Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 journal homepage: www.elsevier.com/locate/csbj Review Computational prediction of protein aggregation: Advances in proteomics, conformation-specific algorithms and biotechnological applications Jaime Santos, Jordi Pujols, Irantzu Pallarès, Valentín Iglesias, Salvador Ventura ⇑ Institut de Biotecnologia i Biomedicina and Departament de Bioquímica i Biologia Molecular, Universitat Autònoma de Barcelona, Bellaterra, Barcelona, Spain a r t i c l e i n f o a b s t r a c t Article history: Protein aggregation is a widespread phenomenon that stems from the establishment of non-native inter- Received 24 March 2020 molecular contacts resulting in protein precipitation. Despite its deleterious impact on fitness, protein Received in revised form 26 May 2020 aggregation is a generic property of polypeptide chains, indissociable from protein structure and func- Accepted 28 May 2020 tion. Protein aggregation is behind the onset of neurodegenerative disorders and one of the serious obsta- Available online 10 June 2020 cles in the production of protein-based therapeutics. The development of computational tools opened a new avenue to rationalize this phenomenon, enabling prediction of the aggregation propensity of individ- Keywords: ual proteins as well as proteome-wide analysis. These studies spotted aggregation as a major force driv- Protein aggregation Bioinformatics ing protein evolution. Actual algorithms work on both protein sequences and structures, some of them Amyloid accounting also for conformational fluctuations around the native state and the protein microenviron- Protein structure ment. This toolbox allows to delineate conformation-specific routines to assist in the identification of Proteomics aggregation-prone regions and to guide the optimization of more soluble and stable biotherapeutics. Evolution Here we review how the advent of predictive tools has change the way we think and address protein Protein production aggregation. Ó 2020 The Author(s). Published by Elsevier B.V. on behalf of Research Network of Computational and Structural Biotechnology. This is an open access article under the CC BY-NC-ND license (http://creative- commons.org/licenses/by-nc-nd/4.0/). Contents 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1404 2. Proteome-wide analysis: A biological framework for protein aggregation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1404 2.1. Defining the interplay between functional and aberrant contacts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1404 2.2. Computational identification of the evolutive strategies constraining protein aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1405 3. Prediction of protein aggregation from different native conformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1406 3.1. Sequence-based predictors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1406 3.2. Structure-based algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1408 3.3. Oligomeric proteins. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1409 4. Aggregation in the biotechnological production of protein-based therapeutics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1409 5. Modulation of the environmental conditions on the prediction of protein aggregation.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1409 6. Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1410 CRediT authorship contribution statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1411 Declaration of Competing Interest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1411 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1411 Abbreviations: A3D, AGGRESCAN3D; APRs, Aggregation-prone regions; DI, Developability index; IAPP, Islet amyloid polypeptide; IDPs, Intrinsically disordered proteins; mAbs, Monoclonal antibodies; SAP, Spatial aggregation propensity; STAP, STructural Aggregation-Prone region. ⇑ Corresponding author: Institut de Biotecnologia i de Biomedicina Parc de Recerca UAB, Mòdul B Universitat Autònoma de Barcelona E-08193 Bellaterra (Barcelona), Spain. E-mail address: salvador.ventura@uab.es (S. Ventura). https://doi.org/10.1016/j.csbj.2020.05.026 2001-0370/Ó 2020 The Author(s). Published by Elsevier B.V. on behalf of Research Network of Computational and Structural Biotechnology. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 1404 J. Santos et al. / Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 1. Introduction of protein variants or even complete proteomes, which has con- tributed substantially to illuminate the connection between pro- Proteins are the ultimate and essential cellular players in almost tein function and aggregation while uncovering aberrant all biological processes, coordinating different functions inherent aggregation as an important constrain of protein evolution [25– to life through the establishment of molecular networks in the 27]. overcrowded cellular milieu [1]. This activity is mediated by speci- In this article, we review some of the most critical biocomputa- fic inter-molecular interactions that finely regulate protein home- tional advances that have contributed to our present understand- ostasis and functioning [2]. In contrast, non-native protein–protein ing of the constraints shaping non-functional protein aggregation interactions can prompt aberrant oligomerization and ultimately in living organisms, helping to provide biological context for the protein aggregation, a process associated with the onset of a wide protein aggregation phenomenon. We define a framework for pre- range of human disorders -including Alzheimer’s, Parkinson’s dis- dicting protein aggregation, taking into account that function and eases and type II diabetes [3–4]. Such detrimental property is not aggregation are often two sides of the same coin. We intend to pro- restrained to disease-related proteins but widespread and is con- vide a comprehensive compendium of strategies that can be sidered a generic trait of polypeptides chains. Indeed, protein adapted to any specific protein of interest. We end up illustrating aggregation constitutes a significant bottleneck in the production the potential of state-of-the-art algorithms to assist in the design of protein-based therapeutics, compromising their recombinant and control of the solubility of proteins of biotechnological expression, downstream processing, and biosafety, precluding the interest. marketing of otherwise promising biotherapeutics [5–6]. It is therefore essential to elucidate the molecular determinants of pro- 2. Proteome-wide analysis: A biological framework for protein tein aggregation, its biological connection with native functions, aggregation and its role in shaping protein evolution since such knowledge would translate into advances in biomedicine and biotechnology. The implementation of predictive tools with the ability to sys- Proteins aggregate in variety of physicochemical diverse tematically analyze extensive collections of proteins has allowed supramolecular assemblies ranging from highly ordered cross-b extending the analysis of aggregation to complete proteomes, amyloid fibrils, formed by a repetitive array of monomers disposed resulting in a deeper understanding of the molecular determinants orthogonally to the fibril axis, to less ordered amorphous deposits that govern protein aggregation while revealing crosstalk between [7]. In between, an array of oligomeric and protofibrillar structures protein evolution and aggregation [28–30]. Different computa- display intermediate structural properties. The same polypeptide tional proteome-wide analyses converged in the identification of chain can form fibrillar or amorphous assemblies depending on aggregation-prone regions (APRs) similar to those identified in its microenvironment [7–10]. Indeed, the molecular determinants disease-linked proteins across all kingdoms of life [31]. The pres- driving the formation of fibrils in amyloidosis and less ordered ence of these sequence stretches is not anecdotic since an average aggregates during biologics production overlap significantly, mak- of one APR per protein was detected, independently of the consid- ing difficult to predict if the assembly of a given protein in a given ered proteome. APRs have been identified in proteins with differ- condition will lead to one or other kind structure, or something in ent conformational properties: intrinsically disordered, globular, between [7]. Additionally, there exist a number of proteins, known transmembrane, or oligomeric proteins. Overall, these computa- as functional amyloids, that exploit the amyloid fold to perform tional studies suggest that APRs are ubiquitously present in nature their physiological activities [4,11–12]. For instance, functional and are an intrinsic trait of proteins, despite being potentially amyloids are involved in curli-mediated biofilm formation in harmful. Of note, the predicted aggregation propensity of pro- E. coli [13], hypersensitive response activation in plants [14], mel- teomes substantially decreases with increasing