A deep learning algorithm that can find features of movement disorders across various species is expected to improve our understanding of Parkinson’s Disease.
In certain illnesses like Parkinson’s and Alzheimer’s disease, there are inherent variables that cannot be directly controlled in humans such as motor and behavioural dysfunctions. Therefore, to understand their pathology and neural underpinnings, scientists have turned to model organisms that are not only easy to breed and maintain, but also allow for better control. By externally analysing and comparing the behavioural repertoires of animal models, researchers hope to gain insight into the underlying mechanisms of these complex diseases and potentially find clues to develop better-targeted therapeutics.
However, conventional statistical analyses have not been able to perform behavioural analyses across different animal species. To this, scientists have brought into play machine learning, specifically deep learning algorithms that use multiple layers of artificial neurons to distinguish different sources of data. Deep learning algorithms are based on artificial neural networks that can be adjusted to perform various functions. For instance, the algorithm can differentiate between species based on the characteristics of its tracks left behind in the snow.
Basing their latest innovation on deep learning, scientists from the Graduate School of Information Science and Technology at Osaka have developed a new algorithm that can automatically detect walking behaviours of movement disorders that are shared across species.
“A central goal of comparative behavioural analysis is to identify human-like behavioural repertoires in animals,” explained first author Takuya Maekawa. To achieve this, the team utilised animal location tracking and artificial intelligence to analyse the locomotion data created by beetle, human, mouse, and worm subjects. This method of comparative analysis would allow them to study human neurological conditions that cause motor dysfunctions including those that result from low dopamine levels.
While the most straightforward method of studying animal behaviour would be to examine animal motion data, the spatial and temporal scales of animal locomotion vary widely among species. This means that motion data of other animals cannot be directly compared with human behaviour. To overcome this issue, the team proposed a novel network architecture: to incorporate a gradient reversal layer. By adding this layer, the team can predict whether or not the input locomotion data came from a diseased animal and which species the input data came from.
Maekawa and his colleagues then further refined their network to specifically study motor diseases by training it to focus on identifying locomotion features to recognise specific diseases, rather than which species the input was gathered. These locomotion features could be considered as defining cross-species hallmarks of the disease, regardless of their body scales and locomotion methods, making them valuable data to collect and assess.
To test the efficacy of their network, the team experimented their network on several dopamine-deficient organisms including beetles, humans, mice, and worms. Their system successfully identified cross-species locomotion features shared by worms, mice, and humans. Despite their evolutionary differences, all three organisms were not able to move while maintaining high speeds. When accelerating, the speed of these animals was also found to be unstable. They also exhibited similar movement disorders when their dopamine levels were low.
Although previous studies have established the link between dopamine deficiency and movement disorders, the team’s research was the first to identify specific, key locomotion features caused by dopamine deficiency that are shared across humans, mice, and worms. Given these results, the team believes that their work will be useful to find other common features for disorders that impact evolutionarily distant species and identify new animal models for a variety of Parkinson’s disease symptoms like akinesia, tremor, and rigidity, among others.
“Our project shows that deep learning can be a powerful tool for extracting knowledge from datasets that appear too different to be compared by human researchers,” said author Takahiro Hara. [APBN]
Source: Maekawa et al. (2021). Cross-species behavior analysis with attention-based domain-adversarial deep neural networks. Nature Communications,12, 5519.