Deep-learning models can predict a gene's mRNA abundance directly from its DNA sequence, using information distributed across the entire gene regulatory structure — promoter, UTRs, terminator, and coding-region codon usage — with combinatorial co-occurrence of sequence motifs explaining most of the expression dynamic range. These findings, however, are correlational. This problem asks for the causal molecular account: which mechanisms make coding-region codon usage a determinant of mRNA abundance in organisms beyond human cells, established through targeted experimental perturbations rather than prediction alone.

Background

How much of a gene's expression level is written into its own DNA sequence? Zrimec et al. (Nature Communications, 2020) trained deep convolutional neural networks — machine-learning models that learn informative sequence features directly from one-hot-encoded DNA — on over 20,000 RNA-seq datasets (sequencing-based measurements of mRNA abundance, reported for instance in transcripts per million, TPM) across seven organisms from bacteria to human. In baker's yeast Saccharomyces cerevisiae the model predicted median mRNA levels from the gene regulatory structure — the promoter, 5' and 3' untranslated regions (UTRs), terminator, and the coding region represented by codon frequencies — with up to 82% of variance explained. Here 'codon usage' means the frequencies of synonymous codons (triplets encoding the same amino acid): genes differ systematically in which synonymous triplets they use. Two findings are central. First, codon usage carried roughly as much information about mRNA level as the flanking regulatory regions. Second, when the authors probed which sequence motifs (short recurring DNA words, typically transcription-factor binding sites) the network used, single motifs could not separate high from low expression, but co-occurrence rules — specific combinations of 2 to 6 motifs located across the gene — recovered about 84% of the expression dynamic range. The authors concluded that the gene regulatory structure is a co-evolving unit. Crucially, these results are correlational: the paper performs no perturbation experiments on codon usage or motif combinations, and lists three mechanisms by which codons might affect transcription only as hypotheses — influence on nucleosome positioning (the packaging of DNA into nucleosomes affects transcriptional elongation), premature transcription termination by polyadenylation-signal mimicry (codon-derived sequence tracts resembling the poly-A signal that normally ends a transcript can trigger early termination), and mRNA toxicity (accumulation of the transcript to levels that are themselves deleterious to the cell) — building on earlier work in Neurospora showing codon usage affects expression largely at the transcription level (Zhou et al., PNAS, 2016) and that codon usage co-evolves with termination machinery to suppress premature cleavage and polyadenylation (Zhou et al., eLife, 2018). The situation has since changed in two ways. On the modeling side, sequence-to-expression predictors now incorporate tissue specificity and long-range context: ExPecto (Zhou et al., Nature Genetics, 2018) predicted tissue-specific expression ab initio; Enformer (Avsec et al., Nature Methods, 2021) enlarged the input window to about 200 kb to capture distal enhancers (regulatory elements far from the gene they control); yet Karollus et al. (Genome Biology, 2023) showed such models still mostly ignore distal enhancers, and EPInformer (Lin et al., Nature Communications, 2026) still reports distal-enhancer effects as an open difficulty, as synthesized in the review of Barbadilla-Martínez et al. (Nature Reviews Genetics, 2025). On the mechanistic side, a genome-wide perturbation screen in human cells (Garg et al., Nature Communications, 2025) identified factors mediating codon-usage effects there: the CCR4-NOT complex subunit CNOT4 acts on nuclear RNA levels via a transcriptional effect (CCR4-NOT is a conserved multi-subunit complex best known for shortening the poly-A tails of mRNAs and repressing gene expression at several steps), while the nuclear RNA exosome (the cell's main molecular machine for degrading RNA in the nucleus) and the PAXT complex (a protein assembly that marks certain nuclear RNAs for exosome-mediated destruction) act through nuclear mRNA surveillance. That resolves much of the codon-usage mechanism question for human cells, but not for the source study's model organism yeast or the other organisms of the source study, where the transcription-level mechanisms hypothesized by Zrimec et al. remain untested by targeted perturbation.

Problem Statement

Provide an experimentally validated causal account of the codon-usage component of the sequence-encoded regulatory grammar of mRNA abundance established correlationally by Zrimec et al. (2020). Concretely: determine, by targeted perturbations (for example, synonymous codon replacements in native genes and constructs, with matched measurements of nascent (newly synthesized, still being transcribed) and mature RNA distinguishing transcriptional output from nuclear and cytoplasmic mRNA decay), the molecular mechanisms by which coding-region codon usage determines mRNA abundance in organisms other than human cells — with Saccharomyces cerevisiae, the source study's model organism, as the primary target — testing among the mechanisms that Zrimec et al. (2020) explicitly list as untested hypotheses: effects on nucleosome positioning, premature termination of transcription by polyadenylation-signal mimicry, and mRNA toxicity. The objective is to convert the deep-learning demonstration that the gene regulatory structure, including coding-region codon usage, encodes mRNA abundance into a mechanistic, cause-and-effect explanation validated by experiment. Coverage is fixed at the source study's model organism: an account that experimentally resolves all three of the listed mechanisms in Saccharomyces cerevisiae satisfies the task, while extension to the source study's other non-human organisms is desirable but not required. Because these perturbations have not yet been performed in that organism, the accepted deliverable is the mechanism account supported by the existing experimental literature together with a decisive, fully specified perturbation experiment — synonymous codon replacement with matched nascent- and mature-RNA measurements — for each mechanism left untested.

