Olive repeats get a genome-wide read: three plant DNA models agree on which STR loci shift their embeddings
Synopsis
Olive short tandem repeats (STRs, stretches whose repeat-unit copy number varies) had been read mainly through small microsatellite panels, leaving genome-wide variation and interpretation open. Across 212 Iranian and Mediterranean olive accessions, the study catalogued 117,119 polymorphic STR loci and passed natural allele series through three plant genomic language models, AgroNT, PlantCAD2 and GPN, measuring how far embeddings move as repeat length departs from the reference allele. On 62,847 loci, the three models shared 7,417 embedding-sensitive loci, and at most tested loci STR responses exceeded length-matched nonrepetitive indels.
Interpretation
Olive repeat variation now has a genome-wide inventory rather than a few dozen microsatellite markers. Earlier olive STR work relied on microsatellite panels assaying only a small number of selected loci, and genome-wide STR variation was poorly characterized in olive. STR genotyping of whole-genome sequencing from 212 Iranian and Mediterranean accessions retained 117,119 polymorphic loci after quality control, of which 65,095 were chromosomal at 125.0 per megabase.
The further a repeat length departs from the reference allele, the further the sequence embeddings of all three plant genomic language models move, so the models read repeat length as a graded change rather than presence or absence. These models had generally been evaluated on nucleotide substitutions, conventional indels or downstream annotation tasks, not on naturally occurring multiallelic STR length series. Among 62,847 loci with at least three non-reference repeat lengths, the L2-distance slope against length difference was positive at 98.1% of loci for AgroNT, 98.8% for PlantCAD2 and 96.5% for GPN.
The three independently trained models agree strongly on which loci are sensitive but not on how large the sensitivity is. A perturbation seen by only one model could reflect architecture-specific behavior rather than a property of the sequence. All three models passed FDR < 0.05 at 7,417 shared loci, with pairwise Jaccard indices of 0.66, 0.61 and 0.62 and overlaps at least fourfold above independence; Pearson correlations of standardized slope magnitude ranged only from 0.047 to 0.167.
Combining embedding evidence with oil fatty-acid associations narrows the candidate set to a handful of loci. Association analysis alone identifies loci whose repeat lengths covary with phenotype but cannot say which repeat changes the sequence models recognize reproducibly. Of 68 phenotype-associated STRs, 46 met the allele-count requirements for all three embedding analyses; 10 were supported by at least one model and four were sensitive in all three, though none of the four passed the trait-wise Bonferroni threshold.
Perspective
The results rest on short-read sequencing of 212 Iranian and Mediterranean olive accessions, the single linear OLEA9 reference genome and three plant genomic models not trained specifically on olive; the oil-trait analysis covers 140 phenotyped accessions.
None of the four cross-model candidates passed the trait-wise Bonferroni threshold, and association plus embedding are framed as complementary lines of evidence rather than causal proof; short-read genotyping remains difficult for long, interrupted and structurally complex repeats, and a single linear reference may introduce mapping and reference-allele bias; the models were not trained on olive, and the zero-shot embeddings came without expression, chromatin or biochemical measurements.