Much of the DNA in the human genome still has no clear meaning. Only about one to two percent of those letters code for proteins. The rest is a mix of junk DNA, and regulatory elements, which are stretches of DNA that decide when genes turn on, where they act, and how strongly they work. Small spelling changes in this DNA have important effects. Scientists still need better ways to tell which of those changes matter.
Researchers at UC Berkeley created a genomic language model named GPN-Star. It is an artificial intelligence (AI) model trained on huge collections of DNA sequences so it can notice repeating patterns, in much the same way a chatbot learns patterns in ordinary language. The model is especially good at judging whether a variant is likely to cause harm, a property called pathogenicity, and at telling functional DNA from DNA that does little. It also needs far less computing time and power than some larger models trained on unaligned genomes from many species.
How the model uses evolutionary maps
Instead of feeding the program raw genomes one by one, the researchers trained it on whole-genome alignments. A whole-genome alignment is a computer-made map that lines up the DNA of many species next to one reference species, such as humans, so that matching and differing letters become easy to see. Alignments for humans were built at three different evolutionary depths: other primates, other mammals, and other vertebrates.
Models trained on longer evolutionary timescales were better at scoring rare changes inside proteins, which tend to stay stable for a very long time. Models trained on closer relatives, such as primates, were better at scoring variants linked to complex traits such as schizophrenia risk. Those traits can involve thousands of small changes, many of them outside protein-coding DNA.
The researchers also released genome-wide predictions that flag variants most likely to influence traits. Other scientists can use those lists to choose which experiments are most worth doing, because it is impossible to test every possible DNA change in the laboratory.
This research is published in Nature.