The identification of protein-protein interaction (PPI) is one of the most important tasks to understand the biological functions and disease mechanisms. Although numerous databases of biological interactions have been published in debt to advanced high-throughput technology, the study of inter-species proteinprotein interactions, especially between human and bacterium pathogens, remains an active yet challenging topic to harness computational models tackling the complex analysis and prediction tasks. In this paper, we comprehensively revisit the prediction task of human-bacterium protein-protein interactions (HB-PPI), which is a first ever endeavour to report an empirical evaluation in learning and predicting HB-PPI based on machine learning models. Firstly, we summarise the literature review of human-bacterium interaction (HBI) study, meanwhile a vast number of databases published in the last decades are carefully examined. Secondly, a broader and deeper experimental framework is designed for HB-PPI prediction task, which explores a variety of feature representation algorithms and different computational models to learn from the curated HB-PPI dataset and perform predictions. Furthermore, a bidirectional LSTM-based model is proposed for the prediction task, which demonstrates a more effective performance in comparison with the others. Finally, opportunities for improving the performance and robustness of machine learning models for HP-PPI prediction are also discussed, laying a foundation for future work.