NNGEOAS: A GEOMETRIC-BASED MULTI-FEATURE NEURAL NETWORK CLASSIFIER FOR ACTIVE SITE SIMILARITY SCREENING.

Abstract

This work introduces NNGeoAS (Neural Network Geometric Active Site), a novel structure-based high-throughput virtual screening method designed to overcome the limitations of existing tools that inadequately incorporate active site geometry and pharmacophore features. A key problem in current virtual screening approaches is the insufficient integration of geometric and physicochemical characteristics of protein active sites into machine learning models, which hampers screening accuracy and generalization across diverse targets. NNGeoAS addresses this gap by combining physicochemical properties, active site geometry, and pharmacophore features to screen ligands against target proteins effectively. It utilizes Convex Hull, Ultrafast Shape Recognition (USR), Euclidean distance, and the Hungarian algorithm to assess active site similarity and enhance screening performance. By autonomously learning structure-informed descriptors, NNGeoAS classifies compounds as active or inactive with high accuracy. The model incorporates 20 diverse features and is optimized using early stopping, class weighting, and hyperparameter tuning to manage data imbalance and reduce overfitting. On known datasets DUD-E (102 targets) and MUV (13 targets), NNGeoAS achieves strong performance with average EF1% scores of 36.39 and 49.30, and ROC AUC values of 0.972 and 0.956, respectively. Comparative evaluations show that NNGeoAS consistently performing good or matches existing virtual screening tools, including LigMate, PyRMD, AutoDock Vina, and machine learning baselines like Logistic Regression and Gradient Boosting. Despite challenges with highly imbalanced targets, NNGeoAS demonstrates robust generalization capability, making it a promising pre-screening tool for accelerating drug discovery through efficient prioritization of active compounds in large-scale virtual screening tasks.