Investigating Unsupervised Learning Techniques on Cell-Image Data

Large-scale image datasets have become increasingly valuable in biomedical research, particularly for identifying patterns that are difficult to detect with traditional methods. This thesis investigates the application of dimensionality reduction techniques to high-dimensional cell image data, aiming to uncover hidden patterns relevant to drug discovery. By extracting features from cell images and applying unsupervised learning techniques like principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), we identify distinct clusters associated with different mechanisms of action (MoA) of various drug compounds. This clustering suggests that dimensionality reduction can effectively capture complex phenotypic relationships in cellular perturbations, potentially linking specific MOAs to observable changes in cellular data. Such clustering is valuable in drug discovery, as it can enable researchers to predict the effects of new compounds based on the similarity to known MOAs, accelerating the identification of promising drug candidates. This research demonstrates that nonlinear dimensionality reduction methods tend to be superior to linear approaches in capturing complex phenotypic relationships for mechanism of action clustering. However, the findings also highlight the critical need to manage batch effects, as they can obscure true biological signals and lead to misleading conclusions. Addressing batch effects rigorously will enhance the accuracy of phenotype-based drug discovery, allowing dimensionality reduction techniques to more reliably identify biologically meaningful patterns and accelerate therapeutic innovation.