cells Article Computational Assessment of Bacterial Protein Structures Indicates a Selection Against Aggregation Anita Carija † , Francisca Pinheiro † , Valentin Iglesias and Salvador Ventura * Institut de Biotecnologia i Biomedicina and Departament de Bioquímica i Biologia Molecular, Universitat Autònoma de Barcelona, 08193 Barcelona, Spain * Correspondence: salvador.ventura@uab.es † These authors contributed equally.  Received: 5 July 2019; Accepted: 6 August 2019; Published: 8 August 2019  Abstract: The aggregation of proteins compromises cell fitness, either because it titrates functional proteins into non-productive inclusions or because it results in the formation of toxic assemblies. Accordingly, computational proteome-wide analyses suggest that prevention of aggregation upon misfolding plays a key role in sequence evolution. Most proteins spend their lifetimes in a folded state; therefore, it is conceivable that, in addition to sequences, protein structures would have also evolved to minimize the risk of aggregation in their natural environments. By exploiting the AGGRESCAN3D structure-based approach to predict the aggregation propensity of >600 Escherichia coli proteins, we show that the structural aggregation propensity of globular proteins is connected with their abundance, length, essentiality, subcellular location and quaternary structure. These data suggest that the avoidance of protein aggregation has contributed to shape the structural properties of proteins in bacterial cells. Keywords: protein aggregation; protein structure; protein function; cell proteostasis; Escherichia coli 1. Introduction Proteins are central components of almost all biological processes, being involved in a variety of complex interactions in the crowded cellular environment [1]. The establishment of non-functional protein–protein interactions has a detrimental impact on cell fitness, both because these contacts sequester proteins into inactive complexes [2] and because it can lead to the aggregation or co-aggregation of proteins into toxic soluble and insoluble assemblies [3]. Importantly, it is increasingly evident that, instead of being an unusual feature of a reduced set of proteins, aggregation is a generic property of many polypeptides [4]. Accordingly, hundreds of unrelated proteins have been reported to aggregate under stress or during ageing [5–7]. It has been suggested that this intrinsic propensity to establish anomalous interactions and aggregate is encoded in the amino acid sequence [8–10] and, therefore, a variety of complementary methods have been developed to predict those propensities from the linear sequence [11]. Large-scale analysis using these algorithms has led to the hypothesis that proteins have evolved sequence adaptations to counteract their natural propensity to aggregate [12–14]. Because in these studies aggregation is analysed along sequences, they mostly measure the aggregation potential of the unfolded state; indeed, the aggregation-prone regions (APRs) that these algorithms identify and evaluate are blocked in properly folded proteins, either because they are buried inside the hydrophobic core or engaged in the series of cooperative non-covalent interactions that sustain the secondary and tertiary protein structure [15]. These sticky sequences might, however, become accessible in case the protein fails to fold due to translational errors. Accordingly, prevention of mistranslation-induced protein misfolding is thought to constraint the evolution of sequences [16,17]. However, it is worth to point out that protein aggregation is not always deleterious and different Cells 2019, 8, 856; doi:10.3390/cells8080856 www.mdpi.com/journal/cells Cells 2019, 8, 856 2 of 20 of mistranslation-induced protein misfolding is thought to constraint the evolution of sequences [16,17]. Cells 2019, 8, 856 However, it is worth to point out that protein aggregation is not always deleterious and 2 of 20 different organisms have exploited the structural/mechanical properties of protein aggregates for functional purposeshave organisms [18,19]. exploited the structural/mechanical properties of protein aggregates for functional Protein purposes [18,19].is not always a requisite for aggregation, and different proteins have been misfolding shown to aggregate Protein from their initial misfolding is not functional conformations always a requisite [20,21], a propensity for aggregation, thatproteins and different would have be been exacerbated shown in tothe crowded aggregate cellular from milieufunctional their initial [22]. Thisconformations suggests that[20,21], many aproteins propensitymight thatbe would be kinetically, but not thermodynamically, stable in their native states under physiological conditions exacerbated in the crowded cellular milieu [22]. This suggests that many proteins might be kinetically, [23,24]. but Indeed, the aggregated state of a protein is usually more stable than its native state [25]. not thermodynamically, stable in their native states under physiological conditions [23,24]. Indeed, According to this view, a solution of proteins that have a certain propensity to aggregate but are the aggregated state of a protein is usually more stable than its native state [25]. According to this view, correctly folded constitutes a system in a transient state, that might ultimately reach the global free a solution of proteins that have a certain propensity to aggregate but are correctly folded constitutes energy minimum of the aggregated state [23]. Therefore, it is conceivable that proteins have evolved a system in a transient state, that might ultimately reach the global free energy minimum of the strategies to delay/prevent this deleterious transition, adjusting the solubilities of their folded states aggregated state [23]. Therefore, it is conceivable that proteins have evolved strategies to delay/prevent to those required for function. this deleterious transition, adjusting the solubilities of their folded states to those required for function. In the present work, we address the relationship between the predicted aggregation propensities In the present work, we address the relationship between the predicted aggregation propensities of protein structures and a set of intrinsic parameters relevant for their function in the cell. To this of protein structures and a set of intrinsic parameters relevant for their function in the cell. To this aim, we exploited a recently developed algorithm, AGGRESCAN3D (A3D) [26–29], to analyse 619 aim, we exploited a recently developed algorithm, AGGRESCAN3D (A3D) [26–29], to analyse 619 high-resolution Escherichia coli protein structures displaying lower than 40% sequence homology. high-resolution Escherichia coli protein structures displaying lower than 40% sequence homology. A3D is an evolution of our previous AGGRESCAN algorithm [30]. The AGGRESCAN amino acid A3D is an evolution of our previous AGGRESCAN algorithm [30]. The AGGRESCAN amino acid aggregation scale was experimentally derived from protein aggregation measurements in vivo in E. aggregation scale was experimentally derived from protein aggregation measurements in vivo in coli cells [31]. This placed AGGRESCAN among the most accurate algorithms for the prediction of E. coli cells [31]. This placed AGGRESCAN among the most accurate algorithms for the prediction of intracellular protein aggregation, especially when analysing E. coli data [32]. A3D also implements intracellular protein aggregation, especially when analysing E. coli data [32]. A3D also implements these in vivo derived intrinsic propensities but they are strongly modulated by the structural context these in vivo derived intrinsic propensities but they are strongly modulated by the structural context in which they are located, allowing prediction of the aggregation properties of proteins in their folded in which they are located, allowing prediction of the aggregation properties of proteins in their folded states [26–29]. By analogy to AGGRESCAN, A3D is expected to be accurate when analysing E. coli states [26–29]. By analogy to AGGRESCAN, A3D is expected to be accurate when analysing E. coli protein structures. Overall, our structural analysis suggests that the avoidance of protein aggregation protein structures. Overall, our structural analysis suggests that the avoidance of protein aggregation has contributed to shape the properties of present proteins in this organism. has contributed to shape the properties of present proteins in this organism. 2. Material and Methods 2. Material and Methods 2.1. Selection Criteria for 2.1. Selection the Analysis Criteria of Bacterial for the Analysis ProteomeProteome of Bacterial We employed a dataseta consisting We employed of 1103 proteins, dataset consisting including, of 1103 proteins, to the best including, of our to the bestknowledge, the of our knowledge, the most complete set of set most complete proteins experimentally of proteins experimentally detected detected in inthe thebacterial bacterial cytosol atatthe thetime timeweweperformed performed the the analysis analysis [33].[33]. ThisThis is the is the same same dataset dataset we used we used previously previously to address to address the relationship the relationship between the betweenaggregation the aggregation propensity propensity of linear of linear sequences sequences and different and different protein protein properties [34],properties facilitating[34], comparison facilitating comparison between both studies. The dataset was used to search through between both studies. The dataset was used to search through the Protein Data Bank (PDB) [35] in the Protein Data Bank (PDB) order [35] inproteins to identify order tofromidentify proteins the dataset from which theavailable have dataset which solved have available crystal/NMR solved E. coli structures. crystal/NMR structures. proteins E. coli three-dimensional with available proteins with available three-dimensional structure in PDB format,structurewith X-ray in resolution PDB format, ≤3.5, were with X-ray resolution considered in this ≤3.5, were Furthermore, analysis. considered in this analysis. available Furthermore, three-dimensional available structures three-<90% of covering dimensional structures protein primarycovering sequence <90% were ofexcluded, protein primary as wellsequence as proteins were excluded,sequence displaying proteins >40%, as well ashomology displaying sequence resulting homology in a set >40%, proteins of 619 bacterial resulting(Table in a S2). set of The 619 bacterial original 1103proteins sequences (Table datasetS2). The was supposed original to1103 sequencesonly correspond dataset was supposed to cytosolic proteinsto [33], correspond but weonly could to identify cytosolicalsoproteins [33], but membrane we proteins in it could identify alsoUnless (Table S2). membrane proteins otherwise in itmembrane indicated, (Table S2). Unlesswere proteins otherwise not taken indicated, into account membrane for the different proteinsanalysis were not taken intoinaccount performed this work,for as thethey different differ analysis performed significantly from the innon-membrane this work, as they differin terms proteins significantly fromand of charge the non-membrane hydrophobicity. proteins in terms of charge and hydrophobicity. 