How it works

Using tumor state gene expression as input, we apply a transformer-based diffusion model conditioned on normal state gene expression to transform tumor gene expression back to a normal state. Additionally, GEMDiff generates synthetic bulk RNA gene expression starting from Gaussian noise. Model pipeline: 1. Data processing: includes replacing the NA values, applying log transformation, and normalization; 2. Cluster quality assessment: examined by silhouette scores to selected well-clustered gene sets; 3. Model training: utilizes selected gene sets for training the diffusion model; 4. Gene augmentation/Gene perturbation: implemented data augmentation or gene perturbation according to the task types and employed UMAP plots for outcome display; 5. Evaluation: Validate the core genes with gene function enrichment analysis..