A closed-form multi-region miRNA target binding affinity atlas covering 3'UTR, 5'UTR, and CDS regions, with computational m6A modulation flag. Each region applies a closed-form biophysical weakening factor: 3'UTR is the canonical post-transcriptional regulatory region; CDS uses a three-state Markov chain steady-state ribosome-displacement model; 5'UTR uses a scanning-ribosome displacement model with cap-proximity exclusion mask. All region pipelines pass their validation gates against literature-curated miRNA-mRNA binding measurements. The atlas also annotates per-site DRACH consensus motif overlap as a computational m6A modulation flag (9.33M DRACH sites across the canonical transcriptome).

What this is

An open dataset, not a defensive publication. The bundle is released under CC-BY-4.0 (atlas database, README, CHANGELOG, citation metadata); the predictor implementation and underlying derivation methodology that produced the binding affinity numbers are held proprietary and are not distributed with the deposit.

The architectural commitment is sharp: for every (mature human miRNA, canonical human protein-coding transcript, region) tuple, the atlas reports a 3'UTR-equivalent binding free energy and dissociation constant, a region-specific weakening factor, an effective Kd after weakening, a tier classification, and per-site m6A overlap counts where DRACH motifs overlap detected seed footprints. There is no machine learning, no curve fitting, and no parameter tuning against any of the validation panels reported below.

What's in the database

Each row carries the 3'UTR-equivalent binding free energy and Kd, the region-specific weakening factor, the effective Kd after weakening, a site count and site-type breakdown, a tier classification (high / medium / low / none / excluded), the m6A overlap count, and the m6A penalty applied to the prediction. The composite primary key is (mirna_id, transcript_id, region).

Validation

3'UTR pipeline regression (M.2 / M.3)

Direct binding affinity Pearson R = 0.847 against the 224-record Wee 2012 + Salomon 2015 single-molecule kinetics panel; M.3 atlas validation AUC = 0.9339 multi-site / 0.8947 single-site against TarBase v9.

CDS pipeline — new in v3

Tested against an n=25 literature panel of measured CDS-targeted miRNA-mRNA fold-change pairs (Schnall-Levin 2011, Hausser 2013, Mayr 2007 HMGA2, Sampson 2007 MYC, Duursma 2008 DNMT3B, Forman 2008 DICER1, Webster 2009 EGFR, Xu 2009 OCT4/SOX2/KLF4, and others).

The CDS weakening model is a three-state Markov chain steady-state with rebinding kinetics under ribosome displacement. Constants are the AGO·miRNA on-rate from Wee 2012 single-molecule kinetics, the per-codon ribosome arrival rate from ribosome profiling consensus, and a geometric ribosome footprint residence time. No parameter tuning against the validation panels.

5'UTR pipeline — new in v3

Tested against an n=18 scanning-zone + n=2 cap-proximity literature panel (Lee 2009, Lytle 2007, Forman 2008, Caporali 2011, Christoffersen 2010 MYC, Ørom 2008 RPL5/RPS19 enhancement special case, Mayr 2007 HMGA2 5'UTR, and others).

The 5'UTR weakening model inherits the CDS three-state structure but with shorter τ_occupy = 0.6 s reflecting the 40S+eIF4F scanning footprint at ~50 nt/s vs the 80S elongating footprint at ~15 nt/s. Sites within 40 nt of the cap are sterically excluded by the cap-binding eIF4F + 40S preinitiation complex and return Kd_eff = ∞ (tier "excluded").

CLIP-seq concordance

Atlas v3 high-tier predictions tested against ENCORI / starBase aggregated CLIP-seq peaks (PAR-CLIP + HITS-CLIP + CLASH + iCLIP, 108 datasets, 37 studies). CLIP-positive labels require clipExpNum ≥ 5 per gene; miRNAs with fewer than 10 CLIP-positive genes in the atlas universe are excluded for power.

