Over the last three decades flow cytometry made enormous progress not only in the number of parameters measured simultaneously but also in the overall analytic workflow. Automated analyses emerge as discovery tools in many areas including hematologic diagnostics.

Duerze et al [1] proposed a fully presenting the automated analysis of standardized 8-color flow cytometry data of patients with myelodysplastic syndrome (MDS). MDS is a group of diseases caused by ineffective hematopoiesis and its diagnosis is often presented as one of the biggest challenges in hematology. Currently, the diagnostic algorithm is based on bone marrow morphology and cytogenetics. Inclusion of flow cytometry was suggested by introduction of several scoring systems evaluating different parameters [1]. Logically, we can expect that the higher the number of scoring parameters, the finer the scoring efficacy in such complex diseases as MDS. However, a higher number of parameters also increases time needed for analysis and quickly becomes too cumbersome for routine diagnostics.

Duetz et al. used standardized data collected in a single center in 2013–2017 and attempted to find the best markers resolving MDS and nonMDS disease categories. The presented study is not the first one which performed the automatic analysis of large flow cytometry datasets. The standardization approaches leading to standardized flow cytometry data were mainly introduced by international consortia (such as Euroflow consortium [3] or European Leukemia Network [4]) and they were further supported by cytometer intrinsic quality check modules which also allowed standardization across different machines using exchange of electronic settings or assays [5]. Alternative approach using data normalization tools can also be applied when being cautious of keeping a good quality of the initial data [6, 7]. Concurrently the advances in computational science offered new approaches in data visualization, automated sample classification, population detection etc. changing the whole field even more. Published data on automated analysis of such data collections showed ability to reduce analysis time and subjectivity [8-11].

Duetz et al. proposed a fully automated diagnostic pipeline by combining existing tools for data preprocessing, feature extraction and classification. Unbiased algorithmic feature extraction techniques not only improve analysis time and reduce subjectivity but also introduce descriptive power unprecedented in manual analysis. The manual analysis is always affected by the prior beliefs and/or knowledge of the investigators impeding them to find descriptors which they were not looking for, on the other hand automated tools treat the dataset as whole using only information explicitly provided for the task at hand. Such a goal oriented approach may result in a large set of features which are highly relevant for the diagnostic and research purposes but would never be considered in manual analysis. Retrospective automated analysis can therefore turn existing (and already thoroughly analyzed) datasets into valuable resources of new knowledge.

Using FlowSOM clustering and machine learning algorithms, Duetz et al. succeeded in differentiating MDS from nonMDS samples using a great variety of markers extracted from 8-color flow cytometric data. Notably, 37% of the features (630 in total) derived from FlowSOM in the single tube and 36% of these features (3780 in total) in the six-tube workflow differed significantly between MDS and nonMDS. Although the use of FlowSOM analysis together with machine learning itself is not a novel approach and was previously successfully applied by the same group as a diagnostic tool for common variable immune disease [12], this study presents yet another example how computational analysis can identify large number of differences when using only 8-color tubes between MDS of various subtypes and nonMDS bone marrow. Only 29% of parameters of the top 50 that significantly differed between MDS and nonMDS cases were previously included in currently used scoring systems for MDS diagnosis (iFS, FCSS, Ogata score, and Red score). Validation of these findings and search for their biological meaning should be the topic for next studies.

Surprisingly, 50% of parameters that differed between these two categories were represented by different scatter characteristics of cell populations, commonly of erythroid populations. Increased side-scatter (SSC) of erythroid population has not been previously included in published scoring systems and may be partially explained by the presence of ring sideroblasts, as the SSC correlated significantly with the percentage of ring sideroblasts. This is certainly a finding worth confirming independently. Nevertheless, we should not forget that scatters are more difficult to standardize than fluorescent parameters[13] and are influenced by adherence to standard operating procedure [10]. Type of erythrocyte lysis influences both: forward and side scatter parameters and abundance of nucleated erythroid cells. Whether we should focus on developing nonlysing methods (either with high flow rate cytometers or with new methods for separation of nucleated cells) remains a question. Similarly, the performance in multicentric settings, possibly using different types of cytometers, is to be validated and promises an exciting area for development. Of note, exclusion of scatter parameters decreased diagnostic accuracy of the current diagnostic workflow by 10%.

