Accurate design of translational output by a neural network model of ribosome distribution

Abstract

Synonymous codon choice can have dramatic effects on ribosome speed and protein expression. Ribosome profiling experiments have underscored that ribosomes do not move uniformly along mRNAs. Here, we have modeled this variation in translation elongation by using a feed-forward neural network to predict the ribosome density at each codon as a function of its sequence neighborhood. Our approach revealed sequence features affecting translation elongation and characterized large technical biases in ribosome profiling. We applied our model to design synonymous variants of a fluorescent protein spanning the range of translation speeds predicted with our model. Levels of the fluorescent protein in budding yeast closely tracked the predicted translation speeds across their full range. We therefore demonstrate that our model captures information determining translation dynamics in vivo; that this information can be harnessed to design coding sequences; and that control of translation elongation alone is sufficient to produce large quantitative differences in protein output.

Figure 1. Design and performance of a neural network model of translation elongation

a, Each ribosome protects an mRNA footprint of approximately 28–29 nt. Sequence coordinates in a neighborhood around a ribosome are indexed relative to the codon in the A site of the ribosome. b, Read count rescaling. For each gene, the counts of footprints assigned to each A site codon are divided by the average counts per codon over that gene. The resulting scaled footprint counts are used for model training and prediction. c, Model performances (Pearson correlations between predicted and true scaled counts over the test set) for neural network and linear regression models over a range of sequence neighborhoods, with and without nucleotide features, as well as correlations for models that also incorporate structure scores of the three 30-nt windows overlapping the footprint region, or the maximum structure score within 59 nt downstream of the ribosome. Bars show the mean of 10 runs of each model; the 10 individual runs for each model are overlaid as gray points. d, True vs. predicted scaled counts for the test set, under a model with codon and nucleotide features spanning codon positions −5 to +4. Color scale shows density of data points. e, True scaled counts (gray bars) and predicted scaled counts (red line) for a highly translated gene.

Figure 2. Performance comparisons on low coverage genes and with competing models

a, Top, per-gene correlations between true and predicted scaled counts, for all 4375 genes in our transcriptome that passed filtering criteria. Training set genes in blue (333/top 500 genes by footprint density). Loess curve on test set genes shown in red. Below, as above, with footprint counts on the top 1000, 2000, 3000, and 4000 genes subsampled to the density of footprint counts on the 1000th, 2000th, 3000th, and 4000th gene, respectively, and ‘true’ scaled counts recomputed. b, Comparison of Iχnos with similar models, RUST and riboshape. Shown are per-gene correlations between true and predicted scaled counts, on 1711 genes passing the filtering criteria from all three methods. Training set genes from Iχnos are excluded. Colored lines are loess curves, which are also compared in the bottom panel.

Figure 3. Interpretation of models of translation elongation rates

a, Predictive value of codon positions in a yeast ribosome profiling dataset. We computed Pearson correlations between true and predicted scaled counts on the test set, for a reference model including codon and nucleotide features from codon positions −7 to +5, and for a series of leave-one-out models, each excluding one codon position. Gray points show differences between Pearson’s r for 10 runs of each leave-one-out model and the mean r of 10 runs of the reference model. Bars represent the mean of these values. b, Mean contributions to scaled counts by codon identity and position. c, P site codon contributions grouped by the codon:anticodon base pair formed by the third nucleotide of each codon. Asterisks indicate p < 0.05 after Bonferroni correction, unpaired two-sided Mann-Whitney U test between each group and all other codons. I:C, p = 0.014. d, Predictive value of codon positions as in A, from a yeast ribosome profiling library we constructed using CircLigase II as described by McGlincy and Ingolia. e, f, Contributions from (e) codon position -5, at the 5′ ends of footprints, and (f) the A site, in human ribosome profiling data versus our yeast ribosome profiling data, both using CircLigaseII. Analysis was limited to 28-nt footprints to avoid frame biases. g, Ligation efficiency of CircLigase II. Oligonucleotide substrates resembling ribosome footprints at the circularization step of the protocol, with different three-nucleotide end sequences, were ligated by both enzymes. Circularization was assayed by qPCR using primers spanning the ligation as compared to primers in a contiguous region of the oligo. Ligation was calculated relative to CircLigase I ligation of the best-ligated substrate. Each point represents the ratio of the means of three qPCR replicates; error bars represent the standard error of that ratio.

Figure 4. Design of synonymous sequences shows elongation rate affects translation output

a, Six reporter constructs with distinct synonymous eCitrine coding sequences were inserted into the his3Δ1 locus of BY4742 yeast, and an equivalent construct with a constant mCherry coding sequence was inserted into the his3Δ1 locus of BY4741 yeast. The haploids were mated to produce diploid yeast with both reporters, whose fluorescence was then measured with flow cytometry. b, The synonymous eCitrine sequences included the fastest and slowest predicted sequences under our model (magenta and red), plus sequences with predicted translation elongation times at the 0th, 33rd, 67th, and 100th percentiles of a randomly generated set of 100,000 synonymous eCitrine sequences (blue, green, yellow, and orange, respectively). The score distribution of 100,000 random eCitrine sequences is shown in lavender. The scores of endogenous yeast genes, rescaled by length to compare with eCitrine, are shown in gray. c, eCitrine:mCherry fluorescence ratio, as measured by flow cytometry of 11,000–18,000 yeast, versus the predicted elongation time of each sequence. Each + symbol represents the median ratio of yellow and red fluorescence from one biological replicate of the given eCitrine strain. Eight biological replicates, each an independent integration of the reporter construct, are included for each strain, except for the strains shown in blue and orange, which have seven, and the strain shown in green, which has three. Colors as in (b). d, Translation efficiency, or median eCitrine:mCherry fluorescence ratio divided by relative eCitrine:mCherry mRNA ratio (ratio of medians of three qPCR replicates), for each eCitrine variant, versus the predicted elongation time of each sequence. Purple, yECitrine sequence; other colors as in (b). Each point represents one biological replicate of the given eCitrine strain; three biological replicates were measured for each strain except two for the strain shown in red.