GenomeDefender: Validated High-Precision Detection of Data Poisoning Attacks in Single-Cell RNA-seq Data using a Multi-Model Ensemble

GenomeDefender: Validated High-Precision Detection of Data Poisoning Attacks in Single-Cell RNA-seq Data using a Multi-Model Ensemble Authors: Renato Álvarez Ramos, Welinton Barrera Mondaca1, Joaquín Araya Bustos1, Claudia Cancino Quiroz1, Victor Escobar Jeria2 , Ana Moya-Beltrán2. Affiliations:1Escuela de Informática, Facultad de Ingeniería, Universidad Tecnológica Metropolitana, Santiago, Chile. 2Departamento de Informática y Computación, Facultad de Ingeniería, Universidad Tecnológica Metropolitana, Santiago, Chile. As omics technologies generate increasingly large-complex datasets, cybersecurity techniques are becoming essential to ensure data integrity in biomedical research. Single-cell RNA sequencing (scRNA-seq), a key tool in modern genomics, is particularly vulnerable to data poisoning attacks that manipulate gene expression matrices to degrade model performance and distort biological interpretation. Common threats include label flips, synthetic cell injection, gene scaling, and structured noise—many of which mimic natural biological variation, making them difficult to detect with conventional QC methods. To address this critical vulnerability, we present GenomeDefender, a high-precision anomaly detection tool specifically designed for scRNA-seq data. GenomeDefender integrates a multi-model ensemble architecture comprising four optimized deep learning components: Variational Autoencoders (VAEs), Graph Neural Network Autoencoders (GNN-AEs), Contrastive Autoencoders (CAEs), and Denoising Diffusion Probabilistic Models (DDPMs). Each is tailored to detect specific poisoning vectors, with model pairings strategically selected for complementary strengths. The pipeline includes preprocessing, parallelized model execution, and a decision fusion module for robust classification of clean vs. poisoned data. GenomeDefender consistently achieves precision above 95% across all evaluated attack types, validated through extensive benchmarking using real-world single-cell RNA-seq datasets (GSE154826). This level of accuracy is enabled by the synergistic design of lightweight, three-layer detection architectures and strategic ensemble fusion. The tool supports incremental training, making it adaptable to evolving data distributions and new attack variants. GenomeDefender brings cybersecurity-driven robustness to scRNA-seq pipelines, delivering a scalable and reproducible solution within HPC frameworks and reinforcing the trustworthiness of genomic discoveries and precision medicine initiatives. Acknowledgement: Departamento de Informática y Computación, UTEM; Escuela de Informática, UTEM; Laboratorio de Investigación Aplicada, Departamento de Informática y Computación, UTEM. This work was supported in part by “Competition for Research Assistant Funding UTEM”, year 2024, code AI24-11, and in part by the “Scientific and Technological Equipment Projects Competition, year 2024, code LE24-03”.