Résumé
Deciphering the DNA cis-regulatory code—critical for gene expression regulation—remains a major challenge in genetics and cancer research. While deep learning and statistical models now predict gene regulation and expression from DNA sequences, they primarily rely on dominant signals often located at promoters and enhancers. Yet, key regulatory features (e.g., open chromatin, transcription factor binding sites, and disease-associated variants) also reside outside these regions, where experimental data are sparse or noisy. As a result, current models struggle to assess the functional impact of genetic variants in these understudied areas. Building on our previous work demonstrating the predictability of transcription beyond canonical regions [Grapotte et al., Nat. Comm 2021], we aim to move beyond epigenetic segmentation (e.g., enhancers/promoters) by developing automated functional annotation methods. These methods leverage machine learning models optimized for specific genomic contexts. Here, we present models predicting CAGE-based transcription in unannotated regions and highlight their divergence from promoter- or enhancer-trained models. We then introduce a clustering approach to group regions by shared features (using different embeddings) and assign each cluster an optimal predictive model. This strategy promises genome-wide variant effect assessment, addressing the limitation of models restricted to canonical regulatory regions.