Interpreting killer mechanisms of gamma-delta T-cells with Machine Learning

Sammanfattning

Gamma-delta (γδ) T-cells represent a unique and promising subset of immune cells with potential applications in cancer immunotherapy. However, the fundamental mechanisms underlying their cellular behavior remain poorly understood. This thesis aims to bridge this gap by studying their killing behavior patterns through machine learning and explainable artificial intelligence techniques. By developing a supervised and unsupervised computational framework, this research aims to identify the key features that predict γδ T-cell fate at the earliest possible time point from a live-cell image sequence and assess their biological relevance to determine the trustworthiness of the model. The methodology integrates specialized image pre-processing to segment the nuclei of the live-cell imagery, classification techniques based on the pretrained models ResNet101 and VGG16, and interpretable machine learning models to uncover previously unknown phenotypic differences. The study employs the AI4CellFate method, which encompasses an adversarial autoencoder and latent space perturbations for visualization of biologically relevant features. Our research reveals that machine learning can effectively distinguish between surviving and dying cells, with classification accuracy significantly improving as the cellular killing process progresses, particularly using ResNet101. Crucially, our explainable AI approach unveils the biological significance of the model’s decision-making process. Two key features are found as primary drivers of cell fate prediction: distinctive morphological changes and fluorescence intensity variations. This work not only advances our understanding of γδ T-cell mechanisms but also validates the potential of machine learning in immunological research by demonstrating that predictive models can extract biologically meaningful insights.