LIMEADE: Local Interpretable Manifold Explanations for Dimension Evaluations

To visualize and analyze high-dimensional biological data, scientists often turn to manifold learning and dimensionality reduction techniques such as tSNE and UMAP. However, these methods (1) are non-projective, which means that new data cannot be projected on the manifold without refitting, and (2) lack the explainability to help practitioners understand which features drive manifold locations and neighborhoods. In practice, scientists often must turn to marginal distributions along a manifold or expert annotations to explain reduced dimension data. Here, we present Local Interpretable Manifold Explanations for Dimension Evaluations (LIMEADE), a surrogate model integrated with a dimensionality reduction method, similar to the LIME surrogate used in classification and regression models. We define LIMEADE as a group lasso-regularized multi-task regression problem that identifies sparse linear projections of the data aligning with local neighborhoods of the manifold space. When applied to single-cell proteomics data, LIMEADE effectively extracts biologically meaningful features, providing a more interpretable approach to feature selection and dimensionality reduction.