Résumé
A publicly available human genome serves as both a valuable resource for researchers and a potential risk to the individual who provided the genome. Many actors with selfish intentions could exploit it to extract information about the donor’s health or that of their relatives. Recent efforts have employed artificial intelligence models to simulate genomic data, aiming to create synthetic datasets with scientific merit while preserving patient anonymity. However, a major challenge arises in dealing with the vast amount of data that constitutes a complete human genome and the resources required to process it. We have developed a dimension reduction method that combines artificial intelligence with our knowledge of in vivo mutation association mechanisms. This approach enables the processing of large amounts of data without significant computational resources. Our genome segmentation follows chromosomal recombination hotspots, closely resembling mutation transmission mechanisms. Training data is sourced from the 1000 Genomes Project, which catalogues over 2500 genomes from diverse ethnic groups. Variational autoencoders, utilising neural networks, serve as an extension to the generative model. Wasserstein Generative Adversarial Networks (WGAN) are a benchmark among generation methods for various data types. After optimisation of our data simulation strategy our pipeline allows the generation of a simulated population meeting several essential criteria. It demonstrates good diversity, closely resembling that found in the reference dataset. It is plausible, as newly generated combinations of mutations do not disrupt the linkage disequilibria found in humans. It also preserves donor anonymity by synthesising combinations of reference genomes that are distant from reference samples.