Review Papers Claim by Claim

 Quick Explanation Copied

 Long Explanation

What MiREx does

MiREx augments the Xpresso CNN backbone—one-hot encoded 10,500 bp TSS-centered sequence plus eight mRNA half-life features Xpresso CNN paper (backbone used)—with a per-gene vector of miRNA expression levels for miRNAs that target that gene, using TargetScan CWCS to select ten high-impact miRNAs per model TargetScan CWCS methodology. Datasets come from TCGA/GDC across four cancer sites: lung (982 samples), kidney (871), breast (1076), corpus uteri (536), restricted to 18,347 protein-coding genes matching Xpresso's split (16,348 train / 1,000 val / 999 test) MiREx original paper.

Key results and caveats

Using all 243 miRNAs (AllMirna) consistently lowered R² vs Xpresso—authors attribute this to sparse per-gene miRNA vectors. The subset-based miREx beat Xpresso with p<0.05 in lung (LUAD, LUSC, mean) and corpus uteri, but was not statistically significant in breast (p=0.06) and mixed in kidney (no improvement for KIRC; mean improvement non-significant) Per-tissue significance results. Selected miRNAs (e.g., miR-23a, miR-506) have known cancer roles, supporting biological plausibility—but this is post-hoc interpretation, not mechanistic validation.

Critical appraisal

• Circular feature selection risk: the ten miRNAs are chosen by correlation with Xpresso residuals on the same task, then evaluated on the same gene split—data leakage into model design is not fully excluded, and no external validation cohort is used.
• Effect size unreported: absolute ΔR² values are not stated in text, only significance claims with 95% CIs in figures; gains appear marginal.
• Cancer-only scope: predictions are tumor bulk means, not per-sample; normal-tissue or pan-cancer generalization is untested.
• Proxy quality: CWCS is a context-score proxy for repression, not measured miRNA–mRNA interaction strength; experimental CLIP-based resources (e.g., miRBench's CLASH/chimeric eCLIP datasets) could refine targeting inputs miRBench experimental miRNA-target datasets.

Overall: a sensible, incremental extension with credible data availability (GitHub + Zenodo), but modest, tissue-inconsistent gains mean the core claim—that selected miRNA features materially improve sequence-based mRNA prediction—should be treated as preliminary. What would change this conclusion: reported ΔR² effect sizes, leakage-free external replication, and per-sample (not per-site-mean) prediction.

 BGPT Paper Review

Study Novelty

Incremental architecture extension of Xpresso with miRNA features; no prior SOTA work incorporated miRNA expression into sequence-based mRNA prediction, but the conceptual step (adding auxiliary regulatory features to a CNN) is familiar.

Scientific Quality

Sound comparison protocol (matched Xpresso split, 95% CIs, multiple runs) but feature selection on the evaluation task risks leakage, absolute effect sizes are absent from text, and gains are inconsistent across tissues; no external validation.

Study Generality

Demonstrated only on four TCGA cancer bulk datasets with tissue-mean targets; normal tissue, per-sample prediction, and other regulatory layers untested.

Study Usefulness

Public code/data (GitHub, Zenodo) and a working recipe for injecting post-transcriptional features into CNN expression predictors; useful for follow-up work but limited clinical or design utility given marginal gains.

Study Reproducibility

Code on GitHub, data on Zenodo/Xpresso repositories, explicit hyperparameters (batch 32, lr 0.0005, SGD, early stopping 20) and documented splits enable replication; exact ΔR² reporting is thin.

Explanatory Depth

Mechanistic reasoning is plausible but shallow: sparsity explains AllMirna failure, residual–CWCS correlation motivates miRNA selection, yet no ablation isolates CWCS vs expression contributions per miRNA.