Modeling chromatin state from sequence across angiosperms using recurrent convolutional neural networks

Abstract

Accessible chromatin regions are critical components of gene regulation but modeling them directly from sequence remains challenging, especially within plants, whose mechanisms of chromatin remodeling are less understood than in animals. We trained an existing deep learning architecture, DanQ, on leaf ATAC-seq data from 12 angiosperm species to predict the chromatin accessibility of sequence windows within and across species. We also trained DanQ on DNA methylation data from 10 angiosperms, because unmethylated regions have been shown to overlap significantly with accessible chromatin regions in some plants. The across-species models have comparable or even superior performance to a model trained within species, suggesting strong conservation of chromatin mechanisms across angiosperms. Testing a maize held out model on a multi-tissue scATAC panel revealed our models are best at predicting constitutively-accessible chromatin regions, with diminishing performance as cell-type specificity increases. Using a combination of interpretation methods, we ranked JASPAR motifs by their importance to each model and saw that the TCP and AP2/ ERF transcription factor families consistently ranked highly. We embedded the top three JASPAR motifs for each model at all possible positions on both strands in our sequence window and observed position-and strand-specific patterns in their importance to the model. With our cross-species "a2z" model it is now feasible to predict the chromatin accessibility and methylation landscape of any angiosperm genome.

Cross-species modeling of plant genomes at single nucleotide resolution using a pre-trained DNA language model

“…DanQ is a hybrid convolutional and recurrent neural network specifically designed for predicting the function of DNA sequences. It has demonstrated impressive performance in predicting chromatin states in plant species, making it a suitable choice for our comparative analysis 49 . For each task, the CNN+LSTM model was trained from scratch using one-hot encoded DNA sequences as input.…”

Current genomic deep learning architectures generalize across grass species but not alleles

“…Though tools like AlphaFold2 (Jumper et al, 2021) have dramatically improved our ability to study coding sequence, similarly performing tools do not yet exist for non-coding regions. Nevertheless, over the last decade deep learning models have rapidly improved performance in predicting non-coding genomic features such as chromatin accessibility (Kelley, 2020; Wrightsman et al, 2022), transcription factor binding (Žiga Avsec, Weilert, et al, 2021; Mejía-Guerra & Buckler, 2019), and RNA abundance (Žiga Avsec, Agarwal, et al, 2021; Linder et al, 2023) directly from DNA sequence. These models can then be queried to highlight functional non-coding sites, which can be useful for filtering large sets of variants down to promising genome editing targets.…”

Exploring the utility of regulatory network-based machine learning for gene expression prediction in maize

“…Generally, deep learning refers to computational methods that aim to learn a hierarchical representation of data by functionally relating the data in many layers [24]. In plant data, these models have learned to predict chromatin state from sequence [25], identify and classify key stress-responsive genes [26], and detect seasonal changes across fields [27]. In addition to these successes, many state-of-the-art models take in not only numerical data but also the structure of the relationships between the data.…”