Translating Brain Anatomy and Disease from Mouse to Human in Latent Gene Expression Space

Summary Background The mouse model is the most widely used animal model in neuroscience, yet translating findings to humans suffers from the lack of formal models comparing the mouse and the human brain. Here, we devised a novel framework using mouse and human gene expression to build a quantitative common space and apply it to models of neurodegenerative disease. Methods We trained a variational autoencoder on mouse spatial transcriptomics, and embedded mouse and human gene orthologs in the model’s latent space. We computed a latent cross-species similarity matrix for translation and compared translated maps to human ground truth evidence. Findings We established the validity of our model based on anatomical homology. Independent of species, brain areas with similar latent patterns clustered together, improving the homology of known anatomical pairs, and preserving principles of brain organisation. Importantly, brain alterations in mouse disease models predicted human patterns of brain changes in Alzheimer’s and Parkinson’s diseases. We further determined the best mouse model for the AD patients, based on how well the translations matched the patient data, across multiple models and timepoints. Interpretation Our work provides i) a quantitative bridge across evolutionary divergence between the human and the predominant preclinical species, ii) a predictive framework to help design and evaluate disease models. By highlighting which models are best suited across stages of disease, we effectively support the understanding of disease mechanisms, assist in the workflow of clinical trials, and ultimately accelerate the transformation of findings into improved human outcomes. Funding Supported by the Biotechnology and Biological Sciences Research Coundil (BBSRC) UK, the Medical Research Council (MRC) UK, the European Research Council, and the NIHR Oxford Health Biomedical Research Centre. Research in Context Evidence before this study Inferences on the human brain based on findings in the mouse are backed by broad neuroanatomical similarities between the two species. However, interventions that were successful in rodents rarely translate into successful outcomes to treat human diseases. One way to address this challenge is to quantify the boundaries of translational possibilities using cross-species common space approaches. Low-level embedding spaces have been used to quantitatively compare across species, but there is currently no framework for effective translation of disease maps between species, and no tool to establish the best animal model for a given patient population. Added value of this study This study is the first to quantitatively link mouse models of neurodegenerative disease with human patient data. After validating our novel VAE-based comparative framework on anatomical homology compared to existing methods, we determine which of three Alzheimer’s Disease (AD) mouse models, and at which timepoint, best models a human patient population with mild AD. In addition, we predicted human brain changes in human Parkinson’s Disease, based on maps from mouse models of this disease. Implications of all the available evidence This approach allows one to directly evaluate mouse models based on the quality of their translation to the human. At the level of preclinical model design, the translation from human to mouse of diverse disease phenotypes will highlight the relevant targets for intervention in the mouse. This will justify the relevance of specific animal models and time courses, and therefore streamline clinical trials. By additionally translating the effect of treatment or intervention longitudinally, we can optimize human treatment plans and life-long disease management. Combined with patient population stratification, this takes an important step towards personalised translational medicine.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.