A multi-component power-law penalty corrects distance bias in single-cell co-accessibility and deep-learning chromatin interaction predictions

Abstract

Scalable proxies for 3D genome contacts - such as single-cell co-accessibility and deep learning predictions - have emerged as powerful alternatives to chromatin capture-based methods, but predictions systematically overestimate long-range interactions. Here we show how to correct this bias using distance-based penalty functions informed by Gaussian mixture modeling and polymer-physics scaling. Using Hi-C datasets from maize, rice, and soybean, we derive tissue-specific and global consensus penalties parameterized by multi-regime power-law exponents. Applying these corrections to scATAC-seq co-accessibility scores improves their distance profiles in concordance with Hi-C and reduces long-range false positives by an average of 73% with tissue-specific penalties and 66% with the global consensus. We provide open-source code and fitted parameters to support adoption in maize, rice, and soybean.