Varying and irregular sea states impose risk to shipping, preventing lifting, or cargo transfer operations, resulting in extreme or unsafe motions. Shipping companies govern the feasibility of operation based on average wave conditions, meaning there are periods in an unsafe sea state where a task could be performed. By predicting the vessel's response, the productivity of operations can be improved, minimising costs. However, traditional methods, including response amplitude operators (RAOs), have limited usability because of linear backbone, static loadings and geometric assumptions. Numerical models are too computationally expensive to be incorporated. Intelligent, data-driven models have increased in popularity, utilising data to inform predictions. Ship motion predictions have been captured using regression and autoregressive models, and recurrent-based neural networks utilising time-series data. However, these models are data-driven, they neglect the physical relationship among data points. To account for this, a transformer-based physics-informed neural network (PINN) is proposed. Utilising a transformer encoder-decoder architecture with sinusoidal positional encoding to reflect the historical features of the time series. While taking advantage of the known physical second-order relationship between vessel roll, and wave excitation as an additional loss function using PINNs; Allowing for the estimation of the vessels added mass, damping and restoring coefficients, in addition to the relationship between the encountered waves and the vessels subsequent roll response. This study aims to utilise a transformer-PINN hybrid model to predict vessel roll response. This hybrid model shows improved accuracy in roll prediction compared to data-driven baseline models. The outcome of this study will inform the applicability of PINNs to the remaining vessel degrees of freedom, enabling rapid prediction for applications across the maritime industry.