Rutting in asphalt pavement is a critical design criterion in the Mechanistic-Empirical Pavement Design Guide (MEPDG). However, many studies have shown that the rutting transfer function in the MEPDG fails to produce reliable predictions in that it calculates simply through a linear combination of the permanent deformations in all susceptible layers. To address this issue, the present study developed two deep neural network (NNs) that can be included in the MEPDG to improve the accuracy of rutting prediction: the first one (NN3) utilized the predicted rutting data by the MEPDG, respectively in the asphalt concrete (AC), granular base and subgrade as the primary inputs, while the other (NN20) further adopted seventeen additional parameters concerning the material, structure, traffic, and climate. To demonstrate the effectiveness of the presented NNs, two multiple linear regression (MLR) models, MLR3 and MLR20, using the same inputs for NN3 and NN20 but developed in the same way with the rutting transfer function in the MEPDG, were employed to act as a performance baseline. The results indicated that both the developed NNs, particularly the NN20, exhibited significantly better predictive performance than the two MLR models, regardless of whether they were in training or testing. As a complement to interpret the NN models, the importance measures from the random forest showed that the transfer function in the MEPDG may have excluded some crucial variables such as the air voids in the AC, and thus caused its unsatisfactory predictive performance.