Skip to main page content
U.S. flag

An official website of the United States government

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2023 Dec 29:23:679-687.
doi: 10.1016/j.csbj.2023.12.044. eCollection 2024 Dec.

DeepCORE: An interpretable multi-view deep neural network model to detect co-operative regulatory elements

Pramod Bharadwaj Chandrashekar  1   2 , Hai Chen  3   4 , Matthew Lee  3 , Navid Ahmadinejad  3   4 , Li Liu  3   4
Affiliations

DeepCORE: An interpretable multi-view deep neural network model to detect co-operative regulatory elements

Pramod Bharadwaj Chandrashekar et al. Comput Struct Biotechnol J. .

Abstract

Gene transcription is an essential process involved in all aspects of cellular functions with significant impact on biological traits and diseases. This process is tightly regulated by multiple elements that co-operate to jointly modulate the transcription levels of target genes. To decipher the complicated regulatory network, we present a novel multi-view attention-based deep neural network that models the relationship between genetic, epigenetic, and transcriptional patterns and identifies co-operative regulatory elements (COREs). We applied this new method, named DeepCORE, to predict transcriptomes in various tissues and cell lines, which outperformed the state-of-the-art algorithms. Furthermore, DeepCORE contains an interpreter that extracts the attention values embedded in the deep neural network, maps the attended regions to putative regulatory elements, and infers COREs based on correlated attentions. The identified COREs are significantly enriched with known promoters and enhancers. Novel regulatory elements discovered by DeepCORE showed epigenetic signatures consistent with the status of histone modification marks.

Keywords: Cooperative regulatory elements; Deep learning; Epigenetics; Gene regulation.

PubMed Disclaimer

Conflict of interest statement

All authors claim no conflict of interest.

Figures

Fig. 1
Fig. 1
The DeepCORE architecture. (A) DeepCORE consists of two components. In the DNN component, genetic and epigenetic signals go through a CNN layer, a BiLSTM layer, an Attention layer, and an FNN layer to predict transcript abundance of a gene. In the Interpreter component, attention scores extracted from the output of the Attention layer is analyzed to identify informative and correlated regions as COREs. (B) The DNN has a genetic view and an epigenetic view, each consisting of a CNN layer and a BiLSTM layer. These two views are joint before fed into an attention layer and subsequently an FNN layer to predict gene transcription level. Nc = 10,000 in both genetic and epigenetic feature matrices.
Fig. 2
Fig. 2
Performance of DeepCORE and other methods. (A, B) Evaluated on two samples, E061 and E071, the boxplots of RMSE (A) and F1-score (B) show DeepCORE has the lowest error rate and the highest accuracy in predicting gene transcription levels, as compared to single-view DNN, ExPecto, and DeepChrome models. (C, D) Evaluated on 25 samples, DeepCORE has consistently the lowest error rate of predicting continuous gene transcription levels (C) and consistently the highest accuracy of predicting binary gene transcription classes (D). (E, F) Evaluated on cross-sample predictions in which a model trained the source sample is applied to predict gene transcription in different target samples, DeepCORE shows consistently lower error rate than ExPecto (E) and higher accuracy than DeepChrome (F). (b) Gray lines denote performance in source samples.
Fig. 3
Fig. 3
Distribution of hPTMs in attended regions. (A) Density plots show distribution of attended bins with hPTMs vs. distribution of attended bins with no hPTMs. (B) Counts of bins with specific hPTMs in attended regions. Data were from randomly selected 25 highly expressed genes and 25 low expression genes. Transcription activating hPTMs are in orange background. Transcription repressing hPTMs are in blue background. (C, D) Heatmaps show the raw hPTM read counts and DeepCORE attention probabilities for the CYFIP2 gene. Transcription of this gene was low in the E007 sample (C) and high in the E058 sample (D). The ± 5kbps TSS-flanking region is encoded into 200 bins each with an attention probability.
Fig. 4
Fig. 4
Attention analysis for regulatory elements (A) Density plot of the frequency of attended bins with known promoters or enhancers across 25 samples in comparison to random bins with high attention scores. (B) In the TMEM88 gene, attended bins matched to known enhancers and promoters. Signals from repressing hPTMs did not receive attention. (C-E) In the ARF5 gene, hPTM signals form two clusters (indicated with blue boxes). The right cluster mapped to the promoter of the ARF5 received attentions. The left cluster mapped to the promoter of another gene GCC1 did not receive attention.
Fig. 5
Fig. 5
COREs in the PSMD8 gene: (A) Heatmaps show attentions in 5 cell lines. (B) Correlation plot shows two blocks (A and B) with significant correlated attentions.
Fig. 6
Fig. 6
CORE in the TMUB1 Gene. The ± 5 kb TSS-flanking region is displayed, which has cis-regulatory roles as annotated in the EPD and GeneHancer databases. DeepCORE identified three blocks (A, B, and C) as putative REs, which received attentions in multiple samples. The correlation matrix of attentions revealed local interaction between B and C, and distal interaction between A and the other two elements. The distal interaction is confirmed in an Hi-TrAC study showing these REs are inside chromatin-chromatin loops.
See this image and copyright information in PMC

Update of

Similar articles

See all similar articles

Cited by

References

    1. Blackwood E.M., Kadonaga J.T. Going the distance: a current view of enhancer action. Science. 1998;281:60–63. - PubMed
    1. Krivega I., Dean A. Enhancer and promoter interactions—long distance calls. Curr Opin Genet Dev. 2012;22:79–85. - PMC - PubMed
    1. Appella E., Anderson C.W. Post-translational modifications and activation of p53 by genotoxic stresses. Eur J Biochem. 2001;268:2764–2772. - PubMed
    1. Carey M., Peterson C.L., Smale S.T. Cold Spring Harbor Laboratory Press,; 2009. Transcriptional regulation in eukaryotes.
    1. Gilad Y., Rifkin S.A., Pritchard J.K. Revealing the architecture of gene regulation: the promise of eQTL studies. Trends Genet. 2008;24:408–415. - PMC - PubMed

LinkOut - more resources