#8 · ML / AI methods

Biologically informed tokenization and patching reshape DNA language model inputs

The dominant thread is input representation: how biological sequence is converted into model inputs before learning. Mut-BPE adds single-nucleotide resolution at variant sites to training-free BPE and improves variant effect prediction in DNABERT-2. PatchDNA replaces tokenization with patches whose boundaries follow evolutionary conservation. TEA derives a 20-letter protein alphabet from contrastive learning on protein language model embeddings, a related move toward redefined units. deepFEPS consolidates k-mer, transformer, autoencoder and graph embeddings in one toolkit for DNA, RNA and protein. A published benchmark of five DNA foundation models on zero-shot genomic tasks supplies comparative context. TIDE and DPCMHC address peptide-MHC and TCR binding with different architectures and are adjacent rather than DNA-specific. What is new is emphasis on biologically informed boundaries and resolution rather than a shared model class.