complexity and anin polymerization in mammalian cells [15], and hormone stor- longevity of organisms, which seems to point to an evolutionary age in humans [16]. Functional and non-functional aggregation pressure acting against protein aggregation [26,31]. Yet, the aggre- have evolved under selective pressures of different signs, which gation phenomena persist, indicating that negative selection can- favored and disfavored them, respectively [17–19]. This review not wholly abrogate it. focus on undesired aberrant protein aggregation and how it can be predicted and modulated. Great efforts have been devoted to the analysis and characteri- 2.1. Defining the interplay between functional and aberrant contacts. zation of specific aggregation-prone proteins, which crystallized into a robust theoretical comprehension of the protein aggregation Cells have evolved a complex network of quality control mech- phenomenon [20–21]. Yet, it remains challenging to translate anisms to mitigate protein aggregation in order to maintain pro- these learned rules to previously uncharacterized proteins, and tein homeostasis. Such strategies consume significant cellular their study requires the dedicated and manual inspection of their energy [32]. Thus, the omnipresence of APRs in proteins does not sequences and folds. only endorse them with the risk of sporadic aggregation but consti- Fortunately, in silico approaches have evolved hand-in-hand tutively drain a substantial amount of cellular resources. It is then with the development of the field, and have become powerful plat- shocking that despite millions of years of evolution sharpening forms to systematically project our empirical and theoretical protein sequences and structures, protein aggregation has not been knowledge into unstudied protein sequences or structures [22– purged out from polypeptides. This resilience against negative nat- 23]. To date, more than 30 algorithms have been implemented to ural selection has been interpreted as an indication of APRs being deal with protein aggregation, allowing to identify aggregation essential to develop certain biological functions [33]. Multiple determinants, predict the effect of disease-related mutations, and experimental and computational studies converge to demonstrate assist in the redesign of protein solubility [23–24]. Each of these that the physicochemical determinants of protein aggregation sub- programs relies on different principles and assumptions and face stantially overlap with those responsible for the establishment of the aggregation conundrum from diverse perspectives. This diver- native intra- and intermolecular contacts (i.e., substrate binding, sity provides us with a versatile toolbox to orthogonally combine protein folding, or protein–protein interactions) [34–37]. This the outputs of conceptually different algorithms and adapt the pre- observation can be easily understood by considering that regions dictive strategy to the intended purpose. Noteworthy, these pre- responsible for native contacts are usually hydrophobic and prone dictive tools allow for the fast evaluation of extensive collections to establishing hydrogen-bonded networks, two features that also J. Santos et al. / Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 1405 Fig. 1. Innate competition between functional interactions and protein non-functional aggregation. Several factors contribute to balancing this subtle equilibrium. favor the non-native interactions that ultimately lead to protein tional analysis identified other submerged regulation mechanisms aggregation. In globular proteins, the evolutionary suppression of to counterbalance the unavoidable aggregation propensity of pro- APRs is often restrained by the need for a densely packed teins [30–31]. These strategies are adapted to the protein size, hydrophobic core to maintain their native fold [38–39]. The inter- half-life, in-cell relative concentration, sub-cellular location, and face between protein sub-units or protein complexes is also translation rate. enriched in hydrophobic residues that stabilize the quaternary Correlations between protein size and foldability report that structure. Accordingly, protein folding, stability, and aggregation longer and multi-domain proteins usually fold slower than short are in a close interplay in globular proteins, being governed by and single-domain polypeptides [44]. Hypothetically, if all pro- the same molecular features, but differently weighted [35,40]. teins harvest equivalent aggregation loads, longer sequences Computational algorithms have contributed substantially to would be more susceptible to aggregate as a result of more pro- clarify the function-aggregation interplay by assisting the experi- longed exposure of their APRs to solvent [31,45]. Computational mental characterization of some archetypical protein examples. studies of bacterial proteomes revealed that single-domain pro- In this way, in the human Josephin domain, the residues with teins (below 20 kDa) could accommodate a higher aggregation higher contributions to its predicted aggregation propensity are load; the average aggregation propensity decreases with increas- also fundamental for the ubiquitin-binding activity of the protein ing protein lengths [26]. Complementarily, the principal bacterial [41]. Solubilizing mutations affecting those residues lead to a con- chaperones GroEL, and DnaK bind preferentially to substrates comitant loss of activity. Likewise, the aggregation of the human above the 20 kDa limit [46]. The same trend has been reported SUMO protein repertoire (SUMO1, SUMO2, and SUMO3) is directed for the human proteome, highlighting the conservation of a con- by specific regions that overlap with SUMO functional interfaces trol mechanism aimed to cope with the theoretical increased risk [42]. All in all, protein regions accounting for the protein function- of aggregation associated with longer proteins [25]. Aggregation ality might -under some circumstances- lead to aberrant contacts propensity is also connected to protein turnover; De Baets and and eventually to protein aggregation, depicting an intrinsic com- co-workers analyzed 611 protein sequences and their lifetimes petition between these two opposed reactions (Fig. 1). with the aggregation predictor TANGO [47]. They show that short-living proteins can accommodate a higher aggregation load 2.2. Computational identification of the evolutive strategies since their time window to misfold and aggregate is smaller than constraining protein aggregation that of proteins with higher lifetimes. The analysis of the aggregation propensity of proteins in differ- As discussed, almost every polypeptide is endorsed with an ent cellular compartments has documented a connection between inherent risk to suffer aggregation and, potentially compromise protein location and aggregation in yeast, bacterial, and human cellular fitness. However, it is also true that proteins remain sol- proteomes [25–26,48]. Tartaglia and co-workers proposed that uble and functional in their natural contexts, and only under cer- protein solubility correlates with the volume of the cellular com- tain conditions (i.e., mutations, gene duplications, or aging), a partment they populate: in smaller subcellular locations proteins reduced set of proteins are found accumulated into insoluble display lower aggregation tendencies, likely a strategy to prevent deposits in vivo [4,43]. Thus, it becomes clear that if APRs cannot abnormal interactions in more crowded environments [49]. be skipped from protein sequences and structures, alternative evo- Sequences of proteins that undergo secretory pathways or residing lutionary strategies must have emerged to cope with their pres- in the periplasm are, on the average, more soluble, probably ence. Chaperones, co-chaperones, and degradation systems because they are more exposed to extracellular stresses and have constitute the first line of defense against aggregation. In addition little access to protective chaperones, which are significantly to this protein quality control machinery, large-scale computa- depleted in those environments [25–26]. 