Single cell transcriptomics enables a high-resolution concurrent measurement of gene expression from tens of thousands of cells We developed a method for determining standardized profiles from SCT data We defined 48 data sets from 13 different studies and developed single-cellderived-class” (SCDC) profiles representing multiple classes and subclasses of peripheral blood mononuclear cells (PBMC) We applied pattern recognition analysis by calculating the distance from each query cell to the SCDC profiles (excluding the profiles of the query cells) Classification of cells by pattern recognition showed excellent performance for PBMC that were isolated, but not further processed by cell sorting

Prediction of PBMC Cell Types Using scRNAseq Reference Profiles

Background Single-cell transcriptome (SCT) sequencing technology has reached the level of high-throughput technology where gene expression can be measured concurrently from large numbers of cells. The results of gene expression studies are highly reproducible when strict protocols and standard operating procedures (SOP) are followed. However, differences in sample processing conditions result in significant changes in gene expression profiles making direct comparison of different studies difficult. Unsupervised machine learning (ML) uses clustering algorithms combined with semi-automated cell labeling and manual annotation of individual cells. They do not scale up well and a workflow used on a specific dataset will not perform well with other studies. Supervised ML classification shows superior classification accuracy and generalization properties as compared to unsupervised ML methods. We describe a supervised ML method that deploys artificial neural networks (ANN), for 5-class classification of healthy peripheral blood mononuclear cells (PBMC) from multiple diverse studies. Results We used 58 data sets to train ANN incrementally – over ten cycles of training and testing. The sample processing involved four protocols: separation of PBMC, separation of PBMC + enrichment (by negative selection), separation of PBMC + FACS, and separation of PBMC + MACS. The training data set included between 85 and 110 thousand cells, and the test set had approximately 13 thousand cells. Training and testing were done with various combinations of data sets from four principal data sources. The overall accuracy of classification on independent data sets reached 5-class classification accuracy of 94%. Classification accuracy for B cells, monocytes, and T cells exceeded 95%. Classification accuracy of natural killer (NK) cells was 75% because of the similarity between NK cells and T cell subsets. The accuracy of dendritic cells (DC) was low due to very low numbers of DC in the training sets. Conclusions The incremental learning ANN model can accurately classify the main types of PBMC. With the inclusion of more DC and resolving ambiguities between T cell and NK cell gene expression profiles, we will enable high accuracy supervised ML classification of PBMC. We assembled a reference data set for healthy PBMC and demonstrated a proof-of-concept for supervised ANN method in classification of previously unseen SCT data. The classification shows high accuracy, that is consistent across different studies and sample processing methods.

https://biorxiv.org/content/10.1101/2021.07.29.454293v1.full.pdf

Artificial Neural Networks for classification of single cell gene expression

An Integrated machine learning and DEA-predefined performance outcome prediction framework with high-dimensional imbalanced data

Background: Single cell transcriptomics is a new technology that enables us to measure the expression levels of genes from an individual cell. The expression information reflects the activity of that individual cell which could be used to indicate the cell types. Chronic lymphocytic leukemia (CLL) is a malignancy of B cells, one of the peripheral blood mononuclear cells subtypes. We applied five analytical tools for the study of single cell gene expression in CLL course of therapy. These tools included the analysis of gene expression distributions: median, interquartile ranges, and percentage above quality control (QC) threshold; hierarchical clustering applied to all cells within individual single cell data sets; and artificial neural network (ANN) for classification of healthy peripheral blood mononuclear cell (PBMC) subtypes. These tools were applied to the analysis of CLL data representing states before and during the therapy.
Results: We identified patterns in gene expression that distinguished two patients that had complete remission (complete response), a patient that had a relapse, and a patient that had partial remission within three years of Ibrutinib therapy. Patients with complete remission showed a rapid decline of median gene expression counts, and the total number of gene counts below the QC threshold for healthy cells (670 counts) in 80% of more of the cells. These patients also showed the emergence of healthy-like PBMC cluster maps within 120 days of therapy and distinct changes in predicted proportions of PBMC cell types.
Conclusions: The combination of basic statistical analysis, hierarchical clustering, and supervised machine learning identified patterns from gene expression that distinguish four CLL patients treated with Ibrutinib that experienced complete remission, partial remission, or relapse. These preliminary results suggest that new bioinformatics tools for single cell transcriptomics, including ANN comparison to healthy PBMC, offer promise in prognostics of CLL.

/pdf/prediction-of-therapy-outcomes-of-cll-using-gene-expression-sk6d49v7re.pdf

Prediction of therapy outcomes of CLL using gene expression intensity, clustering, and ANN classification of single cell transcriptomes

Peripheral blood mononuclear cells (PBMC) are mixed subpopulations of blood cells composed of five cell types. PBMC are widely used in the study of the immune system, infectious diseases, cancer, and vaccine development. Single-cell transcriptomics (SCT) allows the labeling of cell types by gene expression patterns from biological samples. Classifying cells into cell types and states is essential for single-cell analyses, especially in the classification of diseases and the assessment of therapeutic interventions, and for many secondary analyses. Most of the classification of cell types from SCT data use unsupervised clustering or a combination of unsupervised and supervised methods including manual correction. In this chapter, we describe a protocol that uses supervised machine learning (ML) methods with SCT data for the classification of PBMC cell types in samples representing pathological states. This protocol has three parts: (1) data preprocessing, (2) labeling of reference PBMC SCT datasets and training supervised ML models, and (3) labeling new PBMC datasets from disease samples. This protocol enables building classification models that are of high accuracy and efficiency. Our example focuses on 10× Genomics technology but applies to datasets from other SCT platforms. 

Protocol for Classification Single-Cell PBMC Types from Pathological Samples Using Supervised Machine Learning.

We implemented an automated system for single-cell classification using artificial neural networks (ANN). Our system takes single-cell gene expression sparse matrices and trains ANN to classify cell types and subtypes. The assemblies of ANNs predict cell classes by voting. We tested the system in a case study where we trained ANNs with a dataset containing approximately 120,000 single cells and tested the resulting model using an independent data set of 13,000 single cells. The overall accuracy of the 5-class classification was 95%. We trained and tested a total of 100 ANNs in 10 cycles. The prediction system demonstrated excellent reproducibility. The analysis of misclassifications indicated that 2% were likely classification errors, while the remaining 3% were likely due to mislabeled types and subtypes in the test set.

Artificial Neural Network System for Cell Classification using Single Cell RNA Expression

Single cell transcriptomics (SCT) technology reveals gene expression of individual cells. Peripheral blood mononuclear cells (PBMC) are important diagnostic targets in immunology. In this study, we obtained and standardized 27 SCT data sets, derived from healthy PBMC samples using 10x SCT. We used artificial neural networks (ANN) to assess the ability of ANN to classify main PBMC cell types. Incremental learning by the gradual addition of new data sets to ANN training improved classification. The overall prediction accuracy of the final step of incremental learning reached 93% in 4-class classification.

Xin Lin

Papers

Artificial Neural Network System for Cell Classification using Single Cell RNA Expression

Classification of PBMC cell types using scRNAseq, ANN, and incremental learning