spark-cpvs

spark-cpvs accelerates structure-based virtual screening by combining iterative docking, Support Vector Machine models with conformal prediction, and Apache Spark parallelization to prioritize high-scoring ligands from large chemical libraries.

Key Features:

• Iterative Docking and Scoring Strategy: Employs an iterative approach where a subset of ligands is docked against target proteins to create a training set used to predict scores for remaining ligands.
• Model Retraining and Efficiency: Repeated docking and retraining continue until a predefined efficiency level is achieved, enabling exclusion of low-scoring ligands to optimize computational resources.
• Conformal Prediction with SVM: Uses Support Vector Machines (SVM) combined with conformal prediction to provide valid prediction intervals that support reliable ligand ranking.
• Parallelization with Apache Spark: Parallelizes docking and modeling tasks using Apache Spark to scale computations across clusters and accelerate processing.
• Performance Metrics: Empirical evaluation on four protein targets reduced the number of docked molecules by approximately 62.61%, maintained an average accuracy of 94% for identifying top 30 hits, and achieved a 3.7-fold computational speedup.

Scientific Applications:

• Structure-based virtual screening (SBVS): Prioritizes ligands from large chemical libraries to reduce the subset requiring full docking and scoring.
• Lead prioritization in drug discovery: Ranks and selects potential lead compounds for downstream experimental validation.

Methodology:

The methodology performs initial docking of selected ligand subsets to form a training dataset; trains Support Vector Machine models on docked scores and applies conformal prediction to obtain valid prediction intervals; predicts scores for untested ligands and excludes low-scoring compounds; iteratively docks and retrains the model until a predefined efficiency criterion is met; and parallelizes docking and modeling with Apache Spark.