Drug discovery often requires screening huge chemical libraries against protein targets. In the DrugCLIP framework, small molecules and protein pockets are embedded into a shared vector space, so finding candidate drugs becomes a vector retrieval problem. However, typical nearest-neighbor search is built for returning a handful of top results, whereas genome-wide screening needs many matches per query.
The new paper introduces SweepANN, a GPU-accelerated approximate vector search algorithm built for this large-K regime. In a typical campaign, over 20,000 protein pockets are fixed in advance, and the system searches a library of over 500 million molecules. The authors propose using GPU parallelism and approximate methods to make these massive searches practical.
Because this is approximate search, it sacrifices some accuracy for speed, a common trade-off in large-scale retrieval. The abstract emphasizes the problem setup and the proposed solution but does not give specific speed-up or precision numbers. Still, the work points toward making genome-wide drug screening computationally feasible.