[Paper Review] DropLasso: A robust variant of Lasso for single cell RNA-seq data

Beyrem Khalfaoui, Jean‐Philippe Vert|arXiv (Cornell University)|Feb 26, 2018

Single-cell and spatial transcriptomics31 references3 citations

TL;DR

DropLasso is a robust variant of Lasso regression tailored for single-cell RNA-seq data, incorporating dropout regularization to handle high levels of dropout noise. It improves molecular signature identification by selecting biologically relevant genes—particularly in cell cycle and mitosis pathways—outperforming elastic net in both simulated and real scRNA-seq data.

ABSTRACT

Single-cell RNA sequencing (scRNA-seq) is a fast growing approach to measure the genome-wide transcriptome of many individual cells in parallel, but results in noisy data with many dropout events. Existing methods to learn molecular signatures from bulk transcriptomic data may therefore not be adapted to scRNA-seq data, in order to automatically classify individual cells into predefined classes. We propose a new method called DropLasso to learn a molecular signature from scRNA-seq data. DropLasso extends the dropout regularisation technique, popular in neural network training, to esti- mate sparse linear models. It is well adapted to data corrupted by dropout noise, such as scRNA-seq data, and we clarify how it relates to elastic net regularisation. We provide promising results on simulated and real scRNA-seq data, suggesting that DropLasso may be better adapted than standard regularisa- tions to infer molecular signatures from scRNA-seq data.

Motivation & Objective

• Address the challenge of high dropout rates in single-cell RNA-seq data, which distort standard regression methods.
• Develop a supervised method to learn sparse molecular signatures for classifying cells into predefined types.
• Adapt dropout regularization—originally used in deep learning—for feature selection in noisy, zero-inflated scRNA-seq data.
• Improve upon existing regularized methods like elastic net by accounting for data-specific noise patterns in scRNA-seq.
• Enable more biologically interpretable and robust gene signatures by leveraging data augmentation through dropout.

Proposed method

• Extend dropout regularization from neural networks to linear models, applying random feature masking during training to simulate dropout noise.
• Formulate DropLasso as a regularized regression problem where the loss function incorporates stochastic dropout of features during optimization.
• Use Monte Carlo sampling to approximate the expected loss under dropout, enabling optimization via stochastic gradient descent.
• Relate DropLasso to elastic net regularization, showing theoretical parallels while emphasizing its data-dependent, noise-aware nature.
• Apply the method to learn sparse, interpretable gene signatures that predict cell type labels from scRNA-seq profiles.
• Integrate dropout noise patterns that reflect the known gene expression level dependence of dropout in scRNA-seq.

Experimental results

Research questions

• RQ1 Can dropout regularization, originally developed for deep learning, be effectively adapted to improve feature selection in single-cell RNA-seq data?
• RQ2 How does DropLasso compare to elastic net in identifying biologically relevant gene signatures under high dropout conditions?
• RQ3 Does modeling dropout noise as a form of data augmentation improve the robustness and accuracy of molecular signature learning in scRNA-seq?
• RQ4 Can DropLasso recover known functional gene sets (e.g., cell cycle, mitosis) more effectively than standard regularization methods?
• RQ5 What is the impact of data-dependent regularization on signature stability and biological interpretability in scRNA-seq applications?

Key findings

• DropLasso outperformed elastic net in identifying functionally relevant genes, selecting 21 genes related to 'cell division' compared to 10 by elastic net.
• For the 'cell cycle' term, DropLasso selected 30 functionally annotated genes versus 9 by elastic net, indicating superior functional coherence.
• DropLasso identified 22 genes related to 'mitosis' versus 10 by elastic net, further supporting its enhanced ability to recover biologically meaningful signatures.
• On simulated data, DropLasso demonstrated improved stability and accuracy in gene selection under high dropout conditions.
• The method showed consistent performance across multiple real scRNA-seq datasets, with signatures enriched in known biological pathways.
• DropLasso’s performance is attributed to its data-dependent regularization, which better models the structure of dropout noise in scRNA-seq.