Local Entropy and LZ-Complexity of Hi-C Contact Profiles as Independent Predictors of Chromatin Accessibility Across Cell Lines

Background. Three-dimensional chromatin organization is tightly linked to gene regu- lation and genome function, yet quantitative descriptors connecting local contact patterns to functional state remain limited. Standard approaches such as A/B compartments (PC1), topologically associating domains (TADs), and chromatin loops are global or domain-level, and require complex computational pipelines. Methods. We propose two simple local features computed directly from a Hi-C contact profile: Shannon entropy of the contact distribution and normalized LZ-complexity of the binarized contact pattern. Their ability to predict chromatin accessibility (ATAC-seq coverage) was assessed with multiple linear regression, gradient boosting, and graph convolutional networks (GCN). The analysis was performed on two human cell lines—GM12878 (lymphoblastoid) and K562 (erythroleukemic)— for all 23 chromosomes (1–22, X) at 100 kb resolution. Results. Linear regression with entropy, LZ, and PC1 showed that entropy (β = −2.97, p < 10−22) and LZ (β = +9.71, p < 10−8) are significant predictors for GM12878, whereas PC1 is not (p ≈ 0.05). For K562, the analogous model revealed a significant contribution of entropy (β = −1.83, p ≈ 10−6) and a negative coefficient for LZ (β = −6.59, p < 10−15). Stratified correlation analysis for GM12878 revealed a strong negative association between entropy and ATAC signal in the B compartment (Spearman ρ = −0.397, p = 2.5 × 10−5). GCN models with three Hi-C features achieved weighted R2 = 0.319 for GM12878 and 0.299 for K562, outperforming gradient boosting (0.251 and 0.201). Adding the clustering coefficient, also computed from Hi-C, increased R2 to 0.346 for GM12878 and 0.333 for K562, confirming that additional structural information is present in the contact matrix. Permutation importance analysis confirmed the dominance of entropy in both lines (importance > 0.22), whereas the contribution of PC1 was smaller (0.12–0.19) and that of LZ was the smallest (0.05–0.11). Conclusions. Local entropy and LZ-complexity are simple, interpretable, and independent predictors of chromatin accessibility that outperform the standard compartment index in linear models. Graph neural networks provide an additional gain by exploiting local graph topology. The proposed approach is computationally efficient, requires no additional experimental data, and demonstrates stable results across two different cell lines, indicating generalizability. This work opens new avenues for studying the relationship between 3D genome structure and function using simple structural measures.