2.2. Structural 2.2. Structural and Sequential and Sequential Aggregation Aggregation Propensity Propensity Predictions Predictions Predictions Predictions for the structures for the structures set wereset were performed performed using the using A3Dthe A3D algorithm algorithm which implements which implements structure-based structure-based approach approach and predicts and predicts aggregation aggregation propensity propensity of initially of initially folded[26]. folded states states [26]. Selected Selected proteinsproteins were submitted to A3D were submitted in the to A3D ‘Static in the mode’ ‘Static mode’and and1010 Ǻ was wasselected selectedasas a distance a distance for for aggregation aggregation analysis analysis (default (default sphere sphere radius). radius). TheThe average average scores scores forfor each each protein protein in in thethe setset were were obtained. obtained. The The average average score score allows allows comparing comparing thethe solubility solubility of protein of protein structures structures differing differing in size. in size. It allows It also assessing changes in solubility promoted by amino acid substitutions in a particular protein structure; Cells 2019, 8, 856 3 of 20 this value was assumed to reflect the Structural Aggregation Propensity (STAP) of any given protein. To perform sequential aggregation analysis, we used the linear predictor AGGRESCAN [30] on top of the sequences corresponding to the analysed protein structures. Normalized a4v Sequence Sum for 100 residues (Na4vSS) values were obtained, which reflect the average protein aggregation propensities of the sequences once corrected for their size. 2.3. Datasets Protein abundance data for the set of 619 bacterial proteins was taken from PaxDb [36]. The obtained values were then log10 transformed for statistical analysis. Information about protein length, protein subcellular location, number of transmembrane segments and protein active form were obtained from Swiss-Prot Protein knowledgebase [37]. Gene Ontology (GO) enrichment analysis was performed using the Database for Annotation, Visualization and Integrated Discovery (DAVID) [38]. The essentiality of bacterial proteins for cellular fitness was derived from the reported data [39,40] and updated with the data available at server EcoGene 3.0 [41]. The RegulonDB was used to obtain the known E. coli operon structure set [42]. To assess the compositional determinants of STAP in the bacterial proteome, amino acids present on the surface of the 10% least and the 10% most aggregation-prone proteins were identified by A3D, and the frequency of appearance of each individual amino acid among the total set of surface amino acids belonging to the two mentioned groups was calculated. Amino acid frequencies for aggregation-prone and soluble E. coli proteins were compared using Wilcoxon test. 2.4. Definition of the Supersaturation Index (SSI) The structural supersaturation index (SSI), as the sum: (C + A) SSI = (1) 2 where C represents normalized protein concentration and A is the normalized STAP score. The logarithm of the protein abundance levels derived from PaxDb was used in order to determine C for each polypeptide from the dataset. These values were then normalized by rescaling them between 0 and 1. (Ci − min(Ci . . . Cn)) C= (2) (max(Ci . . . Cn) − min(Ci . . . Cn)) Cmin = the minimum value of protein concentration from the dataset; Cmax = the maximum value of protein concentration from the dataset. The propensity of proteins to aggregate represented by A3D average score was normalized in the same manner by rescaling the values between 0 and 1. (Ai − min(Ai . . . An)) A= (3) (max(Ai . . . An) − min(Ai . . . An)) Amin = the minimum A3D score from the dataset; Amax = the maximum A3D score from the dataset. 2.5. Statistical Analysis Statistical significance was determined using Wilcoxon test. Analyses were performed using KaleidaGraph software (Synergy Software, Reading, PA, USA). p-values < 0.01 were considered statistically significant (* statistically significant at p ≤ 0.01; ** statistically significant at p ≤ 0.001; *** statistically significant at p ≤ 0.0001). The cumulative frequencies plots presented in the manuscript give information on the percentage of proteins (y-axis) that display a value equal or lower to a value, x. Fold enrichment indicates how much higher is the proportion of hits in relation to the background sample (E. coli proteome). For every GO term, the fold enrichment is the number of positive hits Cells 2019, 8, 856 4 of 20 Cells 2019, 8, 856 4 of 20 among our list (nl) between the number of annotated proteins in our list (pl); and subsequently divided into the rate of hits of that GO term in background (np) between the number of total proteins among in theour list (nl ) between background l (pb): the number of annotated proteins in our list (p ); and subsequently divided p into the rate of hits of that GO term in background (n ) between the number of total proteins in the b 𝑛𝑙 background (p ): 𝑛 𝑙 𝑝𝑏 𝑝𝑙 𝐹𝑜𝑙𝑑 𝐸𝑛𝑟𝑖𝑐ℎ𝑚𝑒𝑛𝑡 = nl 𝑏 =l b𝑏 𝑙 (4) pl 𝑛 n p𝑛 𝑝 Fold Enrichment = b 𝑝=𝑏 b l (4) n n p b p 3. 3. Results Results 3.1. 3.1. A3D A3D Analysis Analysis Rationale Rationale TheThe A3D A3D algorithm algorithm uses uses the the protein protein 3D structure 3D structure as an input,aswhich an input, which isenergetically is subsequently subsequently energetically minimized using the FoldX force field [43]. Then, an aggregation minimized using the FoldX force field [43]. Then, an aggregation propensity score is calculated for propensity score all is thecalculated spheres withfor all a 10theÅspheres radius inwith theaprotein 10 Å radius in theThe structure. protein structure. variables The variables that contribute that to the A3Dcontribute score to the A3D score are (i) the experimentally-derived individual amino acid aggregation are (i) the experimentally-derived individual amino acid aggregation propensities; (ii) the surface area propensities; (ii) the surface exposure area exposure of the amino acids in ofthethe aminoand sphere; acids (iii)inthe theeffective sphere; and (iii) the distance effective between distance adjacent between residues andadjacent residues the central amino andacid the in central amino acid the sphere. in the sphere. Therefore, the A3D Therefore, the A3D score score is structurally is structurally corrected and, in contrast to sequence-based aggregation predictors, provides information on the STAP of proteins inthe corrected and, in contrast to sequence-based aggregation predictors, provides information on STAP their of proteins functional foldedin their statesfunctional folded [26]. Figure states [26]. 1 illustrates how Figure A3D1allows illustrates how A3D to discard allows to discard the contribution of thethe contribution buried of the buried APRs detected APRs detected by sequence-based by sequence-based algorithms (Figure 1A),algorithms (Figure 1A), and how residues and how scattered in the sequence can come together in the structure upon folding to form an aggregation-prone surface, asan residues scattered in the sequence can come together in the structure upon folding to form aggregation-prone identified by A3D (Figure surface, 1B).as identified by A3D (Figure 1B). Figure 1. Analysis of structural aggregation propensity (STAP) for representative soluble and aggregation-prone proteins. (A) Comparison between the aggregation propensity of representative soluble Figure(PDB: 3T36:A; left 1. Analysis panel) andaggregation of structural aggregation-prone proteins propensity (PDB: (STAP) for2HO9:A; right panel) representative using soluble and sequence-based predictors and A3D. Aggregation-prone residues are indicated in different aggregation-prone proteins. (A) Comparison between the aggregation propensity of representative colours forsoluble each predictor: AGGRESCAN (PDB: 3T36:A; left panel)(black), TANGO (grey) and and aggregation-prone A3D(PDB: proteins (red). 2HO9:A; (B) A3D right analysis of the panel) using above mentioned proteins. The A3D average scores for PDBs 3T36:A and 2HO9:A are –1.061 sequence-based predictors and A3D. Aggregation-prone residues are indicated in different colours and –0.421, respectively. for each predictor:The protein surface AGGRESCAN is coloured (black), TANGO according to A3D (grey) and A3Dscore (red).in(B) a gradient from blue A3D analysis of the (high-predicted solubility) to white (negligible impact on protein aggregation) to red (high-predicted above mentioned proteins. The A3D average scores for PDBs 3T36:A and 2HO9:A are –1.061 and – aggregation propensity). 0.421, respectively. The protein surface is coloured according to A3D score in a gradient from blue (high-predicted solubility) to white (negligible impact on protein aggregation) to red (high-predicted aggregation propensity). Cells 2019, 8, 856 5 of 20 3.2. Relationship Between Protein Abundance and Structural Aggregation Propensity Despite that non-functional interactions might be detrimental, statistically, the large number of non-functional contacts a protein can establish clearly outweighs its functional interactions. This is especially true for abundant proteins, since, according to the law of mass action, the probability to establish a non-functional interaction should be proportional to the protein abundance [2]. Indeed, most protein aggregation processes are strongly dependent on the initial protein concentration [44]. Therefore, an abundant protein with an aggregation-prone surface is expected to be more dangerous than a low-abundant protein with the same surface stickiness [12]. Thus, a negative relationship between protein STAP and protein abundance can be expected. We explored this relationship using E. coli as a model organism. We obtained the abundance data for the 612 proteins in our structural dataset and we proceeded by log10 transforming the reported abundance for statistical analysis, since the number of mRNAs in the bacterial cytosol encoding a given protein can vary greatly from 1 to 100,000 [45]. The negative correlation between STAP and abundance is low (r = 0.21), but highly significative (p = 7.50 × 10−7 ). The comparison of the STAP distribution in the 25% most and least abundant proteins in this dataset is shown in Figure 2A–C and illustrates how highly abundant proteins and proteins present at low concentrations exhibit, indeed, differential aggregation propensities (p = 3.26 × 10−4 ), with low abundant proteins displaying surfaces that can support much higher aggregation load than those of high abundant polypeptides (Figure 2D). The probability density function of abundant proteins illustrates the absence of polypeptides with very high aggregation propensity in this group, with a concomitant enrichment in highly soluble proteins (Figure 2C). In addition, low abundant proteins seem not to be under selection for protein aggregation, as the STAP of the 25% of the least abundant proteins is significantly higher than that of the conjunct of the remaining 75% protein structures (p = 7.77 × 10−5 ). The higher solubility of abundant proteins would likely work to prevent non-functional interactions in these proteins, even if