EARTHMOVER-BASED MANIFOLD LEARNING FOR ANALYZING MOLECULAR CONFORMATION SPACES

Area of Science:

• Computational biology
• Biophysics
• Machine learning

Background:

• Dimensionality reduction is crucial for analyzing complex biological data.
• Manifold learning methods like diffusion maps are effective but can be computationally intensive.
• Standard diffusion maps rely on Euclidean distance, which may not capture complex geometric relationships in biological structures.

Purpose of the Study:

• To develop a novel manifold learning approach combining Earthmover's distance (EMD) with diffusion maps.
• To improve the efficiency and accuracy of dimensionality reduction for learning shape spaces of flexible macromolecules.
• To validate the approach using simulated 3-D density maps of ATP synthase.

Main Methods:

• The proposed method integrates Earthmover's distance (EMD) with diffusion maps for dimensionality reduction.
• A wavelet-based approximation of EMD is employed to reduce computational complexity.
• The approach is tested on simulated 3-D density maps representing macromolecular motion.

Main Results:

• EMD-based diffusion maps require significantly fewer samples to recover intrinsic geometry compared to standard Euclidean distance-based diffusion maps.
• The wavelet approximation effectively reduces the computational burden of pairwise EMD calculations.
• The method successfully learns the shape space of simulated ATP synthase data.

Conclusions:

• The novel EMD-based diffusion maps offer a more sample-efficient and potentially more accurate method for dimensionality reduction in computational biology.
• This approach shows promise for analyzing the conformational dynamics of flexible macromolecules like proteins.
• Wavelet approximation is a viable strategy for enhancing the computational scalability of EMD-based manifold learning.