DNABERT-2

Multi-species genome foundation model using Byte Pair Encoding tokenization. Achieves state-of-the-art on 28 genome understanding benchmarks. Efficient alternative to k-mer tokenization used by DNABERT-1.