature sequences. Negatives at ~8:1, half seed-matched: the lncRNA contains a 7mer site for the miRNA but has no CLIP support. Without those hard negatives, counting seed matches alone scores 0.88 AUROC, because CLIP peaks carry seed matches by construction β€” a circularity that inflates every published number built on this kind of label.

A negative label means "not observed", not "does not interact". Coverage is biased toward well-studied miRNAs and toward whatever cell lines were assayed.

Intended use

Ranking candidates for experimental follow-up. Not a substitute for experimental validation, and not calibrated for clinical or diagnostic use of any kind.

Reproducing

Benchmark, splits and both negative results are reproducible from the repository: cluster-based splits are deterministic given the seed, and every metric is computed by one shared evaluation module with lncRNA-level bootstrap confidence intervals.

Downloads last month
34
Safetensors
Model size
698k params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Space using sawcheet/spongescope 1