The prediction of gene mutation of bacteria when exposed to different conditions is beneficial in the development of drugs and vaccines. However, the existing prediction models are suboptimal in their performance. In this work, we propose a time-series approach where we model mutation occurrence sequentially. In our approach, the most connected cofunctional genes are predicted first, and used as additional input features in subsequent predictions. We train a neural network to predict gene mutations in Escherichia coli from the exposure conditions and the predicted state of the cofunctional genes. We test the performance on adaptive laboratory evolution experiments curated from the literature. Results show the cofunctional network-based features improved the predictive performance by a maximum 33.53% and 45.18% in the area under the receiver operating characteristic curve (AUC) and area under the precision-recall curve (AUPRC), respectively. This study supports the feasibility of gene mutation prediction and demonstrates that the functional relationship between genes can be predicted first and then used as features to boost the predictability of its node genes.