Abstract of the talk// Organisms must protect their genomes against instability to ensure faithful DNA replication. Both the epigenome and DNA sequence affect instability but their relative contributions remain unclear. To disentangle these contributions, Pierre and his colleagues combined the DNA language model GROVER with a panel of epigenomic features and an interpretable machine learning approach.

They show that GROVER can be used to model double strand break (DSB) sensitivity using the DNA sequence alone. A model trained on epigenomic data outperforms this model, however, showing that the epigenome encodes cell-type specific information that is useful for modeling DSB. Combining these two models results in the best performance, indicating that the sequence and the epigenome contain complementary information for modeling genome instability. By analyzing this combined model, the scientists identified histone marks that provide non-redundant information that cannot be learned from the sequence and may shape the cell-type specificity of DSB sensitivity.

Integrating these features directly within the GROVER architecture results in a model that performs similarly to the full-epigenome model and may increase the model’s ability to interpret the genome in a cell-type specific manner. The work demonstrates that DNA language models, combined with interpretable machine learning techniques, can be used to deepen our understanding of how instability is shaped by the genome’s sequence.