A review of machine learning methods to predict the solubility of overexpressed recombinant proteins in Escherichia coli

Background Over the last 20 years in biotechnology, the production of recombinant proteins has been a crucial bioprocess in both biopharmaceutical and research arena in terms of human health, scientific impact and economic volume. Although logical strategies of genetic engineering have been established, protein overexpression is still an art. In particular, heterologous expression is often hindered by low level of production and frequent fail due to opaque reasons. The problem is accentuated because there is no generic solution available to enhance heterologous overexpression. For a given protein, the extent of its solubility can indicate the quality of its function. Over 30% of synthesized proteins are not soluble. In certain experimental circumstances, including temperature, expression host, etc., protein solubility is a feature eventually defined by its sequence. Until now, numerous methods based on machine learning are proposed to predict the solubility of protein merely from its amino acid sequence. In spite of the 20 years of research on the matter, no comprehensive review is available on the published methods. Results This paper presents an extensive review of the existing models to predict protein solubility in Escherichia coli recombinant protein overexpression system. The models are investigated and compared regarding the datasets used, features, feature selection methods, machine learning techniques and accuracy of prediction. A discussion on the models is provided at the end. Conclusions This study aims to investigate extensively the machine learning based methods to predict recombinant protein solubility, so as to offer a general as well as a detailed understanding for researches in the field. Some of the models present acceptable prediction performances and convenient user interfaces. These models can be considered as valuable tools to predict recombinant protein overexpression results before performing real laboratory experiments, thus saving labour, time and cost.


Introduction
In biotechnology, production of recombinant proteins is a crucial process in both biopharmaceutical industries and scientific research. So far, Escherichia coli (E. coli), a bacterium that requires simple conditions to grow is still the favoured host for cloning and overexpressing most proteins which are non-glycosylated and do not have many cysteine residues [1].
Even though logical strategies of genetic engineering are well established, such as strong promoters and codon optimization, protein overexpression is often, still an art. In particular, heterologous expression is often afflicted with low levels of production and insoluble recombinant proteins forming inclusion bodies (protein aggregations). Yet, there is no generic solution available to enhance heterologous overexpression. The use of fusion proteins can sometimes be more successful at the expense of decreased total yield as a result of the fusion partner production. Features that differentiate between proteins in the negative (non-expressed) and positive (expressed) classes might indicate sequence characteristics   that could be modified in optimization, corresponding to what was attained with codon optimization, where sequences of gene are modified to become compatible with the translational apparatus [2]. As the host expresses the proteins, one cause of non-expression is the harmful interaction with the metabolism of the host [3]. For a given protein, the extent of its solubility can indicate the quality of its function. In general, over 30% of recombinant proteins are not soluble [4]. About 33 to 35 percent of all expressed non-membrane proteins are insoluble and about 25 to 57 percent of soluble proteins are prone to aggregate at higher concentrations [5]. For a determined experimental condition (i.e. temperature, expression host, etc.), the solubility of a protein is determined by its sequence [6].
The trial-and-error procedure of protein overexpression can be avoided by identifying the promising proteins to improve the experimental success rate [7]. There are two types of approach for predicting solubility of protein: sequence-based and structure-based. In the structure-based technique, the free energy difference between aggregation and solution phases is computed. This method demands experimentally obtained high resolutions 3D structures which are hard to acquire for aggregation-prone proteins. Hence, the sequence-based technique is a feasible and widely used method. Generally, the computational sequence-based prediction methods investigate the protein overexpression in E. coli at the normal growth temperature of 37°C [8].
The correlation of amino acid sequence and the tendency to form inclusion body was shown for the first time by Wilkinson and Harrison [9]. Later, numerous methods based on machine learning were proposed to predict the solubility of proteins merely from amino acid sequences [10].
Protein solubility prediction can be considered a binary classification task where a classifier should discriminate between soluble proteins (positive samples) and insoluble proteins (negative samples). There are several classification methods (learning algorithm) namely, decision tree (DT) (e.g. C4.5 [11]), k-nearest-neighbour (KNN) [12], neural network (NN) [13,14], support vector machine (SVM) [15], etc.         The learning algorithm (i.e. the classification method) is selected based on numerous factors, such as the number of existing examples in the dataset, the data type to be classified (e.g. symbolic or numeric), and the number of examples probable to be inaccurate or noisy. The level of preferred interpretability of the outcomes is another issue to be considered [16].
The majority of current methods use SVM to build the model of solubility [4]. Appropriate SVM models can often achieve better performance in classification of biological sequence compared to other machine learningbased approaches [1]. Each study employs a different set of features. Considering the model performance, different results are obtained, but 70% is a common accuracy in many studies [4].
To date, all of the prediction approaches examined a single system of protein expression, such as the A. niger or the E. coli system. The works of Hirose et al. [3,10] are exceptions that explored two different systems (E. coli and wheat germ).
In spite of more than two decades of research on the subject, there has been only one report, reviewing seven solubility prediction tools [21]. In their valuable review, the authors have compared seven existing prediction tools based-on the following factors: prediction accuracy, usability, utility, and prediction tool development and validation methodologies. Our aim is to evaluate and investigate all published methods to predict protein solubility, so as to offer a detailed as well as a general understanding for the researchers.
The organization of the paper is as follows. The major protein solubility prediction studies are reviewed in section 2, with emphasis on their datasets, features, feature selection methods, predictor models and performance results. Section 3 presents a discussion on the models details, the best models and the data challenge for solubility prediction task. Lastly, section 4 concludes the paper and proposes some future research directions.

