Computational comparative anatomy: Translating between species in neuroscience

The complexity of the human brain requires that we study its organization at many different levels: from the low level of the genes guiding its development, through the cells, the connections between cells, the regions of the brain, the networks of regions, up to the entire brain. Many of these levels can only be studied in experimental animals, such as rodents or monkeys, because the techniques required are not suitable for use in living humans. Neuroscience research therefore requires us to combine insights obtained in humans with those obtained in non-human animals. Unfortunately, such between-species 'translations' often fail.

The reason many between-species translations are not successful is that we know surprisingly little about how our brains differ from those of the other animals we study. This is particularly striking in the case of the most often-used mammalian study subject, the mouse. Mice are popular experimental animals because they are easy to breed and keep, are clever enough to perform certain tasks, and-and this is increasingly important-there are many genetic variants that can be studied. For neuroscience to benefit society through new insights into brain function, new understandings of disease, and new treatments it is imperative that we better understand how the mouse and human brain relate to one another.

In this project, we will use new insights from artificial intelligence to build a first comparative map of the mouse and human brain. We will train an artificial neural network to learn to recognise different areas of the mouse brain based on different types of data used in neuroscience, including genetic data, tissue properties, and connectivity data. The network will learn which types of data are important to identify areas and how each area can be recognised as a unique combination of genetics, tissue, and connections. Then, we will provide the network with the same types of data from the human brain. The network will then be able to determine which areas of the human brain are organized in ways that it has learned from the mouse and which areas of the human brain it cannot understand based on the mouse. In other works, the network will be able to provide us with a full 'map' of how well each part of the human brain relates to each part of the mouse brain.

Armed with this network, we will be able to examine a number of outstanding questions about mouse-human brain comparisons. For instance, the prefrontal cortex of the human brain is often identified as impaired in many psychiatric diseases. But it is still a matter of fierce scientific debate whether the mouse brain has a similar type of prefrontal cortex. This raises serious issues as to how well we can study psychiatric diseases using mouse models. Our network-based approach will allow us to study such questions in a completely new way.

We will also use our network to test explicitly how well some popular 'mouse models' of disease predict effects on the brains of human patients. We will take four different genetic variants of mice, each of which has been linked to a particular genetic variant in humans. We will study how the brains of these mice have changed relative to healthy controls. Then, using our model, we will predict how the brain of the human patient should look, based on what we found in the mouse model. If our model is capable of predicting how the human patients' brains look, this will provide a first quantitative validation of the mouse model for the brain.

Together, our approach will allow us to establish how much we can rely on knowledge obtained from the mouse brain to achieve the ultimate goal of neuroscience: to understand the human brain.

Untangling the complexity of the brain requires integration of results across scales and species. Unfortunately, translations of results obtained in non-human 'model' species to the human are often unsuccessful. This is largely due to a lack of understanding of the similarities and differences in brain architecture between species. This is particularly true for the most-often used model species, the mouse.

This project will combine approaches from artificial intelligence, newly available big data neuroscience, and anatomy to provide a novel approach to create a between-species mouse-to-human comparative map of brain organization. Our approach acknowledges that mapping brain regions based on genetic, tissue, and connectivity data is a highly unconstrained problem. We therefore propose to train a multilayer perceptron network to classify areas of the mouse brain. Having done so, we will apply the network to similar data from the human brain. This will allow us to assess for each part of the human brain, how similar or distinct it is from any part of the mouse brain.

Having created a mouse-human translational model using this network, we will explore it to answer a number of outstanding questions about mouse-human translatability. Crucially, we will validate the model using data from four mouse strains that have specific genetic polymorphisms that are associated with specific changes in brain morphology. We will use our network to predict how such changes should manifest in the human brain and test these against actual data obtained from human patients. This will provide the first formal validation of a mouse model of human brain alterations.

This project will provide both the tool used to create the between-species model and all comparative maps and results in their entirety freely to the community.