American Journal for Young Scientists · ISSN 3068-9015
Variational Autoencoder Representations Reveal Shared and Distinct Transcriptomic Programs Across Cancers
Authors: Aarian Dasgupta
Published: October 15, 2025
Category: Computational Neuroscience
Abstract
Cancer remains one of the leading causes of death worldwide, responsible for nearly ten million deaths annually and placing a major burden on global health systems. A deeper understanding of its molecular mechanisms is critical for advancing early diagnosis and targeted therapy. High-throughput RNA sequencing provides a rich resource for studying tumor biology, but the high dimensionality of these data presents challenges for analysis and interpretation. Here, we apply a variational autoencoder (VAE) to RNA expression profiles from twelve cancer types to learn low-dimensional, biologically meaningful representations. The VAE embeddings captured distinct tissue-specific patterns, with unsupervised clustering separating tumors by type while also reflecting known biological similarities, such as those between colon and rectal cancers. Supervised classification using these latent features achieved strong accuracy in predicting cancer types, indicating that essential discriminative signals were preserved. Gene set enrichment analysis revealed that individual latent dimensions correspond to established hallmarks of cancer, including proliferative signaling, evasion of apoptosis, angiogenesis, immune evasion, and metabolic reprogramming. These results demonstrate that deep generative models can reduce transcriptomic complexity while retaining interpretable biological structure. By linking machine learning–derived features to cancer hallmarks, our approach provides a framework for integrating computational methods with oncology to guide future precision medicine strategies.