miR‑RadScore: An open, reproducible microRNA signature to predict tumour radio-sensitivity and guide DNA‑damage–response therapy selection

1. Overview

Rationale and importance: Radiotherapy cures many localized solid tumours, yet inter‑patient variability in radio-sensitivity leads to over‑treatment for some and failure for others. MicroRNAs (miRNAs) regulate DNA‑damage–response (DDR) pathways and are measurable in tumor tissue and biofluids. A parsimonious, validated miRNA panel that predicts radio-sensitivity would enable treatment personalization and rational combination with DDR‑targeted agents.

Funding, facilities, and approvals: The project is a computational analysis using open datasets. It requires only compute resources and standard open‑source tooling. No new identifiable human subject data will be generated. Public data with appropriate licenses will be used, so institutional ethics approval is not required. If any controlled‑access dataset is included, only de‑identified records will be accessed under the provider’s data‑use agreement.

Estimated timeline: Five months total, with a sixth month reserved for dissemination and community review.

Data‑sharing commitment: All code, environment files, and analysis notebooks will be released on GitHub under a permissive license, with archived snapshots on Zenodo. A reproducible OSF preregistration and a results repository will host the analysis plan, inclusion criteria, model objects, and documentation. Where licenses permit, harmonized expression matrices and derived scores will be deposited as well.

2. Introduction

Background: Radio-sensitivity is governed by DDR capacity, chromatin state, and cellular stress programs. miRNAs tune these pathways by repressing multiple targets per molecule, creating compact yet robust regulatory signatures. Prior studies have reported miRNA profiles associated with radiation response in several cancers, but results are fragmented across platforms and cohorts. There is no widely accepted, open, cross‑cohort miRNA score with transparent code, prospectively registered methods, and independent validation.

Gap: Existing signatures often overfit to single studies, lack rigorous external validation, or are not reproducible due to closed code and unclear preprocessing. A small, platform‑robust panel that generalizes across cohorts and cancer types would be immediately useful for hypothesis generation and for designing prospective studies.

Objective: Build and openly release a 8–15 feature miRNA score, miR‑RadScore, that predicts adverse outcomes after radiotherapy and correlates with DDR pathway activity. Package the model in an easy‑to‑use web app and Python/R library for research and clinical prototyping.

Research questions and hypotheses.

H1. Pre‑treatment miRNA expression predicts time‑to‑event outcomes after radiotherapy with a concordance index (C‑index) ≥ 0.65 in internal cross‑validation.
H2. Constraining features to miRNAs with credible DDR target sets improves performance over agnostic selection by ΔC‑index ≥ 0.05.
H3. The trained score generalizes to independent external cohorts with hazard ratio per standard‑deviation increase ≥ 1.5 for adverse events.
H4. miR‑RadScore adds incremental prognostic value beyond standard clinical covariates, measured by net reclassification improvement and decision‑curve analysis.

3. Methods

Participants/Sample

Inclusion criteria: Public cancer cohorts that include

miRNA expression measured on microarray or sequencing,
patients treated with radiotherapy as a definitive or adjuvant modality, and
survival or recurrence endpoints with follow‑up.

Primary tumour types: Prostate, head‑and‑neck squamous cell carcinoma, cervical cancer, and rectal cancer. If prostate‑specific datasets are limited, head‑and‑neck and cervical cohorts serve as primary discovery with prostate used for external validation.

Data sources: GEO, ArrayExpress, TCGA legacy resources where permitted, and controlled‑access repositories when feasible under data‑use terms. Clinical covariates will be harmonized where available.

Materials and Procedures

Pre‑registration: The full analysis plan, including dataset inclusion/exclusion, preprocessing rules, and model comparison grid, will be preregistered on OSF before fitting any model to outcome labels.

Data curation and harmonization:

Import raw or processed miRNA expression matrices with consistent gene identifiers.
Apply within‑study normalization suited to platform, then correct batch effects across studies with ComBat or similar methods after blocking by study label.
Map clinical variables to a common schema. Define primary endpoint as time to recurrence or disease‑specific death after radiotherapy.

Feature priors: Construct a DDR‑prior list of miRNAs with experimentally supported targets in ATM/ATR, homologous recombination, non‑homologous end‑joining, chromatin remodelling, p53 signalling, and senescence programs. This list is used for constrained models and pathway interpretation.

