DeepLNE++ leveraging knowledge distillation for accelerated multi-state path-like collective variables.

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Path-like collective variables (CVs) are crucial for modeling complex biomolecular processes in molecular dynamics (MD) simulations.
• DeepLNE (deep-locally non-linear-embedding) was previously introduced as a machine learning-based CV for accurate reaction coordinate approximation.
• Limitations of DeepLNE include computational expense for large systems and difficulty with multi-state reactions.

Purpose of the Study:

• To present DeepLNE++, an enhanced version of DeepLNE designed for accelerated computation of free energy landscapes.
• To improve the efficiency and applicability of path-like CVs for large and complex biomolecular systems.
• To enhance the versatility and effectiveness of DeepLNE through a multitasking framework.

Main Methods:

• Implementation of a knowledge distillation approach to significantly accelerate DeepLNE evaluation.
• Development of a supervised multitasking framework to encode system-specific knowledge.
• Application to large and complex biomolecular systems requiring numerous descriptors.

Main Results:

• DeepLNE++ enables feasible computation of free energy landscapes for realistic biomolecular systems.
• Significant acceleration in the evaluation of DeepLNE is achieved through knowledge distillation.
• Enhanced versatility and effectiveness demonstrated via the multitasking framework.

Conclusions:

• DeepLNE++ represents a significant advancement in computational biomolecular modeling.
• The new method overcomes previous limitations of DeepLNE, enabling broader applications.
• DeepLNE++ facilitates more accurate and efficient analysis of complex molecular dynamics.