they become concentrated at specific sub-cytosolic locations. Different studies have reported a relationship between the abundance of proteins and the aggregation propensity of linear protein sequences [12,34,46,47]. This has been explained in terms of natural selection acting on protein unfolded states in order to minimize aggregation in case of eventual misfolding. In order to test whether this relationship also exists in our dataset, we used our linear aggregation-predictor AGGRESCAN, which uses exactly the same intrinsic amino acid propensities that A3D without taking into account the structural context, to calculate the aggregation properties of the correspondent 612 sequences. As expected, the 25% most abundant proteins display more soluble sequences than the 25% least abundant (Figure 3). However, the significance of the solubility differences between these two subclasses is lower when considering sequences (p = 3.91 × 10−3 ) than when we consider structures (p = 3.26 × 10−4 ). Thus, it seems that, at least in E. coli, in addition to misfolding, structural misassembly, driven by the establishment of non-native interactions between folded protein structures, might also constrain the evolution of proteins [2]. Overall, it can be concluded that proteins’ STAP seems to be adjusted to the gene expression levels required for an optimal cell function. Therefore, mutations that decrease the proteins’ surface solubility, or increases in protein expression levels might exacerbate the probability to aggregate [12]. Cells 2019, 8, 856 6 of 20 Cells 2019, 8, 856 6 of 20 Cells 2019, 8, 856 6 of 20 Figure Figure 2.2.Relationship Relationship between between STAPand STAP andprotein proteinabundance. abundance. (A)(A) Cumulative Cumulative distribution distribution ofof of STAP STAP Figure 2. Relationship between STAP and protein abundance. (A) Cumulative distribution STAP for the forfor the25% 25% most most (red line) line)and andthe the25% 25% least least (blue (blue line) line) abundant abundant proteins. proteins. (B) Box (B)plot Boxshowing plot showing the the 25% most (red line) and the 25% least (blue line) abundant proteins. (B) Box plot showing the the comparison comparison between between the the average A3D A3D average score ofscore the of the 25% most25% (red) most and (red) the 25%andleastthe 25%abundant (blue) least (blue) comparison between the A3D average score of the 25% most (red) and the 25% least (blue) abundant abundant proteins. proteins. proteins. (C) (C) Probability (C)Probability Probability density density density function function function ofof thethe25% 25% ofmost mostthe(red) 25% most (red)andand the (red) the25% 25%and least the least 25%abundant (blue) (blue) least (blue) abundant proteins. abundant proteins. (D) A3D proteins. analysis (D) A3D(D) of representative A3D analysis analysis low-abundance of representative of representative low-abundance (PDB: low-abundance 2WSX:A; (PDB: 2WSX:A; upper (PDB:upper panel) 2WSX:A; and upper panel) high- panel) and high- abundance and abundance proteins high-abundance proteins (PDB: proteins (PDB: 1DFO:A; 1DFO:A; lower (PDB: 1DFO:A; lower panel). lower panel). Views ofof panel). Views opposite Views opposite ofsides ofof opposite sides each sides each protein are of each protein shown. protein are shown. are The shown. A3D average scores for 2WSX:A and 1DFO:A are 0.052 and −0.799, The A3D average scores for 2WSX:A and 1DFO:A are 0.052 and −0.799, while log10 values of theirof The A3D average scores for 2WSX:A and 1DFO:A are 0.052 and while −0.799, log 10 values while log of 10 their values protein their protein abundance protein abundance abundance are are−1.523 are −1.523and −1.523 and3.420, and respectively. 3.420, 3.420, Colour respectively. respectively. Colour code Colour code is code asas is ininFigure is as in1.Figure Figure 1. 1. Figure Figure 3. Relationshipbetween between sequential aggregation propensity and protein abundance. Figure3. 3.Relationship Relationship betweensequential sequentialaggregation aggregationpropensity propensityand andprotein proteinabundance. abundance.(A)(A) (A) Cumulative Cumulative distribution of Na4vSS values for the 25% most (red line) and the 25% least abundant Cumulativedistribution ofofNa4vSS distribution Na4vSSvalues valuesforfor the the25% 25%most most(red (redline) line)and andthe the25% 25%least leastabundant abundant proteins (dashed blue line). (B) Box plot representing the Na4vSS values for the 25% most (red) and the 25% least (blue) abundant proteins. Cells 2019, 8, 856 7 of 20 proteins Cells 2019, 8, 856(dashed blue line). (B) Box plot representing the Na4vSS values for the 25% most (red) and7 of 20 the 25% least (blue) abundant proteins. 3.3. Relationship 3.3. Relationship Between Between Protein Protein Length Length and and Structural Structural Aggregation Aggregation Propensity Previous studies Previous studies have have suggested suggested thatthat the the length length ofof aa given given protein protein might might be be an an important important determinant of determinant of its its aggregation aggregation propensity. propensity. In In this this way, way, it it has been shown shown that that longer longer proteins proteins were were more likely more likely to to co-aggregate co-aggregate and and be be sequestered sequestered in in vivo by artificially designed aggregation-prone Also, itit has polypeptides [48]. Also, hasbeen beenreported reported that that the theleast leastsoluble solubleproteins proteins of ofthree threedistinct distinct eukaryotic eukaryotic organisms share several common traits, one of them being generally organisms share several common traits, one them being generally longer in comparison tolonger in comparison to highly highly soluble proteins soluble proteins [49]. [49]. The The comparison comparison of of the the 25% 25% shortest shortest and and the the 25% 25% longest longest polypeptides polypeptides in in our our protein set, indicates that both groups exhibit protein exhibit different different STAP STAP(p (p== 2.65 10−9 2.65×× 10 −9)) (Figure (Figure 4A–C). 4A–C). Short proteins exhibit a wider distribution of A3D average scores (SD = 0.30), indicating = 0.30), indicating that that the dynamic dynamic range of aggregation range aggregation propensities propensitiesininthisthisgroup groupisislarger when larger when compared compared to to thethe group composed group composed of the of 25%25% the longest polypeptides longest polypeptides = 0.18) (SD(SD (Figure = 0.18) 4C).4C). (Figure Figure 4. Relationship between STAP and protein length. (A) Comparison between cumulative Figure 4. Relationship between STAP and protein length. (A) Comparison between cumulative distribution of STAP for the 25% longest (red line) and the 25% shortest proteins (blue line). (B) Box distribution of STAP for the 25% longest (red line) and the 25% shortest proteins (blue line). (B) Box plot representing the A3D average score of the 25% longest (red) and the 25% shortest (blue) proteins. plot representing the A3D average score of the 25% longest (red) and the 25% shortest (blue) proteins. (C) Probability density function of the 25% longest (red) and the 25% shortest proteins (blue). (C) Probability density function of the 25% longest (red) and the 25% shortest proteins (blue). (D) (D) Cumulative distribution of protein length for the 25% proteins with the highest STAP (solid Cumulative distribution of protein length for the 25% proteins with the highest STAP (solid line) and line) and the 25% proteins with the lowest STAP (dashed line). the 25% proteins with the lowest STAP (dashed line). Cells 2019, 8, 856 8 of 20 Cells 2019, 8, 856 8 of 20 To To confirm confirm thethe relationship relationship between betweensize sizeand andsolubility, solubility, we we compared compared the the size size of of the the 25% 25% most most soluble solubleproteins, proteins,according accordingto totheir theirA3DA3D average averagescore, score, and and the the 25% 25% least least soluble. soluble. These Thesetwo twosubsets subsets correspond correspondto to proteins proteins with different size properties (p = with different = 9.11 × 10−7 −7) )(Figure (Figure4D).4D).As As expected, expected, the the group groupof oflow lowaggregation-prone aggregation-proneproteinsproteinsisis enriched enrichedin insmall smallpolypeptides, polypeptides,whereaswhereasthe theproteins proteinsin in the thehigh highaggregation-prone aggregation-proneset setare areclearly clearlylarger larger (Figure (Figure 4D). 4D). AA visual visual inspection inspection of ofthe the aggregation aggregation surfaces surfacesofof short short and and long long proteins proteins suggests suggests that that the the higher higher number number of of exposed exposed aggregation-prone aggregation-prone structural structuralpatches patchesin inthe thelongest longestproteins proteinscouldcouldbebeananimportant importantdeterminant determinantofoftheir theirhigher higherSTAP. STAP. The presence of these patches likely responds to the fact that these proteins The presence of these patches likely responds to the fact that these proteins can interact simultaneously can interact simultaneously with a larger numberwith ofa partners larger number or through of larger partners or through protein interfaces,larger protein but this interfaces, functionality alsobut this involves functionality a higher risk also involves a higher risk of aggregation. of aggregation. We We used AGGRESCAN AGGRESCAN to tocalculate calculatethe theaggregation aggregation propensities propensities of of all all thethe protein protein sequences sequences and and thenthen grouped grouped them them as described as described above. above. TheThe overall overall aggregation aggregation propensities propensities of 25% of the the 25% shortest shortest and and the longest the longest proteinprotein sequences sequences do notdo not significantly differ (p = 0.72) differ significantly (p (Figure = 0.72) 5).(Figure 5). The differences The differences observed observed when considering when considering structuralstructural or sequential or sequential aggregationaggregation propensitiespropensities are not surprising are not surprising if we thinkif that we think manythat many of the of the sequential sequential aggregation-prone aggregation-prone regions areregions hiddenare hidden inside inside thewhen the structure structure when the protein the protein(Figure is folded is folded1A)(Figure and that1A) and structural many that manyaggregation-prone structural aggregation-prone regions at the regions protein at the protein surface are surface composedare composed of residuesofthat residues are notthat are not consecutive consecutive in the sequence. in the sequence. In any case, In any case,inatour at least least in our bacterial bacterial proteinthe protein dataset, dataset, the relationship apparent apparent relationship between aggregation between aggregation and explained and size is best size is bestifexplained we consider if we consider that the selective pressure acts on protein structures instead that the selective pressure acts on protein structures instead of/in addition to sequences. of/in addition to sequences. Figure 5.5. Relationship Figure Relationship between between sequential sequential aggregation aggregation propensity propensity and and protein protein length. length. (A) (A) Cumulative distribution of Na4vSS values for the 25% longest (red line) and the 25% shortest Cumulative distribution of Na4vSS values for the 25% longest (red line) and the 25% shortest proteins proteins blue (dashed (dashed line).blue (B)line). (B) Box Box plot plot showing showing the comparison the comparison between between the Na4vSS the Na4vSS valuesvalues of theof25% the 25% longest (red) and the 25% shortest (blue) longest (red) and the 25% shortest (blue) proteins.proteins. 