Answers are grounded in this paper and cite their sources.
Shannon entropy is a standard proxy for DNA sequence complexity, but under fixed-length sampling it often exhibits saturation and limited dynamic range, reducing its utility for discriminating sequences that share similar base composition. We introduce the entropy–rank ratio R , a distribution-aware complexity rank defined as the cumulative proportion of all length- T blocks (under non-overlapping n -tuples) whose Shannon entropy is less than or equal to that of the target block. By calibrating entropy against its attainability distribution at fixed $$(T,n)$$ , R∈ [0,1] provides a comparable and expanded complexity scale under a constrained protocol. We operationalize R as a crop-selection signal for fixed-length data augmentation: among candidate crops, we select segments whose R value best matches the full-sequence complexity signature while penalizing excessive displacement from the center. Across two independent classification problems (viral gene classes and human genes with polynucleotide expansions), ratio-guided cropping yields consistent improvements over random, entropy-based, and compression-based crop heuristics when training lightweight CNNs in small labeled-data regimes. Because the method is model-agnostic and the reference distribution can be cached per $$(T,n)$$ , it scales to the sequence lengths used in our benchmarks while preserving a strict fixed-length input protocol.
No takes yet. Share an insight, caveat, or question.
Pastore et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: