CoChat
Sign Up
Cited by 176Genomics and Chromatin DynamicsGene Regulatory Network AnalysisRNA and protein synthesis mechanisms

The DOI matches its Crossref record; 158 of the 158 references it cites were found in a registry or an index.

Deep learning suggests that gene expression is encoded in all parts of a co-evolving interacting gene regulatory structure

Open access
Zrimec, Jan, Börlin, Christoph S., Buric, Filip, Muhammad, Azam Sheikh, Chen, Rhongzen, Siewers, Verena, Verendel, Vilhelm, Nielsen, Jens, Töpel, Mats, Zelezniak, Aleksej·2020·Nature Communications
DOI 10.1038/s41467-020-19921-4PMC PMC7708451

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Abstract Understanding the genetic regulatory code governing gene expression is an important challenge in molecular biology. However, how individual coding and non-coding regions of the gene regulatory structure interact and contribute to mRNA expression levels remains unclear. Here we apply deep learning on over 20,000 mRNA datasets to examine the genetic regulatory code controlling mRNA abundance in 7 model organisms ranging from bacteria to Human. In all organisms, we can predict mRNA abundance directly from DNA sequence, with up to 82% of the variation of transcript levels encoded in the gene regulatory structure. By searching for DNA regulatory motifs across the gene regulatory structure, we discover that motif interactions could explain the whole dynamic range of mRNA levels. Co-evolution across coding and non-coding regions suggests that it is not single motifs or regions, but the entire gene regulatory structure and specific combination of regulatory elements that define gene expression levels.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41467-020-19921-4

