TopU-LBVS: A Realistic Multi Target Benchmark for Ligand Based Virtual Screening

TL;DR - TopU-LBVS is a 93-target benchmark for ligand-based virtual screening that replaces easy random negatives with property-matched, structurally similar decoys. It aims to provide a more realistic and reproducible assessment of molecular models for early-stage drug discovery.

• Built from curated ChEMBL 35 data, covering seven protein classes with a fixed 1:40 active-to-decoy ratio.
• Includes full generalization, low-data learning, and compact seven-target protocols with fixed splits and standardized evaluation.
• Tests ten baselines spanning fingerprints, molecular GNNs, hybrid approaches, and modern molecular models.
• Baseline performance drops sharply when moving from random decoys to hard negatives, highlighting overestimation by conventional evaluations.