Alex Foster-powell1, Dr Guy Meno-Tetang2, Prof Amin Rostami-Hodjegan3, Dr Kayode Ogungbenro, Prof Donald Mager4
1University of Manchester, 2AztraZeneca, 3Certara, 4Enhanced Pharmacodynamics
Introduction: Very few disease-modifying therapies for Alzheimer’s disease (AD) have been developed to date, and as our understanding of the underlying causes of AD evolves, numerous animal models have been generated to study the disease. However, persistent therapeutic failures raise questions about the reasons for these shortcomings, including whether they stem from poor target selection or limitations in replicating key aspects of Alzheimer’s pathology in animal models. Previously, the DAL group from MIT, proposed a computational workflow, called TransComp-R and TransPath-C and applied this to various indications such as inflammatory bowel disease [1], schizophrenia [2], tuberculosis [3] and AD [4]. In this study, this workflow has been extended and applied to three commonly used AD animal models – APP/PS1, 3xTg and 5xFAD – to assess their translatability to human Alzheimer’s. Objective: To establish a computational workflow, able to evaluate the translatability of three commonly used Alzheimer’s animal models; APP/PS1, 3xTg and the 5xFAD. Methods: Microarray data from the literature from the hippocampus of three commonly used Alzheimer’s disease animal models (APP/PS1, 3xTg and the 5xFAD), and from human brains in the diseased and healthy states were collected from the literature. The microarray gene expression data was converted into pathway enrichment scores through gene enrichment analysis. For each animal model, a separate mouse sparse principal component (sPC) model was generated, and then the human data projected into the space of each animal model. For each animal model, a support vector machine learning model (SVM) was trained using the coordinates of each datapoint in the sPC model. Using leave one out cross validation, the SVMs model were trained to optimise for F1 score, and model performance was evaluated through the AUC ROC, F1 score and accuracy. Any overlapping sPCs were identified between each animal model and human dataset. For any overlapping sPCs identified, a separate SVM was trained using just these pathways to reduce this list further. This resulted in the identification of pathways which contained phenotype defining information in both humans, and rodents. The pathways identified were validated through their relevance to AD using literature data. The workflow was also applied to ibuprofen treated mice data, to investigate if the pathways identified through this process could predict clinical trial outcome before they had been performed. Results: SVMs developed for each animal model generally performed well, with a range of 0.67-1, and a mean score of 0.88, 0.81 and 0.80 for AUC ROC, F1 and accuracy respectively. These SVMs, identified 0 overlapping pathways between the APP/PS1 and 3xTg models and the human dataset, at the default regularisation value. However, sPC1 was identified to contain phenotype-defining information for both 5xFAD mice and human data, which was associated with 26 pathways. The final SVM models trained using just these 26 pathways, identified SREBP control of lipid synthesis and Cytotoxic T-Lymphocyte pathways as containing phenotype-defining information for both species. Additionally, 60% of APP/PS1 ibuprofen treated mice were classified as diseased, replicating clinical findings. Conclusion: This workflow suggests that both the APP/PS1 and the 3xTg animal models have low translatability to humans, identifying 0 translatable pathways. The 5xFAD models appear to more faithfully replicate aspects of Alzheimer’s disease in humans, with SREBP control of lipid synthesis and Cytotoxic T-Lymphocyte pathways identified as containing phenotype-defining information for both humans and mice.
[1] Brubaker DK, Kumar MP, Chiswick EL, et al. An interspecies translation model implicates integrin signaling in infliximab-resistant inflammatory bowel disease. NIH Public Access; 2020. [2] Carroll MJ, Garcia-Reyero N, Perkins EJ, Lauffenburger DA. Translatable pathways classification (TransPath-C) for inferring processes germane to human biology from animal studies data: example application in neurobiology. Oxford Academic; 2021. p. 237-245 [3] Pullen KM, Finethy R, Ko SB, Reames CJ, Sassetti CM, Lauffenburger DA. Cross-species transcriptomics translation reveals a role for the unfolded protein response in Mycobacterium tuberculosis infection. NPJ Syst Biol Appl. Feb 15 2025;11(1):19. doi:10.1038/s41540-024-00487-6 [4] Lee MJ, Wang C, Carroll MJ, Brubaker DK, Hyman BT, Lauffenburger DA. Computational Interspecies Translation Between Alzheimer’s Disease Mouse Models and Human Subjects Identifies Innate Immune Complement, TYROBP, and TAM Receptor Agonist Signatures, Distinct From Influences of Aging. Front Neurosci. 2021;15:727784. doi:10.3389/fnins.2021.727784
Reference: PAGE 33 (2025) Abstr 11434 [www.page-meeting.org/?abstract=11434]
Poster: Methodology – AI/Machine Learning