Improving the prediction accuracy of protein abundance in Escherichia coli using mRNA accessibility

Transfer learning for cross-context prediction of protein expression from 5'UTR sequence

Model-guided DNA sequence design can accelerate the reprogramming of living cells. It allows us to engineer more complex biological systems by removing the need to physically assemble and test each potential design. While mechanistic models of gene expression have seen some success in supporting this goal, data-centric, deep learning-based approaches often provide more accurate predictions. This accuracy, however, comes at a cost - a lack of generalization across genetic and experimental contexts that has limited their wider use outside the context in which they were trained. Here, we address this issue by demonstrating how a simple transfer learning procedure can effectively tune a pre-trained deep learning model to predict protein translation rate from 5' untranslated region (5'UTR) sequence for diverse contexts in Escherichia coli using a small number of new measurements. This allows for important model features learnt from expensive massively parallel reporter assays to be easily transferred to new settings. By releasing our trained deep learning model and complementary calibration procedure, this study acts as a starting point for continually refined model-based sequence design that builds on previous knowledge and future experimental efforts.

Protein constraints in genome-scale metabolic models: Data integration, parameter estimation, and prediction of metabolic phenotypes

Genome-scale metabolic models provide a valuable resource to study metabolism and cell physiology. These models are employed with approaches from the constraint-based modeling framework to predict metabolic and physiological phenotypes. The prediction performance of genome-scale metabolic models can be improved by including protein constraints. The resulting protein-constrained models consider data on turnover numbers (kcat ) and facilitate the integration of protein abundances. In this systematic review, we present and discuss the current state-of-the-art regarding the estimation of kinetic parameters used in protein-constrained models. We also highlight how data-driven and constraint-based approaches can aid the estimation of turnover numbers and their usage in improving predictions of cellular phenotypes. Finally, we identify standing challenges in protein-constrained metabolic models and provide a perspective regarding future approaches to improve the predictive performance.

Accurate prediction of in vivo protein abundances by coupling constraint-based modelling and machine learning

Quantification of how different environmental cues affect protein allocation can provide important insights for understanding cell physiology. While absolute quantification of proteins can be obtained by resource-intensive mass-spectrometry-based technologies, prediction of protein abundances offers another way to obtain insights into protein allocation. Here we present CAMEL, a framework that couples constraint-based modelling with machine learning to predict protein abundance for any environmental condition. This is achieved by building machine learning models that leverage static features, derived from protein sequences, and condition-dependent features predicted from protein-constrained metabolic models. Our findings demonstrate that CAMEL results in excellent prediction of protein allocation in E. coli (average Pearson correlation of at least 0.9), and moderate performance in S. cerevisiae (average Pearson correlation of at least 0.5). Therefore, CAMEL outperformed contending approaches without using molecular read-outs from unseen conditions and provides a valuable tool for using protein allocation in biotechnological applications.

DeepRaccess: high-speed RNA accessibility prediction using deep learning

RNA accessibility is a useful RNA secondary structural feature for predicting RNA-RNA interactions and translation efficiency in prokaryotes. However, conventional accessibility calculation tools, such as Raccess, are computationally expensive and require considerable computational time to perform transcriptome-scale analysis. In this study, we developed DeepRaccess, which predicts RNA accessibility based on deep learning methods. DeepRaccess was trained to take artificial RNA sequences as input and to predict the accessibility of these sequences as calculated by Raccess. Simulation and empirical dataset analyses showed that the accessibility predicted by DeepRaccess was highly correlated with the accessibility calculated by Raccess. In addition, we confirmed that DeepRaccess could predict protein abundance in E.coli with moderate accuracy from the sequences around the start codon. We also demonstrated that DeepRaccess achieved tens to hundreds of times software speed-up in a GPU environment. The source codes and the trained models of DeepRaccess are freely available at https://github.com/hmdlab/DeepRaccess.

