Abstract:
1. Generative genomics leverages computational models, particularly transformer-based architectures, to create synthetic genomic data. This study explores their application through a pipeline for gene expression analysis, synthetic data generation, and dataset augmentation, enabling the identification of genes previously missed due to low sample sizes. We compared a transformer-based model, ReALTabFormer, with a statistical model, Gaussian Mixture Model (GMM), for synthetic data generation.
2. The pipeline, implemented in three modular steps—Data Downloader, Synthetic Data Generation, and Data Analysis—supports independent execution or seamless integration using a single GSE_ID. It downloads gene expression data from platforms (GPL10558, GPL570, GPL6480), generates augmented datasets with GMM, and performs enrichment analysis on both original and augmented data. Additionally, an interactive tool was developed to process 384 embeddings derived from abstracts of 15,471 GSE IDs. The tool enables term- or disease-specific searches, retrieves related GSE IDs, visualizes relationships using 3D PCA plots, and generates cosine similarity heatmaps. These functionalities were extended into a fully functional website that integrates data downloading, synthetic data generation, and analysis, though it is not currently hosted.
3. A Retrieval-Augmented Generation (RAG) system was prototyped for Tuberculosis, Ovarian Cancer, and Breast Cancer, integrating KEGG genes, KEGG drugs, and TDT data and capturing genes using NER from Abstracts of Papers available on Pubmed. This system uses both JSON query based approach and Knowledge Graph-based approaches, powered by Llama 3.1 (70B). Current efforts are focused on refining the RAG system only for Tuberculosis to optimize its methodology then further extend it.
This research demonstrates the potential of transformer models in synthetic data generation, addressing challenges of scalability and data quality, and advancing functional genomics and disease research.
