V-SYNTHES2 Enables Physics-Based Virtual Screening of 36 Billion Compounds for Challenging Drug Targets
核心洞察
V-SYNTHES2 is a fully automated platform that extends physics-based virtual screening to the 36-billion-compound Enamine (搜索) REAL Space, achieving over 10,000-fold reduction in computational cost compared to brute-force docking.
The novel CapSelect algorithm automates evaluation of fragment "productivity" by analyzing binding pocket geometry, eliminating the need for expert target-specific manual tuning required in the original V-SYNTHES.
Benchmarking across diverse targets (AT2 receptor (搜索), cPLA2 (搜索) enzyme, rhodopsin) demonstrated enrichment factors up to 5,475.9 (EF100) for three-component reactions, with over 90% pose reproducibility for top-scoring molecules.
A research team has unveiled V-SYNTHES2, a next-generation computational platform that makes physics-based virtual screening of giga-scale chemical libraries feasible for the first time. The fully automated system can screen the entire 36-billion-compound Enamine (搜索) REAL Space—a collection of make-on-demand, synthetically accessible molecules—in approximately 48 hours using 320 CPU cores, a task that would otherwise require an estimated 130 years of wall-clock time and approximately $3.6 million in cloud computing costs using brute-force docking.
The platform, described in Nature Protocols, represents a major advance over the original V-SYNTHES approach by introducing CapSelect, an automated algorithm that evaluates whether docked molecular fragments have sufficient space within a protein binding pocket to "grow" into productive, high-scoring drug-like molecules during hierarchical enumeration.
CapSelect: Automating Fragment Productivity Assessment
A central limitation of the original V-SYNTHES was its reliance on user-defined criteria for assessing whether a docked fragment's binding pose was "productive"—that is, whether the capped attachment point oriented toward open space in the pocket, allowing for productive expansion during subsequent enumeration steps. This required expert knowledge of each target and precluded full automation.
CapSelect addresses this by placing a sphere at the labeled carbon of each capped R-group and iteratively generating non-overlapping spheres along a 120-degree cone, with each sphere maintaining a minimum 2 Å distance from pocket-lining atoms. The algorithm calculates a CapScore based on the number of spheres that can be fitted before steric crowding or exiting the pocket. This CapScore is then integrated with the docking score via a weighted logarithmic MergedScore function to rank fragments for enumeration.
"CapSelect replaces expert-defined heuristics with a geometry- and physics-based assessment of fragment growth potential," the authors note, emphasizing that the primary objective "is not to outperform expert intuition in every isolated case, but to remove the reliance on target-specific manual tuning."
Benchmarking Across Diverse Binding Pocket Geometries
The platform was benchmarked against three targets with distinct binding pocket architectures: the deep, narrow pocket of the angiotensin type 2 (AT2) receptor; the shallow active site of cytosolic phospholipase A2 (cPLA2 (搜索)); and the small, enclosed pocket of rhodopsin (Rho).
For two-component REAL Space reactions, V-SYNTHES2 achieved enrichment factors (EF10,000) of 106.9, 66.0, and 137.3 for AT2, cPLA2 (搜索), and Rho, respectively, relative to randomly selected, property-matched subsets. At more stringent thresholds, EF100 values reached 301.6 (AT2), 232.6 (cPLA2), and 539.9 (Rho). For three-component reactions, enrichment was substantially higher: EF10,000 values of 253.6 (AT2), 843.1 (cPLA2), and 670.7 (Rho), with EF100 values climbing to 367.0, 5,475.9, and 1,024.7, respectively.
Pose reproducibility—the consistency between initial MEL fragment docking poses and corresponding moieties in fully enumerated molecules—was assessed via RMSD analysis. For two-component libraries, all three targets exhibited median RMSD values below 0.5 Å across the top 10,000 ranked compounds. In three-component libraries, cPLA2 (搜索) demonstrated the lowest median RMSD (0.17 Å), while the AT2 receptor (搜索) showed a median of 0.68 Å. The Rho target, with its very small enclosed pocket, reached a median RMSD of approximately 4.65 Å, though reproducibility improved substantially when analysis was restricted to the top 1,000 highest-ranked compounds.
CapSelect significantly outperformed a "greedy" baseline that selected fragments solely on docking scores. For the AT2 receptor (搜索), CapSelect achieved a 1.5-fold reduction in both median and interquartile range of RMSD within the top 10,000 scored compounds in the two-component subset.
Prospective Validation and Therapeutic Applications
V-SYNTHES2 has been prospectively validated in two challenging drug discovery campaigns. For cPLA2 (搜索), a phospholipase enzyme implicated in neuroinflammation, 19 out of 117 synthesized and tested compounds showed significant inhibition (>40%) at 10 µM, representing a 16% hit rate. SAR-by-catalog and optimization of the two best hits yielded BRI-50460 (搜索) and BRI-50469 (搜索) with improved solubility, bioavailability, and brain permeability.
BRI-50460 (搜索) demonstrated 0.88 nM potency at cPLA2 (搜索) with selectivity against the iPLA2 homolog. In cultured astrocytes and neurons derived from human induced pluripotent stem cells, the compound mitigated the effects of amyloid beta 42 oligomers on cPLA2 activation, tau hyperphosphorylation, and synaptic and dendritic reduction. These findings position BRI-50460 as a potential lead candidate for Alzheimer's disease (搜索) and other neurodegenerative disorders.
For the AT2 receptor (搜索), a peptide-activated GPCR for which selective small-molecule agonists have been difficult to identify, screening identified two new agonists out of 40 tested compounds with significant activity at 1 µM concentration. SAR-by-catalog and optimization is ongoing, with potential applications in hypertension (搜索) and cardiovascular diseases (搜索).
Computational Efficiency and Scalability
V-SYNTHES2 operates through an iterative four-step cycle: generation of a minimal enumeration library (MEL) of approximately 1.7 million fragments for two-component reactions and 140,000 for three-component reactions; docking and CapSelect-based selection of productive fragments; synthon enumeration at remaining attachment points; and re-docking of enumerated subsets. The process requires docking approximately 3.8 million compounds—about 1.8 million MEL fragments plus roughly 2 million fully enumerated compounds—representing a greater than 10,000-fold reduction over docking the full 36-billion-compound space.
The platform has already been tested with the experimental xREAL Space containing 173 billion compounds, requiring only a minor increase in CPU cost. The authors note that further scaling beyond trillions of compounds may require additional accelerators, potentially based on deep learning approaches.
Retrospective comparison with the original V-SYNTHES on the CB2 receptor target showed that CapSelect-guided enumeration markedly tightened RMSD distributions relative to greedy selection, reducing median RMSD from 3.185 Å to 0.54 Å in the two-component set and from 1.35 Å to 0.67 Å in the three-component set, while achieving EF100 values of 165.5 and 847.4, respectively.
The complete V-SYNTHES2 pipeline, including CapSelect implementation, is publicly available at https://github.com/katritchlab/V-SYNTHES2.