HyLnc

A transcript assembler hands back a sequence of 1,400 nucleotides with no convincing open reading frame. It is either a long non-coding RNA or a protein-coding transcript whose start codon the assembler clipped, and the two answers send the downstream analysis in opposite directions. Classical coding-potential tools decide from explicit arithmetic — longest ORF, Fickett score, hexamer bias — while nucleotide language models decide from learned context. HyLnc's ablation shows the two are not substitutes for one another: a random forest given only sequence-level descriptors and the transformer's own embeddings reaches 80% accuracy, and adding ORF and untranslated-region features lifts it to 91%. Whatever masked language modelling over RNA learns, it does not absorb open reading frame structure.

HyLnc is the framework built around that observation, from the Computational Structural Biology Lab at Indian Institute of Technology Kharagpur with the plant genomics group at Visva-Bharati. A custom BERT encoder is pretrained by masked language modelling on a broad metazoan RNA corpus, fine-tuned on labelled lncRNA and protein-coding transcripts, and then frozen and used as a feature extractor: its 256-dimensional sequence embedding is concatenated with 348 handcrafted descriptors, pruned by a three-stage selection pipeline, and handed to a random forest. The transformer is not given a separate name in the paper — HyLnc refers to the whole pipeline — though the released code calls the encoder module lincBERT and its class CSEBERT.

That design is the deliberate opposite of lncRNA-BERT, which pursues the same task by pushing everything into the language model and showing that coding potential separates in embedding space before any supervised head is attached. HyLnc keeps the biologically interpretable features in the loop and pays for them with a classical classifier on top.

#Key Features

• Chunked long-transcript encoding: Transcripts are partitioned into fixed 4,000-nucleotide chunks, encoded independently, and their representations aggregated into one sequence-level vector, so full-length transcripts are covered rather than truncated to a context window.
• Convolutional nucleotide embedding: Four embedding layers at a hidden dimension of 512 apply convolutional kernels of width 6 and 9 to raw nucleotides, capturing motifs at two receptive fields without a k-mer or byte-pair vocabulary, then add sinusoidal positional encoding before the transformer encoder.
• Hybrid feature space: Each transcript is described by 604 features — 256 contextual embeddings plus 348 descriptors spanning ORF characteristics, 5' and 3' UTR properties, nucleotide composition and the Fickett score.
• Multi-stage feature selection: Pairwise correlations above 0.8 are dropped, the remainder ranked by mutual information, and LASSO regression at a regularization parameter of 1x10⁻⁷ keeps the non-zero coefficients, yielding five candidate feature sets whose ablation isolates each feature family's contribution.

#Technical Details

Pretraining used 428,823 metazoan RNA sequences drawn from GENCODE, RefSeq and RNAcentral, spanning mRNA, several non-coding classes and uncharacterized transcripts, with roughly 15% of nucleotides masked, for 50 epochs under Adam at an initial learning rate of 1x10⁻⁴. Fine-tuning ran 100 epochs on a labelled set reduced by CD-HIT at 80% identity to 14,455 lncRNA and 13,876 protein-coding sequences, split 80:20 with hyperparameters tuned by 10-fold cross-validation. Among decision tree, k-nearest neighbour, random forest, LightGBM and XGBoost classifiers, the random forest was consistently strongest and is what the released tool runs. On an independent validation set of 10,000 Ensembl transcripts — 5,000 lncRNA and 5,000 coding — it reports 91.30% accuracy, 91.23% F1, MCC 82.60, 90.56% sensitivity and 92.04% specificity, against CPC2, LGC, lncRNA_Mdeep, NCResNet, PLEK, IRSOM and lncRNA-BERT. The comparison is close rather than decisive: lncRNA-BERT posts a higher F1 (91.89%) and much higher specificity (97.28%) at markedly lower sensitivity (87.32%).

#Applications

The intended use is transcriptome annotation — screening assembled transcripts for coding potential before differential expression or co-expression analysis, where a false negative discards a regulatory transcript outright. The authors demonstrate this on RNA-seq from three neurodegenerative disease studies (GEO series GSE213610, GSE181029 and GSE124439 for Alzheimer's, Parkinson's and ALS), assembling transcripts with StringTie, keeping intergenic, antisense and intronic class codes from gffcompare, and classifying the survivors. The resulting candidates appear in diseased samples and not in controls; none emerged from the terminally differentiated neurons of the Parkinson's dataset. These are computational predictions, and the paper is explicit that experimental validation remains outstanding.

#Impact

HyLnc's contribution is a measurement rather than a leaderboard position: by ablating the handcrafted half of its own feature space, it quantifies how much of coding potential an RNA language model fails to recover on its own, which is a more durable result than the accuracy margin over its benchmarks. The practical footprint is a command-line tool and a public web server accepting up to 100 FASTA sequences of more than 200 nucleotides, both released by the lab. The code repository distributes a single packaged archive that is more complete than most: the fine-tuned encoder checkpoint alongside the full pretraining and fine-tuning sequence corpora and the labelled feature matrix. The random forest itself is not shipped trained — it is refit from that matrix on every prediction run. What the archive lacks is a license file, in the repository or inside it, which constrains reuse of any of it.