Genome model

The genetic material of every living being can be written down as a very long chain made up of just four building blocks. They are abbreviated with the letters A, C, G, and T. In humans, that adds up to around three billion letters per cell. A genome model is a computer program that has learned patterns from huge quantities of such letter chains. It works similarly to the programs that complete texts — except its language is genetic material instead of German or English. If you ask it about a section, it predicts what most likely follows there or what significance that spot has.

What biologists get out of it

The human genome has been decoded for over twenty years now. But decoded only means: the sequence of letters is known. What most sections actually do, nobody knows precisely. Only about two percent of the human genome contains direct blueprints for proteins. The large remainder was long considered useless, but in truth controls when which blueprint gets used.

Examining these control regions individually in the lab takes years and costs a lot of money. A genome model can suggest where a closer look is worthwhile. It doesn’t replace the experiment, but it considerably shortens the search. Instead of a thousand possibilities, a lab might then only need to check twenty.

This is especially important for diseases. In many patients, an unusual spot is found in the genome, but it’s unclear whether it’s harmless. Genome models assess such deviations and estimate how strongly they might disrupt a function. This helps doctors classify findings.

Learning without a dictionary

During training, the model is shown genetic sections in which individual letters are hidden. It has to guess what’s there. At the start, it’s almost always wrong. After billions of such attempts, it has internalized the typical patterns. Nobody taught it biology in the process — it worked out the rules from the data itself.

This training method is the same one behind chat programs. The big difference lies in the length of the relationships involved. A control region can be a hundred thousand letters away from the gene it regulates. So the model has to keep track of very distant spots simultaneously. This is precisely what the developers of such systems are working hardest on.

Training usually isn’t done with human genetic material alone. Many models also learn from bacteria, plants, and animals. The reason is simple: whatever has stayed the same across very different organisms is probably important. The model recognizes these constants and transfers them.

From research into the headlines

Well-known examples are Evo from the USA and the Nucleotide Transformer models from Europe. Some of these are freely available, so universities can use them without licensing costs. Google DeepMind has already been very successful with protein structures using AlphaFold and is also working on genetic systems. That’s why these companies regularly show up in business news.

Incidentally, a genome model is not the same thing as a gene-editing program. Gene editing actually alters genetic material in the lab. The model only computes and doesn’t touch anything. But the two are increasingly being combined: first the AI suggests a spot, then an intervention is carried out there in the experiment.

This combination is also what gives rise to the debate about risks. A model capable of designing functioning genetic sections could theoretically also be misused for dangerous pathogens. That’s why some providers withhold certain capabilities or vet users beforehand. Nevertheless, among pharmaceutical and biotech companies, the technology is considered one of the most important fields of the future.