Virtual SYNthon Hierarchical Enumeration Screening (V-SYNTHES) addresses the computational challenge of screening enormous chemical spaces by first docking a Minimal Enumeration Library (MEL) of chemical fragments against a target pocket. Only the top-scoring fragments are then expanded into full ligands for large-scale docking, making gigascale screening tractable. However, under a fixed docking budget many fragments produce comparably good docking scores; only a small subset can be expanded, leaving many similarly promising fragments unexplored. This budget constraint limits the breadth of chemical space that can be effectively sampled within practical computational resources.
PharmacoNet is a deep-learning–based pharmacophore prescreening tool that predicts interaction hotspots from a protein structure and ranks candidate molecules via graph matching against a fixed set of interaction-type weights. Because PharmacoNet uses a general-purpose, target-agnostic set of weights, it does not explicitly account for the unique pocket-specific interaction preferences that emerge for different protein targets. Applying such a generic prescreener to redistribute docking budget across larger pools of fragments may therefore miss target-specific signals present in fragment-docking results.
The authors introduce MEL-Steered PharmacoNet, a parameter-efficient framework to specialize PharmacoNet to a particular target by recovering the signal already present in the initial fragment-docking step of V-SYNTHES. The adaptation uses two composable mechanisms:
Empirical density-map steering of predicted pharmacophore hotspots: fragment-docking of the MEL reveals spatial distributions of favorable interactions in the pocket. These empirical density maps are used to steer PharmacoNet's predicted pharmacophore hotspot locations to better reflect the pocket's observed preferences.
Empirical fine-tuning of interaction-type scoring weights: among fragments that dock well, the types of interactions they form (for example, hydrogen bonds, hydrophobic contacts, ionic interactions) indicate which interaction classes the pocket favors. MEL-Steered PharmacoNet adjusts PharmacoNet's interaction-type scoring weights empirically for the target, rather than relying on the original fixed, general-purpose weights.
Both mechanisms are designed to be parameter-efficient: they adapt PharmacoNet using signals derived from fragment-docking results without requiring full model retraining or additional experimental data.
MEL-Steered PharmacoNet was evaluated on three structurally distinct G protein–coupled receptor (GPCR) targets: CB2, GPR91, and 5-HT2AR. Performance was assessed using enrichment factor at 100 (EF100) relative to both a random baseline and the original, target-agnostic PharmacoNet prescreener. Across these targets, MEL-Steered PharmacoNet produced substantial EF100 gains compared with a random baseline and improved EF100 over PharmacoNet by reported factors of 8.94x for CB2, 6.87x for GPR91, and 1.69x for 5-HT2AR. These results indicate that incorporating MEL fragment-docking signals into the prescreening stage can materially increase the enrichment of true positives among top-ranked candidates.
The empirical fitting of per-target interaction weights yielded interaction profiles that are chemically interpretable and distinct across the tested GPCRs. These target-specific weight fits capture pocket-level preferences that PharmacoNet's generic fixed weights fail to represent. Because the adaptation mechanisms are built from fragment-docking observations, the resulting interaction-weight profiles can be examined to understand which interaction types and spatial hotspots a particular pocket prefers, potentially informing downstream medicinal-chemistry reasoning.
MEL-Steered PharmacoNet demonstrates that the fragment-docking data generated by the standard V-SYNTHES pipeline can be reused to adapt a general-purpose pharmacophore prescreener to individual targets. The approach retains PharmacoNet's ultra-fast screening capability while substantially improving target-specific performance, and it requires no additional experimental data or full neural-network retraining. By reallocating the same docking budget more effectively across a larger pool of fragment-derived full ligands, MEL-Steered PharmacoNet can expand the practical coverage of chemical space in hierarchical synthon-based virtual screening.
Limitations and reporting notes
The source document is a preprint and has not been certified by peer review. Details beyond those reported in the abstract—such as precise dataset sizes, training or adaptation hyperparameters, and full numerical results beyond the reported EF100 fold-improvements—were not provided in the supplied source text.