aceHyLnc: a hybrid deep learning and feature-based approach for long non-coding RNA prediction
Abstract: Long non-coding RNAs (lncRNAs) play important roles in gene regulation, development and disease, yet accurate identification of lncRNAs from transcriptomic data remains a major computational challenge. Existing methods often rely either on handcrafted sequence features or deep learning approaches, each with their inherent limitations in capturing the full complexity of RNA sequences. In this study, we proposed HyLnc, a computational framework that integrates transformer-based contextual embeddings with biologica…
"An atlas of human long non-coding RNAs with accurate 5’ ends" by Hon C is licensed under CC BY-SA 4.0. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/4.0/.
On 2026-09-16, a peer-reviewed study described HyLnc, a framework that merges transformer-based contextual embeddings from a BERT model pre-trained on metazoan RNA with biologically meaningful sequence features for lncRNA prediction.
The work matters because accurate lncRNA identification remains a major computational challenge for gene regulation and disease research, and the reported hybrid approach offers a scalable annotation method, though generalizability beyond the curated validation datasets remains to be tested.
- Custom BERT model pre-trained on large corpus of metazoan RNA sequences with masked language modelling to learn contextual nucleotide dependencies.
- 256-dimensional deep embeddings combined with 348 handcrafted features including ORF characteristics, UTR properties, nucleotide composition and Fickett scores.
- Multi-stage feature selection produced optimized hybrid sets and RF classifier achieved best performance among tested classifiers.
HyLnc combining transformer embeddings with handcrafted biological features improved lncRNA prediction to 91.3% accuracy on independent validation, outperforming existing tools.
The rundown
Researchers pre-trained a custom BERT-based model on metazoan RNA sequences, then fine-tuned on curated lncRNA and protein-coding transcripts to extract 256-dimensional embeddings, while separately computing 348 handcrafted features.
After multi-stage feature selection, multiple classifiers were evaluated and the RF model delivered the reported 91.30% accuracy, 91.23% F1 and 82.60 MCC on independent validation, presented as outperforming several existing tools.
Sources
- Peer-reviewedRNA Biology2026-09-16
How should this claim be treated?
The debate