1.Background:
Currently over 55 million people worldwide are estimated to be living with dementia; this number is projected to triple to 139 million by 2050.
It is thought that the molecular changes that cause dementia begin decades before the diseases first present with cognitive symptoms. However, pathological mechanisms are difficult to fully resolve due to the complex organisational structure of human brains which varies considerably across individuals, masking subtle signatures of disease. Moreover, dementia presents heterogeneously across individuals, leading to suggestions of different disease subtypes, and uncertainty over how each patient should progress.
We urgently need better tools for simulating disease progression in humans – methods that can precisely tailor predictions to the anatomy and physiology of individual human brains – such that the mechanisms and clinical targets may be better understood
2.Novelty and Importance:
The objective of this project is to build personalised models of Alzheimer’s disease progression that are tailored to the anatomy and physiology of individual human brains. These Digital Twins of disease progression will improve understanding of the mechanisms of disease and ultimately support in silico modelling of interventions and treatments.
This project will build from past work in which we have built tailored generative models of brain ageing [1-2] and physics-constrained simulations of neurodegeneration [3]. It will also leverage the recent explosion of work on neural operator learning, in which adapted neural network architectures learn physical models from data [4-5] in order to smoothly simulate biophysical processes.
3.Aims and Objectives:
The aim is to combine these techniques to scaffold learning through:
– Building deep generative models to predict how the brain’s structural connectome and molecular distribution change with ageing (from multimodal MRI and PET).
– Use this to constrain neural operator learning of a general reaction-diffusion model of disease progression, conditioned on age, sex, APOE status and disease state
– Integrate with a biomechanical model of brain atrophy to precisely simulate neurodegeneration at the resolution of individual brain scans.
1. Bass, Cher, et al. NeurIPS (2020): 7697-7709. https://proceedings.neurips.cc/paper/2020/hash/56f9f88906aebf4ad985aaec7fa01313-Abstract.html
2. Xie, Zhenshan, et al. MICCAI Workshop on Deep Generative Models. 2025. https://link.springer.com/chapter/10.1007/978-3-032-05472-2_30
3. Silva, Mariana Da, et al. MLCN, 2021. https://link.springer.com/chapter/10.1007/978-3-030-87586-2_2
4. Li, Z. et al. (2020). https://arxiv.org/abs/2003.03485
5. Xu, M. et al. ICML abs/2401.11037, (2024). https://openreview.net/forum?id=2UlfvGU6rL