机器学习算法优化下的高精度拉曼效应微流体系统用于脑胶质母细胞瘤的分离识别以及侵袭性的实验研究

The main objective of this research is to develop a high-throughput microfluidic biosensor that can measure multi-modal physical attributes of cells. The accurate diagnosis of glioblastoma requires low heterogeneity in biopsy samples, in order to study the metastasis of glioblastoma subtypes precisely. This novel platform combines dielectrophoresis (DEP) separation and isolation of glioblastoma cells, surface enhanced Raman spectroscopy (SERS) with biomechanical indentation stimuli on targeted cells by multi-constriction channel based microfluidics technology, where the data will be analyzed by kernel-based machine learning methods to achieve artificial intelligence. .The DEP separation and isolation of individual cell type can reach over 95% accuracy and purities, which provides a good source for further single cell analysis and experiments. The multi-constriction channels can create velocity variables and provide unique features about the deformability of cells. The progression of cancer is associated with an alteration of the cell structure which in turn causes transformed cells to be softer and hence more deformable than their healthier counterparts to facilitate metastasis. SERS has the unique potential to distinguish different subtypes of cancer cells using the vibrational properties of the specific chemical bonds existing in their membranes, which adds a biochemical dimension to biophysical characterization. Advanced machine learning statistical methods will be applied combining biomechanical and subcellular-level bio-optical properties to derive unique label-free biophysical signatures. By combining these four powerful techniques: DEP separation, biomechanical characteristics, bio-optical characteristics, and kernel-based machine learning methods, the system is able to leverage the advantages of deformability assays and Raman spectroscopy, allowing for enhanced resolution, which makes it possible for unique “biosignatures”to be obtained, as well as the drug treatment efficacy, which can have a direct impact on clinical tumor subtypes classification and treatment protocols planning.

脑胶质母细胞瘤及其亚型的精确识别和诊断可以为精准治疗支持。目前无论是分离胶质母细胞瘤与小胶质细胞，还是各亚型的区分识别都需要通过对混合样品的实验，对亚型识别的精准度和成本都不理想。此项目提出了一套多级微流体芯片系统，能利用介电泳技术分离脑胶质母细胞瘤与小胶质细胞，利用多级微流体形变通道和拉曼光谱扫描识别胶母细胞瘤亚型，最后结合机器学习算法分析得到的高维变量数据，达到快速、精确的数据分析结果。根据最新研究成果，介电泳可以分离不同介电特性的癌细胞，精确度可达到95%以上；多级微流体芯片和拉曼光谱技术可以识别90%以上的癌细胞亚型。不同类型的细胞骨架，细胞核，及细胞表面化学成分可以产生独特的拉曼光学特性。机器学习在细胞生物机械和光学特性数据分析方面，可以采集出细胞独特的“生物指纹”。这项技术和产品可以大幅提高胶质瘤临床诊断分型，并根据分型进行药物筛选，提高脑胶质瘤的治疗效果。

脑胶质母细胞瘤及其亚型的精确识别和诊断可以为精准治疗提供支持。对于分离胶质母细胞瘤与小胶质细胞，还是各亚型的区分识别都需要通过对混合样品的实验，对亚型识别的精准度和成本都不理想。此项目采用多级微流体芯片系统，能利用介电泳技术分离脑胶质母细胞瘤与小胶质细胞，利用多级微流体形变通道和拉曼光谱扫描识别胶母细胞瘤亚型，最后结合机器学习算法分析得到的高维变量数据，达到快速、精确的数据分析结果。根据成果，介电泳可以分离不同介电特性的癌细胞，精确度可达到95%以上；多级微流体芯片和拉曼光谱技术可以识别90%以上的癌细胞亚型。在实验过程中发现，采用渐变宽度为8-12微米的通道可以更好的用于脑胶质瘤的分类和识别，因此将这种新型多级微流芯片命名为循环渐变压缩通道，缩写为CCC芯片。实验不仅验证了经典脑胶质瘤细胞系U-87，U-251, 对比正常细胞系HA-1800，还通过临床获取了多例不同分级的患者脑胶质瘤样本进行实验，结果发现可以精确判别出WHO不同等级的胶质瘤，且精度达到单细胞级。不同类型的细胞骨架，细胞核，及细胞表面化学成分可以产生独特的拉曼光学特性。机器学习在细胞生物机械和光学特性数据分析方面，可以采集出细胞独特的“生物指纹”。实验结果表明拉曼基底不仅可用于精确识别细胞本身，还可以对低剂量抗癌药效果进行精准判定。机器学习方法的开发对不同种类肿瘤的效果也不同，传统的机器学习方法在脑胶质瘤数据的分析上具有较大局限性，此项目对比了NGK，Ridge，和Lasso算法，在kernel模型上并进一步改进，通过热图进行高维数据分析，采用ENet方法对多级位流体芯片结果进行分析，无论是对于胶质瘤细胞系还是患者样本，ENet方法都达到了90%以上的理想效果。这项技术和产品可以大幅提高胶质瘤临床诊断分型，并根据分型进行药物筛选，提高脑胶质瘤的治疗效果。