Pathogens, such as HIV and influenza, evolve in response to the selective pressures of their host environments accumulating changes in their genomes that offer fitness benefits. This selective pressure is characterised by three properties: (1.) it is episodic, tracking changes in the adaptive immune response and drug therapy, (2.) it is directional in that only particular amino acid substitutions are favoured and (3.) it varies between genomic loci. Most previous models have ignored or inadequately addressed some of these phenomena. This work extends recent approaches to modelling episodic directional selection acting on protein-coding sequences. We use inference techniques within the topic model framework to identify loci evolving under natural selection. A notable example of such techniques are the variational Bayesian methods. We show that our approach performs well in terms of specificity and power, and demonstrate its utility by applying it to some real datasets of HIV sequences.