Novel AI-powered computational method using tensor decomposition for identification of common optimal bin sizes when integrating multiple Hi-C datasets.

Identifying the optimal bin sizes (or resolutions) for the integration of multiple Hi-C datasets is a challenge due to the fact that bin sizes must be common over multiple datasets. By contrast, the dependence of quality upon bin sizes can vary from dataset to dataset. Moreover, common structures should not be sought in bin sizes smaller than the optimal bin sizes, below which common structure cannot be the primary structure any more even after increasing the number of mapped short reads per bin. In this case, there are no common structures at finer resolutions, suggesting that individual Hi-C datasets may have to be analyzed separately in the bin sizes smaller than the optimal one. Thus, quality assessments of individual datasets have a limited ability to determine the best bin size for all datasets. In this study, we propose a novel application of tensor decomposition (TD) based unsupervised feature extraction (FE) to choose the optimal bin sizes for the integration of multiple Hi-C datasets. TD-based unsupervised FE exhibit phase transition-like phenomena through which the smallest possible bin size (or the highest resolution) can be automatically estimated empirically, without the need to manually set a threshold value for the integration of multiple Hi-C datasets, retrieved from GEO with GEO ID, GSE260760 and GSE255264. To our knowledge, ours is the first one that can optimize bin sizes over multiple Hi-C profiles without any tunable parameters.