Source-resolved decomposition of coding–noncoding ambiguity in a non-model plant full-length transcriptome

Coding–noncoding classification remains challenging in non-model plant full-length transcriptomes because complete transcript recovery does not resolve coding potential, protein-homology support, open reading frame (ORF) architecture or repeat-associated evidence. Using Arrhenatherum elatius as the primary case study, we analyzed 63,642 previously clustered and polished full-length transcripts with a source-resolved multi-evidence framework integrating PINC and CPC2 predictions, BLASTx/Swiss-Prot homology, TransDecoder ORF annotations and RepeatMasker evidence. The framework separates six mutually exclusive operational evidence states from four overlapping ambiguity-source indicators, with a limited Daucus carota implementation used as a second-species comparative baseline. The operational coding–noncoding grey-zone target contained 20,178 transcripts, representing 31.71% of the A. elatius transcriptome. Short-ORF-associated ambiguity was the largest source, involving 7,383 transcripts, followed by homology-unsupported long ORFs (6,977), prediction conflict (5,710) and repeat-rich ambiguity (4,766). Across ORF-length thresholds of 50–150 aa, 16,738 transcripts remained consistently within the grey-zone union, whereas 4,170 showed threshold-dependent membership. Under the exact interval-union refinement at the 0.5 coverage threshold, 4,405 transcripts met the all-repeat ambiguity-source definition and 2,359 met the TE-only definition. Sequence-derived 3-mer and 4-mer representations showed statistically detectable but modest state-associated structure. Accession-collapsed GO sensitivity tested 137 terms and identified 68 enriched terms after excluding shared best-hit accessions, and all 2,359 TE-supported ambiguity-source targets had valid TE-consensus coordinates. The six-state operational assignment logic was also implemented for 39,936 D. carota transcripts as a limited comparative baseline. Coding–noncoding ambiguity in the A. elatius full-length transcriptome forms a structured computational evidence space rather than a set of validated lncRNAs, confirmed novel protein-coding genes or demonstrated TE-derived transcripts. The source-resolved framework preserves the evidence responsible for uncertainty, while sensitivity analyses and the limited D. carota comparative baseline define its robustness and scope without establishing broad cross-species generalizability. The framework provides a transparent basis for transcript prioritization, manual curation and genome-enabled or experimental follow-up.