Deep learning has demonstrated significant potential in advancing state of the art in many problem domains, especially those benefiting from automated feature extraction. Yet, the methodology has seen limited adoption in the field of ligand-based virtual screening (LBVS) as traditional approaches typically require large, target-specific training sets, which limits their value in most prospective applications. Here, we report the development of a neural network architecture and a learning framework designed to yield a generally applicable tool for LBVS. Our approach uses the molecular graph as input and involves learning a representation that places compounds of similar biological profiles in close proximity within a hyperdimensional feature space; this is achieved by simultaneously leveraging historical screening data against a multitude of targets during training. Cosine distance between molecules in this space becomes a general similarity metric and can readily be used to rank order database compounds in LBVS workflows. We demonstrate the resulting model generalizes exceptionally well to compounds and targets not used in its training. In three commonly employed LBVS benchmarks, our method outperforms popular fingerprinting algorithms without the need for any target-specific training. Moreover, we show the learned representation yields superior performance in scaffold hopping tasks and is largely orthogonal to existing fingerprints. Summarily, we have developed and validated a framework for learning a molecular representation that is applicable to LBVS in a target-agnostic fashion, with as few as one query compound. Our approach can also enable organizations to generate additional value from large screening data repositories, and to this end we are making its implementation freely available at https://github.com/totient-bio/gatnn-vs.