The first part introduces hybrid approaches for data assimilation, where core algorithmic components are replaced with specialized neural networks. Specifically, the Kalman gain matrix within the Kalman filter is substituted with a deep neural operator, and the solution field of the four-dimensional variational method is replaced by a physics-informed neural network. By leveraging the expressivity and trainability of these models, we relax traditional bottlenecks — such as the requirement for large ensemble sizes and the constraints of time-sequential processing — thereby improving computational efficiency. The second part shifts focus to a protein language model (PLM)-based method for screening structure-disrupting mutations, a task inspired by broader computational antigen design initiatives. While machine learning-based protein folding models offer high accuracy, evaluating every possible mutation remains prohibitively expensive. In contrast, PLMs provide a computationally cheap alternative while maintaining the ability to capture nuanced structural information. Various feature extraction techniques will be compared with a case study on the Rift Valley Fever virus.