Samospodbujevano učenje Coxovega regresijskega modela za predstavitev genskih skupin v latentnem prostoru

Enhanced understanding of disease mechanisms on a molecular level leads to more effective treatment decisions. The high-dimensional nature of such data requires dimensionality reduction techniques to extract important patterns. To improve the interpretation and relevance of latent representations, domain knowledge is introduced during modeling by influencing data preprocessing or model architecture, while domain-inspired loss functions are scarcely explored. We propose a novel autoencoder loss function for modeling mRNA concentration based on the first-order differential equation of mRNA dynamics. We decompose the concentration into transcription (synthesis) and mRNA decay and reformulate the problem as survival analysis. By extending the definition of CoxPH partial likelihood, we perform gradient descent through both risk and survival time, which achieves the correct interpretation of both processes. Representations of clinical samples and cell-line data show increased performance on clustering and drug response prediction tasks compared to standard variational autoencoders.