Despite the success of deep neural networks (DNNs) in RNA secondary structure prediction, these models have yet to utilize comparative genomic information. We introduce PhyloDNNfold, an innovative model that encodes characteristic substitution patterns from species evolution into DNNs, aiming to improve prediction accuracy. PhyloDNNfold incorporates this comparative genomic data as a constraint in the post-processing network of E2Efold. Our experiments on synthetic and real-world RNAs show substantial improvements in performance, especially in complex structures, confirming that integrating evolutionary signals enhances prediction accuracy. PhyloDNNfold offers a novel approach to RNA secondary structure prediction, demonstrating the potential of incorporating evolutionary information into DNN architectures for biological data analysis. The model also provides a flexible template that can be applied to a broader range of RNA structure prediction methods.