A Pre-Docking Filter Based on Image Recognition

TL;DR

This paper presents a novel pre-docking filter based on image recognition techniques, specifically tangent distance and principal component analysis, to efficiently reduce the chemical space before molecular docking in drug design.

Contribution

The study introduces a new pre-docking filtering method using tangent distance and PCA, adapted for 3D molecular data, to improve screening efficiency in drug discovery.

Findings

PCA-based filtering achieved 81% sensitivity with 1-4 seconds runtime.

Tangent distance approach was ineffective for ligand orientation optimization.

Enrichment of active compounds increased slightly from 2.64% to 2.82%.

Abstract

Molecular docking is a central method in the computer-based screening of compound libraries as a part of the rational approach to drug design. Although the method has proved its competence in predicting binding modes correctly, its inherent complexity puts high demands on computational resources. Moreover the chemical space to be screened is prohibitively large. Therefore the application of filtering prior to docking is a promising concept. We implemented a pre-docking filter based on the tangent distance algorithm originally conceived for optical character recognition. The challenging transfer of the method from two-dimensional to three-dimensional data was achieved by representing the molecular structure by a set of density maps extracted from different views of the compound. Additionally, our program applies a binary classification using principal component analysis. Ligand and binding pocket are aligned according to their centroidal axes, enabling a size-based filtering for the purpose of enriching the dataset regarding ligands before docking. The evaluation of our program via redocking produced RMSD values between 8{\AA} and 25{\AA}, indicating that the tangent distance approach is not suited for optimizing the orientation of a ligand and binding pocket. Investigating probable explanations lead to the conclusion that a likely cause for these results is the method's known inability to approximate large transformations. A validation of the principal component analysis alone performed better: Tests on a dataset of 170 ligands and 6,435 decoys yielded a sensitivity of 0.81, while keeping the runtime within a reasonable timeframe (1 to 4 seconds). The dataset's enrichment increased from 2.64% to 2.82%.