Available online at www.sciencedirect.com Current Opinion in ScienceDirect Structural Biology Computational methods to predict protein aggregation Susanna Navarro and Salvador Ventura Abstract Intracellular assemblies present different natures In most cases, protein aggregation stems from the establish- ranging from biomolecular condensates, which are ment of non-native intermolecular contacts. The formation of multi-component, dynamic, and reversible assemblages, insoluble protein aggregates is associated with many human often formed via liquideliquid phase separation, to diseases and is a major bottleneck for the industrial production irreversible protein aggregates, like amyloids [2]. During of protein-based therapeutics. Strikingly, fibrillar aggregates protein assembly, native contacts might preserve the are naturally exploited for structural scaffolding or to generate surface conformation that interacts to drive the supra- molecular switches and can be artificially engineered to build molecular structure, in a process known as agglomera- up multi-functional nanomaterials. Thus, there is a high interest tion [3,4]; however, more often, protomers undergo in rationalizing and forecasting protein aggregation. Here, we partial or global unfolding, and native contacts are review the available computational toolbox to predict protein replaced by non-native intermolecular interactions aggregation propensities, identify sequential or structural leading to the formation of non-structured amorphous aggregation-prone regions, evaluate the impact of mutations aggregates or highly ordered amyloid fibrils, character- on aggregation or recognize prion-like domains. We discuss ized by a cross-b conformation formed by the repetitive the strengths and limitations of these algorithms and how they stacking of b-strand monomers perpendicularly to the can evolve in the next future. fibril axis [5]. Addresses The formation of insoluble aggregates is often associ- Institut de Biotecnologia I de Biomedicina, Departament de Bioquímica ated with a loss of protein activity and/or a gain in I Biologia Molecular, Universitat Autònoma de Barcelona, 08193, Bellaterra, Barcelona, Spain toxicity, and is intimately related to a broad array of diseases and aging, including neurodegenerative disor- Corresponding author: Ventura, Salvador (salvador.ventura@uab.es) ders, such as Alzheimer’s and Parkinson’s [6], and non- neuronal localized or systemic diseases [7]. Current Opinion in Structural Biology 2022, 73:102343 Protein aggregation constitutes a major bottleneck in This review comes from a themed issue on Folding and Binding producing protein-based therapeutics and biotechnolog- Edited by Fabrizio Chiti and Anna Sablina ical products. During manufacturing, proteins are exposed For a complete overview see the Issue and the Editorial to unnatural stresses and formulated at concentrations far Available online xxx from their physiological abundance, which exacerbate https://doi.org/10.1016/j.sbi.2022.102343 their aggregation potential. Aggregation directly reduces production yields, final product activities and might trig- 0959-440X/© 2022 The Author(s). Published by Elsevier Ltd. This is an open access ar ticle under the CC BY-NC-ND license (http:// gerunpredictable immune responses in patients [8]. creativecommons.org/licenses/by-nc-nd/4.0/). Although amyloids are best known for their deleterious effects, they have also evolved to play a pivotal role in biological functions that cannot be attained by globular proteins, from bacteria to humans [9e12]. These ac- Introduction tivities, associated with their common cross-b architec- Proteins are the most abundant biomolecules in living ture, have become a source of inspiration for designing systems playing a pivotal role in most biological pro- amyloid-based materials with a broad range of applica- cesses. The establishment of functional native intra- tions in nanotechnology [13e15]. and interchain interactions is a fundamental aspect of protein biology, governing protein folding, binding, and Extensive research has been devoted to understanding activity. In the crowded environment of living cells, the molecular determinants behind uncontrolled and proteins transit through different conformational states aberrant protein oligomerization. However, empirical in the search for a free energy minimum, which can methods are costly, time-consuming, and limited by the correspond to a monomeric state or a wide variety of availability of proper protein models. In this context, assemblies [1], depending on the polypeptide and computational methods have emerged as powerful cellular conditions. complementary tools for studying