Deep learning of the regulatory grammar of yeast 5′ untranslated regions from 500,000 random sequences
Problem
Predicting protein expression from DNA sequence alone remains poor due to limited understanding of cis-regulatory grammar, hampering design of engineered genes for synthetic biology.
Why it matters
Sequence-to-expression prediction is needed for designing engineered genes in synthetic biology.
What the AI did
A convolutional neural network trained on 500,000 random 50-nucleotide yeast 5' UTRs from a massively parallel growth selection predicts translational efficiency, including on native S. cerevisiae UTRs.
Results
Model performs well on held-out random and native yeast 5' UTRs; no numeric metrics stated.
Impact
Quantified effects of Kozak sequence, uORFs and secondary structure on translation and enabled computational evolution of highly translated UTRs (abstract truncated).
Record
| Venue | Genome Research |
|---|---|
| Published | 2017 |
| Work type | preprint |
| Authors | Josh T. Cuperus, Benjamin Groves, Anna Kuchina, Alexander B. Rosenberg, Nebojša Jojić, Stanley Fields, Georg Seelig |
| Citations | 214 (OpenAlex, at import time) |
| References | 82 |
| Model or method | Convolutional neural network (CNN) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1101/gr.224964.117 |
| Open access | green |
| Keywords | CNN, 5' UTR, translational efficiency, yeast, synthetic biology, regulatory grammar |
Importance rating
real but narrower advance
Sequence-to-expression prediction enables engineered gene design in synthetic biology; a genuinely useful capability but abstract frames it as a design tool, not a large-scale impact.