Cites

158 of 158 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 The dynamic range of gene expression levels is encoded in the DNA sequence. All results shown are for S. cerevisiae, except in (h). a Expression levels (transcripts per million, TPM) of protein-coding genes across 3025 RNA-seq experiments. Inset: distribution of genes with a relative standard deviation (RSD = σ/μ) below 1. b Experimental variation of the gene expression levels expressed as RSD. c Distribution of ratios between the variance of expression levels within a genome and the variance of expression levels per gene across the experiments. d Schematic diagram of the explanatory variables used for modeling, with distributions of the sequence lengths in the regions where the lengths varied. “TSS” denotes the transcription start site and “TTS” the transcription termination site. e The optimized deep neural network (NN) architecture, where “conv” denotes convolutional NNs, “FC” fully connected layers, and “max-pool” max-pooling layers. The values denoting sizes of parameters layers, stride, kernels, filters, max-pooling, and dilation are specified. f Experimentally determined (true) versus predicted expression levels with the S. cerevisiae model on the held out test dataset (n = 425). Red line denotes least squares fit. g Comparison of published experimental fluorescence measurements44 with the predicted expression levels (n = 625). Red line denotes least squares fit. h R2 on held out test datasets across 7 model organisms. Red line denotes mean value. Source data are provided as a Source data file.
Fig. 1 The dynamic range of gene expression levels is encoded in the DNA sequence. All results shown are for S. cerevisiae, except in (h). a Expression levels (transcripts per million, TPM) of protein-coding genes across 3025 RNA-seq experiments. Inset: distribution of genes with a relative standard deviation (RSD = σ/μ) below 1. b Experimental variation of the gene expression levels expressed as RSD. c Distribution of ratios between the variance of expression levels within a genome and the variance of expression levels per gene across the experiments. d Schematic diagram of the explanatory variables used for modeling, with distributions of the sequence lengths in the regions where the lengths varied. “TSS” denotes the transcription start site and “TTS” the transcription termination site. e The optimized deep neural network (NN) architecture, where “conv” denotes convolutional NNs, “FC” fully connected layers, and “max-pool” max-pooling layers. The values denoting sizes of parameters layers, stride, kernels, filters, max-pooling, and dilation are specified. f Experimentally determined (true) versus predicted expression levels with the S. cerevisiae model on the held out test dataset (n = 425). Red line denotes least squares fit. g Comparison of published experimental fluorescence measurements44 with the predicted expression levels (n = 625). Red line denotes least squares fit. h R2 on held out test datasets across 7 model organisms. Red line denotes mean value. Source data are provided as a Source data file.CC BY · PubMed Central
Fig. 2 Coding and cis-regulatory regions jointly contribute to gene expression prediction. a Test R2 with different combinations of coding (codon prob.), mRNA stability (see Supplementary Fig. 1c) and cis-regulatory regions, using shallow (black) and deep modeling (red). b Test R2 with deep models trained on different combinations of cis-regulatory regions. Red lines denote mean values. c Schematic depiction of deep neural network (NN) for prediction of codon probabilities (Y) based on regulatory DNA as input (X). d Actual versus predicted distribution of codon probabilities as predicted from non-coding DNA regions based on the model for prediction of codon probabilities based on regulatory DNA as input (n = 11,947). Red line denotes least squares fit. e Evolutionary substitution rates in promoter vs. coding regions in orthologous genes of 14 yeast species (n = 3248). Red line denotes least squares fit. f Evolutionary substitution rates in terminator vs. coding regions in orthologous genes of 14 yeast species (n = 3248). Red line denotes least squares fit. Source data are provided as a Source data file.
Fig. 2 Coding and cis-regulatory regions jointly contribute to gene expression prediction. a Test R2 with different combinations of coding (codon prob.), mRNA stability (see Supplementary Fig. 1c) and cis-regulatory regions, using shallow (black) and deep modeling (red). b Test R2 with deep models trained on different combinations of cis-regulatory regions. Red lines denote mean values. c Schematic depiction of deep neural network (NN) for prediction of codon probabilities (Y) based on regulatory DNA as input (X). d Actual versus predicted distribution of codon probabilities as predicted from non-coding DNA regions based on the model for prediction of codon probabilities based on regulatory DNA as input (n = 11,947). Red line denotes least squares fit. e Evolutionary substitution rates in promoter vs. coding regions in orthologous genes of 14 yeast species (n = 3248). Red line denotes least squares fit. f Evolutionary substitution rates in terminator vs. coding regions in orthologous genes of 14 yeast species (n = 3248). Red line denotes least squares fit. Source data are provided as a Source data file.CC BY · PubMed Central
