Why AI still struggles to read the human genome

The hard part is regulation

The Human Genome Project mapped roughly 3 billion DNA building blocks between 1990 and 2003. Yet genes that encode proteins account for only about 2% of the genome. The bigger challenge is understanding how genes are used: which ones a cell turns on, when it does so, and how that changes with cell type and circumstance.

The conventional picture is familiar. Polymerase copies a gene into messenger RNA, then a ribosome reads that RNA and builds a protein. But the steps around this process are not simple switches.

In humans, transcription factors often act in groups, and the same factor can activate a gene in one cell type while suppressing it in another. Enhancers can influence several genes, while a single gene may be affected by several enhancers. Some sit millions of DNA letters away from the genes they regulate.

Source: wired.com

The genome’s physical arrangement matters too. DNA is wound around proteins into chromatin, whose structure can leave genes accessible or keep them inactive. Chromatin loops and topologically associating domains help bring regions together, while chemical epigenetic marks can alter how the DNA is packaged. These structures and marks change over time and between cells.

Regulation continues after transcription. MicroRNAs can interfere with messenger RNA before it becomes protein, and the spliceosome can process the same pre-mRNA in different ways, producing protein variants with different functions. A model that predicts from DNA sequence alone has to account for a system that changes at multiple stages.

What a sequence model can miss

Genomic foundation models such as Evo 2, Genos and AlphaGenome are trained on large genomic datasets. Their promise is to infer how DNA differences affect biological processes and, eventually, traits such as disease risk. The regulatory machinery is not necessarily represented explicitly: the hope is that the model will learn its patterns from examples.

That can make the models useful without making them explanatory. Wendy Bickmore, a genome biologist at the University of Edinburgh, expects AlphaGenome to help, but says it will have limits. One gap is the diversity of cell types and how they change during development. The data needed to represent that complexity are not yet available.

There is another constraint: an organism’s biology depends on more than its genome. Nutrition, the environment, the microbiome and, in humans, culture can affect health and behavior as well as the genome itself. Adrian Woolfson, co-founder of California biotech company Genyro, calls this broader collection of information the “informiome.” In his book On the Future of Species, published in April 2026, he argues that genomic foundation models cannot predict every consequence of a mutation when relevant information is absent from the sequence.

I think that is the important distinction: predicting patterns in genomic data is not the same as modelling the conditions under which a living cell uses them. A black box can be useful for the first task while leaving the second largely unexplained.

A system, not a blueprint

Evolution did not design the genome from scratch. Bickmore suggests that increasingly complex organisms needed more precise ways to regulate genes across more tissue types. As regulatory sequences accumulated, some had to sit far from the genes they control, making three-dimensional chromatin organization part of the system. Cells also needed additional controls to suppress potentially harmful DNA, including transposons.

That history helps explain why the genome resists the clean input-output model often associated with computation. The regulatory machinery is itself produced by the genome, then responds to the genome’s structure, past state and surrounding conditions. It is a system acting on itself.

Barbara McClintock described the genome as a sensitive organ of the cell in her 1983 Nobel lecture. The metaphor is less tidy than a program, but it fits the evidence better: the genome responds to changing signals and can reorganize itself. AI may help researchers find regularities in that activity. Understanding the principles behind them will still require more than a sequence and a prediction.