The categorization of cancer subtypes is critical for cancer analysis and diagnosis. Deep learning algorithms have centred substantial recognition for cancer subtype detection in recent years; but, It's difficult to create a Neural Network (NN), and the results of deep learning approaches aren't always predictable are mostly reliant on their structure. In gene expression datasets, a learning strategy takes a long time and the output of the representation diminishes due to duplicated genes and the curse of dimensionality. To increase the classifier's performance and address these difficulties, Enhanced Artificial Bee Colony based Flexible Neural Forest (EABC-FNT) and Ensemble Gene Selection (EGS) is proposed in this paper. The EABC algorithm is used to optimise the parameters of the EABC-FNT classifier to help to utilize the cancer subtypes classification. In the EABC algorithm, new modified onlooker bee behaviour is used with the better fitness food source as the middle. FNT is a particular NN through the benefit of formation and parameter tuning that may be utilised for multi-class classification. The Fisher Ratio (FR), Neighbourhood Rough Set (NRS), Correlation Based Gene Selection (CFS) and Greedy Hill climbing method are combined with the EGS algorithm. It is used to select primarily beneficial genes with data on the expression of a known breast cancer gene. The FR is used for the elimination of useless genes and NRS is then introduced to eliminate redundant genes. Experimentation on Ribonucleic Acid (RNA)-seq gene expression data of Breast Invasive Carcinoma (BRCA), Glioblastoma Multiforme (GBM), Lung Cancer (LUNG) show with the purpose of proposed EABC- FNT classifier gives higher accuracy with selected genes for cancer subtypes classification when compared to other methods such as Deep Flexible Neural Forest (DFNForest) and FNT classifier concerning the metrics like precision, recall, f-measure, accuracy, and error rate.