DeepNCI

DeepNCI improves the accuracy of noncovalent interaction (NCI) calculations traditionally performed with density functional theory (DFT) by using a multimodal deep learning model that integrates 3D electron density and quantum chemical descriptors.

Key Features:

• Multimodal Architecture: A three-dimensional convolutional neural network (3D CNN) processes 3D electron density data while a separate neural network models one-dimensional quantum chemical properties.
• Error Reduction: Merging features from both networks reduces root-mean-square error in DFT-calculated NCIs from approximately 1.19 kcal/mol to around 0.2 kcal/mol across a database of over 1000 molecules.
• Feature Visualization: Joint features are visualized with t-distributed stochastic neighbor embedding (t-SNE) to distinguish categorized NCI systems.
• Model Applicability and Transferability: The 3D CNN accepts electron density inputs standardized across varying molecular sizes to maintain consistent performance across systems.
• Application Domain (AD) Definition: An application domain is defined using merged features via the k-nearest-neighbor method to assess prediction reliability on external test sets.
• Transfer Learning Capability: Pretrained parameters from a large NCI database are applied to smaller datasets, such as homolysis bond dissociation energy data, achieving comparable or improved performance.

Scientific Applications:

• Computational Chemistry and Molecular Modeling: Improves accuracy of NCI calculations used to analyze molecular interactions and stability.
• Drug Design: Provides more precise NCI data to support structure-based drug design and interaction assessment.
• Material Science: Enhances prediction of intermolecular interactions relevant to material properties and performance.
• Homolysis Bond Dissociation Energy Prediction: Enables transfer learning for improved predictions on small datasets involving bond dissociation energies.

Methodology:

DeepNCI integrates a 3D CNN for spatial electron density processing with a separate neural network for quantum chemical descriptors, fuses their features for prediction, visualizes joint features with t-SNE, defines an application domain via k-nearest-neighbor on merged features, and applies transfer learning using pretrained parameters from a large NCI database.