Cancer is one of the most significant global health challenges, arising from genetic alterations in gene expression patterns that lead to uncontrolled cell growth an‎d the formation of malignant tumors. Advances in next-generation sequencing technologies—especially at the single-cell level—have enabled more precise identification of cell types an‎d deeper analysis of the tumor microenvironment. However, the scarcity of high-quality data, technical noise, an‎d class imbalance continue to pose serious difficulties in analyzing this information. In this study, a novel method based on Generative Adversarial Networks is proposed for generating synthetic single-cell gene expression data. This approach is designed to compensate for data scarcity an‎d protect patient privacy while preserving key biological characteristics an‎d maintaining statistical consistency with real data. Raw data were preprocessed using the Seurat method an‎d transformed into image-like representations through GenoMap, organizing them in a way that retains biological structure an‎d gene–gene relationships among cells. The proposed model, Contrastive IP3G, consists of two main components: a GAN-based generator an‎d discriminator that produce synthetic gene expression data an‎d align their distribution with that of real data, an‎d a contrastive learning module that defines positive an‎d negative cell pairs to structure latent representations such that similar cells are mapped close together while dissimilar cells are pushed apart. The integration of these two mechanisms allows the model not only to reconstruct the statistical structure of the data but also to automatically an‎d unsupervisedly reveal the biological boundaries between different cell types. Experimental results demonstrate that Contrastive IP3G generates synthetic data with remarkable structural an‎d statistical similarity to real data, effectively reproducing gene–gene correlations. Incorporating the synthetic data produced by this model enhanced the performance of SVM an‎d CNN classifiers, reducing accuracy fluctuations to less than 5% an‎d 10%, respectively. Moreover, the model outperformed the baseline IP3G by approximately 38% an‎d 60% in the KID an‎d MMD eva‎luation metrics, respectively.

عبدالرضا ميرزايي

محمدحسين منشئي , مهران صفاياني

فرزانه شايق بروجني , سمانه حسيني