The missing link of nonlinear manifold learning computation application in bimolecular simulations

BEHROOZ HASHEMIAN; DANIEL MILLÁN; MARINO ARROYO

Casteldefells, Barcelona

Workshop; 12th Workshop on Numerical Methods in Applied Science and Engineering (NMASE 2013); 2013

LaCàN - Universitat Politècnica de Catalunya

Many biological molecules are highly complex and involve many thousand atoms. However in most cases, a substantial amount of redundancy exists due to the variable interdependency that is an obvious result of the constraints on atomic positions due to covalent bonds and energy barriers. Hence, they lie close to a low-dimensional manifold. Nonlinear dimensionality reduction methods have been proved to be very powerful to discover such intrinsic manifolds and have been widely used to visualize the inherent patterns of data. Nevertheless, performing computation on the discovered low- dimensional manifold is not feasible due to the missing link from low-dimensional to high-dimensional space. Here, we proposed a method that enables nonlinear manifold learning techniques to do computations in the low-dimension space and transfer the calculated potential to the high-dimension space.