LUAD-SynthNet: Generative Adversarial Networks for Synthetic Single-Cell Transcriptomics in Lung Adenocarcinoma

CARLA 20 25 v/s LUAD-SynthNet: Generative Adversarial Networks for Synthetic Single-Cell Transcriptomics in Lung Adenocarcinoma Motivation: Data: Data collection was conducted using a highly reliable, publicly accessible source: the National Library of Medicine (NLM). Specifically, the GSE154826 dataset was accessed, which is available through the Gene Expression Omnibus (GEO) database of the National Center for Biotechnology Information (NCBI). This catalog is a global reference for gene expression profiling studies, ensuring that the data used in this research come from a reputable repository with rigorous validation. The scarcity of biological samples and the high variability of tumor microenvironments present significant challenges for building robust models in cancer research. In response, this study proposes LUAD-SynthNet, a generative framework for producing synthetic single-cell RNA-seq (scRNA-seq) gene expression profiles representative of lung adenocarcinoma (LUAD) cells. Goal: The goal is to expand existing datasets with highfidelity synthetic data using Generative Adversarial Network (GAN) to facilitate the training and hypothesis generation of new prediction models in cancer research. Conclusion Methodology Joaquín Araya Bustos , Welinton Barrera Mondaca , Renato Álvarez Ramos , Claudia Cancino Quiroz , Jorge vergara-Quezada , Ana Moya-Beltrán 1 1 1 1 2 2 Escuela de Informática, Facultad de Ingeniería, Universidad Tecnológica Metropolitana, Santiago, Chile. Departamento de Informática y Computación, Facultad de Ingeniería, Universidad Tecnológica Metropolitana, Santiago, Chile. 1 2 Methodological Limitation: The Instability of GAN Architectures The inherent instability of generative models can be faced scaling up the complexity of the methodology. Failure of Standard Models: A standard GAN failed completely due to mode collapse. This forced us to discard simpler approaches and invest time and computational resources into implementing a more advanced and stable architecture, the WGAN-GP. Preprocessing Limitation: The Extreme Nature of Genomic Data The main technical barrier was not the model itself, but the distribution of the input data. Extremely Skewed Data: The high dispersion of gene expression data (with most values being zero) created a "learning landscape" that was nearly impossible for the GAN to navigate. Need for Advanced Techniques: Standard normalization techniques were not enough. The project's success depended on implementing a nontrivial and computationally more expensive solution. The Quantile Transformer was applied to transform the data into a distribution that facilitates effective model learning. Limitation in Evaluation and Validation Finally, a limitation inherent to the entire field of generative models is the lack of a single "success" metric. Multifaceted Validation: It is necessary to combine multiple analyses (PCA, KDE, correlation heatmaps, error metrics, etc.) to build a convincing case for the model's fidelity. Limitations of the Study [email protected] [email protected] Contact: Variance is not simply statistical "noise"; it reflects the diversity of functional states, cell cycles, and phenotypic responses that exist in a population of real cells. The graph not only exhibit a numerical difference; it also expose a fundamental and systematic limitation of the generative model: its inability to replicate cellular heterogeneity. This graph describe a model that has successfully learned the "who's who" of the transcriptome, but still hasn't mastered the "quantum" of its expression, revealing quantitative biases that limit its direct application. 1. 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 AI25-11, and in part by the “Scientific and Technological Equipment Projects Competition, year 2024, code LE24-03” 2. The graph shows the distribution of expression values (logarithmic scale) for several relevant genes, comparing real data (green) with synthetic data (blue). Each point represents an individual cell. Side-by-side visualization allows for visual assessment of the fidelity of synthetic data with respect to real patterns. The high concordance in the shape, range, and dispersion of the distributions suggests that the synthetic data correctly capture the key statistical characteristics of the original data, which is essential for applications in computational analysis and model validation. This work successfully establishes a robust pipeline for generating LUAD genomic data using a WGAN-GP and advanced preprocessing. The results demonstrate high visual fidelity, replicating the global structure, individual distributions, and complex interdependencies between genes (Correlation Heatmaps). However, quantitative metrics reveal subtle overfitting and underestimation of variance in certain genes, limitations expected given the size of the training dataset. These findings, while indicating that the model is "fair" in terms of statistical perfection, are crucial and directly guide our future work. The samples used in this study were classified according to their origin and the patient's diagnosis: a) patients with lung adenocarcinoma (LUAD) and b) patients with lung squamous cell carcinoma (LUSC). Two type of samples were obtained per patient: primary tumor tissue and tissue from a different anatomical site. The study was limited to the analysis of 37 lung adenocarcinoma (LUAD) samples, selected from a total set of 49 samples. Generator Input: A random noise vector (100 dims) that acts as a seed. Function: An expander network that transforms the noise into a synthetic genetic profile (2,404 dims). It uses Linear, BatchNorm1d (for stability), and LeakyReLU layers. The final Tanh layer scales the output to the range [-1, 1]. Critical Input: A genetic profile (2,404 dims), either real or synthetic. Function: A reducer network that evaluates the sample and compresses it into a single realism score. It uses Linear, LeakyReLU, and Dropout (for regularization) layers. Crucially, the output layer has no activation function. Both networks were trained competitively, using the Adam optimizer and the BCEWithLogitsLoss loss function, until the Generator becomes so good that the Discriminator can no longer distinguish fake data from real data. Metric Conflict: Heatmaps showed resounding success. While the more stringent quantitative test revealed the problem of overfitting. This demonstrates the limitation of relying solely on visual inspection and the need for more rigorous testing. Direct Comparison of Mean Expression Variance Comparison per Gene Step 1: Initial Setup Objective: Generate cancer data with a GAN Failure 1: Data loading error 'ValueError: num_samples=0' Diagnosis: dropna() removes all rows Solution: Implement a robust data_loader Step 2: Train Standard GAN Failure 2: Mode Collapse PCA/KDE show separate clusters Diagnosis: The standard GAN is too unstable Step 3: Implement WGAN-GP Step 4: Train WGAN-GP Step 5: Hypothesis: The data distribution must be 'smoothed' Step 6: Train WGANGP with transformed data Failure 3: The failure persists Same visual results as in Failure 2 Diagnosis: The problem is not the model, but the data distribution Solution: Use QuantileTransformer to normalize the distribution Solution: Use QuantileTransformer to normalize the distribution Success! PCA, KDE, and Heatmap show excellent overlap and correlation Reference: 1.- Maier B, Leader AM, Chen ST. A conserved dendritic-cell regulatory program limits antitumour immunity. Nature. 2020;580:257–262. 2.- Godec T, Gvozdenovic E. LncRNAWiki: a comprehensive resource for long noncoding RNAs. Nucleic Acids Res. 2020;48(15):e85. doi:10.1093/nar/gkaa527