Manual gating is the traditional procedure adopted to identify cellular clusters from multi-dimensional datasets generated with flow cytometry, a tool for detecting and monitoring diferent diseases by acquiring single cell features. However, the identification of cellular subpopulations by manual gating is a time-consuming process strongly afected by human expertise. Automated analysis supported by computational systems, such as machine learning approaches, can radically change the way flow cytometry data are elaborated. In this paper we applied a suite of machine learning classifiers for analysing samples of peripheral blood acquired with flow cytometry. The goal was to identify CD4+ lymphocytes population. Four ML classifiers are examined -Support Vector Machine, Random Forest, Multilayer Perceptron and Logistic Regression using stratified 10-fold cross-validation. All the four models perform very well, with a balanced accuracy score > 0.945. We come to the conclusion that all four algorithms classify the events of interests with promising results, paving the way for further investigations.
Flow cytometry (FL) is an experimental technique that enables to measure cellular properties at
a single-cell resolution by quantifying, for instance, antigens expressed on the cell surface and
various physical properties [
The general computational cytometry workflows can be classified into two categories based
on the methods employed: discovery analysis, i.e., the detection of unknown, unique cell
populations, versus focused analysis i.e., the detection of known well-defined ones. Automation
can potentially lessen variability in the data analysis process in both situations. Cell populations
that are neglected in successive manual gating procedures, such as cells gated out in earlier steps,
can be found using automated technologies in discovery mode. When using focused analysis,
the cell populations of interest are precisely specified, and the data analysis procedure adheres
to a set of techniques that is likely to be validated and authorised. Automated technologies can
lessen human efort by automatically classifying cases as healthy or diseased and only raising
questions about certain cases for people to consider [
The computational approaches can be further categorised: automating the manual gating
process based on rules or cell densities (flowDensity [
In this paper, we applied a suite of machine learning classifiers for analyzing samples of peripheral blood acquired with flow cytometry. The goal was to identify the 4+ lymphocyte population. We show the efectiveness of our approach for classifying cells with a series of tests, cross-verifying the trained models on various data files, and comparing the cell classifications with those acquired by manual gating. Four ML classifiers are examined —Support Vector Machine, Random Forest, Multilayer Perceptron and Logistic Regression —using stratified 10fold cross-validation. All four models perform very well, with a balanced accuracy score of > 0.945.
Flow cytometry is a standard method for analysing and quantifying biological data. The
capabilities of cytometry have increased, giving rise to several data-analysis methodologies [
The flowDensity [
Several machine learning approaches have been devised to identify new or
predetermined cellular populations and perform automated analysis of flow cytometry data
[
The FlowSOM [
Similar to flowDensity , another piece of software called flowLearn uses density
characteristics but does not require the user to adjust hyper-parameters to achieve the best results
manually. Instead, it operated in a semi-supervised manner necessitating the establishment
of thresholds by a human expert for gating one or a few distinctive samples. Then, these
criteria are automatically applied to all data using a process known as derivative-based density
alignments. It predicted gates on additional samples using a limited number of manually gated
samples with density alignments. A drawback in this approach could be the gating of a limited
number of samples, and density-bound rules could over-fit results. A supervised learning-based
approach Automated Cell-type Discovery and Classification (ACDC) [
This section is divided into two subsections: (i) details about the data set and pre-processing of the data, and (ii) experiments with machine learning models.
The data exploited in the study have been derived by the routinarian diagnostic activity performed by the Center of Cytometry of Urbino University. Data set were randomly selected and anonymized in order to make impossible the identification of the source.
For every peripheral blood sample, data were acquired on intrinsic parameters and antigen expression displayed by white blood cells. The analyses were performed through a commercial lfow cytometer and focused on the following parameters: i) Forward Scatter (FSC), ii) Side Scatter (SSC), iii) CD3 antigen expression, iv) CD4 antigen expression, v) CD8 antigen expression, vi) CD16 and CD56 combined antigen expression, and vii) C45 antigen expression. In all, 15 subjects were analyzed. For each parameter, data related to the pulse area were considered.
The cytometric files produced by the analytical runs were then stripped of metadata and exported in csv format. Consequently, the dataset consists of fifteen (15) diferent data files, where each row is a cell, and each column contains the corresponding value of one of the 8 parameters as mentioned above (Table 1) with no missing values. Data records are labeled with binary values, i.e., gated and ungated (1 used for gated and 0 for ungated records). The extraction of gated and ungated records was possible due to the combined use of a commercial program to select the clusters of interest (Cytopaint1) and an unpublished program for the management of flow cytometry standard (FCS) files (Wizard) provided by one of us (MDA). In particular, in each experiment, CD4+ T cells (gated for CD4 expression) are identified by manual gating performed by experienced operators. The gating logic performed on each data ifle to filter out the gated records is as follows: 1. The creation of a first gate on parameters CD45+ vs SSC (denoting it as Gate-1) and selecting lymphocytes; 2. The results of Gate-1 are expressed for parameters CD3 and CD19 and a gate was traced on the events CD3+ CD19 and CD3 CD19+ (denoting it as Gate-2) 3. The results of Gate-2 are expressed for parameters CD4 and CD8, and a gate is traced on the events CD4+ CD8 (denoting it as Gate-3 which constitutes the population of T Helper 1http://leukobyte.com/cytopaint-classic/
Lymphocytes CD3+ CD4+ CD8-) and another gate on the events CD4 CD8+ (denoting it as Gate-4).