References

  1. Jan Zrimec, Christoph S. Börlin, Filip Buric, Azam Sheikh Muhammad, Rhongzen Chen, Verena Siewers, Vilhelm Verendel, Jens Nielsen, Mats Töpel, Aleksej Zelezniak (2020), 'Deep learning suggests that gene expression is encoded in all parts of a co-evolving interacting gene regulatory structure', Nature Communications 11:6141, DOI 10.1038/s41467-020-19921-4, https://doi.org/10.1038/s41467-020-19921-4
  2. Zhipeng Zhou, Yunkun Dang, Mian Zhou, Lin Li, Chien-hung Yu, Jingjing Fu, She Chen, Yi Liu (2016), 'Codon usage is an important determinant of gene expression levels largely through its effects on transcription', Proceedings of the National Academy of Sciences 113(41), DOI 10.1073/pnas.1606724113, https://doi.org/10.1073/pnas.1606724113
  3. Zhipeng Zhou, Yunkun Dang, Mian Zhou, Haiyan Yuan, Yi Liu (2018), 'Codon usage biases co-evolve with transcription termination machinery to suppress premature cleavage and polyadenylation', eLife 7:e33569, DOI 10.7554/eLife.33569, https://doi.org/10.7554/eLife.33569
  4. Renu Garg, Pancheng Xie, Jiabin Duan, Huan Liu, Yi Liu (2025), 'Nuclear effects play an important role in determining codon usage-dependent human gene expression', Nature Communications 16:10865, DOI 10.1038/s41467-025-65907-5, https://doi.org/10.1038/s41467-025-65907-5
  5. Robert Buschauer, Yoshitaka Matsuo, Takato Sugiyama, Ying-Hsin Chen, Najwa Alhusaini, Thomas Sweet, Ken Ikeuchi, Jingdong Cheng, Yasuko Matsuki, Risa Nobuta, Andrea Gilmozzi, Otto Berninghausen, Petr Tesina, Thomas Becker, Jeff Coller, Toshifumi Inada, Roland Beckmann (2020), 'The Ccr4-Not complex monitors the translating ribosome for codon optimality', Science 368(6488), DOI 10.1126/science.aay6912, https://doi.org/10.1126/science.aay6912
  6. Jian Zhou, Chandra L. Theesfeld, Kevin Yao, Kathleen M. Chen, Aaron K. Wong, Olga G. Troyanskaya (2018), 'Deep learning sequence-based ab initio prediction of variant effects on expression and disease risk', Nature Genetics 50:1171-1179, DOI 10.1038/s41588-018-0160-6, https://doi.org/10.1038/s41588-018-0160-6
  7. Žiga Avsec, Vikram Agarwal, Daniel Visentin, Joseph R. Ledsam, Agnieszka Grabska-Barwinska, Kyle R. Taylor, Yannis Assael, John Jumper, Pushmeet Kohli, David R. Kelley (2021), 'Effective gene expression prediction from sequence by integrating long-range interactions', Nature Methods 18:1196-1203, DOI 10.1038/s41592-021-01252-x, https://doi.org/10.1038/s41592-021-01252-x
  8. Alexander Karollus, Thomas Mauermeier, Julien Gagneur (2023), 'Current sequence-based models capture gene expression determinants in promoters but mostly ignore distal enhancers', Genome Biology 24:56, DOI 10.1186/s13059-023-02899-9, https://doi.org/10.1186/s13059-023-02899-9
  9. Lucía Barbadilla-Martínez, Noud Klaassen, Bas van Steensel, Jeroen de Ridder (2025), 'Predicting gene expression from DNA sequence using deep learning models', Nature Reviews Genetics 26:666-680, DOI 10.1038/s41576-025-00841-2, https://doi.org/10.1038/s41576-025-00841-2
  10. Jiecong Lin, Zhijian Li, Yajie Zhao, Ruibang Luo, Luca Pinello (2026), 'EPInformer: scalable and integrative prediction of gene expression from promoter-enhancer sequences with multimodal epigenomic profiles', Nature Communications 17:3975, DOI 10.1038/s41467-026-70535-8, https://doi.org/10.1038/s41467-026-70535-8