Rank-Based Representations to Learn from Omics Data
summary
Proposes using rank-based representations—quantile normalization and pairwise comparisons—to change omics feature space, and jointly optimizes the quantile target distribution F with a linear predictor (W) to improve survival prediction from mutation and expression data. Uses PPI diffusion then quantile normalization to boost signal, and implements a Kendall tau kernel so pairwise binary matrices are handled implicitly (no explicit P×P storage). Reports mixed benchmarks: method is competitive on several cancer datasets but not a consistent win; behavior depends on preprocessing and dataset.
AI-written summary · about
chapters (14)
- 0:00Start
- 1:20Motivation
- 2:27Machine learning formulation
- 8:07Full quantile normalization
- 10:32Quantile normalization matters
- 14:25Learning the target distribution
- 15:39SUQAN: supervised quantile normalization
- 19:26Outline
- 20:49An idea: Top scoring pairs (TSP)
- 24:15Practical challenge
- 25:00Kernel trick
- 26:58Related work
- 28:59Application: supervised classification Datasets
- 30:33Conclusion
related talks
30mLeveraging Molecular Data for Drug Discovery
24mLearning "Learning to Rank"
44mNetwork Propagation as a Tool for Deciphering Disease Mechanisms
34mPrediction intervals
32mMulti-Omic Integration for Understanding Disease
35mFrom cells to drug responses - machine learning in cancer research
27mModelling the extreme using Quantile Regression
17mPrecision Medicine for All: Using Tidymodels to Validate PRS in Brazil