Classification accuracy.

We compare the accuracy of the following methods: FSei—fine tuned Sei model that simply replaces the classification head of the original model; FSei* adds a fine-tuning block that includes a convolutional layer; FSei-frozen has the same architecture as FSei, except that all layers except for the classification block are held fixed; the two Sei-targets models trained a logistic regression or LightGBM classifier based on Sei predictions; the DNABERT-2 model uses DNABERT-2 embeddings, while DNABERT-2* uses the same fine-tuning block used in FSei*. The two trained-from-scratch models are based on the architectures used in Basenji [10] and Basset [9]. The accuracy of the two best performing models is highlighted in boldface.