Data, code and results for: Computational Identification and Classification of Non-Coding RNA: A PRISMA-Guided Survey and Controlled Re-Evaluation of Representations, Architectures and Evaluation Practice

This record holds the data, code and results behind the article "Computational Identification and Classification of Non-Coding RNA: A PRISMA-Guided Survey and Controlled Re-Evaluation of Representations, Architectures and Evaluation Practice" by Manish Kumar, Manpreet Kaur and Puneet Garg (manuscript under review). It lets a reader repeat the audit of the nRC non-coding RNA benchmark, the re-evaluation of twelve predictors over four representations and seven partitioning protocols, and the literature counts.Contents:- Literature: the eighteen query strings with their counts and caps, a record-level log of their re-execution on 31 October 2026, the 121 included records with DOIs, the per-study coding sheet, the coding of the eight reporting items with the passages behind each decision, a blind recoding form with an agreement script for a second rater, and the scite.ai audit of the reference list.- Benchmarks: both nRC releases with the script that reconciles them, exhaustive pair audits, cluster assignments at 80% identity and the cluster-disjoint partitions.- New evaluation sets: 4752 Rfam sequences over thirteen classes with no relative above 80% identity in nRC, and 1716 sequences from RNA types outside those classes.- Representations: the frozen RNA-FM, RNA-MSM and RiNALMo embeddings, so results on them can be reproduced without a GPU or the encoders.- Code and results: the analysis scripts, the JSON results they write at full precision, the figures, a record-level log of the RNAcentral queries, a README with every definition, protocol and seed, the package versions, and SHA-256 checksums for every file.Model checkpoints are not included. The classical models are retrained by the scripts and the fine-tuning script does not save weights. The pre-trained encoders are third-party models, listed by name and revision in the README.Running it: the contamination audit, classical ladder, composition control and clustering grid take under two hours on one CPU workstation; the exhaustive pair audit takes about twenty minutes on eight processes; each frozen encoder embeds the benchmark in minutes on a laptop GPU; end-to-end fine-tuning takes several GPU hours; and the pre-training overlap audits need access to the RNAcentral public API. No new sequence data were generated.Licences: code MIT; the two derived sequence sets CC0 1.0; results, figures and coding sheets CC BY 4.0. The nRC files are redistributed as released by their authors (Fiannaca et al., BioData Mining 10:27, 2017).