aggregation at the www.sciencedirect.com Current Opinion in Structural Biology 2022, 73:102343 2 Folding and Binding individual protein and proteome scales [16]. Under- Exploiting sequential features to predict protein standing and anticipating protein aggregation has a aggregation translational impact in five main areas: pathology, evo- Protein sequences encode the information for folding lution, protein manufacturing, development of novel into native globular states but also to remain partially or nanomaterials, and unveiling the physiological role fully unstructured in the case of intrinsically disordered played by functional amyloids. proteins (IDPs). This linear code overlaps with the one accounting for the formation of the cross-b fold common Here, we provide a compendium of computational to amyloids and thus first-generation aggregation pre- methods available to predict aggregation propensity, dictors were aimed to read aggregation propensities identify aggregation-prone regions, evaluate the impact from the primary sequence. of mutations on aggregation and solubility, and for prion- like proteins identification. We also discuss how the To date, more than 20 prediction algorithms exist for the incorporation of yet unattended parameters might recognition of linear APRs in polypeptides using sundry improve such predictions. metrics to derive aggregation propensity, solubility, and thermodynamic stability. All of them use protein se- Prediction of aggregation propensities in quences as an input, but they differ on the strategies protein sequences employed for APRs identification and quantification, Intrinsic determinants behind protein aggregation relying on either phenomenologically or theoretically Many of the molecular determinants governing protein derived parameters. A first set of algorithms exploits scales aggregation are naturally imprinted in the primary of aggregation propensity for amino acids determined sequence and depend on intrinsic properties of the experimentally, exemplified by AGGRESCAN [21] or polypeptide chain, including the amino acid composi- Zyggregator [22]. A second set of methods combines tion or residues patterning. Hydrophobicity is a major different key features of APRs, such as specific physico- driving force in the establishment of intrachain and chemical properties of amino acids, their distribution interchain contacts leading to aggregation, whereas the along the sequence or the feasibility to adopt a b-sheet protein local and net charges often play an opposing role, conformation, employing particular functions to weigh promoting electrostatic repulsion between individual each of these parameters. They include Waltz [23], residues or molecules. Most amyloid aggregates adopt a SALSA [24], PAGE [25], Tango [26], FoldAmyloid [27] cross-b fold, implying that aggregation is favoured by and PASTA [28], among others. For a description of first- amino acids with high b-sheet propensity and counter- generation methods see Table 1. acted by b-sheet breaking residues, such as Pro and Gly. However, the aggregation propensity of a protein is not Given the intricacy of protein aggregation reactions, equally distributed along the sequence and tends to averaging the predictions provided by algorithms relying concentrate in short linear stretches, known as hot spots on distinct strategies seems reasonable. This is the or aggregation-prone regions (APRs). These regions are concept behind AmylPred2, which combines the out- sufficient and necessary to nucleate protein aggregation. puts of eleven different algorithms, identifying an APR They are often enriched in hydrophobic residues and, in when the stretch is predicted by at least n/2 of the globular proteins, they generally map at the hydrophobic methods [29]. In METAMYL, scores are weighted using core. However, exposed APRs involved in catalysis and a logistic regression model arising from the combination binding have also been described. The presence of APRs of four predictive methods [30] (Table 1). is a requirement for protein stability and it seems that positive selection for an amyloid structure might be Artificial intelligence to evaluate sequential behind the emergence of globular folds [17], their po- aggregation propensity tency being modulated by the presence of adjacent The need to train and perfect linear predictors boosted residues of low aggregation propensity, known as gate- the generation of datasets and databases compiling keepers. In addition, aggregation propensities are experimentally validated aggregating regions in peptides balanced by extrinsic factors, such as pH, ionic strength, and proteins [31,32]. Most of these repositories provide