EvoSNR-Prom: Predicting promoters at single-nucleotide resolution with label-aware transfer learning of the pretrained EVO model

Token-level sequence labeling on top of the Evo foundation model, with lexicon-enhanced embeddings and label-aware transfer learning from well-annotated source species.

Most deep learning approaches to promoter prediction operate at whole-sequence level, labeling a DNA segment as "promoter" or "non-promoter", and therefore provide no nucleotide-level resolution. EvoSNR-Prom reframes the task as token-level sequence labeling, by analogy with named entity recognition in natural language processing, building on the Evo foundation model. To offset the limited context available with single-nucleotide tokenization, the authors add a lexicon-enhanced embedding strategy using biologically meaningful DNA lexicons, then a label-aware transfer learning framework aimed at small datasets. Reported evaluations cover several prokaryotic datasets, on which the model performs well according to the authors.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

The methodological idea is sound — moving from one label per sequence to one label per nucleotide is exactly what regulatory variant annotation would require — but the demonstration remains prokaryotic and the abstract reports no quantitative metrics. As it stands nothing transfers to interpreting a human non-coding variant, where the problem is not locating a promoter but predicting the effect of a substitution on its activity. Worth following for the method, not for use.

Analysis by Dr Thibaut Benquey