Knodle: A Support Vector Machines-Based Automatic Perception of Organic Molecules from 3D Coordinates

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Accurate assignment of atom types and bond orders is crucial for understanding molecular structures.
• Existing methods for structure perception in small molecules have limitations in accuracy and efficiency.

Purpose of the Study:

• To develop a robust and accurate prediction model for assigning atom types and bond orders in low molecular weight compounds.
• To implement this model in a user-friendly software library named Knodle.

Main Methods:

• Development of a prediction model utilizing nonlinear Support Vector Machines (SVM).
• Implementation of the model within the KNOwledge-Driven Ligand Extractor (Knodle) software library.
• Training the model on extensive structural data from the PDBbindCN database.

Main Results:

• Knodle demonstrates high accuracy in atom type and bond order perception, comparable to or exceeding established methods like NAOMI, fconv, and I-interpret.
• Performance evaluation on benchmark sets (Labute's, PDBBindCN, Ligand Expo) shows Knodle's low error rates (e.g., 4.5% on Ligand Expo).
• The method exhibits competitive accuracy and running times compared to other popular structure perception tools.

Conclusions:

• Nonlinear SVM is an effective approach for structure perception tasks in computational chemistry.
• Knodle provides an efficient and robust solution for recognizing atomic types, hybridization states, and bond orders in small molecules.
• The developed software library is a valuable tool for researchers in cheminformatics and drug discovery.