Review
The methods to predict solubility of protein based on the machine learning are summarized in Table 1 in a chronological order, descending from the most recent. Due to space limitation, the reported performance of the works and the features used in each work are shown in Table 2 and Table 3 respectively. More detailed descriptions of the works are presented in "Additional file 1".
In the following tables, for an entry which does not have the corresponding column value, symbol "-" is used. For an entry which we could not find its value, but may exist, symbol "N/A" is used (N/A: Not applicable, not available or no answer)." In order to comprehend the details of the works which are presented in Table 1, Table 2 and Table 3, datasets used, feature selection methods and performance measures are described in greater details in Table 4, Table 5 and Table 6 respectively.
It should be mentioned that in some works several modeling techniques are examined and then one or more are selected as the final model(s). In the "Modeling Technique(s)" column of Table 1, only the final models are shown. It is same true for the "Feature Selection  Method(s)" column. In addition, in most of the works, first an initial feature set is considered, and then using feature selection methods, a smaller sub-set is obtained and employed in the modeling. Table 3 presents the final sets used in the modelings. With respect to the data used in each study, some of the authors created a dataset harvested from the literature, some employed public datasets, while others performed experiments to generate their own dataset.

Discussion
This section investigates the works in more depth. In the following paragraph, the most used dataset, features, feature selection methods and learning techniques are presented. Afterwards, the best models based on the obtained accuracies are introduced. Then, the most convenient to use models are mentioned. Lastly, some datarelated challenges are discussed.
In terms of data, eSol is the most widely used dataset in the field. Considering input features, the following features are the most common ones computed from the protein sequence: aliphatic index, amino acid sequence length, charge, amino acid compositions, instability, isoelectric point (pI), hydrophilicity, molecular weight, and predicted secondary structure. Filter methods (described in Table 5) are used more than the other feature selection techniques. Regarding the machine learning method, support vector machine is the most common technique to make prediction; random forest, decision tree and logistic regression are the next most common ones, respectively.
Based on the results, the method reported by Diaz et al. [20] obtained the best prediction accuracy (94%) on their generated dataset. Similar prediction accuracy was also reported by Samak et al. [4] with an accuracy of 90% on the eSol dataset, followed by the works of Xiaohui et al. [7], and Wilkinson and Harrison [9] with a prediction accuracies of 88% based on their generated datasets.
Comparing the different models in terms of convenience and ease of use, the ones with publicly accessible web servers can be considered the most convenient to use and evaluate. They are ProS [5], PROSOII [6], SCM [8], ESPRESSO [10], SOLpro [17], PROSO [19] and the model of Diaz et al. [20].
It seems that by using an appropriate dataset, as well as suitable machine learning techniques, reasonable prediction performance is attainable. In addition, feature selection methods can reveal, to some extent, influential factors on solubility and the sequence characteristics that could be modified in optimization.
Poor generalization ability is one of the limitations of sequence-based methods founded on a small dataset [35]. In general, extracting a reliable dataset, in terms of experimental conditions and expression system is challenging as the majority of databases that deliver the information on the solubility of proteins often do not provide comprehensive information about the experimental particulars of solubility assessment. Furthermore, researchers generally handle imbalanced (i.e. unequal number of soluble and insoluble samples) data when collecting protein solubility records. Consequently, numerous research teams used different methods to collect consistent datasets that divide proteins into insoluble and soluble categories [24,27].
