ICOR: Improving codon optimization with recurrent neural networks

Abstract

In protein sequences—as there are 61 sense codons but only 20 standard amino acids—most amino acids are encoded by more than one codon. Although such synonymous codons do not alter the encoded amino acid sequence, their selection can dramatically affect the expression of the resulting protein. Codon optimization of synthetic DNA sequences is important for heterologous expression. However, existing solutions are primarily based on choosing high-frequency codons only, neglecting the important effects of rare codons. In this paper, we propose a novel recurrent-neural-network based codon optimization tool, ICOR, that aims to learn codon usage bias on a genomic dataset of Escherichia coli. We compile a dataset of over 7,000 non-redundant, high-expression, robust genes which are used for deep learning. The model uses a bidirectional long short-term memory-based architecture, allowing for the sequential context of codon usage in genes to be learned. Our tool can predict synonymous codons for synthetic genes toward optimal expression in Escherichia coli. We demonstrate that sequential context achieved via RNN may yield codon selection that is more similar to the host genome, therefore improving protein expression more than frequency-based approaches. ICOR is evaluated on 1,481 Escherichia coli genes as well as a benchmark set of 40 select DNA sequences whose heterologous expression has been previously characterized. ICOR’s performance across five metrics is compared to that of five different codon optimization techniques. The codon adaptation index -- a metric indicative of high real-world expression -- was utilized as the primary benchmark in this study. ICOR is shown to improve the codon adaptation index by 41.69% and 17.25% compared to the original and Genscript’s GenSmart-optimized sequences, respectively. Our tool is provided as an open-source software package that includes the benchmark set of sequences used in this study.

Deep learning-based codon optimization with large-scale synonymous variant datasets enables generalized tunable protein expression

“…One common approach for codon optimization is a text translation task, where a sentence written in the language of protein amino acids is translated to DNA nucleotides, conditioned on the identity of the host organism to maximize the natural profile of the CDS [39][40][41]. To this end, we repurposed the T5 language model architecture [43] by training on a dataset, named Protein2DNA, consisting of all protein coding and protein-CDS pairs from high-quality genomes of the Enterobacterales order (taxonomic identifier 91347) (Materials and methods).…”

“…Natural language models based on Deep Learning (DL) have emerged as powerful tools for interrogating complex context-dependent patterns in biological sequences across application domains [35][36][37][38]. Although there are recent examples of DL-enabled codon optimization [39][40][41], these studies do not incorporate expression level information. Here we show that language models are able to learn long-range patterns of codon usage and generate sequences mimicking natural codon usage profiles when trained on genome-scale CDS data.…”

Applying the Deep Learning Techniques to Solve Classification Tasks Using Gene Expression Data

“…As was noted hereinbefore, the RNN is widely recognized for its effectiveness in processing sequential data like text, time series, and speech. In [12], the use of LSTM-RNN architecture was explored for optimizing codon usage in protein sequences. The novel tool was developed to learn codon usage bias from a genomic dataset comprising over 7,000 Escherichia coli genes.…”

Conditional Recurrent Neural Networks for broad applications in nonlinear optics

“…Using recurrent neural networks can drastically reduce the simulation time of numerical solutions. RNNs are ideal for learning sequential data, due to the feedback loops, or connections within their layers [21,26], that provide an internal memory of past observations. In particular, RNNs with long short-term memory (LSTM) [27] layers can deal with longer sequences, where the sequential dependence of the data spans over a large number of observations, since these types of network avoid vanishing gradients problems [28].…”

“…In this work, we develop a fully data-driven approach using conditional recurrent neural networks (RNN) [21] to predict the outcomes of nonlinear pulse propagation for a range of input-pulse parameters in different waveguide structures that simultaneously exhibit secondand third-order nonlinearities. The developed approach also allows for the training of a unified network to compute both the spectral and temporal evolution of a pulse via assigning the real and imaginary parts of the pulse complex-envelope as the network condition.…”