Atlas v3 vs CLIP, full transcriptome

Median across n = 602 miRNAs that pass the coverage threshold:

Atlas v3 vs TargetScan v8, apples-to-apples

Restricted to the TS-conserved subset where TargetScan v8 has predictions; n = 254 miRNAs:

TargetScan's context++ score is purpose-built to discriminate within the TS-conserved subset, and on that universe it discriminates better than atlas v3's thermodynamic Kd. Atlas v3's distinct contribution is absolute-Kd predictions across the full transcriptome (~13× the gene coverage), including non-conserved, synthetic, and modified-base sequences that TargetScan structurally cannot evaluate.

Per-miRNA results in clipseq_bench/clipseq_concordance.tsv (602 rows) and clipseq_apples_to_apples.tsv (254 rows). Two figures and a full report (CLIPSEQ_REPORT.md) ship in the bundle alongside re-runnable acquisition + analysis scripts.

Drug-design demonstration

Five clinical-stage miRNA therapeutics were queried against atlas v3. For each candidate, every published on-target gene was checked for recovery at high or medium tier (Kd_eff ≤ 1 µM) in the relevant region.

Three of the five candidates are halted or discontinued in development; the published target lists remain stable and the recovery test is unaffected by clinical status.

The notebook drug_design_demo/drug_design_demo.ipynb walks through each candidate, lists best-region best-Kd_eff per literature target, and reports the top off-target predictions per candidate. Curated candidate list with citations and last-verified date in clinical_candidates.tsv.

Multi-region query patterns

The atlas supports complementary query modes:

-- Mode 1: by miRNA × region (raw effective Kd)
SELECT mirna_name, gene_name, transcript_id, region, Kd_eff_M, n_sites, tier
FROM atlas_v3
WHERE mirna_name = 'hsa-miR-21-5p' AND region = '3UTR'
ORDER BY Kd_eff_M ASC
LIMIT 100;

-- Mode 2: multi-region for one (miRNA, gene)
SELECT region, n_sites, weakening_factor, Kd_eff_M, tier, n_m6a_overlaps
FROM atlas_v3
WHERE mirna_name = 'hsa-let-7a-5p' AND gene_name = 'HMGA2'
ORDER BY Kd_eff_M ASC;

-- Mode 3: m6A-modulated targets
SELECT mirna_name, gene_name, region, Kd_eff_M, n_m6a_overlaps
FROM atlas_v3
WHERE n_m6a_overlaps > 0 AND tier IN ('high', 'medium')
ORDER BY n_m6a_overlaps DESC, Kd_eff_M ASC;

-- Mode 4: strongest CDS-targeted miRNAs per gene
SELECT mirna_name, Kd_eff_M, weakening_factor
FROM atlas_v3
WHERE gene_name = 'MYC' AND region = 'CDS'
ORDER BY Kd_eff_M ASC LIMIT 20;

Tier classification

Independent verification

The atlas SQLite is the authoritative artifact. Its SHA256 is pinned in the README and the deposit metadata, so the predictions are stable and checkable.

1. Direct binding affinity. Download the Wee 2012 and Salomon 2015 panels from the cited papers' supplements, join against Kd_3utr_eq_M for the corresponding (miRNA, target sequence) pairs, compute Pearson R yourself. Expected: 0.847.
2. TarBase v9 discrimination. Pull TarBase v9 from the DIANA-Lab portal, build a matched-negative panel via dinucleotide-shuffled UTRs, score the atlas's dG_3utr_eq column against the positive/negative split, compute AUC. Expected: 0.9339 multi-site / 0.8947 single-site.
3. CDS panel reproduction. Take the 25-target CDS panel published in this Drop's bench scripts, score each (miRNA, gene, CDS) tuple's Kd_eff_M against the panel's measured fold-change, verify all three validation gates pass.
4. 5'UTR panel reproduction. Take the 20-target 5'UTR panel published in this Drop's bench scripts, score each (miRNA, gene, 5'UTR, distance-from-cap) tuple's Kd_eff_M, verify all four validation gates pass including the cap-proximity exclusion.
5. CLIP-seq concordance reproduction. Re-run the CLIP-seq bench scripts (fetch_encori_clipseq.py, run_clipseq_bench.py, compare_targetscan.py) against ENCORI / starBase to regenerate the full-transcriptome and TS apples-to-apples concordance numbers. Expected: median top-100 enrichment 2.83×, median AUC 0.619 across n=602 miRNAs; atlas AUC 0.519 vs TargetScan AUC 0.595 on the n=254 TS-conserved subset.
6. Drug-design recovery reproduction. Run drug_design_demo.py against the bundled atlas SQLite and the curated clinical_candidates.tsv. Expected: 37/37 (100%) literature on-target genes for the five clinical candidates recovered at high or medium tier.

