Chinese team cuts drug virtual screening from 24 hours to 61 seconds

Researchers led by Cao Zhichao, an assistant professor at Arizona State University’s School of Computing and Augmented Intelligence, introduced SweepANN, a GPU acceleration system that compresses the large-scale vector search used in early drug discovery from the roughly 24-hour timeframe reported for prior methods to about one minute per full pocket query round. In the main experiment, the system retrieved the top 500,000 candidate molecules for each of more than 26,000 protein pockets from a database of approximately 500 million molecule vectors in 61.04 seconds, with an average recall of 99.63 percent. The result shifts virtual screening from a batch job researchers must wait on toward an interactive tool for rapidly testing targets and pockets. The advance builds on DrugCLIP, developed by a Tsinghua University team and later applied to genome-scale screening in a 2026 Science study. DrugCLIP encodes protein pockets and small molecules into a shared high-dimensional vector space, allowing approximate nearest-neighbor retrieval instead of running molecular docking for every compound-pocket pair. Cao said the team’s system-oriented goal is to ask not only whether large-scale screening is possible, but how fast and cheaply it can run. When the group saw the earlier 24-hour benchmark, they judged it still too slow for iterative hypothesis-driven screening, especially as public compound libraries such as ZINC-22 expand to tens of billions of entries. SweepANN therefore targets the vector search bottleneck rather than the encoding model. To fit approximately 500 million 768-dimensional molecular vectors into memory, the team quantized FP32 embeddings to INT8 with group scaling, reducing storage from about 1.54 terabytes to roughly 387 gigabytes across four RTX PRO 6000 Blackwell GPUs. Instead of querying pockets one at a time, SweepANN processes 64 pockets per batch, reads each molecular vector shard once, and computes scores for the whole batch through matrix multiplications. For candidate list maintenance, it uses a window strategy that establishes temporary score thresholds from seed blocks, discarding low-scoring molecules without sorting the full score matrix, and employs CUDA Graphs to reduce launch overhead. The reported 61-second query time begins after compressed vectors are resident in GPU memory; quantization and loading take about 502 seconds but can be amortized across repeated searches as long as the molecular library does not need re-encoding. The team, including ASU computer science doctoral student Zhang Zhenyu and collaborators, says the high recall indicates that quantization mainly affects borderline candidates, not the strongest hits. Beyond drug screening, SweepANN’s approach to retrieving very large result sets could also benefit image and text vector search workloads.