Abstract: The following project is to evaluate different deep-learning frameworks for identification of depression using blood DNA methylation data coming from domestic and open-access cohorts. The frameworks are expected to cover fully-connected models, autoencoders, convolutional models and potentially graph-based and transformer-based models or combinations of these. The idea is to develop a framework with good accuracy which is better or comparable with other ML approaches. As of today no similar works have been performed in the field of psychiatry. The training of models is expected to be performed in a bianca secured environment. The primary challenge of this task is that it requires GPU resources to effectively train the models and also involves sensitive personal data (DNA methylation) thus not making it possible to use other NAISS clusters such as Tetralith, Rackham or Alvis. The inputs of the models could be extremely large tensors and thus are difficult to work with on standard workstations. The main project is expected to take one year since we have already run some pilots.