DNA Foundation Models for Molecular Biology: From Genomic Sequence Understanding to Function Prediction and Biological Design

Molecular biology is increasingly shaped by the ability to read, model, and manipulate biological sequences at scale. While protein and RNA sequences have motivated mature branches of biological language modeling, genomic DNA remains a particularly challenging substrate because it encodes coding regions, non-coding regulatory elements, long-range dependencies, evolutionary constraints, population variation, and disease-associated mutations within a single symbolic sequence. This review examines genomic language models and DNA foundation models from a DNA-centered perspective. We first clarify the conceptual boundary between DNA, RNA, and protein language modeling and report the literature-search and model-selection strategy used for this review. We then summarize key methodological dimensions, including tokenization, pretraining objectives, Transformer architectures, long-context alternatives, state space models, reverse-complement-aware inductive biases, training data coverage, and compute cost. Representative systems, including DNABERT, DNABERT-2, Nucleotide Transformer, HyenaDNA, Caduceus, GPN, megaDNA, Evo, Evo 2, Borzoi, and AlphaGenome, are compared through a taxonomy that separates self-supervised genomic language models, evolution-aware models, generative genome models, supervised sequence-to-function DNA foundation models, and broader functional genomics predictors. Rather than treating model scale or benchmark accuracy as sufficient evidence of biological utility, we emphasize biological validity, benchmark leakage, homology overlap, species and population bias, interpretability, experimental validation, responsible design, and clinical translation. This framing provides a critical foundation for evaluating how DNA foundation models can support regulatory genomics, variant effect prediction, genome annotation, cross-species transfer, and biologically grounded sequence design.