temperature, or even gravity [18]. This dependence on a binary classification regarding the capacity of the the environment is often behind the frequently proteins or segments to assemble or not into amyloid observed amyloid polymorphism [19,20]. fibrils. Thus, they constitute the perfect substrate for machine learning classifiers, which can automatically Our increasing understanding of the intrinsic and uncover unrelated features of polypeptide chains. The extrinsic factors governing protein aggregation and their effectiveness of these algorithms heavily depends on the interplay has crystallized in the development of in silico accurate annotation of the reference training data, methods to predict this phenomenon departing both although a certain degree of misannotated instances still from sequences and structures. allows robust predictions, which suggests that, indeed, Current Opinion in Structural Biology 2022, 73:102343 www.sciencedirect.com Protein aggregation prediction Navarro and Ventura 3 Table 1 Computational methods to predict protein aggregation. Algorithm Characteristics URL and Ref Sequence-based methods AGGRESCAN Intracellular aggregation propensity scale for each of the 20 http://bioinf.uab.es/aggrescan [21] amino acids derived after introducing single point mutations on Ab42 peptide Zyggregator Relative propensities for aggregation based on hydrophobicity, [22] secondary structure propensity, hydrophobic/hydrophilic patterning, net charge, the presence of gatekeepers and the influence of structural protection Waltz Statistical method that exploits Position-Specific Substitution https://waltz.switchlab.org/ [23] Matrices (PSSM) obtained from amyloidogenic hexapeptides, combined with physicochemical properties of b-amyloids and conformational features of amyloid backbone structures SALSA b-strand propensity is calculated from a b-strand contiguity http://amypdb.genouest.org/e107_ score based on Chou and Fasman’s secondary structure plugins/amypdb_aggregation/db_ propensity scale and applied using a sliding window prediction_salsa.php [24] PAGE Prediction of parallel or anti-parallel b-sheet organization in [25] fibrils and aggregation rates based on physicochemical properties and computational design of b-aggregating peptide sequences TANGO Statistical mechanics-based method that calculates the partition http://tango.crg.es [26] function of the phase-space considering conformational states and energy terms, together with physico-chemical and protein stability parameters FoldAmyloid Predicts amyloidogenic regions according to the expected (http://bioinfo.protres.ru/fold-amyloid/ [27] probability of formation of backbone–backbone hydrogen bonds and the expected packing Density PASTA 2.0 Estimation of b-strand inter-molecular pairing probability http://protein.bio.unipd.it/pasta2/ [28] between polypeptide segments, based on experimentally- resolved b-sheets structures, and a statistical energy function to determine fibril formation SecStr Combines the prediction of amyloidogenic regions with the http://biophysics.biol.uoa.gr [82] identification of ‘conformational switches’ ArchCandy Detection of amyloidogenic regions based on the propensity to https://bioinfo.crbm.cnrs.fr/index.php? form parallel in register stacking of b-arches (b-strand-loop-b- route=tools&tool=7 [83] strand motif) BetaSerpentine Reconstruction and ranking of b-serpentine arrangements of github.com/stanislavspbgu/ adjacent b-arches predicted by ArchCandy BetaSerpentine [84] BETASCAN Prediction of b-strands and strand pairs in a sequence based on http://betascan.csail.mit.edu [85] pairwise probabilities for each pair of residues to form hydrogen bonds in amphiphilic b-sheets AmyloidMutants Quantifies the energetic effects of sequence mutation on fibril http://amyloid.csail.mit.edu/ [86] conformation and stability using a potential energy scoring function derived from the frequency of specific residue/ residue interactions in PDB protein structures STITCHER Predicts the formation of strand-pairs into complete b-sheets http://stitcher.csail.mit.edu [87] based on a free-energy model accounting for amino acid sidechain stacking contributions, entropic estimation, and steric restrictions for amyloidal parallel b-sheet formation. A dynamic program returns the top scored structures GAP (Aggregation Evaluation of residue pair propensities to occur at adjacent or http://www.iitm.ac.in/bioinfo/GAP/ [88] Proneness) alternate positions in globular proteins, and calculation of thermodynamic energy potentials. This tool can predict whether APRs would form amorphous b-aggregates or amyloid fibrils. 