A deep transfer learning-based protocol accelerates full quantum mechanics calculation of protein.

Area of Science:

• Computational Biology
• Quantum Chemistry
• Artificial Intelligence

Background:

• Full quantum mechanics (FQM) calculations for proteins are crucial for drug discovery and simulations but are computationally intensive.
• Existing quantum mechanics (QM) methods face significant challenges in handling the complexity of large biological systems.

Purpose of the Study:

• To develop an effective and efficient protocol for full quantum mechanics calculations on proteins using deep learning.
• To overcome the computational complexity limitations of traditional QM methods for large protein systems.

Main Methods:

• A transfer-learning-based deep learning (TDL) protocol, TDL-FQM, was designed by integrating transfer learning with deep neural networks (DNNs).
• The protocol utilizes models trained on small, high-precision datasets combined with knowledge from extensive low-level calculations.
• Performance was evaluated using high-level double-hybrid DFT functional and basis sets on 15 proteins.

Main Results:

• TDL-FQM achieved high accuracy with a mean absolute error of 0.01 kcal/mol/atom for potential energy and 1.47 kcal/mol/Å for atomic forces.
• The protocol demonstrated an average acceleration of over thirty thousand times compared to traditional methods.
• Computational efficiency gains increased significantly with larger protein sizes.

Conclusions:

• The TDL-FQM approach effectively overcomes the limitations of standard DNNs by learning knowledge across tasks.
• This method offers a powerful solution for high-precision prediction in large chemical and biological systems.
• TDL-FQM significantly advances the feasibility of FQM calculations in computational biology and drug discovery.