What should not be forgotten are the preanalytical factors. In MDS, especially in hypocellular forms, the dilution of the sample by peripheral blood is not rare. Of note, hypocellularity is also a typical feature of pediatric MDS [14, 15]. Performance of the algorithm that has been developed on bone marrow samples might become questionable once we consider dilution by peripheral blood. There are several methods published on how to evaluate the representativeness of the sample, but the consensus has not been reached so far [16]. On the other hand, it seems clear that peripheral blood is–compared to bone marrow-more standardized material, which might have been a reason why automated gating performed slightly better in peripheral samples than bone marrow samples in leukemia settings [9]. Performance of the presented pipeline on bone marrow samples showed the ability of the algorithm to recognize and describe a great spectrum of bone marrow populations. Whether finding such delicate differences between MDS and nonMDS peripheral blood samples is a way to overcome the preanalytical factor mentioned above, may be a question for next studies. Moreover, analysis of peripheral blood samples as a first line diagnostic method would enrich current diagnostic algorithms, which are based solely on bone marrow analysis.

In conclusion, the paper by Duetz et al. strongly encourages us that implementation of advanced bioinformatics is capable of enriching fully standardized flow cytometry measurements with no need for high-end instruments nor advanced multiparametric data, such as those derived from mass or spectral cytometers. Such data may help to accelerate the diagnostic process and potentially substitute the classical methods in case they are not feasible. Moreover, finding new features may help us to understand the biology of the disease. Finally, the current study might also raise hopes that we can still find more hidden treasures in old data.

MN was supported by Ministry of Health of the Czech Republic, grant nr. NU20J-07-00028. All rights reserved.

在过去的三十年中，流式细胞术不仅在同时测量的参数数量方面取得了巨大进步，而且在整个分析工作流程方面也取得了巨大进步。自动分析作为发现工具出现在许多领域，包括血液学诊断。

Duerze 等人 [ 1 ] 提出了对骨髓增生异常综合征 (MDS) 患者标准化 8 色流式细胞术数据的自动分析。MDS是一组由无效造血引起的疾病，其诊断通常被认为是血液学中最大的挑战之一。目前，诊断算法基于骨髓形态学和细胞遗传学。通过引入几种评估不同参数的评分系统，建议包括流式细胞术 [ 1]]。从逻辑上讲，我们可以预期评分参数越多，对MDS等复杂疾病的评分效果越好。然而，更多的参数也会增加分析所需的时间，并且很快就会变得过于繁琐，无法进行常规诊断。

杜兹等人。使用 2013-2017 年在单个中心收集的标准化数据，并试图找到解决 MDS 和非 MDS 疾病类别的最佳标志物。所提出的研究不是第一个对大型流式细胞术数据集进行自动分析的研究。导致标准化流式细胞术数据的标准化方法主要是由国际联盟（例如 Euroflow 联盟 [ 3 ] 或欧洲白血病网络 [ 4 ]）引入的，它们得到了细胞仪内在质量检查模块的进一步支持，这些模块还允许使用不同机器进行标准化。电子设置或检测的交换 [ 5]。当谨慎保持初始数据的良好质量时，也可以应用使用数据标准化工具的替代方法 [ 6, 7 ]。同时，计算科学的进步提供了数据可视化、自动样本分类、人口检测等方面的新方法，进一步改变了整个领域。已发布的关于此类数据收集自动分析的数据显示能够减少分析时间和主观性 [ 8-11 ]。

