Predicting 3D chromosome architecture with delay differential analysis

The 3D conformation of chromatin is crucial for the precise regulation of gene expression. Enhancers regulate the transcription of genes by interacting with their respective promoters in what is known as enhancer-promoter interaction (EPI). Enhancers can be located thousands to millions of base-pairs away from genes they regulate. Experimental techniques such as HiC and ChIA-PET enable the identification of EPI pairs, many of which are cell-type specific. The high demands and costs of such experiments limit the availability of EPI data across cell-types. This has driven the development of computational learning models which use sequence and/or epigenomic data for predicting EPIs. Purely sequence-based models are advantageous for studying the effects of genetic variants, but face challenges in achieving satisfactory predictive performance. This naturally raises the question to what extent the information underlying valid EPI pairs is encoded within the DNA sequence itself?
We analyze sequences of EPI pairs in the context of delay differential analysis (DDA), a nonlinear time-series classification framework which we have extended for analyzing DNA sequences (DNA-DDA). DDA employs a sparse functional delay differential embedding in a fundamentally different approach to traditional machine learning techniques. DDA model parameters are fixed and not iteratively updated to "learn" through repeated training cycles. DDA can achieve high classification performance using a low dimensional feature set rendering it insensitive to overfitting. DNA-DDA has been used to predict genome-wide A/B compartmentalization, a fundamental feature of 3D chromatin architecture. It competed well with alternative methods while requiring a fraction of the data used for "training". This highlights the method's ability to extract key sequence features making it a promising alternative computational tool for studying various genomic contexts such as the EPI pair classification task.