Adversarial attack of sequence-free enhancer prediction identifies chromatin architecture

TL;DR

This paper shows that enhancers can be accurately predicted using chromatin data alone, and introduces a new method to understand how AI models make these predictions.

Contribution

The novel contribution is a cell-type invariant enhancer prediction platform and the first use of adversarial particle swarm optimization for genomic neural networks.

Findings

Chromatin datasets alone can accurately identify enhancers genome-wide.

A multi-cell-type neural network platform was developed for enhancer prediction without DNA sequence data.

Adversarial particle swarm optimization (APSO) was used to explain genomic neural network predictions.

Abstract

The wide range of cellular complexity created by multicellular organisms is due in large part to the intricate and synergistic interplay of regulatory complexes throughout the eukaryotic genome. These regulatory elements “enhance” specific gene programs and have been shown to operate in diverse networks that are distinct across cell states of the same organism. Attempts to characterize and predict enhancers have typically focused on leveraging information-dense DNA sequence in parallel with epigenomic assays. We examined the viability of enhancer prediction using only a minimal set of epigenomic datasets without direct DNA information.

We demonstrate that chromatin datasets are sufficient to identify enhancers genome-wide with high accuracy. By training networks leveraging data from multiple cell types simultaneously, we generated a cell-type invariant enhancer prediction platform that utilized only the patterns of protein binding for inference. We also showed the utility of swarm-based adversarial attacks [adversarial particle swarm optimization (APSO)] to deconvolute trained genomic neural networks for the first time. Critically, unlike saliency mapping or other game-theory based approaches, APSO is completely network-architecture independent and can be applied to any prediction engine to derive the features that drive inference.

All software and code for data downloading, processing, enhancer inference, eXplainable AI (XAI), and complete figure generation are publicly available on GitHub at https://github.com/EpiGenomicsCode/ChromEnhancer and Zenodo at https://doi.org/10.5281/zenodo.15652797.

Figures