Tensor Train Optimization for Conformational Sampling of Organic Molecules

Exploring the conformational space of molecules remains
a challenge
of fundamental importance to quantum chemistry: identification of
relevant conformers at ambient conditions enables predictive simulations
of almost arbitrary properties. Here, we propose a novel approach,
called TTConf, to enable conformational sampling of large organic
molecules where the combinatorial explosion of possible conformers
prevents the use of a brute-force systematic conformer search. We
employ tensor trains as a highly efficient dimensionality reduction
algorithm, effectively reducing the scaling from exponential to polynomial.
In our approach, the conformational search is expressed as global
energy minimization task in a high-dimensional grid of dihedral angles.
Dimensionality reduction is achieved through a tensor train representation
of the high-dimensional torsion space. The performance of the approach
is assessed on a variety of drug-like molecules in direct comparison
to the state-of-the-art metadynamics based conformer search as implemented
in CREST. The comparison shows significant acceleration of up to an
order of magnitude, while maintaining comparable accuracy. More importantly,
the presented approach allows treatment of larger molecules than typically
accessible with metadynamics.