American Journal for Young Scientists · ISSN 3068-9015
Uncovering Transcriptomic Structure Across Cancers Using Deep Learning and Pareto Task Inference
Authors: Noah Kaleekal
Published: September 18, 2025
Category: Biology
Abstract
The heterogeneity of cancer poses a significant challenge for accurate classification and mechanistic understanding. In this study, we analyze RNA sequencing data from twelve distinct cancer types using a deep generative modeling framework. We employ a variational autoencoder (VAE) to learn low-dimensional representations of gene expression profiles and visualize the resulting latent space using UMAP. Hierarchical clustering of latent centroids reveals biologically coherent groupings among cancer types. To evaluate the discriminability of the learned features, we train a shallow multilayer perceptron (MLP) classifier, which achieves high accuracy in distinguishing tumor types, as evidenced by the confusion matrix. Furthermore, we apply Pareto Task Inference to explore trade-offs between classification objectives and uncover latent biological relationships among cancer types. Our results demonstrate that deep generative models can extract meaningful structure from transcriptomic data and support interpretable analyses for cancer subtype differentiation and multi-task learning in oncology