Daunting estimates of overdose fatalities, costs of healthcare, lost productivity, and criminal justice involvement for the misuse of prescription opioids have converted opioid use disorders into a major national crisis in the United States. Among likely efficacious pharmacological interventions for the treatment of opioid dependence are those that can attenuate brain reward deficits experienced during periods of abstinence. Pharmacological blockade of κ-opioid receptors (KOR) has been shown to abolish brain reward deficits detected in rodents during withdrawal from opioids, as well as to reduce the escalation of opioid use in rats with extended access to opioids. Although KOR antagonists represent promising candidates for the treatment of opioid addiction, very few potent selective KOR antagonists are known to date and most of them exhibit significant safety concerns. Here, we used a deep learning framework for the de novo design of ligands with predefined molecular properties to predict new chemotypes with KOR antagonistic activity. Specifically, we pre-trained a deep generative tensorial model to learn a mapping of the chemical space from a ZINC subset of purchasable drug-like molecules, as well as sets of effective (IC50 < 1 µM) or less effective (IC50 ≥ 1 µM) known KOR inhibitors from the ChEMBL database, while simultaneously encoding the relationship between the molecules and their properties. Subsequently, we biased the resulting generative model towards the development of putative KOR antagonists with a reinforcement learning algorithm that rewarded similarity to the antagonist binding mode revealed by the JDTic-bound KOR crystal structure, as well as novelty, synthetic accessibility, absence of unstable and reactive moieties, drug-like solubility, and blood-brain barrier permeability. The generated molecules were prioritized for chemical synthesis and functional evaluation based on their predicted optimal interactions with the receptor.