Abstract
- Introduction. The minimum flow decomposition (MFD) problem asks to decompose a directed acyclic flow network (G,f) with a unique source s and a unique sink t into the fewest weighted s-t paths whose combined contributions exactly reproduce f. MFD underlies a broad class of multi-assembly tasks in bioinformatics: reference-based RNA assembly from splice graphs [Trapnell et al., 2010; Guttman et al., 2010; Tomescu et al., 2013; Song et al., 2016; Liu et al., 2016; Pertea et al., 2015; Kovaka et al., 2019; Shao and Kingsford, 2017; Zhang et al., 2022; Tung et al., 2019], metagenomic assembly [Shaw et al., 2024], and viral quasi-species inference [Baaijens et al., 2020].
MFD is strongly NP-hard [Vatinlen et al., 2008] and hard to approximate within some fixed constant factor [Hartman et al., 2012]. Exact solvers include an FPT algorithm whose runtime grows exponentially in the solution size [Kloster et al., 2018] and a family of integer linear programming (ILP) formulations [Dias et al., 2022; Grigorjew et al., 2024] capable of handling extensions such as inexact flows [Williams et al., 2019; Dias and Tomescu, 2024], safety and subpath constraints [Williams et al., 2022; Gibney et al., 2022; Khan et al., 2022; Dias et al., 2023], and graphs with cycles [Dias et al., 2025]. However, ILP remains unscalable on large practical instances.
The widely used greedy-width heuristic [Vatinlen et al., 2008] is very efficient but can be exponentially worse than optimal in the worst case [Cáceres et al., 2024]. The state-of-the-art heuristic, catfish [Shao and Kingsford, 2017], substantially improves this efficiency-performance tradeoff by identifying linear equations among edge flow values - structural constraints implied by any optimal decomposition - and resolving them via safe graph transformations. On simpler instances catfish is highly effective, but three interrelated limitations degrade its performance on complex graphs: (1) it cannot distinguish good equations (arising from a true optimal decomposition) from superficial ones that distort the graph when resolved; (2) many good equations cannot be fully resolved due to the absence of suitable closed subgraphs, so catfish discards their information entirely; and (3) when no equation is resolvable catfish falls back to greedy-width, which performs poorly on entangled graphs.
- Method. We introduce catfish-LP, which augments catfish with a lightweight linear programming (LP) formulation based on saturating subflows. For each edge e ∈ E we define continuous variables {x_e(a) : a ∈ E} modeling a valid s-t subflow that saturates e; intuitively, x_e(a) represents the amount of flow on e that must passes through edge a. Five base constraints enforce saturation, symmetry, flow validity, and flow conservation. Two additional equation constraints require that the aggregate subflow through the left-hand-side edges of an equation equals that through the right-hand-side edges. The full LP is polynomial-time solvable, adding only modest overhead over catfish.
The LP plays three complementary roles within a single unified framework: (1) equation filtering: if adding a candidate equation renders the LP infeasible, that equation cannot arise from any minimum decomposition and is discarded, preventing structurally invalid graph transformations; (2) safe edge merging: a feasible LP solution reveals pairs of edges that must carry identical subflow and can therefore be safely contracted; (3) informed greedy extraction: when no further simplification is possible, rather than invoking greedy-width blindly, catfish-LP extracts from the LP solution the heaviest simple path consistent with all surviving equations, deferring the error-prone greedy step as long as possible.
- Experimental Results. We compare catfish-LP against greedy-width, catfish, and the optimized ILP solver [Grigorjew et al., 2024] on two benchmarks, using Gurobi [{Gurobi Optimization, 2024] as the underlying LP/ILP engine. Table 1 reports decomposition quality on four datasets of biologically derived splice graphs, with abundances estimated by Salmon [Patro et al., 2017] or simulated with the Flux-Simulator [Griebel et al., 2012]. Catfish-LP achieves the smallest excess among heuristics on the Salmon dataset and negative excess on the remaining three, matching ILP quality while being orders of magnitude faster. Figure 1 summarizes results on 1,440 simulated graphs spanning 72 complexity configurations. Catfish-LP consistently produces the smallest decompositions and recovers the most ground-truth paths among all heuristics, achieving near-ILP quality in a fraction of its runtime - including a threefold improvement over catfish on the hardest 498 instances where ILP times out on every instance.
- Conclusion. Catfish-LP demonstrates that incorporating a polynomial-time LP oracle into a combinatorial heuristic yields substantial gains in decomposition quality with negligible scalability cost, addressing each of catfish’s core limitations in a unified manner. Future directions include strengthened LP formulations, domain-specific constraints for transcriptome assembly, and probabilistic interpretations of LP-guided decompositions.