Fig. 3 Deep learning identifies the DNA positions predictive of expression levels. a Relevance profiles across cis-regulatory region sequences obtained by querying the deep models (Fig. 1f). Red lines denote cutoff of 2 standard deviations. The decrease in relevance of UTR sequences at left edges was due to the considerable number of sequences that were shorter than the analysed regions (see Fig. 1d). b Clustered relevance profiles across the cis-regulatory region sequences. The clusters 1 through 4 are colored dark red, light red, light blue and dark blue, respectively. Lines and shaded regions represent means and standard deviations, respectively. c GC content in the cis-regulatory regions (n = 368, 383, 834 and 625, respectively). d Median expression levels of genes in the clustered relevance profiles (n = 1385, 1187, 1206 and 460 with increasing cluster size, respectively). e Total amount of regulatory DNA motifs uncovered in the cis-regulatory regions, with the proportional amounts of motif overlap between adjacent regions (black) as well as with the Jaspar database68 (red) highlighted. f Examples of regulatory DNA motifs uncovered across all the cis-regulatory regions that correspond to published motifs and sequence elements (see Supplementary Fig. 1c). “TFBS” denotes transcription factor binding sites. For box plots in c, d, boxes denote interquartile (IQR) ranges, centres mark medians and whiskers extend to 1.5 IQR from the quartiles. Source data are provided as a Source data file.
Fig. 3 Deep learning identifies the DNA positions predictive of expression levels. a Relevance profiles across cis-regulatory region sequences obtained by querying the deep models (Fig. 1f). Red lines denote cutoff of 2 standard deviations. The decrease in relevance of UTR sequences at left edges was due to the considerable number of sequences that were shorter than the analysed regions (see Fig. 1d). b Clustered relevance profiles across the cis-regulatory region sequences. The clusters 1 through 4 are colored dark red, light red, light blue and dark blue, respectively. Lines and shaded regions represent means and standard deviations, respectively. c GC content in the cis-regulatory regions (n = 368, 383, 834 and 625, respectively). d Median expression levels of genes in the clustered relevance profiles (n = 1385, 1187, 1206 and 460 with increasing cluster size, respectively). e Total amount of regulatory DNA motifs uncovered in the cis-regulatory regions, with the proportional amounts of motif overlap between adjacent regions (black) as well as with the Jaspar database68 (red) highlighted. f Examples of regulatory DNA motifs uncovered across all the cis-regulatory regions that correspond to published motifs and sequence elements (see Supplementary Fig. 1c). “TFBS” denotes transcription factor binding sites. For box plots in c, d, boxes denote interquartile (IQR) ranges, centres mark medians and whiskers extend to 1.5 IQR from the quartiles. Source data are provided as a Source data file.CC BY · PubMed Central
Fig. 4 Motif co-occurrence uncovers the regulatory rules of gene expression. a Distribution of the signal-to-noise ratio (SNR) of expression levels across genes carrying co-occurring motifs (co-occurrence rules, see Results text and Methods). Inset: comparison of SNR with single motifs (n = 2098) and with motif co-occurrence rules (n = 116,734). b The range of gene expression levels, relative to the full observed range in the initial RNAseq data (~4 orders of magnitude, Fig. 1a), that could be retrieved on average by genes carrying either single motifs or motif co-occurrence rules. c The amount of genes carrying a given motif co-occurrence rule versus the average expression level across the set of genes carrying the given rule, with increasing statistical significance levels from a chi-squared test76. FDR denotes Benjamini-Hochberg (BH) adjusted p-value and data is colored gray except for significance cutoffs p < 0.05 (black), FDR < 0.05 (dark red) and FDR < 2.7e-6 (equals Bonferroni correction, red). Inset: distribution of the number of co-occurring motifs in significant (FDR < 0.05) rules. d Distribution of motif co-occurrence rules across single or multiple cis-regulatory regions, according to the locations of the co-occurring motifs. e Distribution of gene expression levels with groups of motif co-occurrence rules that have one motif in common (i.e. unchanged). f Illustration of four genes (CDC6, RIO2, NSP1, EXG1) that carry a group of motif co-occurrence rules with a common motif (NHP6B transcription factor binding site, Tomtom67 BH adj. p-value < 0.005, SGD:S000002157104, blue line) in their promoter region, whereas they all diverge in possessing 2 to 4 other DNA motifs (red lines) across the remaining regulatory regions. These genes span a 648-fold range of expression levels. Red lines in the histogram denote the specific expression levels of the genes. g Distribution of gene expression levels with groups of motif co-occurrence rules that differ by a single motif. Source data are provided as a Source data file.
Fig. 4 Motif co-occurrence uncovers the regulatory rules of gene expression. a Distribution of the signal-to-noise ratio (SNR) of expression levels across genes carrying co-occurring motifs (co-occurrence rules, see Results text and Methods). Inset: comparison of SNR with single motifs (n = 2098) and with motif co-occurrence rules (n = 116,734). b The range of gene expression levels, relative to the full observed range in the initial RNAseq data (~4 orders of magnitude, Fig. 1a), that could be retrieved on average by genes carrying either single motifs or motif co-occurrence rules. c The amount of genes carrying a given motif co-occurrence rule versus the average expression level across the set of genes carrying the given rule, with increasing statistical significance levels from a chi-squared test76. FDR denotes Benjamini-Hochberg (BH) adjusted p-value and data is colored gray except for significance cutoffs p < 0.05 (black), FDR < 0.05 (dark red) and FDR < 2.7e-6 (equals Bonferroni correction, red). Inset: distribution of the number of co-occurring motifs in significant (FDR < 0.05) rules. d Distribution of motif co-occurrence rules across single or multiple cis-regulatory regions, according to the locations of the co-occurring motifs. e Distribution of gene expression levels with groups of motif co-occurrence rules that have one motif in common (i.e. unchanged). f Illustration of four genes (CDC6, RIO2, NSP1, EXG1) that carry a group of motif co-occurrence rules with a common motif (NHP6B transcription factor binding site, Tomtom67 BH adj. p-value < 0.005, SGD:S000002157104, blue line) in their promoter region, whereas they all diverge in possessing 2 to 4 other DNA motifs (red lines) across the remaining regulatory regions. These genes span a 648-fold range of expression levels. Red lines in the histogram denote the specific expression levels of the genes. g Distribution of gene expression levels with groups of motif co-occurrence rules that differ by a single motif. Source data are provided as a Source data file.CC BY · PubMed Central