Download

Atlas SQLite is served directly over HTTPS from Coracle Research's Cloudflare R2 bucket. Bundle metadata (README, LICENSE, CHANGELOG, CITATION, and the validation/demo artifact bundles) are deposited at Zenodo with a permanent DOI for citation.

SQLite:  https://atlas.coracleresearch.com/mirna_atlas_v3.sqlite
Size:    69.78 GB
SHA256:  bc8c5ab614a5fab286239bf213e4210a9a62513c225403af7728bdf5f653d2ea
License: CC-BY-4.0
Zenodo:  https://doi.org/10.5281/zenodo.19932285

Direct download supports HTTP range requests for resumable transfers. Verify the SHA256 after download to confirm integrity.

Predictor implementation and underlying methodology held proprietary. Inquire for commercial licensing or for custom-atlas commissions covering mouse / rat orthologs, per-paralog AGO1/3/4 atlases, per-tissue accessibility maps, experimentally-validated m6A peak overlay, ASO/siRNA design panels, or ceRNA network propagation.

Citation

If you use this atlas in published work, please cite:

coracle-mirna v3 (2026). Multi-region human miRNA target binding
affinity atlas with m6A modulation flag.
Zenodo. https://doi.org/10.5281/zenodo.19932285

Foundational references:

• Wee, L. M., Flores-Jasso, C. F., Salomon, W. E., Zamore, P. D. (2012). Argonaute divides its RNA guide into domains with distinct functions and RNA-binding properties. Cell 151(5), 1055-1067. doi:10.1016/j.cell.2012.10.036.
• Salomon, W. E., Jolly, S. M., Moore, M. J., Zamore, P. D., Serebrov, V. (2015). Single-molecule imaging reveals that argonaute reshapes the binding properties of its nucleic acid guides. Cell 162(1), 84-95. doi:10.1016/j.cell.2015.06.029.
• Hausser, J., Syed, A. P., Bilen, B., Zavolan, M. (2013). Analysis of CDS-located miRNA target sites suggests that they can effectively inhibit translation. Genome Research 23(4), 604-615. doi:10.1101/gr.139758.112.
• Lytle, J. R., Yario, T. A., Steitz, J. A. (2007). Target mRNAs are repressed as efficiently by microRNA-binding sites in the 5' UTR as in the 3' UTR. PNAS 104(23), 9667-9672. doi:10.1073/pnas.0703820104.
• Mayr, C., Hemann, M. T., Bartel, D. P. (2007). Disrupting the pairing between let-7 and Hmga2 enhances oncogenic transformation. Science 315(5818), 1576-1579. doi:10.1126/science.1137999.
• Skoufos, G., Kakoulidis, P., Tastsoglou, S., et al. (2024). TarBase-v9.0 extends experimentally supported miRNA-gene interactions to cell-types and virally encoded miRNAs. Nucleic Acids Research 52(D1), D304-D310. doi:10.1093/nar/gkad1071.
• Agarwal, V., Bell, G. W., Nam, J. W., Bartel, D. P. (2015). Predicting effective microRNA target sites in mammalian mRNAs. eLife 4, e05005. doi:10.7554/eLife.05005.