Problems concerned with learning the relationships between molecular s
tructure and activity have been important test-beds for Inductive Logi
c programming (ILP) systems, In this paper we examine these applicatio
ns and empirically evaluate the extent to which a first-order represen
tation was required. We compared ILP theories with those constructed u
sing standard linear regression and a decision-tree learner on a serie
s of progressively more difficult problems. When a propositional encod
ing is feasible for the feature-based algorithms, we show that such al
gorithms are capable of matching the predictive accuracies of an ILP t
heory. However, as the complexity of the compounds considered increase
d, propositional encodings becomes intractable. In such cases, our res
ults show that ILP programs can still continue to construct accurate,
understandable theories. Based on this evidence, we propose future wor
k to realise fully the potential of ILP in structure-activity problem.