Motivation: Simulating protein folding motions is
an important problem in computational biology. Motion planning
algorithms such as Probabilistic Roadmap Methods (PRMs) have
been successful in modeling the protein folding landscape. PRMs
and variants contain several phases (i.e., sampling, connection,
and path extraction). Global machine learning has been applied
to the connection phase but is inefficient in situations with varying
topology, such as those typical of folding landscapes.
Results: We present a local learning algorithm that considers
the past performance near the current connection attempt as
a basis for learning. It is sensitive not only to different types
of landscapes but also to differing regions in the landscape
itself, removing the need to explicitly partition the landscape.
We perform experiments on 23 proteins of varying secondary
structure makeup with 52–114 residues. Our method models the
landscape with better quality and comparable time to the best
performing individual method and to global learning.