Safe Sequences via Dominators in DAGs for Path-Covering Problems

TL;DR

This paper introduces a novel method for identifying safe sequences in DAGs using dominator trees, enabling efficient solutions for path-covering problems and improving ILP-based bioinformatics applications.

Contribution

It provides a simple characterization of safe sequences via cutnodes and links them to dominator trees, enabling linear-time enumeration and practical ILP simplification.

Findings

Safe sequences can be represented with O(n) size.

Enumeration algorithm runs in linear time O(m + o).

Significant speed-ups in ILP solving for RNA transcript assembly.

Abstract

A path-covering problem on a directed acyclic graph (DAG) requires finding a set of source-to-sink paths that cover all the nodes, all the arcs, or subsets thereof, and additionally they are optimal with respect to some function. In this paper we study safe sequences of nodes or arcs, namely sequences that appear in some path of every path cover of a DAG. We show that safe sequences admit a simple characterization via cutnodes. Moreover, we establish a connection between maximal safe sequences and leaf-to-root paths in the source- and sink-dominator trees of the DAG, which may be of independent interest in the extensive literature on dominators. With dominator trees, safe sequences admit an O(n)-size representation and a linear-time output-sensitive enumeration algorithm running in time O(m + o), where n and m are the number of nodes and arcs, respectively, and o is the total length of the maximal safe sequences. We then apply maximal safe sequences to simplify Integer Linear Programs (ILPs) for two path-covering problems, LeastSquares and MinPathError, which are at the core of RNA transcript assembly problems from bioinformatics. On various datasets, maximal safe sequences can be computed in under 0.1 seconds per graph, on average, and ILP solvers whose search space is reduced in this manner exhibit significant speed-ups. For example on graphs with a large width, average speed-ups are in the range 50-250x for MinPathError and in the range 80-350x for LeastSquares. Optimizing ILPs using safe sequences can thus become a fast building block of practical RNA transcript assembly tools, and more generally, of path-covering problems.