Modelling pipeline:

Split each study into folds that respect patient‑level independence.
Fit Cox proportional hazards models with elastic net penalty; compare to gradient boosted survival trees.
Determine the final panel size using nested cross‑validation to avoid leakage.

External validation: Hold out at least one entire study for each tumour type for leave‑one‑study‑out testing. No parameter tuning will touch held‑out cohorts.

Clinical utility analyses:

Fit multivariable models with clinical covariates (stage, grade, age, margin status) alone and with miR‑RadScore added.
Compute ΔC‑index, integrated Brier score change, calibration curves, and decision‑curve analysis.

Mechanistic association:

Quantify DDR pathway activity using gene‑set variation analysis on matched mRNA data when available, or via miRNA target enrichment.
Test correlation between miR‑RadScore and DDR activity, tumour proliferation, and immune infiltration scores.

IF–THEN contingent analysis decisions:

IF within‑study batch correction inflates variance or degrades calibration, THEN revert to study‑specific scaling and perform stacking of study‑specific models.
IF proportional hazards assumption is violated for the final model, THEN fit time‑varying coefficient models and report restricted mean survival time differences.
IF external validation C‑index < 0.60, THEN reduce panel size and refit with stronger regularization and DDR priors only.

Sample size justification

This is secondary analysis of existing cohorts. To limit overfitting, we will target 10–15 effective parameters with ≥10 events per parameter across the discovery set. The nested cross‑validation and leave‑one‑study‑out design further controls optimism bias. We will cap the final panel to 8–15 miRNAs to respect the events‑per‑variable rule if event counts are lower.

Planned Analyses

Primary: Cox elastic net model for event‑free survival after radiotherapy.
Secondary: Competing‑risks regression where non‑cancer deaths are frequent.
Subgroup: Stratified analyses by tumor site and by treatment context.
Sensitivity: Robustness to platform (microarray versus sequencing) and to missing covariates via multiple imputation.
Model explanation: Stability selection, SHAP value summaries, and target‑pathway enrichment for interpretability.

Ethics and Data Management

Ethical status: Only publicly available or controlled‑access de‑identified datasets will be used. No re‑contact, intervention, or new specimen collection is involved.

Confidentiality: All controlled‑access data will remain on encrypted storage with access limited to the project lead. Derived results shared publicly will be aggregate and non‑identifiable.

Data sharing and archiving.

Code and containers on GitHub with Zenodo DOI.
Preregistration and protocol on OSF.
Harmonized matrices and model objects deposited where licenses allow, with reproducible scripts to regenerate otherwise.

4. Pilot Data

Feasibility has been verified by an initial scoping review that identified multiple candidate cohorts in prostate, head‑and‑neck, and cervical cancer with miRNA profiling and radiotherapy outcomes. Test runs of the harmonization code on a small subset demonstrate stable mapping of miRNA identifiers across platforms and successful batch‑aware normalization. No outcome‑guided modeling has been performed prior to preregistration.

5. Budget

Cloud Compute and Storage Credits ($1,500)
Covers CPU/GPU time for preprocessing, cross-validation, and application hosting over five months.
Data Management and Controlled-Access Processing ($600)
Includes secure storage, encrypted transfer, compliance overhead, and EGA/dbGaP processing costs if required.
Independent Code Audit Micro-Grants ($600)
Two small bounties for external reviewers to reproduce results and submit formal audit reports.
Open Science Dissemination ($500)
Preprint formatting, professional figure design, DOI registration fees, and OSF storage extensions.
Web App Deployment ($300)
Domain registration, HTTPS setup, and VPS hosting for 12 months to run the interactive score calculator.
Conference Dissemination ($1,000)
Virtual registration fees or poster printing to present results to the research community.
Contingency – $500
Reserved for unforeseen compute spikes or replacement dataset costs.
Total: $5,000

6. References

References will be compiled later on. They will include foundational reviews on DDR and radio-sensitivity, primary studies linking miRNAs to radiation response, and best‑practice methods for survival modelling, batch correction, and decision‑curve analysis.

miR‑RadScore: An open, reproducible microRNA signature to predict tumour radio-sensitivity and guide DNA‑damage–response therapy selection

miR‑RadScore: An open, reproducible microRNA signature to predict tumour radio-sensitivity and guide DNA‑damage–response therapy selection

Funding Progress

Related Funding

Request for Proposals - Bioinformatics & Genomics Research

Nonprofit Recipient

Topics

DOI