Combination therapies can often outperform single drugs in diseases like HIV and tuberculosis and can be commercialized faster than de-novo molecules. However, it would be prohibitively expensive to explore all possible combinations of approved drugs with high throughput screening and in-silico approaches typically require massive amounts of data. In this paper, MIT CSAIL researchers present a two-part graph-based machine learning model for finding possible combination therapies without very large datasets. The first part determines what biological target a compound might inhibit using data on individual compounds. The second part models the association between the target and disease. With this approach, they reduce data dependencies and can also identify additional latent targets learned from a single compound and combination (which is critical since several disease relevant targets are still unknown).