missing label problem

missing label problem addresses systematic missing-label and mislabel issues in ligand-based virtual screening (LBVS) by detecting flawed dataset assumptions, estimating unobserved bioactivity labels, distinguishing multiple protein binding sites, and mitigating evaluation bias to improve ligand-protein interaction prediction for drug development.

Key Features:

• Interpretation of Absence as Negative: Detects instances where bioactivity database entries that are untested or unvalidated are treated as true negatives in LBVS training data.
• Single Label for Multiple Binding Sites: Identifies cases where proteins with multiple distinct binding sites are represented under a single label, causing inaccurate ligand grouping.
• Statistical Techniques for Missing Labels: Employs statistical methods to estimate missing bioactivity labels and to identify different binding sites, improving the ranking of potential ligands.
• Bias Removal via Data Blocking: Applies a data blocking approach inspired by experimental design theory to mitigate bias during LBVS model evaluation.

Scientific Applications:

• Improved LBVS accuracy: Enhances the accuracy and scalability of ligand-based virtual screening by providing more reliable training labels for LBVS models.
• Better ligand-protein interaction prediction: Enables development of more precise predictive models for ligand-protein interactions relevant to drug development.
• Objective model evaluation: Provides a more objective assessment of LBVS performance by mitigating evaluation bias through data blocking.

Methodology:

Uses statistical methods to estimate missing bioactivity labels and to identify distinct binding sites, and applies data blocking from experimental design theory to mitigate bias during model evaluation.