Artificial intelligence can sift through vast amounts of DNA data and predict how sequence differences might affect biology. But the human genome is not a simple instruction manual: much of its complexity lies in how genes are switched on and off, and that regulation depends on combinations of signals that change with cell type and circumstance.
DNA contains more than protein instructions
The Human Genome Project, carried out between 1990 and 2003, nearly fully deduced the order of the human genome’s 3 billion or so chemical building blocks. The work found that barely 2 percent of the genome consists of genes that code for proteins. Knowing the sequence was a major achievement, but it did not reveal how the genome operates as a whole.
A gene can be read into messenger RNA, which a ribosome then uses to build a protein. That process helps explain how cells make the molecules they need. Yet different cells—including muscle, brain, and skin—use different sets of genes, and those choices shift over time as cells respond to signals from their surroundings.
That makes gene regulation central to understanding the genome. Noncoding DNA, which does not provide protein-making instructions, can still help govern when genes are active. The challenge is to identify how that control works across many overlapping systems.
Regulation depends on context
In bacteria, transcription factors can behave rather like keys matching particular DNA sites, with a signal helping switch a gene on or off. In complex organisms, including humans, the relationships are less direct. Transcription factors may work in groups, and the same factor can activate a gene in one cell type while suppressing it in another.
Human gene regulation therefore resembles an “AND” decision: several signals may need to come together before a gene responds. This gives cells room to react to fine differences in context, but it also makes the rules harder to infer. A model that sees a DNA sequence may not capture all the other factors that determine what happens in a particular cell at a particular moment.
Enhancers add another layer. These DNA regions gather transcription factors and help prompt the machinery that copies DNA into messenger RNA. There may be hundreds of thousands, perhaps millions, of enhancers—far more than genes. A gene can be influenced by multiple enhancers, and an enhancer can affect multiple genes, so connecting each enhancer to its targets is not straightforward.
What genomic AI can—and cannot—show
Genomic foundation models such as Evo 2, Genos, and Google DeepMind’s AlphaGenome are trained on large collections of genomic data. Researchers use them to predict how DNA differences may affect biological processes and, eventually, traits such as disease risk. Their strength is finding patterns across examples that would be difficult to review one by one.
Those predictions can be useful, but a pattern is not the same as an explanation. A model may associate a sequence with an outcome without showing which regulatory signals caused it, how those signals interacted, or why the result changed across cell types. The model’s internal computation can remain a black box even when its prediction is helpful.
The source of the difficulty is biological as well as computational. The genome is the product of around 4 billion years of evolution, not a plan drawn up from scratch. Its regulatory arrangements can be layered and overlapping, and the rules may not reduce to a direct mapping from DNA input to biological output.
Prediction is a step, not the whole answer
AI may help researchers navigate genomic data and generate useful predictions. But anyone seeking an account of how life works needs more than a reliable forecast: they need to understand the mechanisms behind it. That means learning where enhancers are, which genes they influence, and how combinations of transcription factors act in specific cells and situations.
The gap between prediction and understanding matters because the same DNA can support many kinds of cells and respond to changing conditions. A model trained on known examples can identify correlations in those examples, while the broader logic of regulation remains unsettled. Until researchers can connect those patterns to the interacting processes that produce them, genomic AI will be a tool for inquiry rather than a complete explanation of the genome.