DeepNCI: DFT Noncovalent Interaction Correction with Transferable Multimodal Three-Dimensional Convolutional Neural Networks | JoVE Visualize

Area of Science:

Computational Chemistry
Machine Learning
Quantum Chemistry

Background:

Accurate calculation of noncovalent interactions (NCIs) is crucial in chemistry.
Density functional theory (DFT) is a common method for NCI calculations, but can be computationally intensive and prone to errors.
Developing more efficient and accurate methods for NCI prediction is an ongoing challenge.

Purpose of the Study:

To propose DeepNCI, a novel multimodal deep learning model, for improving the accuracy of NCIs calculated using DFT.
To leverage both electron density and quantum chemical properties for enhanced NCI prediction.
To assess the model's applicability, transferability, and reliability in predicting NCIs.

Main Methods:

A three-dimensional convolutional neural network (3D CNN) was employed to extract features from 3D electron density.
A separate neural network was used to model one-dimensional quantum chemical properties.
Features from both networks were merged in a fused model (DeepNCI) for NCI prediction.
t-distributed stochastic neighbor embedding (t-SNE) was used for visualizing feature representativeness.
An application domain (AD) was defined using K-nearest neighbor for reliability monitoring.
Transfer learning was applied to test model transferability on a small homolysis bond dissociation energy dataset.

Main Results:

DeepNCI reduced the root-mean-square error of DFT-calculated NCIs from 1.19 kcal/mol to approximately 0.2 kcal/mol on a large molecular database.
The fused model demonstrated superior performance compared to its individual component networks.
t-SNE visualization confirmed the representativeness of joint features in distinguishing NCI systems.
The defined application domain effectively monitored prediction reliability on external test sets.
Transfer learning with pretrained NCI parameters achieved competitive or better performance on a small bond dissociation energy dataset, indicating model transferability.

Conclusions:

DeepNCI offers a significant improvement in the accuracy of noncovalent interaction calculations.
The multimodal approach effectively combines electron density and quantum chemical properties for robust predictions.
The model exhibits good applicability and transferability, suggesting potential for broader use in computational chemistry and related tasks.
DeepNCI can potentially address challenges associated with small datasets through transfer learning.