Ultradeep characterisation of translational sequence determinants refutes rare-codon hypothesis and unveils quadruplet base pairing of initiator tRNA and transcript

Translation is a key determinant of gene expression and an important biotechnological engineering target. In bacteria, 5'-untranslated region (5'-UTR) and coding sequence (CDS) are well-known mRNA parts controlling translation and thus cellular protein levels. However, the complex interaction of 5'-UTR and CDS has so far only been studied for few sequences leading to non-generalisable and partly contradictory conclusions. Herein, we systematically assess the dynamic translation from over 1.2 million 5'-UTR-CDS pairs in Escherichia coli to investigate their collective effect using a new method for ultradeep sequence-function mapping. This allows us to disentangle and precisely quantify effects of various sequence determinants of translation. We find that 5'-UTR and CDS individually account for 53% and 20% of variance in translation, respectively, and show conclusively that, contrary to a common hypothesis, tRNA abundance does not explain expression changes between CDSs with different synonymous codons. Moreover, the obtained large-scale data provide clear experimental evidence for a base-pairing interaction between initiator tRNA and mRNA beyond the anticodon-codon interaction, an effect that is often masked for individual sequences and therefore inaccessible to low-throughput approaches. Our study highlights the indispensability of ultradeep sequence-function mapping to accurately determine the contribution of parts and phenomena involved in gene regulation.

QRNAstruct: a method for extracting secondary structural features of RNA via regression with biological activity

Recent technological advances have enabled the generation of large amounts of data consisting of RNA sequences and their functional activity. Here, we propose a method for extracting secondary structure features that affect the functional activity of RNA from sequence–activity data. Given pairs of RNA sequences and their corresponding bioactivity values, our method calculates position-specific structural features of the input RNA sequences, considering every possible secondary structure of each RNA. A Ridge regression model is trained using the structural features as feature vectors and the bioactivity values as response variables. Optimized model parameters indicate how secondary structure features affect bioactivity. We used our method to extract intramolecular structural features of bacterial translation initiation sites and self-cleaving ribozymes, and the intermolecular features between rRNAs and Shine–Dalgarno sequences and between U1 RNAs and splicing sites. We not only identified known structural features but also revealed more detailed insights into structure–activity relationships than previously reported. Importantly, the datasets we analyzed here were obtained from different experimental systems and differed in size, sequence length and similarity, and number of RNA molecules involved, demonstrating that our method is applicable to various types of data consisting of RNA sequences and bioactivity values.

Protein Abundance Prediction Through Machine Learning Methods

Proteins are responsible for most physiological processes, and their abundance provides crucial information for systems biology research. However, absolute protein quantification, as determined by mass spectrometry, still has limitations in capturing the protein pool. Protein abundance is impacted by translation kinetics, which rely on features of codons. In this study, we evaluated the effect of codon usage bias of genes on protein abundance. Notably, we observed differences regarding codon usage patterns between genes coding for highly abundant proteins and genes coding for less abundant proteins. Analysis of synonymous codon usage and evolutionary selection showed a clear split between the two groups. Our machine learning models predicted protein abundances from codon usage metrics with remarkable accuracy, achieving strong correlation with experimental data. Upon integration of the predicted protein abundance in enzyme-constrained genome-scale metabolic models, the simulated phenotypes closely matched experimental data, which demonstrates that our predictive models are valuable tools for systems metabolic engineering approaches.

Analysis of 11,430 recombinant protein production experiments reveals that protein yield is tunable by synonymous codon changes of translation initiation sites

Recombinant protein production is a key process in generating proteins of interest in the pharmaceutical industry and biomedical research. However, about 50% of recombinant proteins fail to be expressed in a variety of host cells. Here we show that the accessibility of translation initiation sites modelled using the mRNA base-unpairing across the Boltzmann's ensemble significantly outperforms alternative features. This approach accurately predicts the successes or failures of expression experiments, which utilised Escherichia coli cells to express 11,430 recombinant proteins from over 189 diverse species. On this basis, we develop TIsigner that uses simulated annealing to modify up to the first nine codons of mRNAs with synonymous substitutions. We show that accessibility captures the key propensity beyond the target region (initiation sites in this case), as a modest number of synonymous changes is sufficient to tune the recombinant protein expression levels. We build a stochastic simulation model and show that higher accessibility leads to higher protein production and slower cell growth, supporting the idea of protein cost, where cell growth is constrained by protein circuits during overexpression.

