Cancer is the second leading cause of mortality worldwide, accounting for over 10 million deaths each year. During a healthy immune response, antigens bind to T cell receptors (TCR) on adjacent lymphocytes and trigger TCR oligomerisation. It is well-documented that clustering of TCR amplifies the signal from antigen-presenting cells and serves as a fundamental precursor of the immune response. Therapeutic mediators can directly elicit aggregation of T cell receptors, inducing immunological activation in the presence of cancer cells, to target and destroy tumours. Such therapeutic mediators could target many cancer antigens, but this process would differ on a cell-to-cell and antigen-to-antigen basis. Immunotherapeutic treatment strategies depend heavily on nanoscale receptor organisation in pre- and post-activated T cells and neoantigen prevalence in cancerous targets. In vitro testing is expensive, time-consuming, and limited to mediator availability – the only viable solution is in silico therapeutic design. Protein aggregation dynamics (PAD) simulators show promise in simulating transmembrane receptor motility. By constructing simulations, learned from protein localisation data, it is possible to track receptor dynamics, determine requirements for clustering, and predict signal digitisation. This data can be acquired through single molecule localisation microscopy (SMLM). However, accurate reconstruction requires tens of thousands of image frames, which are expensive to analyse using traditional fitting. For SMLM, deep learning (DL) has proven useful for extracting emitter coordinates and additional properties under conditions too complex and computationally-expensive for traditional fitters. However, most DL frameworks for SMLM have only been tested for engineered PSFs or on simulated data and also require paired training data for supervised learning, which restricts model performance to be at most as good as the ground truth (usually a traditional image reconstruction from acquired localisations). As such, the goal of this project is twofold: first, we will develop in silico methods of optimising treatment parameters for eliciting T cell activation. This will yield an open-source framework for modelling arbitrary transmembrane protein dynamics, with an emphasis on targeting T cell receptors and neoantigens with therapeutic mediators. Second, we will develop a deep learning-based localisation method which can train entirely on experimental data in a fully-unsupervised fashion. This model recapitulates the biophysical process of stochastic blinking observed during SMLM acquisition to reconstruct diffraction-limited images. By directly comparing to true experimental data, this allows for learning higher-resolution PSFs, estimating localisation coordinates and providing more accurate PSF parameterisation. Fundamentally, this framework can be used to optimise DNA origami structures, as nanotherapeutics for immunotherapy, in treating immunogenic cancers.