Fig. 5 Each gene is predicted to contain a large degree of regulatory freedom for expression engineering. a Two global halves of the gene regulatory structure where 5′-UTRs were combined with promoters and 3′-UTRs with terminators. b Maximum increases (red) and decreases (blue) of gene expression levels with combinations of 4238 native promoters with all terminator variants. c Distributions of gene expression levels for combinations of either strong (red) or weak (blue) promoters with all terminator variants (n = 4238 for each construct). Black lines denote median exp. levels. d Comparison of changes in predicted TDH3 expression levels with measured fluorescence intensities when combining the promoter with n = 4005 terminator variants45. Red line denotes least squares fit. e Maximum increases (red) and decreases (blue) of gene expression levels with combinations of 4238 native terminators with all promoter variants. f Distributions of gene expression levels for combinations of either strong (red) or weak (blue) terminators with all promoter variants (n = 4238 for each construct). Black lines denote median expression levels. g Comparison of changes in predicted ADH1 expression levels with measured fluorescence intensities when combining the terminator with n = 625 promoter variants44. Red line denotes least squares fit. For box plots in c, f, boxes denote interquartile (IQR) ranges, centres mark medians and whiskers extend to 1.5 IQR from the quartiles. Source data are provided as a Source data file.
Fig. 5 Each gene is predicted to contain a large degree of regulatory freedom for expression engineering. a Two global halves of the gene regulatory structure where 5′-UTRs were combined with promoters and 3′-UTRs with terminators. b Maximum increases (red) and decreases (blue) of gene expression levels with combinations of 4238 native promoters with all terminator variants. c Distributions of gene expression levels for combinations of either strong (red) or weak (blue) promoters with all terminator variants (n = 4238 for each construct). Black lines denote median exp. levels. d Comparison of changes in predicted TDH3 expression levels with measured fluorescence intensities when combining the promoter with n = 4005 terminator variants45. Red line denotes least squares fit. e Maximum increases (red) and decreases (blue) of gene expression levels with combinations of 4238 native terminators with all promoter variants. f Distributions of gene expression levels for combinations of either strong (red) or weak (blue) terminators with all promoter variants (n = 4238 for each construct). Black lines denote median expression levels. g Comparison of changes in predicted ADH1 expression levels with measured fluorescence intensities when combining the terminator with n = 625 promoter variants44. Red line denotes least squares fit. For box plots in c, f, boxes denote interquartile (IQR) ranges, centres mark medians and whiskers extend to 1.5 IQR from the quartiles. Source data are provided as a Source data file.CC BY · PubMed Central
Fig. 6 Experimental analysis demonstrates the potential of using the models to fine-tune gene expression. a Conceptual diagram of experiment and constructs. b Correlation analysis between model predictions and measured GFP fluorescence levels for constructs with 6 promoters in combination with native (grey) as well as strong (red) and weak (blue) terminators (n = 18). Black line denotes least squares fit. c Relative change in measured GFP fluorescence levels of the constructs, where native terminators were replaced with either strong (red bars) or weak (blue bars) variants with n = 2 technical replicates indicated as grey points (see text and Supplementary Table 13). On average, the GFP fluorescence levels shifted by 20% in the directions of the predicted levels. Red and blue lines denote median values with strong and weak promoters, respectively. d Additional tests with the weak POP6 and strong RPL3 promoters, whose CDS were more dissimilar to GFP. Native terminators were replaced with either strong (red bars) or weak (blue bars) variants with n = 2 technical replicates indicated as grey points. Source data are provided as a Source data file.
Fig. 6 Experimental analysis demonstrates the potential of using the models to fine-tune gene expression. a Conceptual diagram of experiment and constructs. b Correlation analysis between model predictions and measured GFP fluorescence levels for constructs with 6 promoters in combination with native (grey) as well as strong (red) and weak (blue) terminators (n = 18). Black line denotes least squares fit. c Relative change in measured GFP fluorescence levels of the constructs, where native terminators were replaced with either strong (red bars) or weak (blue bars) variants with n = 2 technical replicates indicated as grey points (see text and Supplementary Table 13). On average, the GFP fluorescence levels shifted by 20% in the directions of the predicted levels. Red and blue lines denote median values with strong and weak promoters, respectively. d Additional tests with the weak POP6 and strong RPL3 promoters, whose CDS were more dissimilar to GFP. Native terminators were replaced with either strong (red bars) or weak (blue bars) variants with n = 2 technical replicates indicated as grey points. Source data are provided as a Source data file.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 154 references