3.4. Relationship Between Structural Aggregation Propensity and Protein Function 3.4. Relationship Between Structural Aggregation Propensity and Protein Function The set of genes in an operon share a common gene expression regulation and are generally The set of genes in an operon share a common gene expression regulation and are generally connected by their biological function. As a result, proteins encoded by the same operon are suggested connected by their biological function. As a result, proteins encoded by the same operon are to be present in similar amounts in the cell [33]. The observed association between STAP and abundance suggested to be present in similar amounts in the cell [33]. The observed association between STAP would imply that polypeptides in the same operon should have related aggregation properties. To test and abundance would imply that polypeptides in the same operon should have related aggregation this hypothesis, we first ascribed the proteins in our complete dataset to individual operons. 234 out of properties. To test this hypothesis, we first ascribed the proteins in our complete dataset to individual 619 proteins could be ascribed to a particular E. coli operon (Table S1). We calculated A3D average operons. 234 out of 619 proteins could be ascribed to a particular E. coli operon (Table S1). We scores and, as hypothesized, found out that the standard deviation of the A3D average score between calculated A3D average scores and, as hypothesized, found out that the standard deviation of the proteins regulated by the same operon is lower in 93% of the cases (38 out of 41) than the standard A3D average score between proteins regulated by the same operon is lower in 93% of the cases (38 deviation in the complete set of proteins that could be ascribed to a particular operon (Figure 6). out of 41) than the standard deviation in the complete set of proteins that could be ascribed to a particular operon (Figure 6). Cells 2019, 8, 856 9 of 20 Cells 2019, 8, 856 9 of 20 In the loss of function scenario, the impact of protein aggregation on cellular fitness would be In the loss of function scenario, the impact of protein aggregation on cellular fitness would be ultimately associated with the particular protein activity. Therefore, it is expected that evolution ultimately associated with the particular protein activity. Therefore, it is expected that evolution would select for an overall decreased aggregation propensity in the proteins from operons involved would select for an overall decreased aggregation propensity in the proteins from operons involved in in essential cellular functions [50]. To explore this possibility, the bacterial operons were divided into essential cellular functions [50]. To explore this possibility, the bacterial operons were divided into two two groups according to their A3D average scores, those with lower (LA operons) and higher groups according to their A3D average scores, those with lower (LA operons) and higher aggregation aggregation propensity (HA operons) than the mean propensity of the complete operon protein set propensity (HA operons) than the mean propensity of the complete operon protein set (A3D average (A3D average score = −0.880). The essentiality of approximately half of the proteins in each subset has score = −0.880). The essentiality of approximately half of the proteins in each subset has been annotated been annotated via genetic footprinting or knockout experiments [39,40]. Importantly, considering via genetic footprinting or knockout experiments [39,40]. Importantly, considering only the annotated only the annotated polypeptides, the majority of proteins regulated by LA operons are essential, polypeptides, the majority of proteins regulated by LA operons are essential, whereas most of those whereas most of those in HA operons are non-essential (Table S1, Figure 6B). This supports the view in HA operons are non-essential (Table S1, Figure 6B). This supports the view that, as a trend, the that, as a trend, the structures of essential bacterial proteins suffer a stronger selection against structures of essential bacterial proteins suffer a stronger selection against aggregation than those of aggregation than those of non-essential ones. non-essential ones. Figure 6. The variation in A3D average scores for known bacterial operons. (A) The standard deviation of A3D average scores in 41 analysed operons was calculated and operons were divided Figure 6. The variation in A3D average scores for known bacterial operons. (A) The standard into 10 classes. The dashed line indicates the standard deviation in the complete set of proteins (0.4). deviation of A3D average scores in 41 analysed operons was calculated and operons were divided Low standard deviation within an operon indicates that the aggregation propensity of its proteins is into 10 classes. The dashed line indicates the standard deviation in the complete set of proteins (0.4). similar. (B) Number of essential (green) and non-essential (orange) proteins codified by operons with Low standard deviation within an operon indicates that the aggregation propensity of its proteins is low (left panel) and high (right panel) aggregation propensity. similar. (B) Number of essential (green) and non-essential (orange) proteins codified by operons with low (left panel) and high (right panel) aggregation propensity. 3.5. Effect of Subcellular Location on the Structural Aggregation Propensity Cells 2019, 8, 856 10 of 20 3.5. Effect of Subcellular Location on the Structural Aggregation Propensity It is of great importance that the protein maintains its biological function in its native state no Cells 2019, 8, 856 10 of 20 matter in which subcellular location this protein resides [46,47,51]. Bacterial proteins can populate other subcellular compartments It is of great apart importance that thefrom protein the cytosol,itslike maintains the periplasm biological function inand the inner its native and outer state no matter Presumably, membranes. in which subcellular location this protein their aggregation resides properties [46,47,51]. would Bacterial to be adapted proteins can populate the specific environment in thoseother subcellular subcellular compartments locations. A3Dapart from shows analysis the cytosol, thatlike the periplasm proteins residing andinthe theinner and outer bacterial cytoplasm membranes. Presumably, their aggregation properties would be adapted to the specific environment and the periplasm possess the lowest STAP of the complete dataset (Figure 7A,B). After excluding in those subcellular locations. A3D analysis shows that proteins residing in the bacterial cytoplasm proteinsandwith transmembrane the periplasm domains, possess the lowest STAP the distribution of the completeofdataset the 20% proteins (Figure with excluding 7A,B). After the lowest A3D averageproteins scores withand 20% proteins with the highest A3D average scores in these transmembrane domains, the distribution of the 20% proteins with the lowest A3Dtwo compartments was analysed. We scores average foundand out20% thatproteins soluble proteins with wereA3D the highest enriched averagein bothinthe scores cytosol these and the periplasm, two compartments whereaswas analysed. in enrichment Wethis found lastout that soluble was compartment proteins were enriched not observed among in both the cytosol and theproteins the aggregation-prone (Figureperiplasm, 7C). Thiswhereas enrichment is consistent with in this the last factcompartment was not observed that the periplasm, among to in contrast thethe aggregation- cytosol, lacks a prone proteins (Figure 7C). This is consistent with the fact that the periplasm, in contrast to the sophisticated cellular system able to control protein quality and avoid aggregation [52], and is separated cytosol, lacks a sophisticated cellular system able to control protein quality and avoid aggregation from the outside [52], solutionfrom and is separated by athe highly outsidepermeable solution by aouter highlymembrane thatmembrane permeable outer providesthat limited protection provides againstlimited environmental variations. protection against environmental variations. Figure 7. Relationship between STAP and protein subcellular location. (A) The bacterial proteins Figure 7. Relationship between STAP and protein subcellular location. (A) The bacterial proteins were divided into four groups based on the protein localization according to UniProt. (B) A cumulative were divided into four groups based on the protein localization according to UniProt. (B) A distribution of STAP cumulative of proteins distribution located of STAP in thelocated of proteins cytoplasm (C; purple), in the cytoplasm periplasm (C; purple), (P; dark periplasm grey), outer (P; dark membrane grey),(OM; outerred) and inner membrane (OM;membrane (IM; red) and inner blue). (C) membrane Proteins (IM; with blue). (C) 20% lowest Proteins with 20%(left panel) lowest (leftand 20% panel) proteins and 20% highest STAP proteins (righthighest panel)STAP (right panel) belonging to thebelonging dataset tointhe dataset which in which proteins proteins with with transmembrane transmembrane segments (TS) were segments excluded, (TS)according were excluded, to GO according terms.to Only GO terms. GOOnly GO cellular cellular component component terms with terms with p-value < 0.05 and false discovery rate (FDR) <0.05 were plotted. (D) Diagram for the p-value < 0.05 and false discovery rate (FDR) <0.05 were plotted. (D) Diagram for the analysed IM analysed IM proteins showing the number of TS and STAP. (E) Analysis of the molecular functions proteins showing the number of TS and STAP. (E) Analysis of the molecular functions associated to IM proteins with A3D average score <0 (left panel) and A3D average score ≥0 (right panel) according to GO terms proposed by PANTHER. Cells 2019, 8, 856 11 of 20 Cells 2019, 8, 856 11 of 20 associated to IM proteins with A3D average score <0 (left panel) and A3D average score ≥0 (right Notpanel) surprisingly, inner membrane (IM) proteins exhibit the highest theoretical aggregation according to GO terms proposed by PANTHER. propensities of all bacterial proteins, according to A3D analysis (Figure 7A,B). The gram-negative bacterialNotinner membrane surprisingly, is amembrane inner semipermeable shieldexhibit (IM) proteins that preserves the highest the cytoplasm theoretical environment. aggregation propensities The proteins of all bacterial associated with thisproteins, according membrane to A3D can have analysis (Figure a variable number7A,B). of TS The per gram-negative protein [53]. These bacterial regions inner