杜兹等人。通过结合现有的数据预处理、特征提取和分类工具，提出了一个全自动诊断管道。无偏算法特征提取技术不仅可以缩短分析时间并减少主观性，还可以引入手动分析中前所未有的描述能力。手动分析总是受到调查人员的先验信念和/或知识的影响，阻碍他们找到他们没有寻找的描述符，另一方面，自动化工具仅使用为手头任务明确提供的信息将数据集视为整体. 这种面向目标的方法可能会产生大量的特征，这些特征与诊断和研究目的高度相关，但在手动分析中永远不会被考虑。

Duetz 等人使用 FlowSOM 聚类和机器学习算法。使用从 8 色流式细胞术数据中提取的各种标记，成功地将 MDS 与非 MDS 样品区分开来。值得注意的是，37% 的特征（总共 630 个）源自单管中的 FlowSOM，而六管工作流程中这些特征中的 36%（总共 3780 个）在 MDS 和非 MDS 之间存在显着差异。尽管将 FlowSOM 分析与机器学习本身结合使用并不是一种新方法，并且先前已被同一小组成功用作常见可变免疫疾病的诊断工具 [ 12]]，这项研究展示了另一个例子，即在不同亚型的 MDS 和非 MDS 骨髓之间仅使用 8 色管时，计算分析如何识别大量差异。在 MDS 和非 MDS 病例之间存在显着差异的前 50 名参数中，只有 29% 的参数以前包含在当前使用的 MDS 诊断评分系统（iFS、FCSS、Ogata 评分和 Red 评分）中。验证这些发现并寻找它们的生物学意义应该是下一步研究的主题。

令人惊讶的是，这两个类别之间不同的参数中有 50% 由细胞群（通常是红细胞群）的不同散射特征表示。红细胞群的侧向散射 (SSC) 增加以前未包括在已发表的评分系统中，部分原因可能是环形铁粒幼细胞的存在，因为 SSC 与环形铁粒幼细胞的百分比显着相关。这当然是一个值得独立确认的发现。然而，我们不应忘记，散射比荧光参数更难标准化[ 13 ]，并且受遵守标准操作程序的影响[ 10]]。红细胞裂解的类型影响：前向和侧向散射参数以及有核红细胞的丰度。我们是否应该专注于开发非裂解方法（使用高流速细胞仪或分离有核细胞的新方法）仍然是一个问题。同样，多中心环境中的性能，可能使用不同类型的细胞仪，将得到验证，并有望成为一个令人兴奋的发展领域。值得注意的是，散射参数的排除将当前诊断工作流程的诊断准确性降低了 10%。

不应忘记的是分析前因素。在 MDS 中，尤其是在低细胞形式中，外周血稀释样品的情况并不少见。值得注意的是，细胞不足也是儿科 MDS 的典型特征 [ 14, 15 ]。一旦我们考虑到外周血稀释，在骨髓样本上开发的算法的性能可能会受到质疑。关于如何评估样本的代表性，已经发表了几种方法，但迄今为止尚未达成共识[ 16 ]。另一方面，很明显，与骨髓相比，外周血是更标准化的材料，这可能是为什么在白血病设置中，外周血样本的自动门控比骨髓样本的表现略好的一个原因。9 ]。所呈现的管道在骨髓样本上的性能表明该算法能够识别和描述大量骨髓群体。在 MDS 和非 MDS 外周血样本之间发现这种微妙的差异是否是克服上述分析前因素的一种方法，可能是下一步研究的一个问题。此外，外周血样本分析作为一线诊断方法将丰富目前仅基于骨髓分析的诊断算法。

总之，Duetz 等人的论文。强烈鼓励我们实施先进的生物信息学能够丰富完全标准化的流式细胞术测量，无需高端仪器或先进的多参数数据，例如来自质谱或光谱细胞仪的数据。这些数据可能有助于加速诊断过程，并有可能在经典方法不可行的情况下替代它们。此外，发现新特征可能有助于我们了解疾病的生物学。最后，目前的研究也可能带来希望，我们仍然可以在旧数据中找到更多隐藏的宝藏。

MN 得到了捷克共和国卫生部的支持，授予 nr。NU20J-07-00028。版权所有。