Large-scale DNA-based phenotypic recording and deep learning enable highly accurate sequence-function mapping

Nat Commun. 2020 Jul 15;11(1):3551. doi: 10.1038/s41467-020-17222-4.

Abstract

Predicting effects of gene regulatory elements (GREs) is a longstanding challenge in biology. Machine learning may address this, but requires large datasets linking GREs to their quantitative function. However, experimental methods to generate such datasets are either application-specific or technically complex and error-prone. Here, we introduce DNA-based phenotypic recording as a widely applicable, practicable approach to generate large-scale sequence-function datasets. We use a site-specific recombinase to directly record a GRE's effect in DNA, enabling readout of both sequence and quantitative function for extremely large GRE-sets via next-generation sequencing. We record translation kinetics of over 300,000 bacterial ribosome binding sites (RBSs) in >2.7 million sequence-function pairs in a single experiment. Further, we introduce a deep learning approach employing ensembling and uncertainty modelling that predicts RBS function with high accuracy, outperforming state-of-the-art methods. DNA-based phenotypic recording combined with deep learning represents a major advance in our ability to predict function from genetic sequence.

Publication types

  • Research Support, Non-U.S. Gov't

MeSH terms

  • Binding Sites / genetics
  • Computational Biology / methods*
  • Datasets as Topic
  • Deep Learning*
  • Escherichia coli / genetics
  • Gene Knockout Techniques
  • Genome, Bacterial / genetics
  • High-Throughput Nucleotide Sequencing
  • Molecular Sequence Annotation / methods*
  • Phenotype*
  • Regulatory Sequences, Nucleic Acid / genetics
  • Ribosomes / metabolism
  • Sequence Analysis, DNA / methods*