JEPA-DNA: Grounding Genomic Foundation Models through Joint-Embedding Predictive Architectures
JEPA-DNA is a new training method for genomic foundation models that changes how they learn. Instead of just memorizing DNA sequences (token reconstruction), it uses a Joint-Embedding Predictive Architecture to force the model to learn meaningful, high-level functional features in a latent space. This shift helps models understand biological context better, leading to improved performance on tasks like promoter recognition.