Sparse discriminative latent characteristics for predicting cancer drug sensitivity from genomic features
Fig 1
Joint analysis of cell line sensitivity across multiple drugs has the potential to improve predictive accuracy as well as model interpretability.
a. While cancers are genetically heterogeneous there are phenotypic characteristics shared by many cancers subtypes, some of which are illustrated here. We refer to these as latent characteristics (LCs) because they are not directly observable from genomic data, but we hypothesize that each characteristic will have a detectable genetic signature and a defined influence on drug sensitivity. b. Pearson correlation of drug sensitivity profiles (active area scores) across CCLE, annotated by known inhibition targets. c The Lacrosse model consists of two components, shown here graphically. The first is a sparse linear regression from cell line features, F (gene expression, copy number variation and genomic mutations) to continuous valued latent characteristics, X. The second is a sparse factor analysis (matrix factorization) where the latent characteristics (the factors) explain the observed sensitivity scores through a sparse loading matrix G.