Abstract
The rapid advancement of DNA foundation language models has revolutionized the field of genomics, enabling the decoding of complex patterns and regulatory mechanisms within DNA sequences. However, the current evaluation of these models often relies on fine-tuning and limited datasets, which introduces biases and limits the assessment of their true potential. Here, we present a benchmarking study of three recent DNA foundation language models, including DNABERT-2, Nucleotide Transformer version-2 (NT-v2), and HyenaDNA, focusing on the quality of their zero-shot embeddings across a diverse range of genomic tasks and species through analyses of 57 real datasets. We found that DNABERT-2 exhibits the most consistent performance across human genome-related tasks, while NT-v2 excels in epigenetic modification detection. HyenaDNA stands out for its exceptional runtime scalability and ability to handle long input sequences. Importantly, we demonstrate that using mean token embedding consistently improves the performance of all three models compared to the default setting of sentence-level summary token embedding, with average AUC improvements ranging from 4.3% to 9.7% for different DNA foundation models. Furthermore, the performance differences between these models are significantly reduced when using mean token embedding. Our findings provide a framework for selecting and optimizing DNA language models, guiding researchers in applying these tools effectively in genomic studies.
Competing Interest Statement
The authors have declared no competing interest.
