Clustering Gene Co-expression Using 2D Contour Analysis and Autoencoder-based Embedding

Abstract

Identifying co-expressed genes is crucial for understanding biological processes; however, common methods such as Pearson and Spearman correlation rely on assumptions of linearity and monotonicity, respectively, that may not hold for complex biological data. To address these limitations, we propose a framework that captures visual features of gene expression profiles without relying solely on correlation-based methods. Our approach involves converting 3D gene expression data, which contain rich information on gene envelopes over time, into 2D contours that retain important visual information. We then train an autoencoder on the 2D contour images and generate embeddings from them, followed by clustering on the generated embeddings. We also introduce a new clustering algorithm, TAHC, based on hierarchical clustering, which performs better than existing methods in higher dimensions while using cosine similarity. We apply this framework to the Percellome database, which contains gene expression data with two variables--time and dosage--across various experimental conditions (e.g., tissue type). The resulting clusters exhibit good visual coherence, with an overall average Pearson correlation coefficient of 0.81, demonstrating the effectiveness of our approach.