Preferred Networks (PFN), a Japanese technology company specializing in deep learning, has launched P-ULVS, a contract service for virtual screening designed to accelerate the hit-finding phase of drug discovery. While 10 million compound libraries were historically standard, the industry has shifted toward massive libraries like Enamine REAL, which contains 10 billion compounds. P-ULVS combines Active Learning with docking simulations to rapidly identify promising candidates from these vast chemical spaces.

The efficacy of large-scale library exploration is well-documented, with studies showing that scaling from 100 million to 10 billion compounds improves the discovery rate of highly active agonists by 2–5 times. Past benchmarks using 10 billion compound libraries yielded hit rates of 33% for CB2R and 28.5% for ROCK1. P-ULVS repeats a cycle of machine learning-based selection and docking calculations, executing massive simulations that would normally take centuries to complete in a fraction of the time.

In benchmarks targeting the 5-HT2AR serotonin receptor, the system evaluated 4 million compounds—just 0.04% of the total library—and successfully captured approximately 80% of the top compounds by docking score. Computation time was reduced to approximately 2,000 hours, which is 1/1,500th the time required for a full-scale exhaustive search using one NVIDIA V100 GPU. Similar performance was confirmed for other target proteins including CB2R, JAK3, and ROCK1, consistently achieving recall rates between 72% and 81%.

The service supports flexible selection strategies, such as 'ML Best (Greedy)' for prioritizing high-potential compounds or 'Uncertainty Sampling' to explore unknown chemical spaces. By offering a more cost-effective and faster alternative to traditional high-throughput screening (HTS), P-ULVS assists in the discovery of novel chemical scaffolds. PFN continues to refine this technology to improve overall productivity in the drug discovery pipeline.

Preferred Networks (PFN), a Japanese technology company specializing in deep learning, has launched P-ULVS, a contract service for virtual screening designed to accelerate the hit-finding phase of drug discovery. While 10 million compound libraries were historically standard, the industry has shifted toward massive libraries like Enamine REAL, which contains 10 billion compounds. P-ULVS combines Active Learning with docking simulations to rapidly identify promising candidates from these vast chemical spaces.

The efficacy of large-scale library exploration is well-documented, with studies showing that scaling from 100 million to 10 billion compounds improves the discovery rate of highly active agonists by 2–5 times. Past benchmarks using 10 billion compound libraries yielded hit rates of 33% for CB2R and 28.5% for ROCK1. P-ULVS repeats a cycle of machine learning-based selection and docking calculations, executing massive simulations that would normally take centuries to complete in a fraction of the time.

In benchmarks targeting the 5-HT2AR serotonin receptor, the system evaluated 4 million compounds—just 0.04% of the total library—and successfully captured approximately 80% of the top compounds by docking score. Computation time was reduced to approximately 2,000 hours, which is 1/1,500th the time required for a full-scale exhaustive search using one NVIDIA V100 GPU. Similar performance was confirmed for other target proteins including CB2R, JAK3, and ROCK1, consistently achieving recall rates between 72% and 81%.

The service supports flexible selection strategies, such as 'ML Best (Greedy)' for prioritizing high-potential compounds or 'Uncertainty Sampling' to explore unknown chemical spaces. By offering a more cost-effective and faster alternative to traditional high-throughput screening (HTS), P-ULVS assists in the discovery of novel chemical scaffolds. PFN continues to refine this technology to improve overall productivity in the drug discovery pipeline.