Vertex Assignment for Frequency Chaos Game Representation of Proteins Affects Classification Performance.

Chaos game representation and its frequency matrix variant are popular sequence encodings commonly assumed to be invariant to alternative nodal orientations. In this paper, we systematically test this assumption on 3 protein benchmark datasets with 2 convolutional neural networks and show that different vertex assignments lead to significant differences in classification performance. Evaluating 1,000 random encodings revealed a wide performance range, proving that vertex assignment is a critical parameter rather than an arbitrary choice. Furthermore, in most of the cases, frequency chaos game representation-based data augmentation proved to be effective only under moderate settings, while aggressive augmentation degraded performance due to distributional shifts introduced by permuted encodings.