Deep learning applied to histological image analysis is revolutionizing biomedical research and represents a crucial advancement in diagnosing and studying diseases, such as cancer. One of the challenges in this field is obtaining the detailed local labels required for traditional machine learning models, as this demands significant effort from expert pathologists. Weakly supervised learning, which is based on multiple instance learning (MIL), is employed when only slide-level labels are available. These models must automatically detect regions of interest within the image and focus on them for classification. Traditional MIL methods overlook spatial relationships between different parts of the image, limiting the capture of continuous tumor patterns and potentially emphasizing intratumoral heterogeneity. In this work, we evaluate the impact of integrating spatial information into MIL models. We propose strategies such as graph-based networks, various forms of positional embedding, and an approach that models patches as charges in a potential field while calculating their interaction energy. The results indicate that, in certain cases, incorporating spatial information improves classification, but the associated computational complexity must be taken into account. Through statistical analysis, we assess the relevance of these techniques, identifying the conditions under which their contribution is significant and evaluating the associated computational costs.