Computation Protein Design instances with small tree-width: selection based on coarse approximated 3D average position volume

TL;DR

This paper introduces a set of computational protein design instances with small tree-width, selected based on 3D volume criteria, to evaluate global search algorithms using graph decomposition techniques.

Contribution

It presents larger protein design instances with low tree-width, selected via 3D volume occupancy, for benchmarking graph decomposition-based search algorithms.

Findings

Instances range from 130 to 282 variables.

Tree-width varies from 21 to 68.

Suitable for evaluating graph decomposition search methods.

Abstract

This paper proposes small tree-width graph decomposition computational protein design CFN instances defined according to the model [1] with protocol defined by Simononcini et al [2] . The proteins used in the benchmark have been selected in the PDB (not on their biological interest) to explore the efficiency of global search method based on tree-width decomposition. The instances are bigger than those previously proposed in the paper [2] with one backbone relaxation and the aka Beta November 2016 Rosetta force-field [3]. The benchmark includes 21 proteins selected with a low level of sequences identity (40%) . Those instances have been selected on the basis of 3D criteria by applying a decreasing average coarse volume occupancy filter by Amino Acid (-i.e. by CFN variable) . The instances characteristic (see Table 1) contain from 130 up to n = 282 variables with a maximum domain size from 383 to 438, and between 1706 and 6208 cost functions. The min-fill tree-width ranges from 21 to 68, and from 0.16 to 0.34 for a normalized tree width. Those instances have been used for UDGVNS search algorithm[4] benchmarking. This approach is suitable for evaluation of search methods that exploit the notion of graph decomposition.