miRacle — Inspect the evidence

SEQUENCE ANALYSIS

Start with a sequence.

Paste your sequence, upload a FASTA file, or explore the MIR21 example below.

Your sequence stays on this device.

A hairpin is a starting point. A score does not establish expression, precise processing, targets, function, or disease relevance.

CONTROL COMPARISON

Put a candidate in context.

Run the identical folding, feature, and model pipeline on a known miRNA region and a negative control. A demonstration contrast, separate from the held-out benchmark.

Run an analysis and select a candidate to inspect its evidence chain.

HELD-OUT EVALUATION

A benchmark beyond the example

Positive families, genomic regions, and Rfam ncRNA families are separated across splits. Similarity matching is excluded from the classifier.

Precision0.774

Recall0.889

PR-AUC0.912

Decision threshold81.2/100

Logistic regression was selected by validation PR-AUC. The held-out test set did not determine model selection.

HELD-OUT PERFORMANCE

Model comparison

00.250.500.751.00

Logistic regressionSelected

0.912

Random Forest

0.877

Extra Trees

0.870

Histogram Gradient Boosting

0.902

MFE/nt baseline

0.590

Precision–recall area under the curve, on a shared 0–1 scale. Higher is better. Model selection used the validation split.

CLASSIFICATION OUTCOMES

Confusion matrix — Logistic regression

	Predicted negative	Predicted positive
Actual negative	True negative63497.8% of row	False positive142.2% of row
Actual positive	False negative611.1% of row	True positive4888.9% of row

CorrectErrorColor intensity shows the proportion of each actual class. Threshold 0.8117.

All benchmark measurements

Held-out method	Precision	Recall	F1	PR-AUC	ROC-AUC	FP / 100 negatives
Logistic regression	0.774	0.889	0.828	0.912	0.981	2.160
Random Forest	0.824	0.778	0.800	0.877	0.982	1.389
Extra Trees	0.778	0.778	0.778	0.870	0.977	1.852
Histogram Gradient Boosting	0.828	0.889	0.857	0.902	0.981	1.543
MFE/nt baseline	0.455	0.833	0.588	0.590	0.952	8.333

Negative-set stress tests

Negative subset	Precision	Recall on shared positives	PR-AUC	FP / 100 negatives
genomic negatives	0.787	0.889	0.914	2.407
rfam ncrna negatives	0.980	0.889	0.986	0.926

{
  "positive_total": 492,
  "rfam_candidate_like_total": 3279,
  "rfam_family_total": 441,
  "train": {
    "positive": 371,
    "genomic_negative": 1113,
    "rfam_negative": 742
  },
  "validation": {
    "positive": 67,
    "genomic_negative": 335,
    "rfam_negative": 134
  },
  "test": {
    "positive": 54,
    "genomic_negative": 540,
    "rfam_negative": 108
  },
  "negative_sources": "hg38 chr22 train/validation; hg38 chr21 test; Rfam seed structural RNAs split by family"
}

METHOD & INTERPRETATION

What the evidence means

Hairpin geometry and folding energy identify thermodynamically plausible local stem-loops. Many genomic sequences can form hairpins, so these signals are necessary but not sufficient.

The supervised score combines sequence and structure features learned from curated human precursors, candidate-like genomic negatives, and structurally plausible Rfam non-miRNA decoys. Logistic regression remains the selected model because it achieved the best validation PR-AUC; Random Forest, Extra Trees, and histogram gradient boosting are retained in the benchmark. It is a triage score, not proof or a calibrated probability of biological function.

Curated-reference similarity is displayed separately. A close match supports “known-like”; a distant match does not prove novelty.

Cross-species similarity compares the folded candidate with separate mouse, orangutan, chimpanzee, gorilla, and limited bottlenose-dolphin precursor panels. It can support family-level sequence similarity but is not a locus-level phylogenetic-conservation calculation and does not alter the classifier score.

Protein-region preprocessing is coordinate-aware rather than sequence-guessed. With a length-matched hg38 FASTA interval, the default masks only annotated protein-coding CDS. An optional aggressive mode masks every exon of a protein-coding transcript, but this can remove genuine miRNAs in untranslated exonic regions. Without trustworthy coordinates, no gene/exon claim is made.

Rfam conflict evidence asks whether the sequence closely resembles another structured RNA family. A close match is a warning, while a distant match cannot exclude every non-miRNA RNA class.

Protein-machinery evidence distinguishes canonical roles from candidate-specific experiments. DROSHA, DGCR8, XPO5, DICER1, TARBP2, AGO2, and TNRC6A are not inferred binding partners merely because a candidate forms a hairpin. A “supported interaction” requires RIP or biochemical binding/processing evidence for the matched external control. “Tested—not required” describes pathway dependency and does not prove that physical contact is impossible.

Literature cards are short, manually curated summaries shown only after a strong human reference match. They report evidence for the known miRNA in the cited experimental context; they are not evidence about expression or function of the submitted sample.

Missing evidence includes transcription, tissue expression, precise mature/star processing, locus-level evolutionary conservation, RISC loading, mRNA targeting, pathway effects, and disease causality.

The intended decision is: which few candidate loci should a researcher investigate experimentally first?

A reproducible local analysis

The browser uses ViennaRNA 2.7.2, scanning 60, 70, 90, and 110 nt windows. The native Python application remains the scientific reference and also supports RNALfold. The exported classifier preserves the trained scaling parameters, coefficients, and validation threshold.

Engine: ViennaRNA-2.7.2-RNAfold-windows-v1
Data release: v1-0cf79300764342d9