Genome

Every living organism carries in its cells a kind of blueprint for itself. This blueprint is stored in a thread-like molecule called DNA. DNA consists of four different chemical building blocks, abbreviated with the letters A, C, G, and T. The entirety of this sequence of letters in a living organism is called the genome. In humans, it consists of about three billion letters, distributed across 46 segments called chromosomes. Almost every body cell contains a complete copy of this blueprint.

Why three billion letters are a data problem

A single human genome as a pure sequence of letters takes up about 700 megabytes, roughly the size of a full CD. That sounds harmless. But when read out in the lab, raw data on the order of 100 gigabytes per person is generated. Large research projects sequence hundreds of thousands of people. This makes genome research one of the most data-intensive sciences there is.

The real value lies in comparison. Any two people differ in their genome by only about 0.1 percent of the letters. It is precisely these small differences that help determine how high the risk is for certain diseases or how someone reacts to a medication. To find such patterns, one must analyze millions of variants across many genomes simultaneously. This cannot be done by hand; it requires statistics and machine learning.

For the economy, this represents a multi-billion-dollar market. Companies that build sequencing devices, analyze genome data, or develop drugs based on it regularly appear in financial news. At the same time, genome data is among the most personal data that exists: it reveals things about relatives who never gave their consent.

From the cell nucleus to a readable sequence of letters

No device can read three billion letters in one continuous pass. Instead, the DNA is broken down into millions of short fragments. Each fragment is read individually and yields perhaps 150 letters. A computer must then reassemble the fragments into the correct order. One can imagine this like a puzzle in which a thousand identical books were shredded and one must reconstruct a complete book from the scraps.

For this to work, each position is read multiple times, typically thirty times over. This coverage compensates for reading errors. Afterward, the software compares the result with a reference genome, a kind of standard edition of human genetic material. All deviations from it are recorded as variants. In this way, 100 gigabytes of raw data becomes a manageable list of a few million differences.

A common misconception: the genome is not a program that you simply run. Only about two percent of the letters contain direct blueprints for proteins, the working molecules of the cell. The rest controls, regulates, or has functions that are only partially understood. What effect a variant has is therefore often unclear. This is precisely where AI models come in, learning from known cases which deviations are dangerous.

Genome data in medicine, ancestry research, and headlines

In cancer medicine, genome analysis is already part of everyday practice. The genetic material of a tumor is sequenced and compared with the healthy tissue of the same patient. From the differences, it is sometimes possible to deduce which medication is likely to work. In rare genetic diseases in children too, genome analysis often leads to a diagnosis faster than years of individual examinations.

Commercially, one encounters this topic with providers of saliva tests for ancestry and health risks. Such tests usually do not read the entire genome, but only a few hundred thousand selected positions. This is cheaper, but significantly less informative. The case of the insolvent provider 23andMe showed in 2024 and 2025 what happens when a database containing millions of genetic profiles suddenly becomes part of a bankruptcy estate.

In AI news, the genome comes up when models treat genetic material like a language. Systems such as AlphaFold predict the three-dimensional shape of proteins from gene sequences. Other models are trained directly on DNA sequences, similar to how language models are trained on text. The idea behind this: if a system has seen billions of letters of genetic material, it recognizes patterns that elude humans.