Cancer Drug Screening Scale-up: Combining Biomimetic Microfluidic Platforms and Deep Learning Image Analysis

Abstract

The development of cancer drugs is usually costly and time-consuming, mainly due to growing complexity in screening large number of candidate compounds and high failure rates in translation from preclinical trials to clinical approval. Despite the great efforts, the preclinical screening platforms combing good clinical relevance and high throughput for large-scale drug testing is still lacking. In addition, accumulating evidence suggests that cancer drug response can be altered by tumor microenvironment (TME), which includes not only cancer cells but also physical, and biochemical cues in niches. To improve the current cancer drug screening assays, it is important to mimic local TME to achieve better physiological relevance. In the first part of this dissertation, three TME-mimicking microfluidic platforms were introduced for three different in-vitro TME-mimicking tumor sphere models: spheres in matrix, self-aggregated spheres, and single-cell clonal spheres. First, a 3D gel-island chip investigated the heterogeneity of single-cell drug responses in biomimetic extracellular matrix (ECM). With 1,500 isolated single cell chambers containing ECM, it was demonstrated that ECM support was favorable for some population of cancer cells to maintain stemness and develop drug resistance. This result suggested the importance of drug screening at single-cell resolution in TME-mimicking platforms. Secondly, a drug combination screening chip enabling high-throughput and scalable combinatorial drug screening was demonstrated for the aggregated sphere model. Instead of screening a single drug on each of the tumors, this chip allows the screening of all pairwise drug combinations from eight different cancer drugs, in total 172 different treatment conditions, and 1,032 tested samples in a single microfluidic chip. The presented design approach was easily scalable to incorporate arbitrary number of drugs for large-scale drug screening. Finally, single-cell Hi-Sphere chip enabled high-throughput clonal sphere culture and selective retrieval. Combining fluorescent dye on-situ staining techniques, we identified rare cancer stem-like cell population and confirms its location at the leading edge of spheres. Advance in experimental throughput generates massive data, which demands the corresponding automatic analysis and intelligent interpretation capabilities. The second part of this dissertation focuses on the applications of computer vision and machine learning algorithms to automated biomedical data processing. Image analysis with convolutional neural network was applied for drug efficacy evaluation in a fast and label-free manner. The estimated drug efficacy is highly correlated with the experimental ground truth (R-value > 0.93), while the predicted half-maximal inhibitory concentration is within 8% error range. In addition, metastatic fast-moving cells could be identified after extracting morphological features from the microscope images and applying deep learning algorithm for image analysis, achieving over 99% accuracy for cell movement direction prediction and 91% for speed prediction. In summary, this dissertation presents high-throughput TME-mimicking microfluidics and deep learning image analysis for large-scale drug screening solutions.