Analysis of gene expressions in DNA microarray data aids in the early detection and accurate classification of different categories of cancer for effective management and treatment of patients. Though the literature is rich in several conventional and deep learning techniques using gene expression analysis for cancer classification, the quest for designing efficient models still exists due to several challenges in analyzing the gene expression data. This paper investigates the potential benefits of combining the respective strengths of LSTM and transformer networks in capturing temporal dynamics and global dependencies in the microarray data for improved accuracy and interpretability of cancer classification models. The proposed method involves preprocessing the gene expression data, filtering genes using a stacked autoencoder, and classification of cancer categories using a parallel fusion model consisting of LSTM and a self-attention-based transformer. The model’s performance is evaluated by gene expression datasets using standard evaluation metrics. Comparative analysis indicates that the proposed model performs better than existing approaches on ten gene expression datasets, demonstrating its potential for enhancing cancer classification accuracy. GO term and pathway analysis confirm that the proposed model effectively identifies genes frequently mutated and associated with cancer. Results of the Student’s t-test further support the significance of the improved performance.