Read the knowledge graph
22,456 entities joined by nine relation types: validated targeting, pathway and GO membership, disease association and protein interaction.
miRNA
Mature sequence
Positions 2–8 form the seed, the part that does most of the target recognition.
Hover a base to see its region and partner; click a region to highlight it on the sequence and the hairpin.
Score distribution
Export
Ranked target genes
Held-out genes are pairs that appear in no training, validation or test split, so their scores are true predictions.
Gene
Transcript and site map
Hover the transcript for coordinates. Hover the density track to see which miRNAs have a site at each position; click or drag on it to filter the lanes and the table below.
Ensembl 116
Darker blue marks stronger site types (8mer, then 7mer-m8, 7mer-A1); grey is 6mer. Positions are in the representative Ensembl 116 3′UTR and match the model’s site counts for every pair.
Predicted regulators
What drives this score
Net contribution of each channel to the logit, including its background term. The intercept b₀ is shared by all pairs. Values are the mean of the five models.
Top evidence shared across models
Open the Evidence tab for the full ledger.
How much to trust it
Signals that help judge this prediction before planning an experiment.
Rules behind these signals are listed under Method.
Evidence map
From the miRNA through its other known targets to the candidate gene. Line width follows |φ|; dashed lines lower the score.
Protein interactionPathway / GOShared diseaseSequence
Score ledger
The logit is the exact sum of these lines. Each bar starts where the running total left off. Switch between the mean and each model above.
φ
Independent data
Experimental results that agree or disagree with the prediction. None of them were used for training or model selection.
How a score is built
An HGT encoder reads the knowledge graph and forms a query for the candidate pair. The query decides which evidence to keep and how much each channel weighs, but it has no direct path to the score. The logit is always this sum:
ℓmg = b₀ + Σc ( φbgc + Σi∈c φi ) + φseq
Masking evidence i at inference changes the logit by exactly −φi. Deleting the supporting edges of the top-φ evidence and re-encoding lowers the score more than a matched random edge in 95% of tested positives.
Protein interaction
Other known targets of the miRNA that physically interact with the candidate (STRING v12.0, combined score ≥ 700). Acts mostly as a channel-level calibration.
Pathway / GO
Terms that contain the candidate and at least one other known target. About 93% of this channel's variance traces to single terms.
Shared disease
Diseases associated with both endpoints after one-step parent expansion, weighted by information content. An association, not a cause.
Sequence
Counts of 8mer, 7mer-m8, 7mer-A1 and 6mer sites in the representative 3′UTR, plus UTR length. Folding and accessibility are not used.
Reading the trust signals
Each badge on the site follows a fixed rule.
In k / 5 modelsNumber of the five split models whose evidence budget (K = 16 per channel) includes this item for this pair. The filter keeps items with k ≥ 3.
Top in k / 5Number of models that rank this item first by φ within its channel. Models are not independent replicates, so this measures stability, not significance.
Mean ledgerEach line is averaged over the five models, counting 0 where a model did not use the item, so the lines still sum to the mean logit.
Calibrated precision (planned)Share of validation pairs in the same score band that are known targets, under the standard and degree-matched protocols.
Popularity cautionShown when an endpoint is above the 80th percentile of training degree, since part of its score may reflect how well studied it is.
Data and citation
Release v1.0. Download files will be listed here once the release is published.
Downloads
Precomputed scores
PlannedMean probability and per-model scores for all 883 × 3,303 miRNA–gene pairs.
Score ledgers
PlannedPer-evidence φ, background terms and the sequence term for every scored pair, as JSON.
Splits and degree-matched sets
PlannedThe fixed test set, five validation splits and their degree-matched counterparts.
Arm-annotation flags
Planned137 validated labels flagged as likely arm-annotation ambiguity, with site and CLASH support.
Code
PlannedModel, training and audit scripts.
Cite MechKG
Author list and venue will be filled in on publication.
@misc{mechkg2026,
title = {MechKG: Interpretable miRNA--Gene Interaction
Prediction via Additive Evidence Readout
on a Knowledge Graph},
author = {To be added},
year = {2026},
note = {Manuscript in preparation}
}Source data in this release
Versions as used in the paper.
| Layer | Source | Version or filter |
|---|---|---|
| miRNA–gene targeting | miRTarBase, miRATBase | strong experimental evidence only |
| Pathways | MSigDB C2:CP | v2025.1 |
| GO annotation | GOA human, GO ontology | released 1 June 2025 |
| miRNA–disease | HMDD | v4 |
| Gene–disease, disease hierarchy | CTD, MEDIC | curated direct associations |
| Protein interaction | STRING | v12.0 physical subnetwork, combined score ≥ 700 |
| Mature miRNA sequences | miRBase | release 22 |
| 3′UTRs | Ensembl canonical protein-coding transcripts | release 116; 3,291 genes with a usable 3′UTR |
| Expression | TCGA PanCanAtlas | EB++AdjustPANCAN RNASeqV2, pancanMiRs EBadj |
Structure layer
Shown on miRNA, gene and record pages to help read a prediction. None of it enters the score.
| Layer | Source and method | Coverage |
|---|---|---|
| Precursor hairpins | miRBase human precursor sequences; RNAfold 2.7.2 minimum-free-energy structure at 37 °C | 709 precursors |
| Mature positions on the hairpin | Exact match of each mature sequence; arm from the 5p/3p name, otherwise from the main terminal loop | all 883 miRNAs, 1,002 positions |
| Transcript models | Ensembl 116 GTF, canonical transcript; CDS includes the stop codon; 3′UTR length checked against the model input | 3,260 genes |
| 3′UTR accessibility | RNAplfold 2.7.2 unpaired probability, W 80, L 40, u 1, 37 °C | 3,291 genes |
| Local folds | RNAfold 2.7.2 on the strongest canonical site ± 40 nt | 50,344 ledger pairs |