Alzheimer’s disease (AD) affects grey and white matter as a consequence of neuronal degeneration processes, but such mechanisms are still poorly understood¹. Combining computational modelling and graph theoretical approaches has been proposed as a way to shed light on AD pathogenesis². In this perspective, a generative disease model could isolate the involved areas and generate hypothesis for new investigations. The aim of this study is to identify the white matter pathways whose impairment could lead to AD structural patterns by simulating neurodegeneration and using graph theory.Forty Alzheimer’s disease patients (AD, 11M/29F, 70+/-5yo) and forty healthy controls (HC, 18M/22F, 60+/-10yo) were scanned with a 3T MRI scanner using a 3D modified driven equilibrium Fourier transform (MDEFT) sequence (TR=1338 ms, TE=2.4 ms, matrix=256x224, number of slices=176, thickness=1 mm) and diffusion tensor imaging (DTI) (TR=7000 ms, TE=85 ms, 61 directions, b factor=1000 smm^-2, resolution=2.3 mm³). Using a previously described approach³, the cortical surface was reconstructed and parcellated into 68 regions (Desikan-Killiany atlas), then the DTI images were used to reconstruct the white matter pathways and realigned to overlap the parcellation scheme. Representing grey matter regions as nodes and white matter tracts that connect them as edges, brain structural networks were represented as weighted graphs⁴. A group threshold was used in order to reduce false positives and negatives⁵. The networks were characterized using global efficiency and overall connection strength³. We then simulated neurodegeneration to discover which pathways, once lesioned, could drive the controls’ network to show patterns of AD. We modelled the simulation as an optimization problem, using the sum of the average global efficiency and global strength values as objective function. Using a greedy approach, we iteratively evaluated the outcome of lesioning each edge and chose the greatest impact edge. The algorithm was stopped when the average metrics were equal to the average for the AD group. Since greedy strategies often lead to only one of the possible solutions of an optimization problem, we repeated the simulation using a genetic algorithm, where 50 random sets of 20 reduced weight edges were used as the population. At each iteration, we performed: selection - sets with the highest global efficiency and strength values were discarded, crossover - new sets were creating selecting edges from the sets with the lowest metrics, mutation - sets were randomly rearranged using edges not present in the initial population. Using the rich-club definition from previous studies⁶, we classified the lesioned connections in rich-club, feeder and local. We used permutation test (10000 permutations) and network-based statistic⁷ (NBS) (threshold=3.1, alpha=0.05, 10000 permutations) for group comparisons.Both global efficiency and overall connectivity strength showed a significant decrease (p=0.0094, p=0.0150) when comparing AD group with controls (fig. 1). The NBS analysis also showed a damaged subnetwork with 12 affected edges (fig. 2), that included frontal and parietal connections as well as the cingulum and the precuneus. The simulation based on the greedy algorithm led to damaged control group (DC) with significantly decreased global efficiency and strength (p=0.0036, p=0.0128) in comparison to the original HC group and, by contrast, there was no significant difference with the AD group. The edges that were damaged by the simulation involved the frontal and parietal networks, the insula and the precuneus as well as temporal and orbital areas (fig. 3). Ten of these connections were classified as feeder ones, three as rich-club and four as local. In the simulations based on the genetic algorithm, the obtained damaged control group showed also a significant decrease in both metrics (p=0.0032, p= 0.0211) when compared to the controls and no significant differences with patients. However, the genetic algorithm showed a more spread out disruption pattern, that mainly included local and feeder connections.Consistently with previous studies on structural networks^8,9, we found reduced global efficiency in the AD group compared to the controls, and we were able to reproduce this aspect in the HC group using lesion simulations. The simulation results highlight the relevant role of the hub regions in the structural disruption of the white matter pathways, and particularly of the feeder connections, which constituted a large fraction of the edges whose damage led to AD disruption pattern. Finally, the genetic algorithm has proven more suitable than the greedy one.To the best of our knowledge, this is the first study that explores the damage process of white matter pathways in AD using computational simulations on the connectome. Using this combined approach, our findings show the relevant role of the pathways between hubs and peripheral regions.