1406 J. Santos et al. / Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 Elevated protein expression has been traditionally linked with given proteome are limited, and their analysis requires significant the formation of aberrant protein deposits, either in conforma- computation time. Despite these limitations, in a recent work, we tional disorders or during recombinant expression [50–52]. analyzed a fraction of the structurally characterized Escherichia coli According to the law of mass action, the probability of establishing proteome to explore whether, in addition to sequence properties, non-functional interactions scales with the protein concentration structural aggregation might also influence the evolution of bacte- [53]. Certainly, in contrast to protein folding, aggregation is a sec- rial proteins [61]. Our analysis revealed that the aggregation fea- ond or higher-order reaction, being strongly dependent on protein tures of protein surfaces and interfaces in folded states are abundance. Several proteomics analyses revealed that there exists constrained according to the protein abundance, length, essential- an anti-correlation between gene-expression and/or protein abun- ity, subcellular location, and function. This observation indicates dance and predicted aggregation propensity in bacteria, nema- that protein structures would have also evolved to minimize the todes, and humans [49,54–56]. In essence, protein abundance in risk of aggregation in their natural environments. the cell is tightly regulated to attain optimal levels, sufficient for proteins to remain soluble and functional, but not more than that. 3. Prediction of protein aggregation from different native Vendruscolo and co-workers have extended this hypothesis by conformations proposing that ‘‘supersaturated” proteins -a subset of proteins liv- ing above their solubility limit- conform a metastable subpro- The previous section illustrates how protein aggregation cannot teome inherently exposed to aggregation [57–59]. They suggest be understood without considering the folding, functional purpose, that after the primary aggregation of the disease-causing amyloid and cellular environment of a protein. In each conformational proteins -i.e., Ab42 or a-synuclein-, ‘‘supersaturated” proteins are state, the risk of aggregation stems from different sources; globular collaterally more exposed to aggregation, expanding the dysfunc- proteins, IDPs, and oligomeric proteins pose different challenges tion to other unrelated biological pathways and prompting a gen- that need to be addressed with dedicated tools. Therefore, in order eral collapse of cellular functions. to anticipate protein aggregation successfully, we need to adapt Computational analyses also revealed a link between protein our computational scheme to the particular properties of the pro- aggregation and function [26,48,60]. Chen and co-workers demon- tein under study. Such a task can be difficult for untrained users strated that essential proteins from three eukaryotic organisms - since an in-depth knowledge of the available computational tools yeast, fly, and nematode- are the subject of a higher evolutionary is needed. In this section, we apply the insights provided by pressure against protein deposition [60]. Similarly, in bacteria, proteome-wide analysis to classify and review a state-of-the-art operons that encode essential proteins or functions display lower collection of predictive tools. The aim is to establish a systematic aggregation loads [26]. It is not surprising that proteins in the same framework for evaluating protein aggregation that can be adapted operon display similar aggregation propensities since they share to the intended predictive purpose (Fig. 2). common gene-expression regulation and thus abundance, all working in related functions. The above-described relationships have been established by 3.1. Sequence-based predictors analyzing aggregation over protein sequences, without taking into account the modulation of these properties by the structural con- The first generation of computational algorithms designed to text in the folded states in which proteins spend the majority of predict protein aggregation is based on the identification of linear their lifetime. Performing an equivalent structure-based proteomic APRs across the polypeptide sequence. The conceptual pillars of analysis is not trivial since the available protein structures for a these algorithms are the theoretical and experimental studies that Fig. 2. Computational strategies to predict protein aggregation. In each folding state, aggregation is driven by different molecular determinants, delimiting the best- performing predictive strategy in each particular case. Aggregation-prone residues are colored in red and solubilizing amino acids in blue. APR and STAP designate Aggregation-Prone Regions and STructural Aggregation-prone Regions, respectively. PDB structures correspond to monomeric and tetrameric transthyretin (PDB: 1F41). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) J. Santos et al. / Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 1407 Table 1 Sequence based-prediction methods, according to the rationale behind their analysis. *Registration prior to analysis is required. Method* Underlying rationale Webserver, software or equation Phenomenological methods AGGRESCAN [63,64] Prediction is assayed against an aggregation propensity scale for the 20 http://bioinf.uab.es/aggrescan/ proteinogenic amino acids derived from in vivo experiments. Zyggregator [65] Prediction of a 21-residue sliding window from an equation http://www-mvsoftware.ch.cam.ac.uk/index.php/login* accounting for hydrophobicity, secondary structure propensity, and net charge built upon changing aggregation rate on mutations.It also considers the presence of gatekeeper residues or hydrophobic patches Theoretical methods TANGO [7,117] Evaluation of the population of random coil, native conformation or http://tango.crg.es/ * aggregated species from empirically and statistically derived conformational amino acidic preferences, along with physico-chemical variables. PASTA 2.0 [66,118] Energetic function derived from high-resolution protein structures, http://protein.bio.unipd.it/pasta2/ which considers interaction potential and H-bond formation between all non-consecutive residues for parallel and anti-parallel b-pairing. FoldAmyloid [67] A protein structure derived scale; from the notion that hydrophobic http://bioinfo.protres.ru/fold-amyloid/ stretches exhibit higher ‘‘packing density” and H-bonding propensity. WALTZ [68] Application of a position specific matrix derived from a large group of https://waltz.switchlab.org/ hexapeptides, for predicting amyloid-like formation. Pafig [119] Analysis of six-residue sliding window for a scale derived from Code can be downloaded from their web page http:// machine supervised learning over 531 physicochemical properties, www.mobioinfor.cn/pafig/ (Requires MS Windows) which led to best discrimination using 41 of them. Betascan [120] Evaluation of b-strand pairing propensity, obtained from probabilities http://cb.csail.mit.edu/cb/betascan/ hosts the web server and of residues to be H-bonded in amphiphilic b-sheets. allows download of the Perl scipt. GAP [121] Discriminates amyloid-like or b-amorphous hexapeptides from https://www.iitm.ac.in/bioinfo/GAP/ position-specific pairing frequencies. 3D Profile [122] Energetic impact on the spatial accommodation to the backbone of the http://services.mbi.ucla.edu/zipperdb/ * fibril forming Sup35 hexapeptide is assessed. Machine learning methods APPNN [71] Machine learning approach based on the analysis of seven http://cran.r-project.org/web/packages/appnn/index.html physicochemical and biochemical features such as b-sheet frequency, hydrophobic moment, helix termination parameters or isoelectric point. NetCSSP [72] Analysis of contact-dependent secondary structure prediction to http://cssp2.sookmyung.ac.kr/ identify hidden b-propensities. FiSH Amyloid [73] Classification of amyloidogenic stretches based on co-ocurrence http://www.comprec.pwr.wroc.pl/COMPREC_home_page. patterns in protein sequences. html Consensus methods AmylPred2 [74] Generates consensus predictions over 11 algorithms but allows user- http://aias.biol.uoa.gr/AMYLPRED2/ * customized predictions as some methodologies can have a certain degree of redundancy, thus biasing the consensus prediction. MetAmyl [75] Score is obtained applying a linear combination of four predictors’ http://metamyl.genouest.org (which showed lower redundancy) outcome, weighting the individual contribution of each method. * This list intends to be illustrative and not to provide an extensive enumeration and description of all available methods. Programs in this list are not necessarily more accurate than those absent. allowed the definition of the main molecular determinants of second class, and they evaluate the tendency of a sequence to aggregation. To date, more than 20 sequential algorithms have adopt a defined b-enriched conformation, the packing density of been developed [23,62]. Their design exploits the evidence that proteins, the composition and patterning of their residues or the protein aggregation is driven by short and well-defined sequential suitability to adopt the topologically restricted conformations that stretches -referred to as APRs- characterized by a high hydropho- characterize amyloid-like states [66–69]. A growing number of bicity, low net charge, and a remarkable preference to adopt b- machine learning methods are being developed. They exploit sheet secondary structure [23]. However, each algorithm relies the potential of neural networks to identify sequential features on different interpretations and weighting of the essential features highly correlated with aggregation [70]. Machine learning driving aggregation, which allows the orthogonal combination of approaches attain performances comparable -or even higher- to conceptually different algorithms to reduce