Sequence-Only Prediction of Super-Enhancers in Human Cell Lines Using Transformer Models.

Abstract

The study discloses the application of transformer-based deep learning models for the task of super-enhancers prediction in human tumor cell lines with a specific focus on sequence-only features within studied entities of super-enhancer and enhancer elements in the human genome. The proposed SE-prediction method included the GENA-LM application at handling long DNA sequences with the classification task, distinguishing super-enhancers from enhancers using H3K36me, H3K4me1, H3K4me3 and H3K27ac landscape datasets from HeLa, HEK293, H2171, Jurkat, K562, MM1S and U87 cell lines. The model was fine-tuned on relevant sequence data, allowing for the analysis of extended genomic sequences without the need for epigenetic markers as proposed in early approaches. The study achieved balanced accuracy metrics, surpassing previous models like SENet, particularly in HEK293 and K562 cell lines. Also, it was shown that super-enhancers frequently co-localize with epigenetic marks such as H3K4me3 and H3K27ac. Therefore, the attention mechanism of the model provided insights into the sequence features contributing to SE classification, indicating a correlation between sequence-only features and mentioned epigenetic landscapes. These findings support the potential transformer models use in further genomic sequence analysis for bioinformatics applications in enhancer/super-enhancer characterization and gene regulation studies.