3D Profile Profile generated by mutating the side chains in the cross-b www.rosettacommons.org [89] spine of NNQQNY crystal structure and employing ROSETTADESIGN to evaluate the sequence energetic fit Machine-learning methods ANuPP (Aggregation Ensemble-classifier that identifies potential aggregation- https://web.iitm.ac.in/bioinfo2/ANuPP/ [34] Nucleation Prediction in nucleating regions trained on an experimental dataset of Peptides and Proteins) amyloidogenic and nonamyloidogenic hexapeptides (continued on next page) www.sciencedirect.com Current Opinion in Structural Biology 2022, 73:102343 4 Folding and Binding Table 1 (continued ) Algorithm Characteristics URL and Ref PATH (Prediction of The input sequence is threated into templates of different Scripts available: https://github.com/ Amyloidogenicity by structural amyloid classes and the model with the lowest KubaWojciechowski/PATH [35] Threading) energy value score according to PyRosetta is used as an input for machine learning classifiers NetCSSP Detects non-native secondary structure propensities based on http://cssp2.sookmyung.ac.kr/ [36] the calculation of contact-dependent secondary structure propensity (CSSP), and the search for chameleon sub- sequences using a PDB structures collection FiSH amyloid Machine learning method trained to recognise and classify http://www.comprec.pwr.wroc.pl/ amyloidogenic segments based on position-specific amino COMPREC_home_page.html [37] acid co-occurrence patterns in protein sequences RF Amyloid A random forest protein classifier based on composition and http://server.malab.cn/RFAmyloid/ [38] physicochemical features from protein sequences Budapest Linear Support Vector Machine (SVM)-based predictor for https://pitgroup.org/bap/ [39] hexapeptides trained and tested with the experimental hexapeptide Waltz dataset. CORDAX Logistic regression approach that detects APRs and predicts the https://cordax.switchlab.org [40] structural topology and architecture of the fibril core. It exploits a curated amyloid template structural database generated from the WALTZ-DB 2.0 repository AgMata Machine learning-based classifier not trained on aggregation https://bitbucket.org/bio2byte/agmata [90] data. It uses manually selected parameters accounting for predicted secondary structure propensities, side-chain and backbone dynamics, and a b-pairing energy function Pre-Amyl-MLP Machine learning-based prediction using a multilayer http://106.12.83.135:8080/amyWeb_ perceptron-based classification that exploit a selected Release/index.jsp [91] combination of amyloid associated features AbAmyloid Random Forest classifier adopted to evaluate amino acid http://iclab.life.nctu.edu.tw/abamyloid [92] composition, dipeptide composition and physicochemical properties Pafig (Prediction of Identification of amyloid fibril-prone hexapeptides with a scale [93] amyloid fibril-forming derived from machine supervised learning of >500 segments) physicochemical properties. Suitable for large-scale analysis APPNN Machine learning approach that analyses features correlated http://cran.r-project.org/web/packages/ (Amyloidogeniciy with self-assembly of peptides and proteins into amyloids: appnn/index.html) [94] Propensity Prediction frequency of b-sheet, isoelectric point, atom-based Neural Network) hydrophobic moment, helix termination parameters and DG values for peptides extrapolated in 0 M urea Amylogram N-grams and random forest machine learning classifiers that http://smorfland.uni.wroc.pl/shiny/ recognise sequential patterns in the amyloids considering AmyloGram/ [95] hydrophobicity, tendency to form b-sheets, and low flexibility of amino acid residues Consensus methods Amylpred 2 It combines 11 individual methods (Aggrescan, http://aias.biol.uoa.gr/AMYLPRED2/ [29] AmyloidMutants, Amyloidogenic, Pattern, Average Packing Density, Beta-strand contiguity, Hexapeptide Conformational Energy, NetCSSP, Pafig, SecStr, Tango and Waltz) to identify amyloid-forming regions. Consensus between the output from at least n/2 of n selected algorithms MetAmyl Uses a logistic regression model after combining and weighting http://metamyl.genouest.org/ [30] the output of 4 popular predictors (SALSA, PAFIG, Waltz and FoldAmyloid). It can handle large sequence datasets 3D structure-based methods Solubis Identifies mutations that reduce protein aggregation employing http://solubis.switchlab.org [55] the statistical thermodynamics algorithm TANGO to evaluate the sequence intrinsic aggregation propensity. This is projected onto high-resolution 3D structures, and the thermodynamic stability of the variant assessed using FoldX Aggscore Trained on a dataset of 31 adnectin proteins with varying Schrödinger’s BioLuminate Suite as of aggregation propensities. It uses the