<p>For choice problems in transportation, machine learning and deep learning are alternative methods to traditional choice models. While several works explored the potential of this technology for modelling mode choice, lower attention is given to route choice, especially in public transport. In this work, we propose a deep learning model designed specifically for route choice in public transport. The model can estimate a nonlinear utility function, allowing complex interactions among the variables; it can easily include non-alternative specific variables, such as weather or socio-demographic information. Moreover, compared to the traditional choice models, it numerically outperforms the Path Size Logit Model in prediction performance, and does not require pre-specification of the model by an experienced human modeler. These properties are particularly useful for route choice analyses, to capture possible heterogeneities or complex behavior, which are difficult to model a priori. We evaluated the interpretability of the model observing the marginal rates of substitution and applying Accumulated Local Effects, showing meaningful effects of the variables on the probability to choose an alternative. We tested the proposed model on a large-scale dataset based on GPS tracking. We considered both synthetic choices, to demonstrate the model properties, and real choices, to evaluate the model in practice. The results showed moderately better performance of the deep learning model compared to the Path Size Logit, confirming the possibility of using it for modeling and predicting route choice.</p>