Gradient-based surrogate optimization usually has a fast convergence capability. However, it can be easily stuck in local minima, especially when the electromagnetic (EM) response of the starting point is far away from the design specification. This article proposes a novel feature and EM sensitivity coassisted neuro-transfer function (TF) surrogate optimization for microwave filter design. The proposed technique introduces EM sensitivity information into the pole-zero-based neuro-TF with feature parameters for the first time. New formulations are derived for establishing the adjoint neuro-TF model with poles and zeros as the transfer function parameters. More accurate gradients of the neuro-TF outputs with respect to design variables are subsequently achieved by the training with EM sensitivity. Two sets of feature parameters, i.e., feature frequencies and feature heights, are used in the proposed technique. The adjoint feature frequencies are proposed as the gradients of feature frequencies, which are calculated using the trained adjoint neural network outputs. New formulations are further derived for the gradients of feature heights using both trained adjoint neuro-TF and adjoint neural network outputs. To improve the robustness of the optimization process, the trust region algorithm is also introduced. By the coassistance of feature parameters and EM sensitivities, the proposed technique can achieve a further acceleration over the existing feature-assisted techniques. This article utilizes three microwave filter examples to demonstrate this technique.