method-specific traditional predictors. APPNN, netCSSP or FISH Amyloid are some biases. These algorithms can be divided into four subclasses examples of this kind of algorithms [71–73]. Finally, a last group according to their underlying principles, further described in of software is the one based on the combination and weighting Table 1. Briefly, the first class of algorithms, known as phe- of the outputs of other predictors (either phenomenological or the- nomenological, employs experimental data to define the determi- oretical) into one single output. These consensus predictors nants of aggregation and provide empiric aggregation scales. They assemble the different concepts behind each predictor to increase include software such as AGGRESCAN, and Zyggregator, both based robustness while reducing method-specific biases. AMYLPRED 2 on the rationalization of experimentally determined factors influ- and MetAmyl exemplify this kind of software [74–75]. encing protein aggregation [63–65]. The second class of computa- The aforementioned programs are just our lab selection among tional approaches relies on the theoretical assessment of sequence the more than 20 available algorithms that have demonstrated properties known to be associated with aggregation. TANGO, their efficacy in the study of disease-related proteins, allowing to PASTA 2.0, FoldAmyloid, Waltz and Amyloid Mutants belong to this discretize the experimentally relevant sequence stretches driving 1408 J. Santos et al. / Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 their aggregation [4,23,29,31]. They are these tools that allowed for ously discussed for the Josephin domain and the SUMO proteins. most of the above aforementioned proteome-wide analysis [31]. This is also the case of antigen-recognition elements in antibodies However, they do not evaluate the modulation of sequential aggre- that exhibit a relatively high exposed hydrophobicity required for gation determinants imposed by the three-dimensional conforma- target binding [87]. Of course, STAPs cannot be identified by linear tion of the protein. This drawback defines the particular scenarios predictors. in which their predictions are particularly accurate: (i) Intrinsically Structure-based algorithms use three-dimensional protein disordered proteins; these proteins lack a defined three- coordinates in order to evaluate the aggregation of proteins in their dimensional structure and fluctuate between multiple transients native fold and overcome the limitations mentioned above. Herein, unfolded or partially folded conformations. During these dynamic we briefly describe the principles of four of the more popular fluctuations, APRs are accessible to the solvent, without significant structure-based algorithms. Their applications to the redesign of structural protection. (ii) After being synthesized in the ribosome, therapeutic proteins will be further reviewed in section 4. proteins depart from an extended conformation that transits SolubiS constitutes one of the most instinctive evolutions of a towards the folded state. During this process, APRs are also linear predictor to evaluate the structural context [88]. SolubiS exposed to the solvent. This situation is of particular relevance dur- identifies linear APRs -using the TANGO sequence-based ing protein overexpression, where the transient concentration of algorithm- and applies the FOLDX force field to evaluate their con- unfolded polypeptide chains increases dramatically (iii) Dynamic tribution to global protein stability [89]. The result is an algorithm fluctuations or destabilization of protein structures may result in able to analyze the structural context and the relative shielding of a the partial unfolding, exposing previously hidden APRs. given APR and provide an estimation of the tendency of such APR Therefore, the straightest application of sequence-based algo- to be solvent-exposed and thus become aggregation competent. rithms is the prediction of IDPs aggregation. IDPs are particularly SolubiS has successfully forecasted the behavior of model globular depleted in APRs since the compositional bias of disordered pro- proteins in vitro and is an excellent tool to evaluate the impact of teins inherently protects them from aggregation. In essence, they APR sequential variations in the solubility of the protein. Nonethe- contain a significant proportion of protective residues (Asp, Arg, less, since SolubiS was build using a sequence-based predictor, this Glu, Gly, Lys, and Pro), so-called gatekeepers, that are difficult to software is still blinded to the emergence of STAPs. SAP (spatial accommodate in a b-sheet aggregated state [7,76–78]. However, aggregation propensity) was the first algorithm designed with this IDPs are not entirely protected against aggregation, and this stems objective, being able to identify hydrophobic patches exposed to from their innate biological functions. IDPs are generally involved solvent in globular proteins [90]. SAP uses a structurally corrected in multiple protein–protein interactions, acting as signal integra- hydrophobicity scale as a proxy for protein aggregation and con- tors and master regulators of diverse biological processes [79]. siders that the aggregation contribution of a given side chain is Such activity entails the presence of short molecular recognition modulated for the residues in the vicinity. Under that premise, motifs that must be at least transiently exposed to the solvent to SAP defines a 5 Å radius sphere centered in the analyzed atom find their suitable binding partner. Those motifs retain an intrinsic and evaluates the spatial contribution of each nearby residue as hydrophobicity, and under pathogenic conditions, they may act as the product of the solvent-accessible area of the atoms within cryptic APRs, triggering aberrant interactions and finally, protein the sphere. In this way, SAP analyzes the local hydrophobicity of aggregation. Accordingly, computationally identified APRs in IDPs solvent-exposed residues regardless of their sequence positions. tend to overlap with interaction motifs; in a recent work, we have The Developability Index algorithm (DI) is an adaptation of SAP computationally identified and characterized the aggregation of to predict the aggregation of monoclonal antibodies (mAbs) based one of such sequence stretches [80]. Ab42, a-synuclein or IAPP on their structure in a faster and more accurate way [91]. DI are some examples of IDPs whose aggregation is associated with includes the effect of electrostatic interactions, which have an human disorders [4,81]. In these proteins, different algorithms essential role in solubility, counterbalancing the contribution of consistently predict APRs that overlap with the regions identified hydrophobicity. Aggrescan3D (A3D) is another example of to drive the aggregation in vitro and lately to be part of the core structure-based algorithms, implementing the sequence-based that sustains the respective amyloid fibrils structure [82–85]. AGGRESCAN aggregation scale to assess the aggregation of protein structures [92–93]. A3D is conceptually similar to SAP, but it intro- 3.2. Structure-based algorithms duces an experimentally derived aggregation scale instead of using hydrophobicity. A3D calculates the aggregation propensity of each Structure-based algorithms were born as a second generation of residue by computing its intrinsic aggregation propensity, which is software designed to translate the predictive potential of also corrected by its solvent-exposure and properties of the side sequence-based algorithms to globular proteins. In folded proteins, chains in the vicinity. The main particularity of A3D, when com- the three-dimensional arrangement of the polypeptide chain sig- pared with other structure-based algorithms, is its capacity to nificantly reshapes the molecular determinants of aggregation, assess the impact of dynamic structural fluctuations of protein weighting the contribution of linear APRs [86]. Sequence-based structures and its effect on the exposition of APRs. The A3D ‘‘dy- algorithms are blinded to these effects, which usually result in namic mode” uses the CABS-flex force field -based on high resolu- overprediction when they are used to forecast the aggregation tion coarse-grained molecular dynamics simulations- to reproduce propensity of folded proteins. In this way, the architecture of glob- the dynamism and plasticity of protein structures in their native ular proteins buries the hydrophobic residues inside the protein states [94–95]. For each particular conformer generated by CABS- core, blurring the exposition of linear APRs to solvent. Thus, once flex, A3D computes its structural aggregation propensity. In this a protein is folded into its compact tertiary structure, those APRs way, A3D allows the identification of dynamic aggregation-prone do not contribute to protein aggregation, although sequence- regions that are otherwise protected in the static PDB deposited based algorithms would predict the contrary. Moreover, in a folded structure [96]. Another structure-based