DeepExoMir: A Reproducible RNA Language Model Framework for CLIP-seq-Supported MicroRNA Target-Site Prioritization

Read full text Summarise with VIRAT
Authors: linwenh09, Chia‐Ni Hsiung, Wen-Yu Lien, Martin A. Sieber
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

DeepExoMir v1.1.0 is the IJMS submission snapshot for the manuscript “DeepExoMir: A Reproducible RNA Language Model Framework for CLIP-seq-Supported MicroRNA Target-Site Prioritization”. It replaces the placeholder v1.0.0 deposit (GitHub auto-archive) with a complete reproducibility package. What this deposit contains Source code at the GitHub master tag matching this release. Public reproducibility surface: deepexomir.config, deepexomir.predict, deepexomir.benchmark, and the CLI scripts scripts/reproduce_table1.py, scripts/reproduce_table3.py, scripts/calibration_analysis.py. Pytest smoke tests with pinned references (tests/test_reproduce_table3.py, tests/test_calibration.py, tests/expected_outputs/). Reproduces Table 3 + Figure 10 in roughly three minutes on a CPU-only laptop with only six packages installed (numpy, pandas, scikit-learn, matplotlib, pyyaml, pytest). Per-sample probability score files for v19 main, DeepExoMir-Lite, three retrain-from-scratch ablations (no-RiNALMo, no-conservation, no-structure), and eight miRBench baselines on the three CLIP-seq test sets (Hejret CLASH, Klimentová eCLIP, Manakov eCLIP). The four trained checkpoints used in Tables 1 and 3 (v19, v19_noStructure / DeepExoMir-Lite, v19_noConservation, v19_noRNALM). A CPU-only smoke-test Dockerfile at the repository root. Performance (paper Table 1) DeepExoMir v19 mean miRBench AU-PRC = 0.855 (Hejret 0.848, Klimentová 0.868, Manakov 0.848); beats all eight retrained miRBench baselines at paired sample-level bootstrap p < 0.001. DeepExoMir-Lite mean miRBench AU-PRC = 0.863, with ~35% fewer trainable parameters and no ViennaRNA dependency. Not included PCA-reduced RiNALMo embeddings cache (~32 GB) -- regenerable from the published RiNALMo-giga checkpoint via the cached PCA fit shipped under deepexomir/data/. Raw miRBench training/test parquet files -- available through the miRBench Python package. See RELEASE_NOTES.md inside the zip for the full diff against v1.0.0. License: MIT. Source: https://github.com/linwenh09/DeepExoMir

Fields of Study & Subjects

MicroRNA in disease regulation RNA Research and Splicing RNA and protein synthesis mechanisms

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN