Unfortunately, similar fossils are very rare. For example, the understanding of Denisova people is based on DNA extracted from a phalanx. Although those combinations from early hybrids and other ancestral combinations are easy to find, they may be difficult to prove when it comes to physical evidence. The clues they have appeared may only exist in some people's DNA. Even so, they may be more subtle than the genes of Neanderthals and Denisovans. Statistical models help scientists infer the existence of these populations without fossil data: for example, the genetic variation pattern of 20 13 ancient humans and modern humans shows that an unknown human population has crossed with Denisova (or their ancestors). But experts believe that these methods also inevitably ignore many details.
Who else has contributed to the human genome today? What do these people look like? Where do they live? How often do they interact and mate with other human species? In a paper published in Nature Communication, researchers showed the potential of deep learning technology, which can help fill some missing parts, and some experts may not even realize it. Through in-depth study, they picked out the evidence of the existence of another population: an unknown human ancestor in Eurasia, possibly a hybrid of Neanderthals and Denisova, or a relative of Denisova. This research work points out the future use of artificial intelligence in paleontology, which can not only identify unforeseen traces, but also reveal the missing parts of our evolution.
At present, statistical methods involve simultaneously detecting the * * * identical features of four genomes, which is a similarity test, but not necessarily the test of actual ancestors; Because many different methods can explain the small amount of gene mixture it reveals. For example, these analyses may show that the genomes of modern Europeans and Neanderthals have some common features, but they are different from those of modern Africans. However, this does not mean that these genes came from the cross between Neanderthals and European ancestors. The latter may be closely related to the population reproduction of Neanderthals, not Neanderthals themselves. Due to the lack of physical evidence to show when, where and how these ancient hypothetical genetic variations originated from people, it is difficult to clearly point out which of many hypothetical ancestors.
John Hawks, a paleoanthropologist at the University of Wisconsin-Madison, said: This technology is simple and powerful, but there are still many problems in understanding evolution. Deep learning methods try to explain the level of gene flow. Although the level of gene flow is too small compared with statistical methods, it provides a broader and more complex model to explain it. Through training, neural networks can learn to classify patterns in genomic data according to the population history that is most likely to produce patterns, without being told how to establish these relationships.
The use of deep learning technology can find traces of ancient humans that researchers have never suspected. First of all, we have no reason to think that Neanderthals, Denisovans and modern people are the only three populations in the historical context of mankind. According to Hawkes, there may be dozens of such populations. Jason lewis, an anthropologist at the State University of New York at Stony Brook, agrees. He said: Our imagination is always limited, because we always pay attention to the living or fossils found in Europe, Africa and West Asia. Deep learning technology refocuses these possibilities in a strange way, no longer limited by our imagination.
Deep learning seems unlikely to solve the problem of paleontologists, because this method usually requires a lot of training data. Take its most common image classifier as an example. When the expert trains the model to recognize the image of the cat, the expert has thousands of pictures to train, and the expert himself knows whether it is effective or not, because he knows what the cat should look like. Due to the lack of relevant anthropological and paleontological data, researchers who want to use deep learning technology have to create their own data to make it more intelligent. Oscar Lao, a researcher at the National Genome Analysis Center in Barcelona, said: We are playing a dirty trick. We can use unlimited data to train deep learning engines because we use simulation.
Researchers have generated thousands of simulated evolutionary histories based on different demographic details: the number and size of ancestral populations, the mixed-race rate when they separated from each other, and so on. From these simulated histories, scientists have created a large number of simulated genomes for modern life. They train these genomes in deep learning algorithms to understand which evolutionary model is most likely to produce a given genetic model. Then, the research team released artificial intelligence to infer the history that best conforms to the actual genome data. Finally, the system concludes that a previously unidentified human group also contributed to the ancestors of Asian descendants. Judging from the gene patterns involved, these people themselves may be a unique group produced by the hybridization between Denisova and Neanderthals 300 thousand years ago
Or a group evolved from the descendants of Denisova shortly after that. This is not the first time that deep learning has been used in this way. Some laboratories in this field have applied similar methods to solve other clues of evolutionary research. Andrew from the University of Oregon? A research team led by Andrew Kern uses simulation-based methods and machine learning techniques to distinguish various models of how species, including humans, evolved. It has been found that most adaptations favored by evolution do not depend on the emergence of beneficial new mutations in the population, but on the expansion of existing genetic variations. Applying deep learning to these new problems is producing exciting results.
There are some problems. First of all, if the actual human evolutionary history is different from the simulation model trained by deep learning method, then this technology will produce wrong results. This is a problem that Cohen and others have been trying to solve, and there is still a lot of work to be done to improve the accuracy. Joshua Akey, an ecologist and evolutionary biologist at Princeton University, said: I think the application of artificial intelligence in genomics has been greatly exaggerated. Deep learning technology is a wonderful new tool, but it is only a method and cannot solve all the mysteries and complexities we want to understand in human evolution.
Some experts are even skeptical. David Pulby, a paleontologist at Harvard University and Peabody Museum, wrote in an e-mail: My judgment is that the density and quality of data are not ideal except for thoughtful, intelligent and manual analysis. However, in the eyes of other paleontologists and geneticists, this is a good progress and can be used to predict possible fossil discoveries in the future and genetic variations that should have existed thousands of years ago. I think deep learning will really promote the development of population genetics, and it may be the same for other fields where we can access data but can't produce data.
At about the same time that Cohen and other population geneticists and evolutionary biologists develop artificial intelligence technology based on simulation to solve problems, physicists are also studying how to filter the massive data generated by the Large Hadron Collider and other particle accelerators. Geological research and earthquake prediction methods have also begun to benefit from deep learning methods. Nick Patterson, a computational biologist at MIT and MIT Broad Institute, said: I really don't know what will happen, but it's always good to have new methods. If it can answer our questions well, we will try our best to develop it!
Boko Park-Popular Science | Reference Journals: Nature, Nature Newsletter
Text: jordana cepelewicz/ quantum magazine/quantum newsletter
DOI: 10. 1038/s 4 1586-0 18-0455-x
DOI: 10. 1038/nature 12886
DOI: 10. 1038/s 4 1467-0 18-08089-7
Boko Garden-Passing the Beauty of Cosmic Science