Review Papers Against Their Own Results

 Quick Explanation Copied

 Long Explanation

Core evidence: what PACES achieved, with what data

PACES trains six random-forest encodings (one-hot, PSNSP, PSDSP, KNF, KSNPF, PseKNC; 800 trees) on motifs with ≥5 consecutive CXX repeats extracted from 2135 acRIP-seq peaks across 2134 genes: 1160 positive / 10855 negative training samples, 469 / 4343 for independent testing PACES dataset construction. The PSDSP+KNF meta-classifier reached AUROC 0.8855 in 5-fold cross-validation and 0.8741 (AUPRC 0.4852, BEP 0.4666 for PSDSP) on the independent set — notably lower than its cross-validated AUPRC, and the AUPRC of ~0.49 on a 10:1 imbalanced set means precision at high recall is modest. Feature importance hinted at G-rich context (GG/GGG/GGGG/GGAG/GGA among top KNF features) but the authors themselves flag that this "requires further experimental study" — correctly refraining from causal claims.

The successor XG-ac4C, trained on the identical PACES benchmark, raised AUPRC by ~9.4% (CV) and ~9.6% (independent) via EIIP-based features plus SHAP/LIME interpretability XG-ac4C benchmark comparison — confirming that sequence context beyond dinucleotide profiles carries signal, while also exposing PACES' ceiling.

Blindspots the paper honestly states — and ones it does not

Declared limitations are real: motif-level (not site-level) resolution, an ambiguous CXX motif definition awaiting a DRACH-like refinement, and no cross-tissue/species generalization PACES stated limitations. Undeclared concerns: negatives (motifs outside peaks) are not verified true negatives — some may simply fall below detection; only one training dataset (Arango 2018) is used, so overfitting to one antibody-based assay is unquantified; the auROC 0.87 headline looks strong while the AUPRC 0.49 is the more honest metric on a 10:1 imbalanced problem. Later work quantified this ambiguity: base-resolution and peak-level ac4C datasets overlap only partially (e.g., acRIP vs ac4C-seq: 60.67%), and total/poly(A)+ RNA ac4C abundance is only ~0.15%/0.01% by LC-MS/MS, making high-confidence negatives hard to define OpenAc4C landscape and method overlap.

Legacy and usefulness

Within its 2019 context PACES was the first motif-level ac4C predictor and its datasets became the de-facto community benchmark (reused by XG-ac4C and several deep-learning successors such as iRNA-ac4C, DPNN-ac4C, Caps-ac4C) Review of ac4C detection and predictors. It has been applied downstream, e.g., to map KLF9 ac4C motifs in adipogenesis PACES applied in KLF9 adipogenesis study. Verdict: an above-average, honestly reported engineering contribution whose empirical claims survive, but whose biological insight (what makes a cytidine acetylated) remains unvalidated. Falsification path: test predicted motifs against base-resolution RedaC:T/ac4C-seq data — if high-scoring PACES motifs show no base-resolution ac4C enrichment, the model's motif assumption collapses.

 BGPT Paper Review

Study Novelty

First motif-level ac4C predictor, but built on an established framework (feature encodings + random forest) mirroring the authors' own SRAMP/RNAm5Cfinder precedents; novel target, conventional method.

Scientific Quality

Transparent methodology, parameter optimization, independent test, public data/server, and honest stated limitations are strengths; weaknesses include unverified negatives, single-dataset dependence, modest AUPRC (0.485), and no external replication dataset.

Study Generality

Human-mRNA-specific, single-dataset, motif-level predictor; authors explicitly state cross-tissue and cross-species prediction is impossible with available data, limiting general scientific reach.

Study Usefulness

Served as the community benchmark dataset and a practical web server used in later mechanistic studies (e.g., KLF9 ac4C mapping), though superseded by XG-ac4C and deep-learning tools.

Study Reproducibility

Full dataset publicly downloadable, hyperparameters and encodings specified in detail, scikit-learn tooling standard; but no public code release per se, and the server's longevity is uncertain.

Explanatory Depth

Feature-importance findings (G-rich context, proximal positions) are correlational and explicitly flagged as requiring experimental confirmation; no mechanistic model of NAT10 sequence preference is offered.