MolGpKa

MolGpKa predicts acid dissociation constants (pKa) of small molecules using a graph-convolutional neural network to support lead optimization and assessment of ionization-dependent properties.

Key Features:

• Graph-convolutional neural network: Learns chemical patterns related to pKa directly from molecular graphs.
• Data-driven feature learning: Constructs predictors from features learned directly from data without reliance on human-engineered fingerprints.
• Training dataset: Trained on ACD/pKa data for 1.6 million compounds sourced from the ChEMBL database.
• Performance versus traditional models: Demonstrates better performance than traditional machine learning models that use manually crafted chemical descriptors.
• Substitution-effect learning: Accurately captures substitution effects on pKa.
• Relevance to physicochemical properties: Enables estimation of ionization states that influence biological activity, solubility, membrane permeability, metabolism, and toxicity.

Scientific Applications:

• Lead optimization: Predicts pKa values to guide medicinal chemistry decisions during lead optimization.
• Ligand design: Supports ligand design by estimating pKa and modeling substitution effects on ionization.
• Ionization-dependent ADMET assessment: Assesses how ionization state affects solubility, membrane permeability, metabolism, and toxicity.

Methodology:

MolGpKa uses a graph-convolutional neural network that learns chemical patterns related to pKa, builds predictors from learned features rather than engineered fingerprints, and was trained on ACD/pKa data for 1.6 million compounds from the ChEMBL database.