Some biological sequences contain subsequences of unusual composition; e.g. some proteins contain DNA binding domains, transmembrane regions and charged regions, and some DNA sequences contain repeats. The linear-time Ruzzo-Tompa (RT) algorithm finds subsequences of unusual composition, using a sequence of scores as input and the corresponding ‘maximal segments’ as output. In principle, permitting gaps in the output subsequences could improve sensitivity. Here, the input of the RT algorithm is generalised to a finite, totally ordered, weighted graph, so the algorithm locates paths of maximal weight through increasing but not necessarily adjacent vertices. By permitting the penalised deletion of unfavourable letters, the generalisation therefore includes gaps. The program RepWords, which finds inexact simple repeats in DNA, exemplifies the general concepts by outperforming a similar extant, ad hoc tool. With minimal programming effort, the generalised Ruzzo-Tompa algorithm could improve the performance of many programs for finding biological subsequences of unusual composition.