Google’s AI system called AlphaGenome was developed to assess the possible functions of non-protein-coding DNA in the human genome and predict the effects of single-base changes. The protein-coding portion accounts for less than 3% of the human genome. The remainder includes centromeres, which help chromosomes divide equally between cells; structures that protect chromosome ends; sequences that regulate gene activity and the processing of messenger RNAs; and regions that control how DNA is packaged inside the cell. However, much of the non-protein-coding DNA is thought to consist of remnants of ancient viral infections, molecular parasites, and genes rendered inactive by mutations.
It is difficult to distinguish functional regions because proteins that bind to DNA are not precisely tied to specific sequences, different cell types contain different proteins, and sometimes the presence of many sites together within a narrow region matters more than a single binding site. AlphaGenome analyzes processes such as gene expression, transcription, chromatin accessibility, histone modifications, transcription factor binding, and RNA splicing. For now, the system has been trained only on human and mouse sequences and a limited number of extensively studied cell types. Nevertheless, it can provide clues about the significance and possible causes of changes in the non-coding regions of a gene; its predictions generally produce results equal to or better than those of specialized tools.
Google evaluated all possible single-base changes in the human genome using the system. It calculated the effects that could result from replacing each base in the reference sequence with the other three bases. However, most people differ from this reference sequence at millions of bases and may carry insertions, deletions, duplications, or inverted sequences; because most changes occur in nonfunctional regions, they do not produce any effects. The study could enable researchers to examine possible mutations in advance and scan the genome for specific effects. However, some of this information could also be obtained by examining the ENCODE data used to train the model. It remains unclear whether the system can perform reliable analyses on Neanderthal and Denisovan genomes or on cell types for which no data are available.
Why it matters
This study shows that regions of the genome that do not produce proteins cannot be treated merely as nonfunctional remnants, and that their possible effects in processes such as gene activity and RNA processing can also be systematically screened. This could allow researchers to narrow down single-base differences that were previously difficult to prioritize before conducting experimental studies and obtain initial clues about their possible effects. However, the results produced by the model do not mean that every change in DNA has biological consequences; differences from the reference sequence and structural changes in humans make the assessment more complex. The key unresolved issue is the extent to which the system, trained on a limited number of cell types, will provide reliable results in cells for which no data are available and in Neanderthal and Denisovan genomes.
Background
Google is not a new name in the FikirPilot archive: we have published 31 news stories mentioning the name in the past 90 days; the most recent was dated 9 September 2026.