distribution of software release 2018–1 [56] hydrophobic and electrostatic patches on the surface together with their intensity and orientation to implement and aggregation propensity scoring function Current Opinion in Structural Biology 2022, 73:102343 www.sciencedirect.com Protein aggregation prediction Navarro and Ventura 5 Table 1 (continued ) Algorithm Characteristics URL and Ref SAP (Spatial Based on full antibody atomistic simulations, it measures the [57] aggregation effective dynamically exposed hydrophobicity of a certain propensity) patch on the protein surface to identify aggregation prone regions. It implements a mutation option on predicted regions. SAP has been applied either in high-throughput developability screening of therapeutic protein candidates or to enhance stability at later stages of manufacturing AGGRESCAN3D 2.0 The method integrates the 3D-structural information of PDBs http://biocomp.chem.uw.edu.pl/A3D2/ and evaluates the contribution of solvent-exposed APRs [58,59] using the amino acid aggregation scale from Aggrescan method. Predictions can be run in static and dynamic modes. An extended option allows the evaluation of stable mutations using FoldX method Camsol Aimed to design protein variants with enhanced solubility by the http://www-vendruscolo.ch.cam.ac.uk/ calculation of the intrinsic solubility profile with a structural camsolmethod.html [60] correction accounting for the residues’ structural environment and for their solvent exposure. The structurally corrected solubility profile is used to identify the most suitable solubilizing mutations these programs can be used to evaluate the quality of acids, which are known as prion or prion-like domains the experimental data [33]. (PrDs, PrLDs). Their enrichment in hydrophilic resi- dues contrasts with the eminent hydrophobic nature of Several protein aggregation predictors employ machine classical APRs and precludes identifying these aggre- learning classifiers to recognise sequence-specific fea- gating regions with pre-existent algorithms. Thus, the tures and position-specific patterns, or they use energy development of new tools was required (Table 2). A first functions of cross-b pairings; they include ANuPP [34], class of methods compares target proteins’ amino acid Amylogram [35], netCSSP [36], FISH amyloid [37], RF composition with the one of PrDs of similar length in amyloid [38], and Budapest [39] (Table 1). The most bona fide yeast prions using different metrics and recently developed PATH [35] and CORDAX [40] al- scanning procedures. They include DIANA [45], LPS gorithms exploit crystallographic structures of amyloid [46], PAPA [47], PLAAC [48], PrionScan [49], and the steric zippers [41] to combine threading and machine machine learning method pRANK [50]. All these learning approaches, providing insights into the forces methods consider that the transition towards the that govern the formation of amyloid assemblies by short aggregated state is mediated by a large number of weak peptides (Table 1). This class of algorithms has interactions distributed along the PrD, according to the succeeded in protein redesign [42], solubility fore- view that the PrD composition and not its specific casting [43], or ranking aggregation rates [44]. Further sequence is relevant for prion conversion. A second class improvements would require connecting their outputs of algorithms, like pWALTZ [51] and PrionW [52], relies with understandable physical, chemical, or structural on identifying and ranking soft-amyloid cores within parameters to feedback the predictive model. PrD, defined as short linear stretches of low complexity with a moderate but significant amyloid propensity that Prion-like proteins can be identified from their nucleates the aggregation reaction. The predictions of sequences the two kinds of methods are complementary, and their Prions and prion-like proteins are a particular subset of joint application has allowed the identification of thou- amyloids that can interconvert between a soluble sands of new prion-like candidates in proteomes conformer and a cross-b structure. Both states can be belonging to all kingdoms of life [53]. functional, and the conversion to the amyloid state can either switch off the initial function or switch on a new Prediction of aggregation propensities in activity. The aggregated state is often transmissible, globular protein structures templating homologous polypeptides conversion. They Overall, linear algorithms have demonstrated to be fast differ from classical amyloids since, in most cases, this and cost-effective tools to predict sequences aggrega- activity is located in disordered regions of low sequence tion