It is worth mentioning that the datasets employed to build SOLpro [17] and PROSOII [6] were gathered by Table 5 Description of feature selection methods used in machine learning [37] Method Description Examples

Filter
Filter methods evaluate the relatedness of features by looking at the inherent properties of the data. Usually a feature relevance score is computed, and the features with low scores are discarded.

Wrapper
In wrapper methods various subsets of features are evaluated by training and testing a specific classification model, so a search algorithm is 'wrapped' around the classification model. This approach adapted to a specific classification algorithm.
Sequential forward selection [42] Sequential backward elimination [42] Beam search [43] ReliefF [44] Embedded Embedded methods, build the search for an optimal subset of features into the classifier construction, so they are specific to a given learning algorithm.
Random forest [45] SVM recursive feature elimination (SvmRfe) [46] One attribute rule [47] integrating different search results of TargetDB, Protein Data Bank (PDB), and Swiss-Prot database. Then, the proteins were categorized into insoluble and soluble samples according to the proteins' annotations. Although these methods were best working when an appropriate experimental dataset did not exist, they might not be reliable completely. A soluble protein without appropriate annotation, for example, may be incorrectly categorized as an insoluble protein and vice versa. Furthermore, annotations from diverse databases may not be consistent. Clearly, it is desirable to have a large protein set with solubility determined based on experiment by a single reliable protocol [5].

Conclusions
In this paper, the works to predict protein solubility prediction are reviewed in details. They are assessed and classified with regards to the datasets used, features used, feature selection methods, machine learning algorithms and performance results.
Since the early work of Wilkinson and Harrison [9], models later proposed became more complex in terms It is an important performance measure that quantifies how much better the decision is in comparison with random drawing of instances [6]. It indicates the correlation between the classifier assignments and the actual class in the two-class case. It is a good measure of classifier performance even when classes are unbalanced [6]. The MCC ranges between −1 and 1, and a large positive value indicates a better prediction [10]. 10 ROC Curve ROC Plotting the "FP-rate" against the "TPrate", while the probability is increased from 0 to 1.0 with 0.01 increments.
The receiver-operator characteristic curve, showing the trade-off between the ratio of false positives and false negatives in testing a classifier [48]. A larger area value indicates a more robust prediction method [10]. of dataset size, number and types of features employed, feature evaluation techniques and machine learning methods to make prediction. In general, the performances of the models have improved greatly as well. Some of the models provide acceptable prediction performance (e.g. in terms of accuracy). Especially the ones with convenient user interfaces (e.g. web applications), can be considered valuable tools to anticipate recombinant protein overexpression results before performing real laboratory experiments. This capability will lead to significant reduction of labour, time and cost.
Generating larger and more accurate datasets, working on organisms other than E. coli and discovering other influential features, are some considerations for future directions in the protein solubility prediction field.

Additional file
Additional file 1: In detailed descriptions of 24 studies to predict protein solubility during 1991-2014 (February).