The olfactory system employs an ensemble of odorant receptors (ORs) to sense molecules in the environment. Molecular features leading to specific olfactory percepts are not fully understood. We argued that olfactory receptors can be viewed as 3D filters that sense a set of molecular spatio-chemical features relevant to olfactory perception. As such, these filters can be represented by neural network used in artificial intelligence algorithms, such as convolutional neural networks. Using large-scale datasets of molecular structures and human olfactory percepts, we trained an artificial neural network called DeepNose to predict human perceptual qualities based on molecules’ 3D structures. We designed the structure of our network to represent the information flow in the human olfactory system. Using DeepNose, we can infer the molecular features leading to particular perceptual qualities. Thus, we find that unpleasantness can be traced to well-localized features of molecules. The spatio-chemical definition of pleasantness, on the other hand, is more diverse, suggesting a non-linear mapping between molecular properties and percepts along the valency axis. Overall, we find a diverse set of definitions for various semantic perceptual qualities and their molecular substrates ranging from well-defined chemical groups and specific spatial shapes to delocalized combinations of chemical features. Our framework helps identify the molecular underpinnings of human olfactory percepts and may facilitate the design of molecules with given qualities.