approach has been devel- protein, neighbor residues do not need to be consecutive in oped by Sormanni and coworkers [97] The CamSol method applies sequence, and structural clustering of non-consecutive hydropho- the physicochemical principles implemented in the sequence- bic residues in the surface or interface of the protein might occur. based predictor Zyggregator and performs structural corrections Those solvent-exposed hydrophobic patches, known as STructural similar to those of SAP or A3D to compute the solubility of native APRs (STAP), are usually crucial for the protein activity, as previ- protein structures [65]. J. Santos et al. / Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 1409 3.3. Oligomeric proteins characterized protein candidates producing more soluble, stable, and long-lasting therapeutic proteins (Fig. 3). Several of the algo- The computational analysis of native oligomeric assemblies rithms cited in the previous section have been exploited to screen revealed that, as a general trend, protein–protein interfaces display for more soluble protein variants and/or introduce solubilizing higher aggregation propensities than the solvent-exposed surfaces mutations. Human a-galactoside or Bacillus anthracis protective of globular proteins [37]. Indeed, a remarkable overlap between antigen are examples of biotherapeutic proteins whose solubility interaction surfaces and APRs has been identified in oligomeric pro- has been successfully redesigned by using SolubiS, reducing their teins associated with conformational disorders, indicating that the aggregation propensity while maintaining their functional activity functional interaction of the monomeric subunits is associated with [106]. Such approximation has also been translated to the redesign an intrinsic risk of aggregation. This exposed hydrophobicity in of mAbs, which currently represent the faster-growing class of bio- monomeric subunits becomes masked once they are incorporated therapeutic molecules. SolubiS has successfully ranked a set of in the quaternary structure. Accordingly, in disease-linked oligo- mAbs according to their aggregation, also being able to apply this meric proteins, pathological mutations usually impact the complex predictive potential to the redesigning of these complex macro- stability, favoring the dissociation of the aggregation-prone mono- molecules into more soluble variants [87]. As a dedicated tool for meric constituents. This is the case of transthyretin and SOD1, for antibody predictions, the DI algorithm also allows the screening which the disentangle of the quaternary structure is the rate- of the aggregation propensity of mAbs in the early phases of drug limiting step in the downhill polymerization process that drives discovery; thus, the development of pre-screened candidates can their aggregation; once the tetramer is destabilized, aggregation be prioritized attending to their solubility. becomes highly favorable [98–99]. Besides, upon disassembly, the Finally, A3D offers a user-friendly platform for the design of monomeric subunits become ‘‘supersaturated,” in comparison with protein solubility by integrating structural aggregation predictions the respective multimeric protein, which exacerbates the probabil- with stability prediction since, eventually, solubilizing mutations ity of spurious intermolecular contacts. For these proteins, in addi- at the protein surfaces may have destabilizing effects on the native tion to STAP detection and scoring, accurate predictions of structure, even if the involved residues are fully exposed to solvent aggregation should also evaluate the thermodynamic stability of [107]. Towards the optimization of such a feature, we have the assembled protein and how amino acid substitutions impact recently updated A3D with an ‘‘Automated mutations” mode that both factors. In this context, the FOLDX force field is useful for the automatically ranks solubilizing mutations in the protein surfaces identification of both stabilizing and destabilizing mutations asso- according to their effect on aggregation and stability. This new ciated with the aggregation of oligomeric proteins [89]. module was tested for the redesign of a human variable heavy chain of the human antibody germline, allowing the identification of solubilizing mutations that do not perturb the stability of the 4. Aggregation in the biotechnological production of protein- antibody fragment. This routine will significantly reduce the time based therapeutics. dedicated to the visual inspection of protein structures and manual selection of mutations [107]. Of note, A3D users can pre-exclude The use of recombinant proteins for therapeutic applications the complementarity-determining regions (CDRs) from the auto- offers a compelling alternative to small molecules. Proteins are mated round of analysis. This is important because several previ- capable of performing highly specific and intricate functions, ous phage display initiatives aimed to increase Abs solubility which is impossible for small molecule drugs. The high specificity have resulted in variants in which the mutations cluster in or close of proteins also results in less drug toxicity through interference to the hydrophobic CDRs, with the subsequent risk of hampering with normal body processes. Aggregation is a significant concern antigen-recognition [108]. The automated A3D selected mutations in the production of protein therapeutics, and it may jeopardize lay sequentially and structurally far from the CDRs, yet they signif- the viability of the complete biotechnological process [5,100– icantly increased the solubility of the Ab of interest when this was 101]. First, because aggregation reduces production yields, but incubated under harsh conditions. The main advantage of this most importantly because aggregates have the potential to trigger upgrade is that it permits its use to academic and industrial fellows immunogenic responses upon administration, threatening that did not have extensive training in protein redesign. patients’ health [102–103]. Consequently, biotech companies allo- cate extensive funding and resources to mitigate protein aggrega- tion in their production pipelines; often by undertaking expensive 5. Modulation of the environmental conditions on the trial/error screenings of conditions that may or may not result in a prediction of protein aggregation. significant improvement. On top of that, aggregation can occur at every step of the process, from expression and purification to for- The vast majority of proteins reside and develop their functions mulation and storage [104]. These aggregation-related issues stem in highly complex and overcrowded cellular microenvironments from a simple principle: proteins are not selected to remain soluble exposed to unpredictable stresses that may impact their solubility out of their evolutionary context. In an external environment, the [52,109]. However, in the absence of external stressors, cellular natural competition between functional and aberrant contacts is conditions tend to remain relatively constant and the extrinsic fac- imbalanced by the absence of native binding partners, quality con- tors impacting protein aggregation are not expected to change sig- trol systems, and deregulation of protein concentration, leading to nificantly over time. Thus, even if the prediction of aggregation uncontrolled intracellular protein deposition. Indeed, because in vivo is far from being trivial, for the computational evaluation protein-based drugs should be administered in most cases par- of this property, it can be considered that we face a defined and enterally, protein concentration in the final formulation can exceed constant environment. In stark contrast, during industrial manu- 200 mg/mL, being several orders of magnitude above their natural facturing, proteins are often exposed to very different conditions abundances [105]. Besides, during industrial processes, polypep- that affect their physical stability [110]. For instance, more than tides are exposed to unnatural stresses such as pH-changes, shear- 65% of antibodies and related constructs are formulated at ing effects, or temperature fluctuations, impacting both their pH < 6.5 [5,111]. Yet, only a small set of in silico tools have been solubility and stability. designed to address the protein background, and many of them In this framework, the computational prediction of protein only collaterally with barely parametrized functions. The aggregation offers an avenue to work with pre-selected and well- sequence-based predictors TANGO and Zyggregator can evaluate 1410 J. Santos et al. / Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 Fig. 3. Comparison between a computationally guided pipeline for optimizing protein-based biotherapeutics and currently used strategies. The computational analysis of a candidate pool and/or the reduction of their aggregation by introducing solubilizing mutations offers a powerful alternative to expensive and blinded trial/error approaches, being cost-effective strategies to