Reverse target fishing is the task of identifying which proteins a given molecule may act on. It is central to drug discovery and repurposing, as it helps researchers understand a compound's potential effects and find new uses for existing drugs.
The paper argues that discriminative approaches, which are often adapted from virtual screening, struggle with sparse and long-tailed associations between molecules and targets. These limitations can reduce their effectiveness in real-world settings where some targets have very few known ligands.
To address this, the authors propose a generative retrieval approach for reverse target fishing. They also introduce a proteome-scale benchmark to evaluate methods on a larger and more realistic set of protein targets. The preprint does not include full results in the abstract, so the performance of the new method remains to be assessed.