SPDRank

SPDRank applies a learning-to-rank algorithm to improve ligand-based virtual screening by ignoring meaningless ranking orders caused by experimental error in compound activity values.

Key Features:

• Learning-to-Rank Technique: Adapts learning-to-rank methods from information retrieval for ligand-based virtual screening.
• Stochastic Pairwise Descent RankSVM: Implements a stochastic pairwise-descent RankSVM to optimize ranking predictions on large compound datasets.
• Ignoring Meaningless Ranking Orders: Excludes pairwise rankings between compounds with similar activity levels (not statistically significant) and rankings among inactive compounds.
• Improved Virtual Screening Accuracy: Filters irrelevant orderings to reduce false positives and false negatives in rank-based hit identification.

Scientific Applications:

• PubChem BioAssay HTS Evaluation: Validated using five high-throughput screening (HTS) assay datasets from the PubChem BioAssay database.
• Ligand-Based Virtual Screening: Prioritizes meaningful compound orderings to improve hit identification in drug discovery workflows.

Methodology:

Employs a learning-to-rank framework with a stochastic pairwise-descent RankSVM that optimizes pairwise order relations while excluding comparisons between similar-activity compounds and among inactive compounds; evaluated on five PubChem BioAssay HTS datasets.