TorsionNet

TorsionNet predicts small molecule torsion energy profiles at density functional theory (DFT) accuracy using a deep neural network for rapid assessment of dihedral energetics in medicinal chemistry.

Key Features:

• Deep Neural Network (DNN): Uses a deep neural network to model torsion energy profiles derived from quantum mechanical data.
• Quantum Mechanics-Level Accuracy: Achieves DFT-level accuracy in predicted torsion energy profiles for small molecules.
• Active Learning: Employs active learning to select nearly 50,000 fragments from a corporate library covering elements H, C, N, O, F, S, and Cl.
• Massively Parallel DFT Scans: Performs density functional theory (DFT) torsion scans on selected fragments using massively parallel cloud computing to generate training data.
• Large Training Dataset: Trained on a dataset comprising approximately 1.2 million DFT energies from torsion scans.
• Rapid Prediction Capability: Once trained, predicts torsion energy profiles of typical drug-like fragments much faster than performing new QM torsion scans.
• Uncertainty Estimation: Provides direct estimates of uncertainty in predicted torsion profiles without additional calculations.
• Benchmark Dataset (TorsionNet500): Includes a benchmark of 500 chemically diverse fragments with DFT torsion profiles comprising 12,000 DFT-optimized geometries and energies.

Scientific Applications:

• Conformational analysis: Predicts preferred dihedral geometries and torsion energy profiles relevant to observed crystal structures.
• Lead discovery and design: Incorporates DNN-based strain energy assessments into lead discovery and optimization workflows in medicinal chemistry.

Methodology:

Active learning selected ~50,000 fragments; massively parallel cloud computing was used to perform DFT torsion scans on those fragments to generate ~1.2 million DFT energies for training a deep neural network; the trained model predicts torsion profiles and provides uncertainty estimates, and the TorsionNet500 benchmark contains 500 fragments with 12,000 DFT-optimized geometries and energies.