Year: 2026
Type: preprint
Abstract: Most genomic foundation models are pretrained on independent linear assemblies and therefore do not explicitly represent population-level segment sharing or local graph connectivity. We developed TomatoPGFM, a graph-conditioned model pretrained on 54.65 Gb of sequence from 66 tomato ( Solanum spp.) accessions. Sequence tokens were conditioned on pangenome node attributes and local adjacency, and the model was optimised using masked language modelling and graph-feature reconstruction. To evaluate model responses to graph-conditioned input, we compared aligned, shuffled and disabled graph inputs in 25, 000 windows from the training panel. Sequence-aligned graph input produced lower masked language modelling loss than graph-off at all five curriculum stages in both training-panel strata, while the shuffled perturbation generally yielded intermediate losses. We then assessed sequence-only transfer in Solanum sitiens LA1974 and S. lycopersicum MicroTom, neither of which was used for graph construction or pretraining. Frozen-probe AUROC values for gene-versus-intergenic and coding-sequence-versus-intergenic classification ranged from 0.8489 to 0.9593. TomatoPGFM produced higher AUROC point estimates than DNABERT-2 in all four comparisons. Enabling the zero-feature GraphAdapter pathway with adjacency messaging disabled changed throughput by less than 1% at 512–2, 048 positions under the tested configuration. Together, these results show that TomatoPGFM responds consistently to sequence-aligned pangenome context in training-panel sequences and provides informative sequence representations for genic-region classification in accessions excluded from graph construction and pretraining.
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7205000823