This article is more than 1 year old

How archaeologists can use AI to date our ancestors

Human remains from millennia ago analyzed by the machines their descendants built

AI algorithms can be used to date ancient human remains by analyzing their DNA, just-published research has proposed.

A team led by Lund University in Sweden developed a method that uses machine learning to identify genomes in dead organisms to determine how far they date back.

"There has been no major development in ancient dating remains since the development of radiocarbon dating 80 years ago," Eran Elhaik, lead author of the study, published in Cell Reports Methods, and an associate professor at the department of biology at Lund University, told The Register.

"The only alternative is archaeological dating, which is far more subjective, but was the first method to be used.

This is the first method that offers a radically different approach in 80 years

"This is the first method that offers a radically different approach in 80 years. It is based on the DNA sequence, which can now be sequenced more easily and accurately than before using very little amount of organic material. Because radiocarbon dates are subjected to many biases, archaeologists can use our methods freely and rapidly to get another estimate."

Radiocarbon dating is a widely used method for estimating the age of carbon-based stuff that is up to 50,000 years old. Living organisms are made up of different forms of carbon that are continuously replenished, and when these creatures die, one radioactive isotope known as carbon-14 within them starts to decay. 

Archaeologists use the half-life of carbon-14 as a yardstick. Every 5,730 years, the amount of carbon-14 in a sample decreases by half. By measuring the level of carbon-14 present in a material and comparing it with other stable carbon isotopes, they can estimate how old it must be.

But radiocarbon dating isn't always accurate, and results depend on the quality and quantity of the material available to date. Pristine samples, such as the bones of a mummy preserved for thousands of years in a tomb, fare better than tiny bits and pieces of leftover remains that have been contaminated by the environment.

The Lund AI technique, known as Temporal Population Structure (TPS), looks for genetic markers that have mutated over time to estimate age. A tiny change in the DNA for the LCT gene, for example, meant that humans could produce the enzyme to digest lactose. The mutation only started spreading after the neolithic period, Elhaik explained. Detecting this mutation in a sample means the organism cannot be older than about 12,000 years. 

TPS looks for tens of thousands of these types of markers to narrow down the age of a person's remains. The machine-learning-powered software was trained on these mutational patterns from remnants that have been dated already. Given a new DNA sample, it can predict how old it must be. Specifically, the team trained their model on a public dataset of human genomes that have been mostly dated using radiocarbon and archaeological methods. 

When TPS was tested, 75 per cent of the samples it dated were within 445 years of the mean radiocarbon date; 16 per cent of the samples dated over 1,000 years from the mean radiocarbon date. In short, the results can be iffy and the method performs worse for samples that are particularly old and where there isn't as much training data. The team also tested the model by getting it to date remains of people within the same family.

"The most important evidence for accuracy was applying TPS to date family members," Elhaik explained. "Family members must be dated to the same time. Father and son cannot be 1,000 years away from each other. TPS was able to date family members to within the same time zone without explicit training."

Genetic markers vary over different geographies. The current model was trained on thousands of genomes collected from Europe, Western Asia, and Eastern-Central Asia, and it won't be very accurate at dating specimens found outside these regions. The researchers want to improve their model by training it with more diverse datasets.

When archaeologists find human remains, it may have a limited amount of organic material, which may not be enough for both radiocarbon dating and DNA sequencing

"We would like to expand it to be applicable to Africans and Americans, which is more challenging to the small number of samples available," Elhaik said. "We would also like to improve the predictions for very ancient samples. The model struggles only with the very old samples – 8,000 to 10,000 years before the present – due to the scarcity of the data, again, a problem that would be resolved with time."

TPS gives archaeologists another method to date specimens, and it could help in cases where there isn't enough of a sample to perform radioactive carbon dating and could help verify age estimates.

"When archaeologists find human remains, it may have a limited amount of organic material, which may not be enough for both radiocarbon dating and DNA sequencing," Elhaik said. "DNA sequencing can tell us a lot more, but without the date, this information is useless, this is the case of half of the skeletons that are not radiocarbon dated."

"If they decide to approach an ancient DNA lab, like the one at the GeoGenetic Center at the University of Copenhagen, they can use TPS to date these remains. Likewise, they can use TPS to test and validate existing samples," he concluded. ®

More about


Send us news

Other stories you might like