Centre for Biological Engineering, Loughborough University, Loughborough LE11 3TU, Leicestershire, UK.
National Measurement Laboratory, LGC, Queens Road, Teddington TW11 0LY, Middlesex, UK.
Int J Mol Sci. 2022 Mar 17;23(6):3224. doi: 10.3390/ijms23063224.
Flow cytometry is widely used within the manufacturing of cell and gene therapies to measure and characterise cells. Conventional manual data analysis relies heavily on operator judgement, presenting a major source of variation that can adversely impact the quality and predictive potential of therapies given to patients. Computational tools have the capacity to minimise operator variation and bias in flow cytometry data analysis; however, in many cases, confidence in these technologies has yet to be fully established mirrored by aspects of regulatory concern. Here, we employed synthetic flow cytometry datasets containing controlled population characteristics of separation, and normal/skew distributions to investigate the accuracy and reproducibility of six cell population identification tools, each of which implement different unsupervised clustering algorithms: Flock2, flowMeans, FlowSOM, PhenoGraph, SPADE3 and SWIFT (density-based, -means, self-organising map, -nearest neighbour, deterministic -means, and model-based clustering, respectively). We found that outputs from software analysing the same reference synthetic dataset vary considerably and accuracy deteriorates as the cluster separation index falls below zero. Consequently, as clusters begin to merge, the flowMeans and Flock2 software platforms struggle to identify target clusters more than other platforms. Moreover, the presence of skewed cell populations resulted in poor performance from SWIFT, though FlowSOM, PhenoGraph and SPADE3 were relatively unaffected in comparison. These findings illustrate how novel flow cytometry synthetic datasets can be utilised to validate a range of automated cell identification methods, leading to enhanced confidence in the data quality of automated cell characterisations and enumerations.
流式细胞术在细胞和基因治疗的制造中被广泛应用，用于测量和表征细胞。传统的手动数据分析严重依赖于操作人员的判断，这是一个主要的变异来源，可能会对给予患者的治疗的质量和预测潜力产生不利影响。计算工具具有最小化流式细胞术数据分析中操作人员变异和偏差的能力；然而，在许多情况下，这些技术的可信度尚未得到充分确立，这反映了监管方面的一些关注。在这里，我们使用了含有受控分离种群特征和正常/偏斜分布的合成流式细胞术数据集，来研究六种细胞群体识别工具的准确性和重现性，每种工具都实现了不同的无监督聚类算法：Flock2、flowMeans、FlowSOM、PhenoGraph、SPADE3 和 SWIFT（基于密度的、-means、自组织映射、-最近邻、确定性- means 和基于模型的聚类，分别）。我们发现，分析相同参考合成数据集的软件的输出差异很大，并且当聚类分离指数低于零时，准确性会下降。因此，随着聚类开始合并，flowMeans 和 Flock2 软件平台在识别目标聚类方面比其他平台更困难。此外，存在偏斜的细胞群体导致 SWIFT 的性能较差，而 FlowSOM、PhenoGraph 和 SPADE3 相对受影响较小。这些发现说明了如何利用新的流式细胞术合成数据集来验证一系列自动化细胞识别方法，从而增强对自动化细胞特征和计数数据质量的信心。
