An inductive transfer learning force field (ITLFF) protocol builds protein force fields in seconds.

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Accurate protein folding simulation is crucial for protein design and drug discovery.
• Current molecular dynamics force fields lack accuracy, applicability, and efficiency.
• Advanced force fields are essential for reliable protein folding simulations.

Purpose of the Study:

• To develop a novel machine learning protocol for constructing accurate and efficient protein force fields.
• To address the limitations of existing force fields in molecular dynamics simulations.
• To accelerate the process of protein folding simulation and protein design.

Main Methods:

• Developed the inductive transfer learning force field (ITLFF) protocol using deep neural networks.
• Incorporated an inductive transfer learning algorithm to learn from large datasets of low-level calculations.
• Utilized double-hybrid density functional theory (DFT) as a case functional, applicable to other high-precision functionals.

Main Results:

• ITLFF constructs protein force fields in seconds from small datasets with high accuracy.
• Achieved a mean absolute error of 0.0039 kcal/mol/atom for energy and RMSE of 2.57 kcal/mol/Å for force.
• Demonstrated over 30,000 times speedup compared to fragment-based double-hybrid DFT, with increasing efficiency for larger systems.

Conclusions:

• ITLFF offers a promising solution for accurate and efficient protein dynamic simulations.
• The protocol significantly advances the field of protein folding simulation.
• ITLFF's knowledge transfer capability makes it applicable to diverse problems in biology, chemistry, and material science.