GE LDA Survival

GE LDA Survival performs gene expression–based survival prediction in cancer by combining discretized Latent Dirichlet Allocation (dLDA) for dimensionality reduction with Multi-Task Logistic Regression (MTLR) for time-to-event modeling.

Key Features:

• Dimensionality reduction via dLDA: Uses a discretized Latent Dirichlet Allocation model to reduce high-dimensional gene expression data into a topic-based representation.
• Patient representation as cancer-topics: Represents each patient as a low-dimensional mixture distribution over cancer-specific topics derived from gene expression.
• Handling of real-valued gene expression: Extends dLDA by transforming real-valued gene expression into a discrete format suitable for topic modeling.
• Survival prediction with MTLR: Applies Multi-Task Logistic Regression (MTLR) to patient topic distributions for survival/time-to-event prediction.
• Performance evaluation on multiple datasets: Evaluated on METABRIC (1,981 breast cancer samples, microarrays) and Pan-kidney (KIPAN, 883 samples, mRNAseq) datasets.
• Model calibration assessment: Assesses calibration using the D-calibrated measure and reports performance using the Concordance metric.

Scientific Applications:

• Personalized medicine: Provides more precise individual survival estimates to inform tailored treatment planning.
• Cross-cancer applicability: Demonstrates applicability across different cancer types and gene expression modalities (microarrays and mRNAseq).

Methodology:

Discretized Latent Dirichlet Allocation (dLDA) applied to discretized real-valued gene expression to learn cancer-topics, patient topic distributions used as inputs to Multi-Task Logistic Regression (MTLR) for survival prediction, and evaluation performed on METABRIC and Pan-kidney (KIPAN) using Concordance and D-calibrated measures.