membrane are stable is a semipermeable in the hydrophobic shield that environment preserves of this lipid the cytoplasm bilayer due to environment. The their enrichment in proteins associated with this membrane can have a variable number of TS apolar residues. A visual inspection of their structures shows that the highly aggregation-prone per protein [53]. These regions are stable in the hydrophobic environment of this lipid bilayer due to their enrichment in surfaces identified by A3D in many of these proteins sharply coincide with regions embedded in the apolar residues. A visual inspection of their structures shows that the highly aggregation-prone membrane, since the membrane width could be traced in their A3D coloured structure (Figure 8A,B). surfaces identified by A3D in many of these proteins sharply coincide with regions embedded in the Interestingly, membrane, when sincethe the A3D average membrane scores width could ofbe inner membrane traced in their A3D proteins colouredwere plotted structure as a distribution, (Figure 8A,B). the existence of two protein groups became evident (Figure 7B). Interestingly, when the A3D average scores of inner membrane proteins were plotted We found that the main difference as a between these two distribution, thegroups existenceis of the number two proteinof TS (Figure groups became7D). In fact, evident the inner (Figure membrane 7B). We found thatgroup the maincontains a number of proteins difference betweenwithout these two TSgroups and a low is theSTAP. number An of analysis of their TS (Figure 7D).structures In fact, theindicates that they are inner membrane group devoid contains of large a number of proteins aggregation-prone without patches TS and that suggesting a low STAP. they An analysis associate to theof their structures membrane transiently indicates that they are devoid of large aggregation-prone patches suggesting that (Figure 8C,D). Proteins with a single TS also display significant predicted solubility, likely because they associate to the membrane transiently (Figure 8C,D). Proteins with a single TS also display apart from the short TS anchored to the membrane, usually an α-helix, they also exhibit globular significant predicted solubility, likely because apart from the short TS anchored to the membrane, usually an α-helix, they domains facing either the cytosol or the periplasm (Figure 7D). Proteins exhibiting two or more TS also exhibit globular domains facing either the cytosol or the periplasm (Figure 7D). Proteins account for the most aggregation prone structures in the analysed sub-proteome. However, it is exhibiting two or more TS account for the most aggregation prone structures in the analysed sub- worth to mention proteome. that itthese However, is worthhighto theoretical mention thatpropensities do not propensities these high theoretical necessarilydo translate into a high not necessarily aggregation translate ininto theacell, high where IM proteins’ aggregation in thehydrophobic cell, where IM stretches remain proteins' protected hydrophobic inside the stretches lipophilic remain membrane protectedenvironment. inside the lipophilic membrane environment. Figure 8. A3D Figure analysis 8. A3D ofofselected analysis selectedrepresentative inner representative inner membrane membrane proteins. proteins. (A) Aquaporin (A) Aquaporin Z (PDB: Z (PDB: 1RC2:A; A3D 1RC2:A; average A3D score average score= =0.417); 0.417); (B) (B) Nitrate/nitrite transporter Nitrate/nitrite transporter NarK NarK (PDB: (PDB: 4JR9:A; 4JR9:A; A3D A3D average average = 0.051); scorescore = 0.051); (C)(C) Acyl-CoAthioester Acyl-CoA thioester hydrolase hydrolase YbgC YbgC(PDB: 1S5U:A; (PDB: A3D 1S5U:A; A3Daverage scorescore average = −0.664); = −0.664); and (D) Septum site-determining protein MinD (PDB: 3Q9L:A; A3D average score = −1.031). The dashed line indicates the existence of two protein groups among inner membrane proteins, exhibiting different STAPs. Lateral view and top view of each protein together with the views of opposite sides are shown. Colour code is as in Figure 1. and (D) Septum site-determining protein MinD (PDB: 3Q9L:A; A3D average score = −1.031). The dashed line indicates the existence of two protein groups among inner membrane proteins, exhibiting different STAPs. Lateral view and top view of each protein together with the views of opposite sides are shown. Cells 2019, 8, 856 Colour code is as in Figure 1. 12 of 20 An analysis of the molecular functions associated to IM proteins with lower and higher An analysis predicted aggregationof the molecularindicates propensity functionsthatassociated they seemtotoIM proteins play differentwith lowerroles, cellular and the higher low predicted aggregation propensity indicates that they seem to play different aggregation-prone ones displaying preferentially catalytic activity, and the high aggregation-prone cellular roles, the low aggregation-prone proteins possessingones mainlydisplaying the role preferentially of transporterscatalytic (Figure 7E).activity, and the high aggregation-prone proteins possessing As IM proteins,mainly the role of transporters outer membrane proteins (OM) (Figure 7E). in a hydrophobic environment, and are located As IM proteins, consequently, they are outer membrane usually thought proteins to have(OM) a high areaggregation located in a tendency. hydrophobic environment, However, their STAPand consequently, lies between thosethey are usually thought of cytoplasmic and to IMhave a high proteins aggregation (Figure 7A,B). tendency. In fact, theHowever, their STAP outer membrane actslies as between a permeablethose of cytoplasmic barrier to hydrophilic andsubstances. IM proteinsIn(Figure general,7A,B). outer In fact, the outer membrane membrane proteins display aacts as a β-barrel permeable structure thatbarrier to hydrophilic encloses a hydrophilic substances. In general, outer cavity surrounded membrane proteins by a hydrophobic display outer layer a β-barrel embedded in structure that encloses the membrane. a hydrophilic A3D structural cavity surrounded predictions are able to by a hydrophobic capture outer architecture. this particular layer embedded In thein the membrane. predictions it canA3D structural be easily seen apredictions are able to capture high aggregation-prone fringe inthis theparticular architecture. outside, restrained to theInexact the predictions it can be easily boundary embedded in theseen a high aggregation-prone membrane, fringe in the outside, flanked by low aggregation-prone restrained regions to the exact that protrude out boundary embedded of the membrane in the9).membrane, (Figure Seen from flanked above, orbybelow, low aggregation-prone it can be observed regions that that the protrude residues out of flanking the the membrane cavity at both(Figure 9). Seen inner and from outerabove, sides or below, itsurfaces conform can be observed that the residues of low aggregation flankingThis propensity. the cavity at both particular inner and assembly is outer achievedsidesbyconform surfaces alternating of low aggregation hydrophobic propensity. and hydrophilic This particular segments in the assembly sequence is[54]. achieved This by alternating explains why hydrophobic when linear and hydrophilic predictors weresegments used to in the sequence analyse [54]. This the aggregation explains why when propensities of theselinear predictors proteins, theywere were used to analyse wrongly the aggregation predicted as being propensities of these than even more soluble proteins, the they were wrongly cytoplasmic proteins predicted as beingineven [34]. In reality, more these soluble proteins, thethan the cytoplasmic scattered proteins [34]. aggregation-prone regionsIn reality, in the in these proteins, sequence the scattered come together in theaggregation-prone structure to allowregions in theinsertion their stable sequence in come the together membrane, in the structure where they to allowprotected remain their stable frominsertion in the membrane, where they remain protected from aggregation. aggregation. Figure 9. A3D analysis of selected representative outer membrane proteins. (A) Usher protein FimD Figure3RFZ:B; (PDB: 9. A3DA3D analysis of selected average representative score = −0.565); outer (B) Outer membrane membrane proteins. protein (A)2F1C:X; G (PDB: Usher protein FimD A3D average (PDB: 3RFZ:B; A3D average score = −0.565); (B) Outer membrane protein G (PDB: 2F1C:X; A3D score = −0.275); (C) Long-chain fatty acid transport protein (PDB: 1T16:A; A3D average score = −0.350) and (D) Fe(3+) dicitrate transport protein FecA (PDB: 1KMO:A; A3D average score = −0.296). Lateral view and top view of each protein together with the views of opposite sides are shown. Colour code is as in Figure 1. average score = −0.275); (C) Long-chain fatty acid transport protein (PDB: 1T16:A; A3D average score = −0.350) and (D) Fe(3+) dicitrate transport protein FecA (PDB: 1KMO:A; A3D average score = −0.296). Lateral view and top view of each protein together with the views of opposite sides are shown. Colour code Cells 2019, is as in Figure 1. 8, 856 13 of 20 3.6. Compositional Determinants of Protein Structural Aggregation Propensity 3.6. Compositional Determinants of Protein Structural Aggregation Propensity To understand whether there is any compositional bias in the surfaces of the 10% most and 10% To understand whether least aggregation-prone E. there is any compositional coli structures, we compared biasthein frequency the surfaces of of thethe2010% mostamino natural and 10% acids least aggregation-prone E. coli at their surfaces. Eleven out of twenty amino acids did not show any significant enrichmentacids structures, we compared the frequency of the 20 natural amino at of in any their surfaces. these two groupsEleven out of10). (Figure twenty amino we However, acids did not found out show any significant that charged residuesenrichment in any (glutamic acid, of lysine these two groups (Figure 10). However, we found out that charged residues (glutamic and arginine) were significantly enriched in the soluble structures’ subset (Figure 10). It is generally acid, lysine and arginine) accepted were that significantlyinteractions charge–charge enriched inare theimportant soluble structures’ for proteinsubset (Figure solubility 10). It is generally by inducing long-range accepted that charge–charge interactions are important for protein solubility repulsion between alike-charged species. This has been clearly demonstrated by introducing multipleby inducing long-range repulsion chargedbetween side chainsalike-charged in proteins,species. Thisthat a strategy hasprovides been clearly demonstrated an increased by introducing resilience multiple towards aggregation, charged side chains in proteins, a strategy that provides an increased resilience particularly at elevated temperatures [55,56]. Conversely, neutralization of surface charges is towards aggregation, particularly suggestedattoelevated temperatures be required [55,56]. for efficient Conversely, amyloid fibril neutralization formation [57]ofand surface chargesand chemical is suggested mutational to neutralization be required for of efficient charges has indeed been shown to promote aggregation [58]. Importantly,ofthe amyloid