bioRxivGenomicsPreprint
Dilated Convolutions for Modeling Long-Distance Genomic Dependencies
Abstract
We consider the task of detecting regulatory elements in the human genome directly from raw DNA. Past work has focused on small snippets of DNA, making it difficult to model long-distance dependencies that arise from DNAs 3-dimensional conformation. In order to study long-distance dependencies, we develop and release a novel dataset for a larger-context modeling task. Using this new data set we model long-distance interactions using dilated convolutional neural networks, and compare them to standard convolutions and recurrent neural networks. We show that dilated convolutions are effective at modeling the locations of regulatory markers in the human genome, such as transcription factor binding sites, histone modifications, and DNAse hypersensitivity sites.
- organism: human
- assay: DNA methylation
- assay: whole-genome sequencing
- research focus: gene regulatory network inference
- research focus: differential expression
- research focus: variant calling
- approach: deep learning
- approach: statistical model
Abstract summary
Claude Opus 5.5 · xhigh effort · 18 Oct 2026 · from the abstract · CC BY-NC 4.0
Dilated convolutions let models of raw DNA span longer stretches; on a newly released large-context dataset they locate regulatory marks, with standard convolutions and recurrent networks as comparisons.
Sequence models of regulatory elements have mostly been trained on short DNA snippets, which leaves out the long-distance dependencies that follow from the three-dimensional folding of the genome. This work supplies both a dataset framed around longer context and an architecture, dilated convolutions, meant to use it.
Key results
ClaimedDilated convolutional networks can predict where regulatory marks sit in the human genome from sequence alone, covering transcription factor binding, histone modifications and DNase hypersensitivity.
Authors report the architecture is effective on their larger-context task, but the abstract gives no accuracy, AUROC or other metric for any of the three marker types. · Abstract
ClaimedDilated convolutions are set against two baselines, ordinary convolutional networks and recurrent networks, on the same task.
The comparison is stated without any reported margin, metric or per-architecture result in the abstract. · Abstract
ClaimedA dataset built specifically for modelling with a larger sequence context is developed and released alongside the models.
Described as a new dataset for a larger-context modelling task; the abstract gives no size, genome build, label source or download location. · Abstract
Methods, data and reproducibility
Regulatory-element detection is framed as prediction from raw DNA over a longer input window; dilated convolutional neural networks are trained on a purpose-built dataset and compared with standard convolutional and recurrent architectures.
- New dataset released by the authors for modelling regulatory marks with a larger sequence context · human
- Methods & tools
- dilated convolutional neural networks, convolutional neural networks, recurrent neural networks
- Reproducibility gaps
- The dataset is said to be released, but the abstract gives no URL, repository or accession, and no code availability statement.
Limitations (2)
- The evaluation runs on a dataset the authors themselves constructed, so the task definition and the method are from the same group and no external benchmark is involved. · noted in this summary
- Abstract only: no metrics are reported for any architecture, so the claim that dilated convolutions work well cannot be weighed against the convolutional and recurrent baselines. · noted in this summary
Read this if you build sequence-to-function models of regulatory DNA and are weighing architectures that widen input context beyond short windows.
- Citations
- 60
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- DNALONGBENCH: A Benchmark Suite for Long-Range DNA Prediction Tasks
Wenduo Cheng, Zhenqiao Song, … Jian Ma·Nature CommunicationsPeer-reviewed·Nov 18, 2025
- Self-supervised learning for DNA sequences with circular dilated convolutional networks
Lei Cheng, Tong Yu, … Zhirong Yang·Neural NetworksPeer-reviewed·Dec 2, 2023
- Genome Functional Annotation using Deep Convolutional Neural Networks
Ghazaleh Khodabandelou, Etienne Routhier, Julien Mozziconacci·bioRxiv·Jun 7, 2019
- ChromDL: A Next-Generation Regulatory DNA Classifier
Chris Hill, Sanjarbek Hudaiberdiev, Ivan Ovcharenko·BioinformaticsPeer-reviewed·May 24, 2023
- Modeling positional effects of regulatory sequences with spline transformations increases prediction accuracy of deep neural networks
Žiga Avsec, Mohammadamin Barekatain, … Julien Gagneur·BioinformaticsPeer-reviewed·Nov 15, 2017
- Projection layers improve deep learning models of regulatory DNA function
Alex Hawkins‐Hooker, Henry Kenlay, John E. Reid·F1000ResearchPeer-reviewed·Feb 5, 2019