TAPIOCA: Topological Attention and Predictive Inference of Chromatin Arrangement Using Epigenetic Features

Abstract

Chromatin conformation is an important characteristic of the genome which has been repeatedly demonstrated to play vital roles in many biological processes. Chromatin can be characterized by the presence or absence of structural motifs called topologically associated domains. The de facto strategy for determination of topologically associated domains within a cell line is the use of Hi-C sequencing data. However Hi-C sequencing data can be expensive or otherwise unavailable. Various epigenetic features have been hypothesized to contribute to the determination of chromatin conformation. Here we present TAPIOCA, a self-attention based deep learning transformer algorithm for the prediction of chromatin topology which circumvents the need for labeled Hi-C data and makes effective predictions of chromatin conformation organization using only epigenetic features. TAPIOCA outperforms prior art in established metrics of TAD prediction, while generalizing across cell lines beyond those used in training. Availabilitythe source code of TAPIOCA and training and test datasets are available at https://github.com/Max-Highsmith/TAPIOCA Author SummaryIn this paper we outline a machine learning approach for predicting the topological organization of chromosomes using epigenetic track data as features. By utilizing an architecture inspired by the sequence transduction transformer network we are able to effectively predict multiple metrics used to characterize topologically associated domains. Our experimental results demonstrate that once trained our algorithm can effectively predict topological organization on novel cell lines all without any exposure to original Hi-C data in test datasets.