fibril formation [57] and chemical and mutational neutralization charges has indeed enrichment in thesebeen shown residues charged to promote aggregation is not associated [58]. withImportantly, differencesthe in enrichment in these the pI of soluble and charged residues is not associated aggregation-prone proteins (p = 0.228). with differences in the pI of soluble and aggregation-prone proteins (p = 0.228). Figure 10. Comparison of amino acid composition in soluble and aggregation-prone protein Figure 10. Comparison of amino acid composition in soluble and aggregation-prone protein surfaces. Plot of p-values correlating surface amino composition for two protein groups: 10% surfaces. Plot of p-values correlating surface amino composition for two protein groups: 10% least least and 10% most aggregation-prone proteins. In black: amino acids that do not show significant and 10% most aggregation-prone proteins. In black: amino acids that do not show significant differences. In blue: amino acids that are more abundant in soluble proteins. In red: amino acids that differences. In blue: amino acids that are more abundant in soluble proteins. In red: amino acids that are more abundant in aggregation-prone proteins. are more abundant in aggregation-prone proteins. The aromatic residues (phenylalanine, tyrosine and tryptophan) were enriched in aggregation-prone The (Figure structures aromatic10). residues This (phenylalanine, can be related tyrosine to the factandthat tryptophan) exposed were enriched aromatic residuesin aggregation- facilitate prone structures (Figure 10). This can be related to the fact that exposed aromatic protein–protein interactions and indeed these three residues are more frequent at protein interfacesresidues facilitate protein–protein interactions and indeed these three residues are more frequent at protein than at their surfaces [59]. This property can be partially explained by their ability to establish both π–π interfaces orthan at their π–cation surfaces [59]. interactions, theirThis propertyand flat surfaces can the be partially explained entropic benefit by their of hiding ability them from to water establish both inside π–π or π–cation interactions, their flat surfaces and the entropic benefit of hiding interfaces. However, these properties also imply an increased probability of establishing competing them from water inside interfaces. non-functional However, interactions. Thethese properties trade-off between alsoproper implyand an anomalous increased probability interactionsof willestablishing explain competing non-functional interactions. The trade-off between proper and anomalous why we did not find any hydrophobic aliphatic residue enriched in aggregation-prone structures, since interactions willwill they explain whythe increase weaggregation did not find any hydrophobic potential aliphaticaresidue without providing enriched significant in aggregation-prone counteracting functional structures, since they will increase the aggregation potential without providing advantage. We also identified glycine and proline enriched in the surface of aggregation-prone a significant proteins counteracting functional advantage. We also identified glycine and proline (Figure 10). These two residues are bad β-sheet formers and their presence might counteractenriched in the surface the of aggregation-prone proteins (Figure 10). These two residues are bad β-sheet formers enrichment in aromatic residues by diminishing the protein probability to form intermolecular β-sheets and their contacts leading to the formation of protein aggregates [60]. Cells 2019, 8, 856 14 of 20 presence might counteract the enrichment in aromatic residues by diminishing the protein probability to form intermolecular β-sheets contacts leading to the formation of protein aggregates Cells 2019, 8, 856 14 of 20 [60]. 3.7.3.7. Relationship Relationship Between Between Functional Functional Protein Protein Assemblies Assemblies andand Protein Protein Structural Structural Aggregation Aggregation Propensity Propensity AsAs wewedescribed describedabove,above, there there is is aanegative negativecorrelation correlation between between protein proteinabundance abundance and andSTAP. Indeed, STAP. it has Indeed, it hasbeen been shown shownthat that“supersaturated” “supersaturated”proteins,proteins, which which are are maintained maintainedatata high a high concentration relative to their solubility are among the most aggregation susceptible proteins in in concentration relative to their solubility are among the most aggregation susceptible proteins different different proteomes proteomes [61]. [61]. TheThe presence presence of of supersaturated supersaturated proteins proteins in in thethe cell cell maymay respond respond to tothethe requirements requirements to exert to exert theirtheir biological biological functions. functions. The stoichiometry The stoichiometry of multimeric of multimeric proteinsproteins might imply might that their monomeric subunits would be supersaturated relative to proteins that are active in the cell as in imply that their monomeric subunits would be supersaturated relative to proteins that are active the cell asWe monomers. monomers. calculatedWe calculated a new a new parameter, theparameter, the structural supersaturation structural supersaturation index (SSI), for index (SSI), for the subunits of the subunits oligomeric andof for oligomeric monomeric andproteins for monomeric proteins in our dataset in our11A). (Figure dataset This(Figure parameter 11A). This parameter reflects the risk reflects the risk of a folded protein subunit to aggregate at its physiological of a folded protein subunit to aggregate at its physiological concentration in the cell. The structures concentration in theof cell. The structures monomeric of monomeric proteins proteins displayed displayed significantly lower SSI significantly values than lower that SSI values thanprotein of oligomeric that ofsubunits oligomeric protein subunits (p = 3.29 × 10 −7). Next, we analysed the oligomeric state of protein subunits displaying (p = 3.29 × 10 ). Next, we analysed the oligomeric state of protein subunits displaying the 25% −7 the 25% lowest and lowest highestand SSI highest SSI (Figure (Figure 11B). The low 11B). SSI The group low SSI group contained contained 33.0% 33.0% of the of the monomers monomers and 21.0% of and 21.0% of the the oligomers in oligomers the dataset, inrespectively. the dataset, respectively. In contrast, theIn contrast, high SSIthe grouphighencompassed SSI group encompassed 14.6% of the14.6% of the monomers monomers and 30.0% of and the30.0% of thein oligomers oligomers in the the dataset, dataset, respectively. respectively. Thus, it becomesThus, itclear becomes that the clear that the formation of multimeric proteins comes at a higher risk of aggregation. formation of multimeric proteins comes at a higher risk of aggregation. Then, why these proteins do Then, why these notproteins aggregate doinnot vivoaggregate in vivo under under physiological physiological conditions? conditions? Mainly, because Mainly, the sticky because interfaces the sticky needed for the assembly of the monomeric subunits into a quaternary structure are protected in the native stateare interfaces needed for the assembly of the monomeric subunits into a quaternary structure andprotected in the accordingly thenative state and aggregation risk accordingly the aggregation is only transitory. This implies, risk is onlythat however, transitory. cellular This implies, conditions however, that cellular conditions or genetic changes that favour multimers or genetic changes that favour multimers dissociation would also favour non-functional interactions, dissociation would also favour non-functional interactions, as it occurs in the case of human transthyretin as it occurs in the case of human transthyretin or superoxide dismutase 1, whose dissociation leads to or superoxide dismutase their 1, whose aggregation dissociation and the leads to their onset of amyloidosis aggregation and the onset of amyloidosis [62,63]. [62,63]. Figure 11. Relationship between structural supersaturation index and protein active form. (A) A comparison between cumulative distributions of SSI values for the subunits of proteins active as monomers (dashed black line) or as oligomers (solid red line). (B) A bar graph representing the percentage of monomers (black) or oligomers (red) from the dataset that display low (left panel) and high (right panel) structural supersaturation indexes (SSI). Cells 2019, 8, 856 15 of 20 Figure 11. Relationship between structural supersaturation index and protein active form. (A) A comparison between cumulative distributions of SSI values for the subunits of proteins active as Cells 2019,monomers 8, 856 (dashed black line) or as oligomers (solid red line). (B) A bar graph representing the 15 of 20 percentage of monomers (black) or oligomers (red) from the dataset that display low (left panel) and high (right panel) structural supersaturation indexes (SSI). 3.8. Protein Structural Aggregation Propensity and Bacterial Symmetric Complexes Self-Assembly 3.8. Protein Structural Aggregation Propensity and Bacterial Symmetric Complexes Self-Assembly A recent study has shown that, even when they are in the native and properly folded states, bacterialA symmetric recent study has shown protein that, even complexes are atwhen risk they are in the native of aggregation [64]. Theandauthors properly folded states, introduced point bacterial symmetric protein complexes are at risk of aggregation [64]. The authors mutations at the surface of 12 distinct symmetric complexes from E. coli, resulting in 73 different introduced point mutations at the surface of 12 distinct symmetric complexes from E. coli, resulting in 73 different variants. Some of these mutants resulted in the formation of aggregates when they were expressed variants. Some of these mutants resulted in the formation of aggregates when they were expressed intracellularly. Biophysical measurements and electron microscopy revealed that the aggregated intracellularly. Biophysical measurements and electron microscopy revealed that the aggregated mutants self-assembled in their folded states. Because these mutations were introduced at the surface mutants self-assembled in their folded states. Because these mutations were introduced at the surface of the symmetric protein complexes they could indeed impact their STAP. Therefore, we analysed of the symmetric protein complexes they could indeed impact their STAP. Therefore, we analysed whether A3D predictions were able to detect any difference between the wild type (wt) proteins and whether A3D predictions were able to detect any difference between the wild type (wt) proteins and those mutants those that mutants formed that formedself-assembled self-assembled entities entities intracellularly. intracellularly. AsAsititisisshown shownininFigure Figure12,12, ininallall cases, the average STAP of protein complexes forming aggregates was higher cases, the average STAP of protein complexes forming aggregates was higher than the STAP of thethan the STAP of the correspondent correspondent wt forms. ThisThis wt forms. suggests that the suggests