Machine Learning Algorithms Accelerate Throughput of a Flow - Sequencing Cell Based Assay for an Acute Myeloid Leukemia (AML) Therapeutic Discovery Platform

Introduction

Optimal therapeutic cancer discovery pipelines rely on high throughput, reproducibility and minimal bias for candidate discovery and clinical trials. Acute Myeloid Leukemia (AML) cell-based assays, combining flow cytometry (FCM) immunophenotyping with genomic sequencing, provide comprehensive characterization. However, sequential manual gating by expert flow analysts has been shown to have interlaboratory variation, weaknesses with higher dimensional feature spaces due to increased fluorescent channels, and limited scalability for throughput due to the necessary expert interpretation. Automated analysis of Flow Cytometry Standard (FCS) data using a combination of unsupervised and supervised algorithms that combine quality control, gating and clustering strategies offers a more scalable and sustainable option. Given these established breakthroughs, we investigated the use of these informatic tools, combined with sequencing data, to create a therapeutic discovery platform.

Methods

Peripheral blood and bone marrow samples with abnormal myeloid population and normal controls were exposed to multiple chemotherapeutic small molecules at varying concentrations. Samples were flow cytometrically evaluated with an 8-color tube of antibodies targeting various antigens allowing for apoptosis, viability and blast characterization using standardized lab protocols with a flow cytometer, followed by hematopathologist review. Furthermore, the samples underwent molecular sequencing and annotation with the MyAML® assay to interpret results in the context of AML molecular classification.

The computational pipeline utilized Bioconductor version 3.11 open-source packages for quality control, Logicle transformation, apoptosis and viability analyses utilizing a supervised gating strategy. Once viability was established, a Self-Organizing Map (unsupervised dimensionality reduction) algorithm was used to cluster cells into different cell populations. Moreover, a neural network utilizing TensorFlow combined with the API, Keras, was used to detect cell populations from these clusters. The neural network classifier was trained, validated and tested on these clusters compared to expert interpretation.

Results

Six bone marrow and three peripheral blood samples with the appropriate controls were evaluated via the discovery pipeline. Each sample was exposed to various chemotherapeutic molecules for a total of 203 samples for the pipeline with 50,000 cells per sample. Sequencing demonstrated these samples reflected various European LeukemiaNet (ELN) molecular risk categories.

The correlation between the total viable cells in 59 samples as determined by manual and automated gating processes was 98% (r ²=0.98) (Figure 1a). Heatmaps were generated for data visualization of therapeutic efficacy using this viable cell data as demonstrated in one example, Figure 1b, which was consistent with biological hypotheses.

Beyond total cell viability detection, cell populations were determined. The validation and test accuracy of our lymphocyte classification model were both 99%, and F1 statistic for model's testing performance was 0.98. Figure 2a displays the ROC curve. The classification of lymphocyte cell population using our combined unsupervised learning/neural network approach for 59 samples had a correlation of 97% (r ²=0.97) (Figure 2b).

Conclusions

Based on a preliminary comparison with manual gating by an expert flow analyst, we show an automated scalable protocol gives comparable results; therefore, it would be an appropriate alternative to use to screen multiple cells for downstream therapeutic analysis with increased throughput for both cell viability and cell population identification.

Figure 1

View largeDownload slide

Figure 1

View largeDownload slide

Close modal