Background: Selecting feature genes to predict phenotypes is one of the typical tasks in analyzing genomics data. Though many general-purpose algorithms were developed for prediction, dealing with highly correlated genes in the prediction model is still not well addressed. High correlation among genes introduces technical problems, such as multi-collinearity issues, leading to unreliable prediction models. Furthermore, when a causal gene (whose variants have an actual biological effect on a phenotype) is highly correlated with other genes, most algorithms select the feature gene from the correlated group in a purely data-driven manner. Since the correlation structure among genes could change substantially when condition changes, the prediction model based on not correctly selected feature genes is unreliable. Therefore, we aim to keep the causal biological signal in the prediction process and build a more robust prediction model. Method: We propose a grouping algorithm, which treats highly correlated genes as a group and uses their common pattern to represent the group's biological signal in feature selection. Our novel grouping algorithm can be integrated into existing prediction algorithms to enhance their prediction performance. Our proposed grouping method has two advantages. First, using the gene group's common patterns makes the prediction more robust and reliable under condition change. Second, it reports whole correlated gene groups as discovered biomarkers for prediction tasks, allowing researchers to conduct follow-up studies to identify causal genes within the identified groups. Result: Using real benchmark scRNA-seq datasets with simulated cell phenotypes, we demonstrate our novel method significantly outperforms standard models in both (1) prediction of cell phenotypes and (2) feature gene selection.