STAP that the ofSTAP protein of complexes is an important protein complexes is an determinant important of their assembly and accordingly that mutations impacting this structural property determinant of their assembly and accordingly that mutations impacting this structural property might trigger folded functional complexes to self-assemble into higher-order structures in the cell. might trigger folded functional complexes to self-assemble into higher-order structures in the cell. Figure 12.12. Figure STAP predictions STAP predictionsfor forbacterial bacterialself-assembled homomerswith self-assembled homomers withdihedral dihedralsymmetry symmetry and and their point their mutants point that aggregate. mutants (A) Bar that aggregate. (A)graphs for indicated Bar graphs supramolecular for indicated complexescomplexes supramolecular comparing A3D average scores comparing for wt proteins A3D average scores and their for wt mutantsand proteins forming either fibres their mutants or focis. forming Corresponding either PDB fibres or focis. accession numbers for analysed proteins are indicated on top of the bar graphs. A3D analysis of representative supramolecular assemblies: (B) dihedral octamer (PDB: 1YAC; A3D average score = −0.567) and its aggregating mutant (D92L, E94L, K98L, K101L; A3D average score = −0.409) and (C) dihedral decamer (PDB: 2IV1; A3D average score = −0.690) and its foci-forming mutant (K24L, K25L, D26L; A3D average score = −0.535). Top-bottom views of each protein are shown. Colour code is as in Figure 1. Cells 2019, 8, 856 16 of 20 4. Discussion It is now well established that, in addition to stability and function, the avoidance of aggregation is an important driver of the evolution of proteins. However, the large majority of studies that provide support to this hypothesis have measured aggregation along protein sequences [34,46,47]. Therefore, they essentially evaluate the aggregation potential of unfolded or at least partially unfolded states. These conformations are of course relevant for aggregation during or immediately after translation or in the case of eventual misfolding; nevertheless, the time that globular proteins expend in these transient states is relatively short in comparison to the one in which they remain in the native state. For a long time, it has been assumed that the attainment of a stable folded conformation in which aggregation-prone sequences are shielded from the solvent constitutes a permanent protection against aggregation. However, it is now clear that proteins can aggregate from their native states, being thus kinetically, but not thermodynamically, stable [23,24]. The present study suggests that, as for protein sequences, aggregation might influence the evolution of protein structures. We acknowledge that the structures we studied here represent only a fraction of the soluble E. coli proteome. However, we expect that the conclusions we delineate would remain valid when more structures are available, since the significance of the observed differences was almost independent of the proportion of proteins selected to build up the subsets in the different analysis and, indeed, those relative to protein abundance are totally consistent with the ones obtained previously analysing a dataset of 397 E. coli structures, from which only 172 had abundance data [2]. It is important to note, however, that our analyses assume minor structural differences to exist between the used static structures and the conformation the respective proteins might adopt in the in vivo environment. Thus, it will be worth re-evaluating them when computational means would allow to model these functional conformations. It has been suggested that, in the cell, globular proteins can remain soluble for long periods of time only because of the presence of high kinetic barriers that separate the native and the aggregated states [24]. Despite speculative, it is tempting to propose that the STAP of proteins has been shaped to keep these barriers high enough to allow them functioning at the concentrations required for an optimal cell fitness (Figure 13A,B). In this context, amino acid changes that, despite not affecting protein stability, would increase STAP, without providing a compensatory fitness benefit, would be purged out by natural selection, since they would decrease the energy barrier between folded and aggregated states and therefore potentially allow aggregation in a biologically relevant time (Figure 13C). The study of the constraints imposed by protein aggregation on natural protein structures according to their abundance, function or tertiary/quaternary organization may uncover novel protein design principles with different potential applications. In this context, our analysis explains why the in vitro concentration of therapeutic globular proteins, like antibodies, above their natural abundance levels results, in many cases, in slow, but irreversible aggregation, precluding their marketing. Designed mutations that decrease the aggregation propensity of their protein surface, without impacting their stability, are anticipated to allow them to remain soluble and functional in the same conditions [65]. Cells 2019, 8, 856 17 of 20 Cells 2019, 8, 856 17 of 20 Figure Figure 13. Schematic representation 13. Schematic of the representation ofenergy diagrams the energy diagramsfor aggregation for aggregation processes.The processes. Theensemble of folded proteins ensemble in solution of folded proteins inis solution considered as one as is considered thermodynamic one thermodynamic system. system. The aggregated state The aggregated state in represents represents all casesin all the cases the global global minimum minimum of of Gibbsenergy. Gibbs energy. However, However, to reach this state, to reach thisproteins state, proteins have tohave to overcome kinetic barriers that differ in their activation energies (Eact). (A) Proteins with low overcome kinetic barriers that differ in their activation energies (Eact ). (A) Proteins with low STAP, such as highly abundant, short, essential and proteins active as monomers are protected from STAP, such as highly abundant, short, essential and proteins active as monomers are protected from aggregation by a high energy barrier. (B) Proteins with higher STAP, such as low abundant, long, aggregation by a high non-essential energyactive and proteins barrier. (B) Proteins in oligomeric formswith higher have to cross aSTAP, such as lower energy low and barrier abundant, thus, long, non-essential and proteins active in oligomeric forms have to cross a lower energy they can potentially access the aggregated state more frequently. (C) Mutations that increase the STAP barrier and thus, they can potentially without access conferring the aggregated additional state morewould functional advantages frequently. be purged(C)out Mutations by naturalthat increase selection, the STAP since withoutthey decrease the conferring energy barrier additional for aggregation functional from the advantages initially would folded state. be purged out by natural selection, since they decrease the energy barrier for aggregation from the initially folded state. Supplementary Materials: The following are available online at www.mdpi.com/2073-4409/8/8/856/s1, Table S1: List of low-aggregation and high-aggregation propensity operons, Table S2: List of the analysed structures together with the information about protein length, abundance, the presence of transmembrane segments, active Supplementary Materials: The following are available online at http://www.mdpi.com/2073-4409/8/8/856/s1, form and essentiality. Table S1: List of low-aggregation and high-aggregation propensity operons, Table S2: List of the analysed structures togetherAuthor with the information Contributions: about protein length, Conceptualization, abundance,S.V., S.V.; Methodology, the A.C. presence of transmembrane and F.P.; segments, Formal Analysis, A.C. F.P. active form andandessentiality. V.I.; Writing-Original Draft Preparation, S.V and A.C.; Writing-Review & Editing, S.V. and F.P.; Supervision, S.V.; Funding Acquisition, S.V. Author Contributions: Conceptualization, S.V.; Methodology, S.V., A.C. and F.P.; Formal Analysis, A.C., F.P. and V.I.; Writing-Original Funding: PleaseDraft Preparation, add: This S.V.funded research was and A.C.; Writing-Review by MINECO & Editing, grant number S.V. and and BIO2016-78310-R F.P.;by Supervision, ICREA S.V.; Fundinggrant Acquisition, S.V. number ICREA-Academia-2016. Funding: Pleaseof Conflicts add: ThisThe Interest: research authorswas funded declare by MINECO no conflict grant of interest. numbersponsors The founding BIO2016-78310-R had no role and in theby ICREA grant design numberofICREA-Academia-2016. the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the decision to publish the results. Conflicts of Interest: The authors declare no conflict of interest. The founding sponsors had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the decision to publish the results. References 1. Deeds, E.J.; Ashenberg, O.; Gerardin, J.; Shakhnovich, E.I. Robust protein protein interactions in crowded cellular environments. Proc. Natl. Acad. Sci. USA 2007, 104, 14952–14957. [CrossRef] [PubMed] Cells 2019, 8, 856 18 of 20 2. Levy, E.D.; De, S.; Teichmann, S.A. Cellular crowding imposes global constraints on the chemistry and evolution of proteomes. Proc. Natl. Acad. Sci. USA 2012, 109, 20461–20466. [CrossRef] [PubMed] 3. Hartl, F.U.; Hayer-Hartl, M. Converging concepts of protein folding in vitro and in vivo. Nat. Struct. Mol. Biol. 2009, 16, 574–581. [CrossRef] [PubMed] 4. Chiti, F.; Dobson, C.M. Protein Misfolding, Amyloid Formation, and Human Disease: A Summary of Progress Over the Last Decade. Annu. Rev. Biochem. 2017, 86, 27–68. [CrossRef] [PubMed] 5. Chapman, E.; Farr, G.W.; Usaite, R.; Furtak, K.; Fenton, W.A.; Chaudhuri, T.K.; Hondorp, E.R.; Matthews, R.G.; Wolf, S.G.; Yates, J.R.; et al. Global aggregation of newly translated proteins in an Escherichia coli strain deficient of the chaperonin GroEL. Proc. Natl. Acad. Sci. USA 2006, 103, 15800–15805. [CrossRef] [PubMed] 6. O’Connell, J.D.; Tsechansky, M.; Royall, A.; Boutz, D.R.; Ellington, A.D.; Marcotte, E.M. A proteomic survey of widespread protein aggregation in yeast. Mol. Biosyst. 2014, 10, 851–861. [CrossRef] [PubMed] 7. David, D.C.; Ollikainen, N.; Trinidad, J.C.; Cary, M.P.; Burlingame, A.L.; Kenyon, C. Widespread protein aggregation as an inherent part of aging in C. elegans. PLoS Biol. 2010, 8, 47–48. [CrossRef] [PubMed] 8. Chiti, F.; Stefani, M.; Taddei, N.; Ramponi, G.; Dobson, C.M. Rationalization of the effects of mutations on peptide andprotein aggregation rates. Nature 2003, 424, 805–808. [CrossRef] 9. Ventura, S. Sequence determinants of protein aggregation: Tools to increase protein solubility. Microb. Cell Fact. 2005, 4, 11. [CrossRef] 10. Reumers, J.; Maurer-Stroh, S.; Schymkowitz, J.; Rousseau, F. Protein sequences encode safeguards against aggregation. Hum. Mutat. 