Evaluating the representational power of pre-trained DNA language models for regulatory genomics | Synapse