In this paper, a neuromusculoskeletal (NMS) solver-informed artificial neural network (ANN) is proposed to estimate ankle joint torques in seven movements, including walking at fast, slow and self-selected speeds, ankle isokinetic dorsi- and plantarflexion at 60 and 90 degrees/s. The NMS solver-informed ANN model is an extension of a standard ANN model with additional features from an NMS solver, namely ankle joint torque and muscle forces. The standard ANN, the NMS solver-informed ANN and a muscle-driven NMS model, were used to predict ankle torque. Prediction accuracy were compared, based on data capture in 10 subjects. In all methods, we trained the models with measured ankle joint angle and electromyography signals as inputs. Seven different cases were investigated, using trials at different speeds across three movement types (walking, isokinetic plantarflexion and dorsiflexion) to calibrate/train models in the same movement types. The NMS solver-informed ANN model predicted ankle joint torque better than both the NMS and standard ANN models, which indicates benefit gained from integrating NMS features into standard ANN models. The proposed NMS solver informed-ANN model thus shows promise in assistance-as-needed rehabilitation exoskeleton controller design.