Artists' impressions of the feathered troodontid Jianianhualong tengi (left) and the large clawed Therizinosaurus cheloniformis (right).

Dinosaurs like Jianianhualong tengi and Therizinosaurus cheloniformis (images not to scale) are members of the maniraptorans. Jianianhualong image © Julius T. Csotonyi/Xu et al., licensed under CC BY 4.0 via Nature Communications and Therizinosaurus image © PaleoNeolitic licensed under CC BY 4.0 via Wikimedia Commons

Read later


During Beta testing articles may only be saved for seven days.

First British fossils of therizinosaur and troodontid dinosaurs are world's oldest

Machine learning models can be trained to identify individual dinosaur teeth.

The technique, being pioneered by scientists from the Museum, can be used to address some of the biggest mysteries in palaeontology even when the fossil record is patchy. 

Isolated teeth found in southern England could represent some of the earliest relatives of birds ever discovered.

Teeth found in Oxfordshire, Gloucestershire and Dorset are believed to belong to the maniraptorans, a group of dinosaurs, including Velociraptor, which include birds and their closest relatives.

These dinosaurs evolved into numerous species during the Middle Jurassic, but because fossils during this time are scarce, knowledge of their origins is lacking.

Researchers from the Museum and Birkbeck College used pioneering machine learning techniques to train computer models to identify the mystery teeth, which push back the origin of some of the group's members by almost 30 million years.

Simon Wills, a PhD student at the Museum who led the research, says, 'Previous research had suggested that the maniraptorans were around in the Middle Jurassic, but the actual fossil evidence was patchy and disputed. Along with fossils found elsewhere, this research suggests the group had already achieved a global distribution by this time.'

'The teeth we analysed include what are currently the only troodontid and therizinosaur fossils ever recorded from the UK and are the oldest evidence of these dinosaurs anywhere in the world.'

The findings of the study were published in the journal Papers in Palaeontology.

Researchers in high visibility clothing search for dinosaur teeth at Woodeaton.

Middle Jurassic fossils are relatively rare, with Woodeaton quarry in Oxfordshire a hotspot for finds. Image © The Trustees of the Natural History Museum, London.

What are the maniraptorans?

The maniraptorans are a group of dinosaurs which contains all living birds, as well as extinct dinosaurs such as Deinonychus and Archaeopteryx.

Historically, they have been vitally important in shaping how we understand the dinosaurs. Until the 1960s, it was generally considered that these ancient reptiles were slow, lumbering animals, but the description of Deinonychus in 1969 started to change that.

Its remains had the characteristics of a fast-running hunter, while also sharing similarities with living birds. Some maniraptorans are preserved with feathers, and as time passed, more evidence began to build that birds were the direct descendants of this group of dinosaurs.

Among the closest relatives of birds are groups such as the dromaeosaurids, which contains Velociraptor and Vectiraptor, as well as related groups such as the troodontids and therizinosaurs.

An artist's impression of Velociraptor, covered in brown and white feathers.

A vareity of maniraptorans, such as Velociraptor, are thought to have been covered in feathers, based on fossil evidence and their close relationship with living birds. Image © Fred Wierum, licensed under CC BY-SA 4.0 via Wikimedia Commons

The troodontids are an extinct, bird-like group, while the therizinosaurs are a clade of dinosaurs with unusual characteristics such as extremely large claws measuring over 50 centimetres long.

While their fossils are common in the Late Jurassic and Cretaceous, these maniraptorans are not well known from the Middle Jurassic. Some species, such as Eshanosaurus deguchiianus, have tentatively been identified from even earlier, but this is still being debated.

'While the family tree of the maniraptorans suggests they evolved during the Middle Jurassic, there aren't many fossils from this period to prove this,' Simon explains.

'As a result, any potential maniraptoran remains from the Middle Jurassic are important, because they can help us to understand dinosaur evolution. These remains are often isolated teeth, which can be difficult to assign to a particular dinosaur group, and so are left out of evolutionary models.'

The holotype of Eshanosaurus, consisting of a lower jawbone.

The enigmatic Eshanosaurus may be an Early Jurassic therizinosaur, but as it was described based on a single jawbone this is difficult to confirm. Image © Zanno et al, licensed under CC BY 4.0 via PeerJ

Unlike in humans, dinosaurs had teeth which were continuously shed and replaced throughout their life. They are also highly resistant to erosion and degradation, making it more likely they will survive as fossils.

While previous studies have attempted to classify isolated teeth based on a variety of statistical methods, they've not always been particularly successful.

The researchers behind the current study have been working to improve this, having demonstrated that machine learning models can achieve up to 96% accuracy in identifying isolated teeth from known taxa.

Similar models have now been applied to teeth from the Middle Jurassic held in the collections of the Museum and the Museum of Gloucester in an attempt to identify them.

'The use of machine learning in vertebrate palaeontology is still in its infancy,' Simon adds. 'This is partly because of the need to have a comprehensive training dataset for the models to learn how to identify the teeth of different dinosaurs, and partly because the models themselves can also be quite complex.'

'In our study we are fortunate that there is already a relatively large dataset of dinosaur tooth measurements available that we could use to train the models.' 

Scans of four faces of a troodontid tooth (left) and four faces of a therizinosaur tooth (right).

The troodonitd and therizinosaur teeth are the oldest found anywhere in the world. Image © The Trustees of the Natural History Museum, London and the Museum of Gloucester

How can machine learning identify dinosaur fossils?

To translate the information contained in the fossils into data that could be used in the machine learning models, the researchers first had to produce a 3D model of each tooth from CT scanning.

The measurements of thousands of teeth from known dinosaur species were used to train three different models. Each model analyses the data in a different way, with the results of each combined to give the most likely identity of each tooth.

When compared to other statistical methods, the machine learning models gave more accurate results and increased the researchers' confidence that they would be able to classify unidentified teeth.

The models were then applied to the unidentified, isolated teeth, which found that most of the teeth belonged to maniraptorans, and dromaeosaurids in particular. These teeth were split into three distinct groups based on their size and shape.

One additional tooth was also identified as a troodontid, and another as a therizinosaur. However, these had to be additionally qualified by comparing their shape to known teeth, as they could not be identified by the machine learning models alone.

However, identifying these two teeth does push back the origins of these groups by 27 million years.

By confirming the presence of these dinosaur groups in the Middle Jurassic, the paper also lends support to theories that maniraptorans had already achieved a worldwide distribution before the supercontinent of Pangaea began to break up around 170 million years ago.

This continental breakup would have exposed these dinosaurs to a variety of different environmental conditions, driving their diversification into a range of new species.

As technological innovations continue, and digitisation projects make more information available to create training datasets, it is likely that machine learning will become more common as a way to investigate questions in palaeontology.