Hydrologic signatures are quantitative metrics that describe a streamflow time series. Examples include annual maximum flow, baseflow index and recession shape descriptors. In this paper, we use machine learning (ML) to learn an optimal equivalent of hydrologic signatures, and use the learnt signatures to build rainfall-runoff models in otherwise ungauged watersheds. Our model has an encoder-decoder structure. The encoder is a convolutional neural net mapping historical flow and climate data to a low-dimensional vector encoding describing watershed function. The encodings are analogous to hydrological signatures. The decoder uses a process-informed network structure to predict streamflow based on current climate data, stored watershed state, static watershed attributes and the encoding. The decoder structure includes stores and fluxes similar to a classical hydrologic model. For each timestep, the decoder predicts coefficients that distribute precipitation between stores and store outflow coefficients. The model is trained end-to-end on the U.S. CAMELS watershed dataset to minimize streamflow error . Using learnt signatures as input to the process-informed model improves prediction accuracy over benchmark configurations that use classical signatures or no signatures. Median NSE performance on 100 watersheds excluded from the training set was 0.69. The process-informed model structure simulates hydrologic dynamics such as snow accumulation and melt, quickflow and baseflow. We interpret learnt signatures by correlation with classical signatures, and by using sensitivity analysis to assess their impact on modeled store dynamics. We conclude that process-informed ML models and other applications using hydrologic signatures may benefit from replacing expert-selected signatures with learnt signatures.