increase the success rate in the development of protein-based therapeutics. the effect of pH on protein net charge, but their performance in endeavors. We expect that this model would inspire the imple- predicting the pH-dependent solubility of experimentally charac- mentation of novel structure-based algorithms, including this fea- terized proteins is relatively low; mostly because they do not com- ture, to develop more robust and versatile software. pute the partial charge of the side chains [7,65,112–113]. The first algorithm in which pH was implicitly considered is DI. DI can eval- uate the net charge -accounting for partial ionization- of mAbs in a 6. Conclusions pH-dependent way. In such a way, DI can evaluate if mAbs with similar solubility in neutral conditions would have different solu- The widespread nature of aggregation influences every aspect of bility in a particular step of the purification, thus extending the protein function and applications either in vivo or in vitro, from dis- pre-selection and optimization of candidates to a more realistic ease onset to the production of biotherapeutics. Accordingly, it level. One of the main disadvantages of this approach and other becomes fundamental to develop new tools to systematically related applications is that they only evaluate the effect of pH in address protein aggregation issues in a fast and reliable way, with terms of its modulation over protein net charge. However, it is the ultimate objective of transforming this arbitrary and unpre- well-known that the protonation/deprotonation of the ionizable dictable process into an anticipatable variable. Nowadays, in silico residues also has a significant effect in the hydrophobicity, as illus- approximations are being progressively integrated into the routi- trated by their shifts in the partition coefficients upon changes on nes of many laboratories, being cost-effective tools to assist and the solution pH [114–115]. On that basis, we have recently built a nurture experimental efforts. Likewise, it can be expected that they novel phenomenological model to predict the effect of pH on the will also gain relevance in the biotechnological arena, replacing or aggregation of IDPs considering both fluctuations in hydrophobic- complementing the actual costly and limited experimental meth- ity and global net charge [116]. Our algorithm calculates the pH- ods used to optimize protein solubility. In the near future, we could dependent hydrophobicity along the sequence at a given pH and expect that novel automated and more robust algorithms account- computes the effect of the global protein net charge in order to ing from conditions extrinsic to the protein sequence and static profile the pH-dependent solubility of a given protein. Such an structure would be progressively implemented in the streamlined approach has demonstrated a higher predictive potential than workflow of companies as an easy and fast optimization step for mere charge-dependent approaches stressing the importance of protein-based therapeutics; impacting both their marketing and including the changes in hydrophobicity in future predictive clinical application. J. Santos et al. / Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 1411 CRediT authorship contribution statement Structure to Function with Bioinformatics (ed. Rigden, D.J.) 205-263 (Springer, 2017).doi: 10.1007/978-94-024-1069-3_7 [24] Santos J, Iglesias V, Ventura S. Computational prediction and redesign of Jaime Santos: Data curation, Visualization, Writing - original aberrant protein oligomerization. Prog Mol Biol Transl Sci 2020;169:43–83. draft. Jordi Pujols: Writing - original draft. Irantzu Pallarès: Writ- https://doi.org/10.1016/bs.pmbts.2019.11.002. [25] Monsellier E, Ramazzotti M, Taddei N, Chiti F. Aggregation propensity of the ing - original draft, Visualization, Supervision. Valentín Iglesias: human proteome. PLoS Comput Biol 2008;4:. https://doi.org/10.1371/journal. Writing - original draft, Visualization, Software. Salvador Ventura: pcbi.1000199e1000199. Conceptualization, Supervision, Funding acquisition, Writing - [26] de Groot NS, Ventura S. Protein aggregation profile of the bacterial cytosol. PLoS ONE 2010;5:. https://doi.org/10.1371/journal.pone.0009383e9383. review & editing. [27] Pallares I, Ventura S. Understanding and predicting protein misfolding and aggregation: Insights from proteomics. Proteomics 2016;16:2570–81. Declaration of Competing Interest https://doi.org/10.1002/pmic.201500529. [28] Redler RL et al. Computational approaches to understanding protein aggregation in neurodegeneration. J Mol Cell Biol 2014;6:104–15. https:// The authors declare that they have no known competing finan- doi.org/10.1093/jmcb/mju007. cial interests or personal relationships that could have appeared [29] Buck PM, Kumar S, Singh SK. On the role of aggregation prone regions in to influence the work reported in this paper. protein evolution, stability, and enzymatic catalysis: insights from diverse analyses. PLoS Comput Biol 2013;9:. https://doi.org/10.1371/journal. pcbi.1003291e1003291. References [30] Monsellier E, Chiti F. Prevention of amyloid-like aggregation as a driving force of protein evolution. EMBO Rep 2007;8:737–42. https://doi.org/10.1038/sj. [1] Huttlin EL et al. Architecture of the human interactome defines protein embor.7401034. communities and disease networks. Nature 2017;545:505–9. https://doi.org/ [31] Castillo V, Grana-Montes R, Sabate R, Ventura S. Prediction of the aggregation 10.1038/nature22366. propensity of proteins from the primary sequence: aggregation properties of [2] Yamada T, Bork P. Evolution of biomolecular networks: lessons from proteomes. Biotechnol J 2011;6:674–85. https://doi.org/10.1002/ metabolic and protein interactions. Nat Rev Mol Cell Biol 2009;10:791–803. biot.201000331. https://doi.org/10.1038/nrm2787. [32] Sherman MY, Goldberg AL. Cellular defenses against unfolded proteins: a cell [3] Eisenberg D, Jucker M. The amyloid state of proteins in human diseases. Cell biologist thinks about neurodegenerative diseases. Neuron 2001;29:15–32. 2012;148:1188–203. https://doi.org/10.1016/j.cell.2012.02.022. https://doi.org/10.1016/s0896-6273(01)00177-5. [4] Chiti F, Dobson CM. Protein Misfolding, Amyloid Formation, and Human [33] Pastore A, Temussi PA. The two faces of Janus: functional interactions and Disease: A Summary of Progress Over the Last Decade. Annu Rev Biochem protein aggregation. Curr Opin Struct Biol 2012;22:30–7. https://doi.org/ 2017;86:27–68. https://doi.org/10.1146/annurev-biochem-061516-045115. 10.1016/j.sbi.2011.11.007. [5] Roberts CJ. Therapeutic protein aggregation: mechanisms, design, and [34] Linding R, Schymkowitz J, Rousseau F, Diella F, Serrano L. A comparative control. Trends Biotechnol 2014;32:372–80. https://doi.org/10.1016/j. study of the relationship between protein structure and beta-aggregation in tibtech.2014.05.005. globular and intrinsically disordered proteins. J Mol Biol 2004;342:345–53. [6] Roberts CJ. Protein aggregation and its impact on product quality. Curr Opin https://doi.org/10.1016/j.jmb.2004.06.088. Biotechnol 2014;30:211–7. https://doi.org/10.1016/j.copbio.2014.08.001. [35] Castillo V, Ventura S. Amyloidogenic regions and interaction surfaces overlap [7] Rousseau F, Schymkowitz J, Serrano L. Protein aggregation and amyloidosis: in globular proteins related to conformational diseases. PLoS Comput Biol confusion of the kinds?. Curr Opin Struct Biol 2006;16:118–26. https://doi. 2009;5:. https://doi.org/10.1371/journal.pcbi.1000476e1000476. org/10.1016/j.sbi.2006.01.011. [36] Castillo V, Chiti F, Ventura S. The N-terminal helix controls the transition [8] Alam P, Bousset L, Melki R, Otzen DE. alpha-synuclein oligomers and fibrils: a between the soluble and amyloid states of an FF domain. PLoS ONE 2013;8:. spectrum of species, a spectrum of toxicities. J Neurochem 2019;150:522–34. https://doi.org/10.1371/journal.pone.0058297e58297. https://doi.org/10.1111/jnc.14808. [37] Pechmann S, Levy ED, Tartaglia GG, Vendruscolo M. Physicochemical [9] Frare E et al. Characterization of oligomeric species on the aggregation principles that regulate the competition between functional and pathway of human lysozyme. J Mol Biol 2009;387:17–27. https://doi.org/ dysfunctional association of proteins. Proc Natl Acad Sci U S A 10.1016/j.jmb.2009.01.049. 