After the hierarchical gating process, a subset of records is obtained based on CD4+ cells which will be part of class 1. The resulting dataset is unbalanced between the two classes. The percentage of CD4+ manually gated samples is presented in Table 2 against the total sample size of each data file. Due to the critical nature of medical data sets, data balancing techniques are not often recommended. Thus, we applied the k-fold stratified cross-validation technique for training and testing the ML classifiers because it maintains the same class ratio across the K folds as the ratio in the original dataset.
To conclude, our approach avoids information loss in the cell gating stage by directly using the labelled flow cytometry data.
Our study compares the outcomes of allocating cell events to discrete cell populations (gated and ungated cells) using automated gates with the results from manual gates produced by expert analysis. In particular, the classification goal is the automatic identification of T Helper Lymphocytes CD4+, which constitute the gated population, against the ungated ones.
We adopted supervised ML models, and trained the algorithms with gates supplied by experts. A suite of ML classifiers is employed for classification. The reason is to use various types of classifiers to observe the accuracy and over-fitting issues under the diferent classification mechanisms, i.e., decision tree-based, gradient-based, neural network-based and able to classify non-linearly separable data. These classifiers, with their brief descriptions, are listed below (the mathematical equations for these classifiers are omitted as these are well-understood methods): 1. Random Forest (RF)—is a meta estimator that averages the results of many decision tree classifiers by fitting diferent sub-samples of the dataset to increase predicted accuracy and reduce over-fitting. 2. Logistic Regression (LR)—is a parametric regression technique that involves fitting a line (or a curve) to the data and then using the gradient descent function to distinguish between diferent output classes. In this study, we utilized LR with gradient descent, which fits the dataset with a curve.
3. Support Vector classifier (SVC)—similar to LR, it also fits a curve on the dataset; however,
the curve itself tends to maintain a maximum margin on both sides [
For the training and validation phase, we mainly designed two experiments (each maintains details about the sub-experiments) to examine the classification of gated and ungated samples. Exp. 1 The first experiment applies the ML classifiers listed above to each single data file, by extracting the train and test-set (80-20 split); Exp. 2 The second experiment is conducted by training the classifiers on 10 subjects randomly selected (accumulated training data) and testing the rest on the remaining 5 subjects. Figure 2 illustrates the general ML pipeline adopted for the flow cytometry data. The flow cytometry standard (FCS) data files are used for preprocessing. The processed FCS data then subjected to apply the hierarchical gating (the gating process is described in section 3.1). After the gating process, cell-annotation is performed and CD4+ T cells are separated from the rest of cell populations. The supervised machine learning pipeline is used to train and validate the models. A stratified 10-fold cross-validation (CV) technique is employed to validate the classification performance of trained models. The results for CD4+ T cells are compared with results of trained ML models to evaluate the performance of all models. The training data for each fold were obtained in equal amounts to equalize the class occurrence frequencies. These data were then used to train a model. The validation set is then used to validate the model.
The experiments are conducted using a Python Collaboratory environment. Experimental results are described in Section 4.
In this section, we present the results for the two types of experiments.
Exp. 1 Results of the first experiment are reported in Table 3 for only 3 data files, since all classifiers’ results showed similar performances on the diferent data files, which likely means that the datasets’ distributional variations weren’t significant. The average scores for both metrics are presented.
It can be observed from Table 3 that RF outperformed other classifiers for both = .985 and 1 = .998 metrics for data file-1. LR performed better for data file-2, and all the classifiers achieved similar results on data file-3 for BA and F1 (Balanced Accuracy and F1-Score). In general, all the four ML classifiers performed very well, with a balanced accuracy score of > 0.945.
Exp. 2 We examined the same four classifiers in the second experiment. The training data from 10 data files are added to a new file, making a more extensive training set. The obtained training set contains the same features and data distribution with corresponding labels as the source data and is split into 10-folds. The ML classifiers are trained and validated with stratified 10-fold cross-validation. Then, the resulting trained classifiers are subjected to evaluation on the testing sets of other 5 files. The performance of ML classifiers on each file for 4+ cell classification is presented in Table 4, in which average scores for both metrics are presented. It can be observed from Table 4 that SVC and RF maintained their performance the same as for the first experiment. The MLP and LR have shown a slight downfall in results. The least score of BA for MLP on file-12 and file-15 was recorded at .925 and .897, respectively. The least score of BA for LR on file-14 and file-15 was recorded at .952 and .968, respectively. Generally, the performance for all classifiers is promising, with a score of BA > .897.
The field of flow cytometry witnesses a significative progress in blood samples analysis that brought the acquisition of a huge amount of data. Given these premises, automatic data analysis, carried out using supervised learning techniques that automatically categorize samples according to clinical protocols, can provide enormous benefits. Such analysis is possible through automated methods without human subjectivity and gating bias.
We conducted two experiments to automate the manual gating procedure for classifying CD4+ T cells from flow cytometry data. Four ML supervised algorithms have been trained with samples manually gated by experts in the pre-processing phase. The current study demonstrates our method’s capacity to distinguish the T Helper Lymphocytes CD4+ among all the types of cells present in the dataset, with high precision in terms of balanced accuracy and f1-score. This result suggests that, with training data available as gated examples, supervised classification ofers an efective technique for automatic analysis of flow cytometry data, enabling to extract and compute the size of diferent cellular populations.
Despite its apparent simplicity, this approach is of particular importance, as it constitutes a replicable mechanism in a series of increasingly complex contexts, which can be exploited for the realization of algorithms aimed at the automatic diagnosis of haemato-oncological pathologies with characteristic phenotypes.
As the accuracy of all the models is very high, the future work is to explore the importance of the features and evaluate if some straightforward relationship between features and output is present.
This work is a part of a collaboration project to automate the gating process of clinical flow cytometry at University of Urbino.