Research works printed but not linked: 2

  1. Kuhn, M. & Johnson, K. Applied predictive modelling. (Springer, New York, Heidelberg, Dordrecht, London, 2013).
  2. Shalizi, C. Advanced data analysis from an elementary point of view. (Cambridge University Press, 2013).

Cited by

Sorted by citations

See all 176

Related

Metadata

Authors
Zrimec, J., Börlin, C. S., Buric, F., Muhammad, A. S., Chen, R., Siewers, V., Verendel, V., Nielsen, J., Töpel, M., & Zelezniak, A.
Affiliations
  • Chalmers University of Technology (SE)
  • University of Gothenburg (SE)
  • Science for Life Laboratory (SE)
Corresponding author
Zelezniak, Aleksej
ORCID
Venue
Nature Communications
Publisher
Springer Science and Business Media LLC
Published
2020-12-01
Volume
11
Issue
1
Pages
6141-6141
ISSN
2041-1723 (ISSN-L)
DOI
10.1038/s41467-020-19921-4
PMC
PMC7708451
OpenAlex
W3108101741
Cited by
174
References
154
Citations by year
2026: 19 · 2025: 35 · 2024: 42 · 2023: 28 · 2022: 27 · 2021: 21 · 2020: 2
FWCI
6.32
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • Chalmers Research (Chalmers University of Technology) (submittedVersion)
Funders
  • Nvidia
  • European Commission (722287)
  • Chalmers Tekniska Högskola
  • Science for Life Laboratory
  • Novo Nordisk Fonden (NNF10CC1016517)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Genomics and Chromatin Dynamics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Gene Regulatory Network Analysis
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › RNA and protein synthesis mechanisms
Keywords
gene expression regulation, mRNA levels, DNA sequences, regulatory motifs, coevolution, non-coding regions, coding sequences, deep learning, transcriptome, model organisms
MeSH terms
Deep Learning, Animals, Bacteria, Base Sequence, Drosophila melanogaster, Gene Expression Regulation, Humans, Regulatory Sequences, Nucleic Acid, RNA, Messenger, Saccharomyces cerevisiae, Evolution, Molecular, Mice

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review