2009;106:10159–64. https://doi.org/10.1073/pnas.0812414106. [10] Vetri V et al. Amyloid fibrils formation and amorphous aggregation in [38] Castillo V, Espargaro A, Gordo V, Vendrell J, Ventura S. Deciphering the role of concanavalin A. Biophys Chem 2007;125:184–90. https://doi.org/10.1016/j. the thermodynamic and kinetic stabilities of SH3 domains on their bpc.2006.07.012. aggregation inside bacteria. Proteomics 2010;10:4172–85. https://doi.org/ [11] Avni A, Swasthi HM, Majumdar A, Mukhopadhyay S. Intrinsically disordered 10.1002/pmic.201000260. proteins in the formation of functional amyloids from bacteria to humans. [39] Fraga H, Grana-Montes R, Illa R, Covaleda G, Ventura S. Association between Prog Mol Biol Transl Sci 2019;166:109–43. https://doi.org/10.1016/bs. foldability and aggregation propensity in small disulfide-rich proteins. Antioxid pmbts.2019.05.005. Redox Signal 2014;21:368–83. https://doi.org/10.1089/ars.2013.5543. [12] Otzen D. Functional amyloid: turning swords into plowshares. Prion [40] Yan C, Wu F, Jernigan RL, Dobbs D, Honavar V. Characterization of protein- 2010;4:256–64. https://doi.org/10.4161/pri.4.4.13676. protein interfaces. Protein J 2008;27:59–70. https://doi.org/10.1007/s10930- [13] Chapman MR et al. Role of Escherichia coli curli operons in directing amyloid 007-9108-x. fiber formation. Science 2002;295:851–5. https://doi.org/ [41] Masino L, Nicastro G, Calder L, Vendruscolo M, Pastore A. Functional 10.1126/science.1067484. interactions as a survival strategy against abnormal aggregation. FASEB J [14] Oh J et al. Amyloidogenesis of type III-dependent harpins from plant 2011;25:45–54. https://doi.org/10.1096/fj.10-161208. pathogenic bacteria. J Biol Chem 2007;282:13601–9. https://doi.org/ [42] Sabate R, Espargaro A, Grana-Montes R, Reverter D, Ventura S. Native 10.1074/jbc.M602576200. structure protects SUMO proteins from aggregation into amyloid fibrils. [15] Fowler DM et al. Functional amyloid formation within mammalian tissue. Biomacromolecules 2012;13:1916–26. https://doi.org/10.1021/bm3004385. PLoS Biol 2006;4:. https://doi.org/10.1371/journal.pbio.0040006e6. [43] Stroo E, Koopman M, Nollen EA, Mata-Cabana A. Cellular Regulation of [16] Maji SK et al. Functional amyloids as natural storage of peptide hormones in Amyloid Formation in Aging and Disease. Front Neurosci 2017;11:64. https:// pituitary secretory granules. Science 2009;325:328–32. https://doi.org/ doi.org/10.3389/fnins.2017.00064. 10.1126/science.1173155. [44] Ivankov DN et al. Contact order revisited: influence of protein size on the [17] Maury CP. The emerging concept of functional amyloid. J Intern Med folding rate. Protein Sci 2003;12:2057–62. https://doi.org/10.1110/ 2009;265:329–34. https://doi.org/10.1111/j.1365-2796.2008.02068.x. ps.0302503. [18] Houben, B. et al. Autonomous aggregation suppression by acidic residues [45] Watters AL et al. The highly cooperative folding of small naturally occurring explains why chaperones favour basic residues. EMBO J, e102864 (2020).doi: proteins is likely the result of natural selection. Cell 2007;128:613–24. 10.15252/embj.2019102864 https://doi.org/10.1016/j.cell.2006.12.042. [19] Sanchez de Groot, N. et al. Evolutionary selection for protein aggregation. [46] Ellis RJ. Chaperone substrates inside the cell. Trends Biochem Sci Biochem Soc Trans 40, 1032-7 (2012).doi: 10.1042/BST20120160 2000;25:210–2. https://doi.org/10.1016/s0968-0004(00)01576-0. [20] Nooren IM, Thornton JM. Diversity of protein-protein interactions. EMBO J [47] De Baets G et al. An evolutionary trade-off between protein turnover rate and 2003;22:3486–92. https://doi.org/10.1093/emboj/cdg359. protein aggregation favors a higher aggregation propensity in fast degrading [21] Eichner T, Radford SE. A diversity of assembly mechanisms of a generic proteins. PLoS Comput Biol 2011;7:. https://doi.org/10.1371/journal. amyloid fold. Mol Cell 2011;43:8–18. https://doi.org/10.1016/ pcbi.1002090e1002090. j.molcel.2011.05.012. [48] Tartaglia, G.G. & Caflisch, A. Computational analysis of the S. cerevisiae [22] Pallares I, Ventura S. Advances in the Prediction of Protein Aggregation proteome reveals the function and cellular localization of the least and most Propensity. Curr Med Chem 2019;26:3911–20. https://doi.org/10.2174/ amyloidogenic proteins. Proteins 68, 273-8 (2007).doi: 10.1002/prot.21427 0929867324666170705121754. [49] Tartaglia GG, Vendruscolo M. Correlation between mRNA expression levels [23] Ricardo Graña-Montes, J.P.-P., Carlota Gómez-Picanyol & Ventura, a.S. and protein aggregation propensities in subcellular localisations. Mol Biosyst Prediction of Protein Aggregation and Amyloid Formation. in From Protein 2009;5:1873–6. https://doi.org/10.1039/b913099n. 1412 J. Santos et al. / Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 [50] Conrad B, Antonarakis SE. Gene duplication: a drive for phenotypic diversity [78] De Baets G, Van Durme J, Rousseau F, Schymkowitz J. A genome-wide and cause of human disease. Annu Rev Genomics Hum Genet 2007;8:17–35. sequence-structure analysis suggests aggregation gatekeepers constitute an https://doi.org/10.1146/annurev.genom.8.021307.110233. evolutionary constrained functional class. J Mol Biol 2014;426:2405–12. [51] Hardy J. Amyloid double trouble. Nat Genet 2006;38:11–2. https://doi.org/ https://doi.org/10.1016/j.jmb.2014.04.007. 10.1038/ng0106-11. [79] Wright PE, Dyson HJ. Intrinsically disordered proteins in cellular signalling [52] Wang W, Nema S, Teagarden D. Protein aggregation–pathways and and regulation. Nat Rev Mol Cell Biol 2015;16:18–29. https://doi.org/ influencing factors. Int J Pharm 2010;390:89–99. https://doi.org/10.1016/j. 10.1038/nrm3920. ijpharm.2010.02.025. [80] Pujols J, Santos J, Pallares I, Ventura S. The Disordered C-Terminus of Yeast [53] Levy ED, De S, Teichmann SA. Cellular crowding imposes global constraints on Hsf1 Contains a Cryptic Low-Complexity Amyloidogenic Region. Int J Mol Sci the chemistry and evolution of proteomes. Proc Natl Acad Sci U S A 2018;19. https://doi.org/10.3390/ijms19051384. 2012;109:20461–6. https://doi.org/10.1073/pnas.1209312109. [81] Coskuner O, Uversky VN. Intrinsically disordered proteins in various [54] Vecchi G et al. Proteome-wide observation of the phenomenon of life on the hypotheses on the pathogenesis of Alzheimer’s and Parkinson’s diseases. edge of solubility. Proc Natl Acad Sci U S A 2020;117:1015–20. https://doi. Prog Mol Biol Transl Sci 2019;166:145–223. https://doi.org/10.1016/bs. org/10.1073/pnas.1910444117. pmbts.2019.05.007. [55] Castillo V, Grana-Montes R, Ventura S. The aggregation properties of [82] Li Y et al. Amyloid fibril structure of alpha-synuclein determined by cryo- Escherichia coli proteins associated with their cellular abundance. electron microscopy. Cell Res 2018;28:897–903. https://doi.org/10.1038/ Biotechnol J 2011;6:752–60. https://doi.org/10.1002/biot.201100014. s41422-018-0075-x. [56] Tartaglia GG, Pechmann S, Dobson CM, Vendruscolo M. Life on the edge: a [83] Fitzpatrick AWP et al. Cryo-EM structures of tau filaments from Alzheimer’s link between gene expression levels and aggregation rates of human proteins. disease. Nature 2017;547:185–90. https://doi.org/10.1038/nature23002. Trends Biochem Sci 2007;32:204–6. https://doi.org/10.1016/j. [84] Paravastu AK, Leapman RD, Yau WM, Tycko R. Molecular structural basis for tibs.2007.03.005. polymorphism in Alzheimer’s beta-amyloid fibrils. Proc Natl Acad Sci U S A [57] Ciryam P, Tartaglia GG, Morimoto RI, Dobson CM, Vendruscolo M. 2008;105:18349–54. https://doi.org/10.1073/pnas.0806270105. Widespread aggregation and neurodegenerative diseases are associated [85] Fitzpatrick AW, Saibil HR. Cryo-EM of amyloid fibrils and cellular aggregates. with supersaturated proteins. Cell Rep 2013;5:781–90. https://doi.org/ Curr Opin Struct Biol 2019;58:34–42. https://doi.org/10.1016/j. 10.1016/j.celrep.2013.09.043. sbi.2019.05.003. [58] Ciryam P, Kundra R, Morimoto RI, Dobson CM, Vendruscolo M. [86] Navarro S, Ventura S. Computational re-design of protein structures to Supersaturation is a major driving force for protein aggregation in improve solubility. Expert Opin Drug Discov 2019;14:1077–88. https://doi. neurodegenerative diseases. Trends Pharmacol Sci 2015;36:72–7. https:// org/10.1080/17460441.2019.1637413. doi.org/10.1016/j.tips.2014.12.004. [87] van der Kant R et al. Prediction and Reduction of the Aggregation of [59] Kundra R, Ciryam P, Morimoto RI, Dobson CM, Vendruscolo M. Protein Monoclonal Antibodies. J Mol Biol 2017;429:1244–61. https://doi.org/ homeostasis of a metastable subproteome associated with Alzheimer’s 10.1016/j.jmb.2017.03.014. disease. Proc Natl Acad Sci U S A 2017;114:E5703–11. https://doi.org/ [88] Van Durme J et al. Solubis: a webserver to reduce protein aggregation 10.1073/pnas.1618417114. through mutation. Protein Eng Des Sel 2016;29:285–9. https://doi.org/ [60] Chen Y, Dokholyan NV. Natural selection against protein aggregation on self- 10.1093/protein/gzw019. interacting and essential proteins in yeast, fly, and worm. Mol Biol Evol [89] Schymkowitz J et al. The FoldX web server: an online force field. Nucleic Acids 2008;25:1530–3. https://doi.org/10.1093/molbev/msn122. Res 2005;33:W382–8. https://doi.org/10.1093/nar/gki387. [61] Carija A, Pinheiro F, Iglesias V, Ventura S. Computational Assessment of [90] Chennamsetty N, Voynov V, Kayser V, Helk B, Trout BL. Design of therapeutic Bacterial Protein Structures Indicates a Selection Against Aggregation. Cells proteins with enhanced stability. Proc Natl Acad Sci U S A 2019;8. https://doi.org/10.3390/cells8080856. 