The Human Genome Proves More Complex Than AI Models Assume
A Wired article explains that gene regulation in human cells is highly intricate and three-dimensional. As a result, AI models built to analyze genomes cannot yet fully decode it.
More than two decades after the international Human Genome Project finished, biologists still do not understand how human DNA is regulated. Wired reports this, drawing on a Quanta Magazine story. Researchers hope artificial intelligence will help, but the genome's complexity makes that hard.
More than genes
The Human Genome Project, which ran from 1990 to 2003, worked out the order of roughly 3 billion chemical building blocks in human DNA. It showed that only about 2 percent of the genome consists of genes, the sequences that code for information. Humans have around 20,000 of them.
The harder question is how genes are switched on and off. This is how different cell types, such as muscle, brain or skin, arise from the same material. Karen Adelman, who studies gene regulation at Harvard Medical School, says genes are probably not the most interesting part of the genome.
How regulation works
Transcription factors are proteins that start the reading of a gene. In humans they do not bind to very specific DNA sites, they often work in pairs or groups, and the same factor can activate a gene in one cell type and suppress it in another. Adelman compares this to an "AND" logic, where many signals are weighed together.
Transcription factors attach to enhancers, DNA segments that give the signal to read a gene. There are hundreds of thousands, perhaps millions, of enhancers, more than there are genes. Some sit millions of nucleotides from the gene they control. Wendy Bickmore, a genome biologist at the University of Edinburgh, says it is still unknown where all the enhancers are and what they do. Loops of chromatin, formed by the protein cohesin, bring distant enhancers close to their genes.
Regulation is also shaped by the three-dimensional structure of chromatin and by epigenetic marks, which alter how DNA is packaged.
Limits of AI
Genomic "foundation models" such as Evo 2, Genos and Google DeepMind's AlphaGenome are trained on large volumes of data. They predict how differences in DNA sequence affect an organism's traits, including disease risk. They bypass the regulatory complexity and instead find correlations in cases already known. The article's authors consider this useful but not enough for real understanding, and suggest the genome may not suit such a simple input-output approach.
