The success of machine learning depends heavily on the choice of features on which the algorithms are applied. For that reason, much of the efforts go into engineering of informative features. In this talk, I describe our efforts in learning deep representations that are actionable and allow endpoint users to ask what-if questions and receive robust predictions that can be interpreted meaningfully. These methods specify deep graph neural functions that map entities from a rich, interconnected dataset to points in a compact vector space, termed embeddings. Importantly, these graph neural methods are optimized to embed entities such that performing algebraic operations in the embedding space reflects the structure of the data. I will describe how these methods enabled repurposing of drugs for an emerging disease where our predictions were experimentally verified in human cells (Gysi et al., 2021). The methods also enabled discovering dozens of drug combinations safe for patients with considerably fewer unwanted side effects than today's treatments. The graph neural network approach can successfully prioritize ultra high-order combinations of drugs despite extreme scarcity of labeled data instances (Huang et al., 2020). Last, I will highlight Therapeutics Data Commons (https://tdcommons.ai), a platform with AI/ML-ready datasets and tasks for therapeutics together with an ecosystem of tools, libraries, leaderboards, and community resources.