2009;106:11937–42. https://doi.org/10.1073/pnas.0904191106. [62] Z, L.A. & R, M.M.B. Structure and Aggregation Mechanisms in Amyloids. [91] Lauer TM et al. Developability index: a rapid in silico tool for the screening of Molecules 25(2020).doi: 10.3390/molecules25051195 antibody aggregation propensity. J Pharm Sci 2012;101:102–15. https://doi. [63] Conchillo-Sole O et al. AGGRESCAN: a server for the prediction and evaluation org/10.1002/jps.22758. of ‘‘hot spots” of aggregation in polypeptides. BMC Bioinf 2007;8:65. https:// [92] Kuriata, A. et al. Aggrescan3D (A3D) 2.0: prediction and engineering of doi.org/10.1186/1471-2105-8-65. protein solubility. Nucleic Acids Res 47, W300-W307 (2019).doi: [64] Sanchez de Groot, N., Pallares, I., Aviles, F.X., Vendrell, J. & Ventura, S. 10.1093/nar/gkz321 Prediction of ‘‘hot spots” of aggregation in disease-linked polypeptides. BMC [93] Kuriata A, Iglesias V, Kurcinski M, Ventura S, Kmiecik S. Aggrescan3D Struct Biol 5, 18 (2005).doi: 10.1186/1472-6807-5-18 standalone package for structure-based prediction of protein aggregation [65] Tartaglia GG, Vendruscolo M. The Zyggregator method for predicting protein properties. Bioinformatics 2019;35:3834–5. https://doi.org/10.1093/ aggregation propensities. Chem Soc Rev 2008;37:1395–401. https://doi.org/ bioinformatics/btz143. 10.1039/b706784b. [94] Jamroz M, Kolinski A, Kmiecik S. CABS-flex: Server for fast simulation of [66] Walsh, I., Seno, F., Tosatto, S.C. & Trovato, A. PASTA 2.0: an improved server protein structure fluctuations. Nucleic Acids Res 2013;41:W427–31. https:// for protein aggregation prediction. Nucleic Acids Res 42, W301-7 (2014).doi: doi.org/10.1093/nar/gkt332. 10.1093/nar/gku399 [95] Kuriata, A. et al. CABS-flex 2.0: a web server for fast simulations of flexibility [67] Garbuzynskiy SO, Lobanov MY, Galzitskaya OV. FoldAmyloid: a method of of protein structures. Nucleic Acids Res 46, W338-W343 (2018).doi: prediction of amyloidogenic regions from protein sequence. Bioinformatics 10.1093/nar/gky356 2010;26:326–32. https://doi.org/10.1093/bioinformatics/btp691. [96] Pujols J, Pena-Diaz S, Ventura S. AGGRESCAN3D: Toward the Prediction of the [68] Maurer-Stroh S et al. Exploring the sequence determinants of amyloid Aggregation Propensities of Protein Structures. Methods Mol Biol structure using position-specific scoring matrices. Nat Methods 2018;1762:427–43. https://doi.org/10.1007/978-1-4939-7756-7_21. 2010;7:237–42. https://doi.org/10.1038/nmeth.1432. [97] Sormanni P, Aprile FA, Vendruscolo M. The CamSol method of rational design [69] O’Donnell, C.W. et al. A method for probing the mutational landscape of of protein mutants with enhanced solubility. J Mol Biol 2015;427:478–90. amyloid structure. Bioinformatics 27, i34-42 (2011).doi: 10.1093/ https://doi.org/10.1016/j.jmb.2014.09.026. bioinformatics/btr238 [98] Hurshman AR, White JT, Powers ET, Kelly JW. Transthyretin aggregation [70] Stanislawski, J., Kotulska, M. & Unold, O. Machine learning methods can under partially denaturing conditions is a downhill polymerization. replace 3D profile method in classification of amyloidogenic hexapeptides. Biochemistry 2004;43:7365–81. https://doi.org/10.1021/bi049621l. BMC Bioinformatics 14, 21 (2013).doi: 10.1186/1471-2105-14-21 [99] Nordlund A, Oliveberg M. SOD1-associated ALS: a promising system for [71] Familia C, Dennison SR, Quintas A, Phoenix DA. Prediction of Peptide and elucidating the origin of protein-misfolding disease. HFSP J 2008;2:354–64. Protein Propensity for Amyloid Formation. PLoS ONE 2015;10:. https://doi. https://doi.org/10.2976/1.2995726. org/10.1371/journal.pone.0134679e0134679. [100] Shah M. Commentary: New perspectives on protein aggregation during [72] Kim C, Choi J, Lee SJ, Welsh WJ, Yoon S. NetCSSP: web application for Biopharmaceutical development. Int J Pharm 2018;552:1–6. https://doi.org/ predicting chameleon sequences and amyloid fibril formation. Nucleic Acids 10.1016/j.ijpharm.2018.09.049. Res 2009;37:W469–73. https://doi.org/10.1093/nar/gkp351. [101] den Engelsman J et al. Strategies for the assessment of protein aggregates in [73] Gasior P, Kotulska M. FISH Amyloid - a new method for finding amyloidogenic pharmaceutical biotech product development. Pharm Res 2011;28:920–33. segments in proteins based on site specific co-occurrence of aminoacids. BMC https://doi.org/10.1007/s11095-010-0297-1. Bioinf 2014;15:54. https://doi.org/10.1186/1471-2105-15-54. [102] Ratanji KD, Derrick JP, Dearman RJ, Kimber I. Immunogenicity of therapeutic [74] Tsolis AC, Papandreou NC, Iconomidou VA, Hamodrakas SJ. A consensus proteins: influence of aggregation. J Immunotoxicol 2014;11:99–109. https:// method for the prediction of ’aggregation-prone’ peptides in globular proteins. doi.org/10.3109/1547691X.2013.821564. PLoS ONE 2013;8:. https://doi.org/10.1371/journal.pone.0054175e54175. [103] FDA. Guidance for Industry Immunogenicity Assessment for Therapeutic [75] Emily M, Talvas A, Delamarche C. MetAmyl: a METa-predictor for AMYLoid Protein Products. (2014).doi: proteins. PLoS ONE 2013;8:. https://doi.org/10.1371/journal. [104] Cromwell ME, Hilario E, Jacobson F. Protein aggregation and bioprocessing. pone.0079722e79722. AAPS J 2006;8:E572–9. https://doi.org/10.1208/aapsj080366. [76] Uversky VN, Gillespie JR, Fink AL. Why are ‘‘natively unfolded” proteins [105] Schermeyer MT, Woll AK, Kokke B, Eppink M, Hubbuch J. Characterization of unstructured under physiologic conditions?. Proteins 2000;41:415–27. highly concentrated antibody solution - A toolbox for the description of https://doi.org/10.1002/1097-0134(20001115)41:3<415::aid-prot130>3.0. protein long-term solution stability. MAbs 2017;9:1169–85. https://doi.org/ co;2-7. 10.1080/19420862.2017.1338222. [77] Dyson HJ. Making Sense of Intrinsically Disordered Proteins. Biophys J [106] Ganesan A et al. Structural hot spots for the solubility of globular proteins. 2016;110:1013–6. https://doi.org/10.1016/j.bpj.2016.01.030. Nat Commun 2016;7:10816. https://doi.org/10.1038/ncomms10816. J. Santos et al. / Computational and Structural Biotechnology Journal 18 (2020) 1403–1413 1413 [107] Gil-Garcia M et al. Combining Structural Aggregation Propensity and Stability [116] Santos J et al. pH-Dependent Aggregation in Intrinsically Disordered Proteins Predictions To Redesign Protein Solubility. Mol Pharm 2018;15:3846–59. Is Determined by Charge and Lipophilicity. Cells 2020;9. https://doi.org/ https://doi.org/10.1021/acs.molpharmaceut.8b00341. 10.3390/cells9010145. [108] Sidhu SS. Phage display in pharmaceutical biotechnology. Curr Opin [117] Fernandez-Escamilla AM, Rousseau F, Schymkowitz J, Serrano L. Prediction of Biotechnol 2000;11:610–6. https://doi.org/10.1016/s0958-1669(00)00152-x. sequence-dependent and mutational effects on the aggregation of peptides [109] Breydo L, Redington JM, Uversky VN. Effects of Intrinsic and Extrinsic Factors and proteins. Nat Biotechnol 2004;22:1302–6. https://doi.org/10.1038/ on Aggregation of Physiologically Important Intrinsically Disordered Proteins. nbt1012. Int Rev Cell Mol Biol 2017;329:145–85. https://doi.org/10.1016/bs. [118] Trovato A, Chiti F, Maritan A, Seno F. Insight into the structure of amyloid ircmb.2016.08.011. fibrils from the analysis of globular proteins. PLoS Comput Biol 2006;2:. [110] Zapadka, K.L., Becher, F.J., Gomes Dos Santos, A.L. & Jackson, S.E. Factors https://doi.org/10.1371/journal.pcbi.0020170e170. affecting the physical stability (aggregation) of peptide therapeutics. [119] Tian, J., Wu, N., Guo, J. & Fan, Y. Prediction of amyloid fibril-forming segments Interface Focus 7, 20170030 (2017).doi: 10.1098/rsfs.2017.0030 based on a support vector machine. BMC Bioinformatics 10 Suppl 1, S45 [111] Wang W, Singh S, Zeng DL, King K, Nema S. Antibody structure, instability, (2009).doi: 10.1186/1471-2105-10-S1-S45 and formulation. J Pharm Sci 2007;96:1–26. https://doi.org/10.1002/ [120] Bryan Jr AW, Menke M, Cowen LJ, Lindquist SL, Berger B. BETASCAN: probable jps.20727. beta-amyloids identified by pairwise probabilistic analysis. PLoS Comput Biol [112] Chiti F, Stefani M, Taddei N, Ramponi G, Dobson CM. Rationalization of the 2009;5:. https://doi.org/10.1371/journal.pcbi.1000333e1000333. effects of mutations on peptide and protein aggregation rates. Nature [121] Thangakani AM, Kumar S, Nagarajan R, Velmurugan D, Gromiha MM. GAP: 2003;424:805–8. https://doi.org/10.1038/nature01891. towards almost 100 percent prediction for beta-strand-mediated aggregating [113] DuBay KF et al. Prediction of the absolute aggregation rates of amyloidogenic peptides with distinct morphologies. Bioinformatics 2014;30:1983–90. polypeptide chains. J Mol Biol 2004;341:1317–26. https://doi.org/10.1016/j. https://doi.org/10.1093/bioinformatics/btu167. jmb.2004.06.043. [122] Thompson MJ et al. The 3D profile method for identifying fibril-forming [114] Simm S, Einloft J, Mirus O, Schleiff E. 50 years of amino acid hydrophobicity segments of proteins. Proc Natl Acad Sci U S A 2006;103:4074–8. https://doi. scales: revisiting the capacity for peptide classification. Biol Res 2016;49:31. org/10.1073/pnas.0511295103. https://doi.org/10.1186/s40659-016-0092-5. [115] MacCallum JL, Tieleman DP. Hydrophobicity scales: a thermodynamic looking glass into lipid-protein interactions. Trends Biochem Sci 2011;36:653–62. https://doi.org/10.1016/j.tibs.2011.08.003.