Biologists have long spoken of DNA as a blueprint or a program, metaphors borrowed from engineering and computing. The trouble is that the protein.">genome behaves like neither. It is a physical object, folded into loops and tangles inside the nucleus, and the distance between two sequences along the strand often matters far less than how close they sit in three-dimensional space.
This physicality is a serious obstacle for artificial intelligence trained on sequence data alone. A model can learn the statistical patterns of DNA letters, and it can predict a great deal about function, but it has no representation of the fact that a regulatory element a million bases away may sit right beside the gene it controls once the fibre folds. Genomes are also dynamic, rearranging as cells differentiate, so any static picture is incomplete.
Researchers are working on ways to encode this geometry, combining imaging, chromosome conformation data and sequence models. The goal is not to abandon machine learning but to give it the right kind of information. Whether current architectures can absorb that kind of physical structure, or whether they will need something new, remains an open question.
