Label-free melanoma phenotype classification using AI-based morphological profiling

Melanomas are the deadliest skin cancers, in part due to cellular plasticity and heterogeneity. Within tumors, cells coexist in different mutable phenotypes that exhibit differential functional properties and drug responses. The definition of these phenotypic states has been challenging to rigorously define with conventional marker-based methods, and more high-parameter molecular methods are cell-destructive, labor-intensive, and can take days to weeks to obtain a readout. To overcome these technical and practical limitations, we utilized the Deepcell platform to perform real-time classification of unlabeled melanoma cells into Melanocytic and Mesenchymal phenotypes. We used 19 patient-derived cell lines with known Melanocytic or Mesenchymal transcription scores to develop the 'Melanoma Phenotype Classifier' to phenotype melanoma cells based on morphology alone. A Classifier accuracy of >88% was achieved, and morphology analysis of the images revealed distinct morphotypes for each phenotype, highlighting distinct morphological differences. To further link phenotypic state with multi-dimensional morphological profiles, we performed genetic and chemical perturbations known to shift the phenotypic state. The AI Classifier successfully predicted shifts in phenotype driven by the perturbations. These results further demonstrate how phenotype is linked to distinct morphological changes that are detectable by AI. Lastly, we applied the Melanoma Phenotype Classifier to dissociated biopsy samples, which revealed phenotypic heterogeneity that was confirmed by single cell RNASeq. This work establishes a link between morphology and Melanoma phenotype, and lays the groundwork for the use of morphology as a label-free method of phenotyping viable melanoma cells combined with additional analyses. Overall design: Cell line samples and dissociated biopsy samples were processed with the Deepcell device and analyzed using scRNA-sequencing

黑色素瘤（Melanomas）是致死性最高的皮肤癌，其诱因之一在于细胞可塑性与肿瘤异质性。肿瘤内部的细胞以多种可变表型共存，不同表型具备差异化的功能特性与药物响应能力。传统基于标记物的方法难以精准界定这些表型状态，而高参数分子检测手段则存在细胞破坏性强、操作耗力耗时，且结果获取需耗时数日至数周的局限。
为克服上述技术与实践层面的瓶颈，我们依托Deepcell平台（Deepcell platform）实现了未标记黑色素瘤细胞的实时分类，将其划分为黑素细胞型（Melanocytic）与间质型（Mesenchymal）两种表型。我们采用19株携带已知黑素细胞型或间质型转录评分的患者来源细胞系，开发了“黑色素瘤表型分类器（Melanoma Phenotype Classifier）”，仅通过细胞形态学特征即可完成黑色素瘤细胞的表型分型。该分类器准确率超过88%，图像形态学分析显示两种表型各自具有独特的形态特征，凸显了显著的形态学差异。
为进一步关联表型状态与多维形态学特征谱，我们实施了已知可诱导表型转变的遗传与化学扰动实验。该AI分类器成功预测了扰动所驱动的表型转变。上述结果进一步证实，表型与AI可检测到的特异性形态学变化存在明确关联。
最后，我们将“黑色素瘤表型分类器（Melanoma Phenotype Classifier）”应用于解离的活检样本，揭示了样本中的表型异质性，该结果经单细胞RNA测序（single cell RNASeq）验证。本研究确立了形态学与黑色素瘤表型之间的关联，为将形态学作为无标记方法用于活体细胞黑色素瘤表型分型，并结合多维度后续分析奠定了基础。
整体实验设计：细胞系样本与解离的活检样本均通过Deepcell设备进行处理，并采用单细胞RNA测序开展分析。