2009, 30, 431–437. [CrossRef] 11. Trainor, K.; Broom, A.; Meiering, E.M. Exploring the relationships between protein sequence, structure and solubility. Curr. Opin. Struct. Biol. 2017, 42, 136–146. [CrossRef] [PubMed] 12. Tartaglia, G.G.; Pechmann, S.; Dobson, C.M.; Vendruscolo, M. Life on the edge: A link between gene expression levels and aggregation rates of human proteins. Trends Biochem. Sci. 2007, 32, 204–206. [CrossRef] [PubMed] 13. Gsponer, J.; Babu, M.M. Cellular Strategies for Regulating Functional and Nonfunctional Protein Aggregation. Cell Rep. 2012, 2, 1425–1437. [CrossRef] [PubMed] 14. Rousseau, F.; Serrano, L.; Schymkowitz, J.W.H. How evolutionary pressure against protein aggregation shaped chaperone specificity. J. Mol. Biol. 2006, 355, 1037–1047. [CrossRef] [PubMed] 15. Espargaró, A.; Castillo, V.; de Groot, N.S.; Ventura, S. The in vivo and in vitro aggregation properties of globular proteins correlate with their conformational stability: The SH3 case. J. Mol. Biol. 2008, 378, 1116–1131. [CrossRef] 16. Drummond, D.A.; Wilke, C.O. Mistranslation-Induced Protein Misfolding as a Dominant Constraint on Coding-Sequence Evolution. Cell 2008, 134, 341–352. [CrossRef] 17. Drummond, D.A.; Wilke, C.O. The evolutionary consequences of erroneous protein synthesis. Nat. Rev. Genet. 2009, 10, 715–724. [CrossRef] 18. Falsone, A.; Falsone, S.F. Legal but lethal: Functional protein aggregation at the verge of toxicity. Front. Cell. Neurosci. 2015, 9, 1–16. [CrossRef] 19. Bruce, J.B.; West, S.A.; Griffin, A.S. Functional amyloids promote retention of public goods in bacteria. Proc. R. Soc. B 2019, 286. [CrossRef] 20. Marcon, G.; Plakoutsi, G.; Canale, C.; Relini, A.; Taddei, N.; Dobson, C.M.; Ramponi, G.; Chiti, F. Amyloid formation from HypF-N under conditions in which the protein is initially in its native state. J. Mol. Biol. 2005, 347, 323–335. [CrossRef] 21. Garcia-Pardo, J.; Graña-Montes, R.; Fernandez-Mendez, M.; Ruyra, A.; Roher, N.; Avilés, F.X.; Lorenzo, J.; Ventura, S. Amyloid formation by human carboxypeptidase D transthyretin-like domain under physiological conditions. J. Biol. Chem. 2014, 289, 33783–33796. [CrossRef] [PubMed] 22. McGuffee, S.R.; Elcock, A.H. Diffusion, crowding & protein stability in a dynamic molecular model of the bacterial cytoplasm. PLoS Comput. Biol. 2010, 6, e1000694. 23. Gazit, E. The “correctly folded” state of proteins: Is it a metastable state? Angew. Chem. Int. Ed. 2002, 41, 257–259. [CrossRef] 24. Ciryam, P.; Kundra, R.; Morimoto, R.I.; Dobson, C.M.; Vendruscolo, M. Supersaturation is a major driving force for protein aggregation in neurodegenerative diseases. Trends Pharmacol. Sci. 2015, 36, 72–77. [CrossRef] [PubMed] Cells 2019, 8, 856 19 of 20 25. Buell, A.K.; Dhulesia, A.; White, D.A.; Knowles, T.P.J.; Dobson, C.M.; Welland, M.E. Detailed Analysis of the Energy Barriers for Amyloid Fibril Growth. Angew. Chem. Int. Ed. 2012, 51, 5247–5251. [CrossRef] [PubMed] 26. Zambrano, R.; Jamroz, M.; Szczasiuk, A.; Pujols, J.; Kmiecik, S.; Ventura, S. AGGRESCAN3D (A3D): Server for prediction of aggregation properties of protein structures. Nucleic Acids Res. 2015, 43, W306–W313. [CrossRef] [PubMed] 27. Pujols, J.; Peña-Díaz, S.; Ventura, S. AGGRESCAN3D: Toward the Prediction of the Aggregation Propensities of Protein Structures. In Computational Drug Discovery and Design; Gore, M., Jagtap, U.B., Eds.; Springer: New York, NY, USA, 2018; pp. 427–443. ISBN 978-1-4939-7756-7. 28. Kuriata, A.; Iglesias, V.; Pujols, J.; Kurcinski, M.; Kmiecik, S.; Ventura, S. Aggrescan3D (A3D) 2.0: Prediction and engineering of protein solubility. Nucleic Acids Res. 2019, 47, 300–307. [CrossRef] 29. Kuriata, A.; Iglesias, V.; Kurcinski, M.; Ventura, S.; Kmiecik, S. Structural bioinformatics Aggrescan3D standalone package for structure-based prediction of protein aggregation properties. Bioinformatics 2019, btz143. 30. Conchillo-Solé, O.; de Groot, N.S.; Avilés, F.X.; Vendrell, J.; Daura, X.; Ventura, S. AGGRESCAN: A server for the prediction and evaluation of “hot spots” of aggregation in polypeptides. BMC Bioinform. 2007, 8, 65. [CrossRef] 31. de Groot, N.S.; Pallarés, I.; Avilés, F.X.; Vendrell, J.; Ventura, S. Prediction of “hot spots” of aggregation in disease-linked polypeptides. BMC Struct. Biol. 2005, 5, 18. [CrossRef] 32. Belli, M.; Ramazzotti, M.; Chiti, F. Prediction of amyloid aggregation in vivo. EMBO Rep. 2011, 12, 657–663. [CrossRef] [PubMed] 33. Ishihama, Y.; Schmidt, T.; Rappsilber, J.; Mann, M.; Hartl, F.U.; Kerner, M.J.; Frishman, D. Protein abundance profiling of the Escherichia coli cytosol. BMC Genom. 2008, 9, 102. [CrossRef] [PubMed] 34. De Groot, N.S.; Ventura, S. Protein aggregation profile of the bacterial cytosol. PLoS ONE 2010, 5, e9383. [CrossRef] [PubMed] 35. Berman, H.M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat, T.N.; Weissig, H.; Shindyalov, I.; Bourne, P. The Protein Data Bank. Nucleic Acids Res. 2000, 28, 235–242. [CrossRef] [PubMed] 36. Wang, M.; Weiss, M.; Simonovic, M.; Haertinger, G.; Schrimpf, S.P.; Hengartner, M.O.; von Mering, C. PaxDb, a Database of Protein Abundance Averages Across All Three Domains of Life. Mol. Cell. Proteom. 2012, 11, 492–500. [CrossRef] 37. Boeckmann, B.; Bairoch, A.; Apweiler, R.; Blatter, M.C.; Estreicher, A.; Gasteiger, E.; Martin, M.J.; Michoud, K.; O’Donovan, C.; Phan, I.; et al. The SWISS-PROT protein knowledgebase and its supplement TrEMBL in 2003. Nucleic Acids Res. 2003, 31, 365–370. [CrossRef] [PubMed] 38. Dennis, G.; Sherman, B.T.; Hosack, D.A.; Yang, J.; Gao, W.; Lane, H.; Lempicki, R.A. DAVID: Database for Annotation, Visualization, and Integrated Discovery. Genome Biol. 2003, 4, R60. [CrossRef] 39. Gerdes, S.; Scholle, M.D.; Campbell, J.W.; Balázsi, G.; Ravasz, E.; Daugherty, M.D.; Somera, A.L.; Kyrpides, N.C.; Anderson, I.; Gelfand, M.S.; et al. Experimental determination and system level analysis of essential genes in Escherichia coli MG1655. J. Bacteriol. 2003, 185, 5673–5684. [CrossRef] 40. Baba, T.; Ara, T.; Hasegawa, M.; Takai, Y.; Okumura, Y.; Baba, M.; Datsenko, K.A.; Tomita, M.; Wanner, B.L.; Mori, H. Construction of Escherichia coli K-12 in-frame, single-gene knockout mutants: The Keio collection. Mol. Syst. Biol. 2006, 2. [CrossRef] 41. Zhou, J.; Rudd, K.E. EcoGene 3.0. Nucleic Acids Res. 2013, 41, D613–D624. [CrossRef] 42. Huerta, A.M.; Salgado, H.; Thieffry, D.; Collado-Vides, J. RegulonDB: A database on transcriptional regulation in Escherichia coli. Nucleic Acids Res. 1998, 26, 55–59. [CrossRef] [PubMed] 43. Schymkowitz, J.; Borg, J.; Stricher, F.; Nys, R.; Rousseau, F.; Serrano, L. The FoldX web server: An online force field. Nucleic Acids Res. 2005, 33, W382–W388. [CrossRef] [PubMed] 44. Meisl, G.; Kirkegaard, J.B.; Arosio, P.; Michaels, T.C.T.; Vendruscolo, M.; Dobson, C.M.; Linse, S.; Knowles, T.P.J. Molecular mechanisms of protein aggregation from global fitting of kinetic models. Nat. Protoc. 2016, 11, 252–272. [CrossRef] [PubMed] 45. Selinger, D.W.; Cheung, K.J.; Mei, R.; Johansson, E.M.; Richmond, C.S.; Blattner, F.R.; Lockhart, D.J.; Church, G.M. RNA expression analysis using a 30 base pair resolution Escherichia coli genome array. Nat. Biotechnol. 2000, 18, 1262–1268. [CrossRef] [PubMed] 46. Monsellier, E.; Ramazzotti, M.; Taddei, N.; Chiti, F. Aggregation propensity of the human proteome. PLoS Comput. Biol. 2008, 4, e1000199. [CrossRef] [PubMed] Cells 2019, 8, 856 20 of 20 47. Tartaglia, G.G.; Vendruscolo, M. Correlation between mRNA expression levels and protein aggregation propensities in subcellular localisations. Mol. Biosyst. 2009, 5, 1873–1876. [CrossRef] 48. Olzscha, H.; Schermann, S.M.; Woerner, A.C.; Pinkert, S.; Hecht, M.H.; Tartaglia, G.G.; Vendruscolo, M.; Hayer-Hartl, M.; Hartl, F.U.; Vabulas, R.M. Amyloid-like aggregates sequester numerous metastable proteins with essential cellular functions. Cell 2011, 144, 67–78. [CrossRef] 49. Albu, R.F.; Chan, G.T.; Zhu, M.; Wong, E.T.C.; Taghizadeh, F.; Hu, X.; Mehran, A.E.; Johnson, J.D.; Gsponer, J.; Mayor, T. A feature analysis of lower solubility proteins in three eukaryotic systems. J. Proteom. 2015, 118, 21–38. [CrossRef] 50. Chen, Y.; Dokholyan, N. V Natural selection against protein aggregation on self-interacting and essential proteins in yeast, fly, and worm. Mol. Biol. Evolut. 2008, 25, 1530–1533. [CrossRef] [PubMed] 51. Linding, R.; Schymkowitz, J.; Rousseau, F.; Diella, F.; Serrano, L. A comparative study of the relationship between protein structure and β-aggregation in globular and intrinsically disordered proteins. J. Mol. Biol. 2004, 342, 345–353. [CrossRef] [PubMed] 52. Dougan, D.A.; Mogk, A.; Bukau, B. Protein folding and degradation in bacteria: To degrade or not to degrade? That is the question. Cell. Mol. Life Sci. 2002, 59, 1607–1616. [CrossRef] [PubMed] 53. Santoni, V.; Molloy, M.; Rabilloud, T. Membrane proteins and proteomics: Un amour impossible? Electrophoresis 2000, 21, 1054–1070. [CrossRef] 54. Cowan, S.W.; Schirmer, T.; Rummel, G.; Steiert, M.; Ghosh, R.; Pauptit, R.A.; Jansonius, J.N.; Rosenbusch, J.P. Crystal structures explain functional properties of two E. coli porins. Nature 1992, 358, 727–733. [CrossRef] [PubMed] 55. Lawrence, M.S.; Phillips, K.J.; Liu, D.R. Supercharging proteins can impart unusual resilience. J. Am. Chem. Soc. 2007, 129, 10110–10112. [CrossRef] [PubMed] 56. Miklos, A.E.; Kluwe, C.; Der, B.S.; Pai, S.; Sircar, A.; Hughes, R.A.; Berrondo, M.; Xu, J.; Codrea, V.; Buckley, P.E.; et al. Structure-based design of supercharged, highly thermoresistant antibodies. Chem. Biol. 2012, 19, 449–455. [CrossRef] [PubMed] 57. Jeppesen, M.D.; Westh, P.; Otzen, D.E. The role of protonation in protein fibrillation. FEBS Lett. 2010, 584, 780–784. [CrossRef] [PubMed] 58. Morshedi, D.; Ebrahim-Habibi, A.; Moosavi-Movahedi, A.A.; Nemat-Gorgani, M. Chemical modification of lysine residues in lysozyme may dramatically influence its amyloid fibrillation. Biochim. Biophys. Acta-Proteins Proteom. 2010, 1804, 714–722. [CrossRef] [PubMed] 59. Mohamed, R.; Degac, J.; Helms, V. Composition of overlapping protein-protein and protein-ligand interfaces. PLoS ONE 2015, 10, e0140965. [CrossRef] [PubMed] 60. Villar-Piqué, A.; Ventura, S. Modeling amyloids in bacteria. Microb. Cell Fact. 2012, 11, 166. [CrossRef] [PubMed] 61. Ciryam, P.; Tartaglia, G.G.; Morimoto, R.I.; Dobson, C.M.; Vendruscolo, M. Widespread Aggregation and Neurodegenerative Diseases Are Associated with Supersaturated Proteins. Cell Rep. 2013, 5, 781–790. [CrossRef] [PubMed] 62. Hörnberg, A.; Eneqvist, T.; Olofsson, A.; Lundgren, E.; Sauer-Eriksson, A.E. A comparative analysis of 23 structures of the amyloidogenic protein transthyretin. J. Mol. Biol. 2000, 302, 649–669. [CrossRef] [PubMed] 63. Chattopadhyay, M.; Durazo, A.; Sohn, S.H.; Strong, C.D.; Gralla, E.B.; Whitelegge, J.P.; Valentine, J.S. Initiation and elongation in fibrillation of ALS-linked superoxide dismutase. Proc. Natl. Acad. Sci. USA 2008, 105, 18663–18668. [CrossRef] [PubMed] 64. Garcia-Seisdedos, H.; Empereur-Mot, C.; Elad, N.; Levy, E.D. Proteins evolve on the edge of supramolecular self-assembly. Nature 2017, 548, 244–247. [CrossRef] [PubMed] 65. Gil-Garcia, M.; Bañó-Polo, M.; Varejão, N.; Jamroz, M.; Kuriata, A.; Díaz-Caballero, M.; Lascorz, J.; Morel, B.; Navarro, S.; Reverter, D.; et al. Combining Structural Aggregation Propensity and Stability Predictions To Redesign Protein Solubility. Mol. Pharm. 2018, 15, 3846–3859. [CrossRef] [PubMed] © 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).