popV

Welcome to the popV framework. We provide state-of-the-art performance in cell-type label transfer using an ensemble of experts approach. We provide here pre-trained models to transfer cell-types to your own query dataset. Cell-type definition is a tedious process. Using reference data can significantly accelerate this process. By using several tools for label transfer, we provide a certainty score that is well calibrated and allows to detect cell-types, where automatic annotation has high uncertainty. We recommend to manually check transferred cell-type labels by plotting marker or differentially expressed genes before blindly trusting them. This is an open science initiative, please contribute your own models to allow the single-cell community to leverage your reference datasets by asking in our GitHub repository to add your dataset.

---

Model Overview

popV trains up to 9 different algorithms for automatic label transfer and computes a consensus score. We provide an automatic report. To learn how to apply popV to your own dataset, please refer to our tutorial

Algorithms

Currently implemented algorithms are:

• K-nearest neighbor classification after dataset integration with BBKNN
• K-nearest neighbor classification after dataset integration with SCANORAMA
• K-nearest neighbor classification after dataset integration with scVI
• K-nearest neighbor classification after dataset integration with Harmony
• Random forest classification
• Support vector machine classification
• OnClass cell type classification
• scANVI label transfer
• Celltypist cell type classification

---

Key Applications

The purpose of these models is to perform cell-type label transfer. We provide models with (CUML support)[collection] for large-scale reference mapping and (without CUML support)[collection] if no GPU is available. PopV without GPU scales well to 100k cells. PopV has three levels of prediction complexities:

• retrain will train all classifiers from scratch. For 50k cells this takes up to an hour of computing time using a GPU.
• inference will use pretrained classifiers to annotate query as well as reference cells and construct a joint embedding using all integration methods from above. For 50k cells this takes in our hands up to half an hour of computing time using a GPU.
• fast will use only methods with pretrained classifiers to annotate only query cells. For 50k cells this takes 5 minutes without a GPU (without UMAP embedding).

---

Publications

• Original popV paper:
  • Published in Nature Genetics, this paper introduces popV and benchmarks it.

Contact

• GitHub: https://github.com/YosefLab/popV
• User questions: Discourse