[Paper Review] CloudPred: Predicting Patient Phenotypes From Single-cell RNA-seq

Bryan He, Matthew Thomson+4 more|arXiv (Cornell University)|Oct 13, 2021

Single-cell and spatial transcriptomics20 references4 citations

TL;DR

CloudPred is a novel, interpretable machine learning framework that predicts patient phenotypes—such as lupus status—from single-cell RNA-seq data by modeling each patient's cell population as a differentiable Gaussian mixture. It outperforms existing methods with an AUROC of 0.98 on a lupus dataset, identifying a specific CD4+ T cell subpopulation as highly predictive of disease, while handling variable cell counts and preserving single-cell resolution without pseudo-bulk transformation.

ABSTRACT

Single-cell RNA sequencing (scRNA-seq) has the potential to provide powerful, high-resolution signatures to inform disease prognosis and precision medicine. This paper takes an important first step towards this goal by developing an interpretable machine learning algorithm, CloudPred, to predict individuals' disease phenotypes from their scRNA-seq data. Predicting phenotype from scRNA-seq is challenging for standard machine learning methods -- the number of cells measured can vary by orders of magnitude across individuals and the cell populations are also highly heterogeneous. Typical analysis creates pseudo-bulk samples which are biased toward prior annotations and also lose the single cell resolution. CloudPred addresses these challenges via a novel end-to-end differentiable learning algorithm which is coupled with a biologically informed mixture of cell types model. CloudPred automatically infers the cell subpopulation that are salient for the phenotype without prior annotations. We developed a systematic simulation platform to evaluate the performance of CloudPred and several alternative methods we propose, and find that CloudPred outperforms the alternative methods across several settings. We further validated CloudPred on a real scRNA-seq dataset of 142 lupus patients and controls. CloudPred achieves AUROC of 0.98 while identifying a specific subpopulation of CD4 T cells whose presence is highly indicative of lupus. CloudPred is a powerful new framework to predict clinical phenotypes from scRNA-seq data and to identify relevant cells.

Motivation & Objective

• To develop a machine learning framework that directly predicts clinical phenotypes from single-cell RNA-seq data, overcoming limitations of standard methods.
• To address the challenge of variable cell counts across patients and high data heterogeneity in scRNA-seq without relying on pseudo-bulk averaging.
• To identify biologically relevant cell subpopulations predictive of disease status without prior annotation, enabling interpretable phenotype prediction.
• To create a scalable, end-to-end differentiable model that leverages geometric structure in scRNA-seq data as 'point clouds' for improved generalization.
• To validate the method on real-world datasets and provide a systematic simulation platform for benchmarking future scRNA-seq phenotype prediction methods.

Proposed method

• CloudPred models each patient's scRNA-seq data as a mixture of Gaussians, where each component represents a putative cell subpopulation with learned mean and covariance.
• It uses end-to-end differentiable optimization via stochastic gradient descent to jointly learn cell subpopulation parameters and phenotype prediction weights.
• The algorithm computes patient-specific cell population abundances as continuous bag-of-features, which serve as input to a classifier for phenotype prediction.
• Unlike standard clustering, CloudPred optimizes mixture components not just for data fit but for predictive power on the phenotype, integrating clinical labels during training.
• The method is initialized with unsupervised clustering but is fine-tuned using phenotype labels to identify biologically salient subpopulations.
• A synthetic data generation framework is developed to simulate patient-level scRNA-seq data with controlled disease signatures and variability for method evaluation.

Experimental results

Research questions

• RQ1 Can a machine learning model accurately predict patient phenotypes such as disease status from single-cell RNA-seq data despite variable cell counts and high heterogeneity?
• RQ2 Does an end-to-end differentiable mixture model outperform standard pseudo-bulk or clustering-based approaches in phenotype prediction accuracy?
• RQ3 Can the model automatically identify biologically relevant cell subpopulations that are predictive of disease without relying on prior annotations?
• RQ4 How does model performance scale with the number of patients and cells in the training set?
• RQ5 Can the model identify specific cell types whose presence or abundance correlates strongly with clinical outcomes, such as lupus or race?

Key findings

• CloudPred achieved an AUROC of 0.98 in predicting lupus status from scRNA-seq data, significantly outperforming alternative methods across multiple evaluation settings.
• The model identified a specific subpopulation of CD4+ T cells whose abundance was highly predictive of lupus, consistent with known immunological roles of T cells in autoimmunity.
• On the lupus dataset, CloudPred maintained high performance even with as few as 20 training patients, demonstrating strong generalization from limited data.
• Performance stabilized after approximately 1,000 cells per patient, suggesting that sequencing more patients may be more beneficial than increasing per-patient sequencing depth.
• CloudPred made phenotype predictions in under one second per sample on a standard CPU, even with 100,000 cells, indicating strong computational scalability.
• Systematic simulations showed that CloudPred consistently outperformed alternative methods, including DeepSets and pseudo-bulk baselines, across diverse synthetic data configurations.