From the notebook to recombinant protein production in Escherichia coli: Design of expression vectors and gene cloning

Research in recombinant protein expression in microorganism hosts spans half a century. The field has evolved from mostly trial-and-error approaches to more rational strategies, including careful design of the expression vectors and the coding sequence for the protein of interest. It is important to reflect on many aspects about vector construction, such as codon usage, integration site, coding sequence mutagenesis and many others. In this chapter, we overview methods and considerations to generate a suitable construct and anticipate possible experimental roadblocks.

TISIGNER.com: web services for improving recombinant protein production

Experiments that are planned using accurate prediction algorithms will mitigate failures in recombinant protein production. We have developed TISIGNER (https://tisigner.com) with the aim of addressing technical challenges to recombinant protein production. We offer three web services, TIsigner (Translation Initiation coding region designer), SoDoPE (Soluble Domain for Protein Expression) and Razor, which are specialised in synonymous optimisation of recombinant protein expression, solubility and signal peptide analysis, respectively. Importantly, TIsigner, SoDoPE and Razor are linked, which allows users to switch between the tools when optimising genes of interest.

Learning the Regulatory Code of Gene Expression

Data-driven machine learning is the method of choice for predicting molecular phenotypes from nucleotide sequence, modeling gene expression events including protein-DNA binding, chromatin states as well as mRNA and protein levels. Deep neural networks automatically learn informative sequence representations and interpreting them enables us to improve our understanding of the regulatory code governing gene expression. Here, we review the latest developments that apply shallow or deep learning to quantify molecular phenotypes and decode the cis-regulatory grammar from prokaryotic and eukaryotic sequencing data. Our approach is to build from the ground up, first focusing on the initiating protein-DNA interactions, then specific coding and non-coding regions, and finally on advances that combine multiple parts of the gene and mRNA regulatory structures, achieving unprecedented performance. We thus provide a quantitative view of gene expression regulation from nucleotide sequence, concluding with an information-centric overview of the central dogma of molecular biology.

Universal Constraints on Protein Evolution in the Long-Term Evolution Experiment with Escherichia coli

Although it is well known that abundant proteins evolve slowly across the tree of life, there is little consensus for why this is true. Here, I report that abundant proteins evolve slowly in the hypermutator populations of Lenski's long-term evolution experiment with Escherichia coli (LTEE). Specifically, the density of all observed mutations per gene, as measured in metagenomic time series covering 60,000 generations of the LTEE, significantly anticorrelates with mRNA abundance, protein abundance, and degree of protein-protein interaction. The same pattern holds for nonsynonymous mutation density. However, synonymous mutation density, measured across the LTEE hypermutator populations, positively correlates with protein abundance. These results show that universal constraints on protein evolution are visible in data spanning three decades of experimental evolution. Therefore, it should be possible to design experiments to answer why abundant proteins evolve slowly.

Ligand-dependent tRNA processing by a rationally designed RNase P riboswitch

We describe a synthetic riboswitch element that implements a regulatory principle which directly addresses an essential tRNA maturation step. Constructed using a rational in silico design approach, this riboswitch regulates RNase P-catalyzed tRNA 5′-processing by either sequestering or exposing the single-stranded 5′-leader region of the tRNA precursor in response to a ligand. A single base pair in the 5′-leader defines the regulatory potential of the riboswitch both in vitro and in vivo. Our data provide proof for prior postulates on the importance of the structure of the leader region for tRNA maturation. We demonstrate that computational predictions of ligand-dependent structural rearrangements can address individual maturation steps of stable non-coding RNAs, thus making them amenable as promising target for regulatory devices that can be used as functional building blocks in synthetic biology.