propensities, with a remarkable overlay between complexity enriched in polar residues, like glutamine predicted and experimentally validated APRs [54]. and asparagine, and depleted in hydrophobic amino They are of particular interest when dealing with short www.sciencedirect.com Current Opinion in Structural Biology 2022, 73:102343 6 Folding and Binding Table 2 Computational methods to identify proteins bearing prion domains. Method Description URL and Ref Composition-based predictors DIANA (Defined Interval Amino Inspired by the length and composition of Sup35p and [45] acid Numerating Algorithm) Ure2p yeast prions and the length of pathogenic polyQ expansions. The algorithm searches for consecutive 80 residue-long sequence and retrieves the most Q/N- rich stretch containing at least 30 Q and/or N residues LPS (Lowest-Probability Identifies compositional-biased regions by defining the http://libaio.biol.mcgill.ca/lps-annotate. Subsequences) lowest-probability sub-sequences (LPSs) for a given html [46] amino acid composition in a defined proteomic context PAPA (Prion Aggregation Combination of analysis of disordered regions using http://combi.cs.colostate.edu/ Prediction Algorithm) FoldIndex and calculation of the prion propensity of supplements/papa/ [47] each amino acid using a scale obtained experimentally from a randomly mutated Sup35p segment PLAAC (Prion-like amino acid Implements a Hidden Markov model derived from 28 http://plaac.wi.mit.edu/ [48] composition) characterised yeast prion proteins. The program employs the HMM-derived residue log-likelihood for every position, and calculates the probability of each amino acid to be part of a PrLD, considering the background frequencies in the selected proteome PrionScan Bimodal method incorporating an open-source database http://webapps.bifi.es/prionscan [49] of prion predictions for all the proteins in UniProt KB and a sequence analysis tool to test prionogenicity of user’s protein sequences, relying on the amino acid frequencies of validated prion and non-prion of similar composition pRank Employs supervised multiple-instance learning to solve http://faculty.pieas.edu.pk/fayyaz/prank. the problem of inaccurately annotated data. Trained on html [50] top of 22 known Q/N-rich yeast prions against the rest of the yeast proteome, it ranks and classifies prion sequences Soft Amyloid Core-based predictors pWaltz Employs the amyloid propensity scoring matrix of the http://bioinf.uab.es/pWALTZ/ [23,51] Waltz linear predictor and scans for soft-amyloid cores (SACs) by implementing a 21-residue window and calculating its average amyloid load. It identifies SACs as the highest scoring stretches above a calibrated threshold PrionW Intrinsically disordered segments of at least 80 residues http://bioinf.uab.cat/prionw/ [52] are identified with FoldIndex and their Q/N content computed. Q/N enriched sequences are then scored using pWaltz. The Q/N and pWaltz identification thresholds can be adjusted for each specific protein dataset. peptides, IDPs, or fluctuating protein regions, where The above-described limitations, together with the ev- APRs are permanently or transiently exposed to solvent. idence that many of the biotechnologically relevant However, they mistakenly overestimate the aggregation proteins are globular (e.g., antibodies or enzymes for propensity of globular proteins, and native oligomers, for industrial and therapeutic applications), have prompted which the structural context strongly modulates the the development of a generation of structure-based contribution of sequential determinants. Over- prediction methods to address the location of solvent- prediction occurs because, in native states, APRs are accessible aggregation hot-spots, like Solubis [55] and usually embedded within the hydrophobic core, being Aggscore [56], or the identification of spatially close sheltered from the solvent and thus not contributing aggregation-prone regions, such as SAP [57], significantly to the aggregation propensity. On the con- AGGRESCAN3D 2.0 [58,59], and CamSol [60] trary, the clustering of sequentially distant hydrophobic (Table 1). Globular proteins are flexible in solution, and residues at protein surfaces may generate structural this impacts the degree of exposure of their STAPs. aggregation-prone regions (STAP) in the native state, to Accordingly, molecular dynamics have been incorpo- which linear predictors remain blind. rated in SAP and AGGRESCAN3D 2.0 predictions to Current Opinion in Structural Biology 2022, 73:102343 www.sciencedirect.com Protein aggregation prediction Navarro and Ventura 7 simulate structural fluctuations under native conditions Conclusions and perspectives [58,59], using full atomistic calculations and the coarse- In silico aggregation predictors have contributed to guide grained CABS-flex approach [61], respectively. Note- and assist experimental endeavours in elucidating the worthy, several of these tools allow for predicting molecular mechanisms underlying protein aggregation- structural aggregation propensity and thermodynamic related diseases [68]. They have also boosted the stability upon automatically introducing multiple mu- design of engineered protein variants with improved tations, which is extremely useful for selecting and solubility and stability [69,70], saving time and costs in producing highly soluble and stable protein vari- therapeutic proteins production pipelines. Each of the ants [62]. discussed approaches has its own pros and cons (Table 3), and users should bear in mind the specific The predictions of structure-based algorithms are ac- problem they want to address. This is because they curate when a high-resolution experimental structure is capture different aspects of protein aggregation, that available, being less precise when working with molec- might or might not be relevant for the application ular models, and this has restricted their use to the of interest. subset of structures deposited in the protein data bank. However, the apparition of programs like AlphaFold2 Protein aggregation reactions are complex processes in [63] and RoseTTaFold [64] providing high fidelity which, besides the primary sequence, multiple param- structural models for natural or mutated sequences eters might impact the stability, structure, cooperativity, should bypass this limitation. We foresee a future in solubility, kinetics, and dynamics of polypeptides. Thus, which structure and aggregation prediction algorithms the binary classification into aggregating and non- would be sequentially implemented in protein produc- aggregating proteins or protein regions provided by tion pipelines, making it fast and economically afford- predictors might be accurate only in conditions close to able to redesign therapeutic products with increased those in which the initial experiments or calculations solubility [65]. that feed the training data sets were performed. Finally, later advances in solid-state NMR (ssNMR) The influence of factors intrinsic to the sequence is [66] and cryo-EM [67] are providing an increasing evident because we need different programs to predict repertoire of fibrillar structures of pathological and archetypical amyloids and prion-like proteins. Indeed, functional amyloids, which can be used to train two recent works [40,71] indicate that the amyloid structure-based methods in order to assist the redesign sequence space is much larger than previously thought, of amyloid-like structures for therapeutic and nano- including highly soluble sequences with low aliphatic technology applications [65]. content and/or high net charge. At the core of most of Table 3 Pros and cons of the different classes of computational methods to predict protein aggregation. Method Pros Cons Sequence Based - Low computational demand - Dismiss the 3D-structural context - User-friendly - Disregard the impact of mutations on protein stability - Many webservers available - Do not account for protein micro-environment - Fast analysis of complete proteomes - Perform well on small peptides and IDPs - Consensus methods available Machine learning - Produce outputs from apparently unrelated protein - Accurate annotation of the reference experimental sequence features training data is required - Attain high accuracy - Do not calculate full-length proteins average - Implemented in webservers propensities - Suitable for fast and wide-range analyses - Disconnection between the output and understandable physicochemical interpretations Structure Based - Predict on the native, functional state of proteins - Dependence on high-resolution structures - Consider the quaternary structure availability - Analyse surface-exposed hydrophobic and hydro- - Not suitable for IDPs philic patches - High computational cost - Incorporate protein dynamics calculations - Difficulty to perform proteome-wide analysis - Accurate in predicting globular proteins solubility - Protein chemistry/engineering knowledge required - Consider the impact of mutations on protein stability - Do not account for protein micro-environment - Allow automated design of more soluble protein variants www.sciencedirect.com Current Opinion in Structural Biology 2022, 73:102343 8 Folding and Binding the discussed algorithms is the concept that low solu- resolution cryo-EM structures of amyloid fibrils extrac- bility and aggregation propensity are interchangeably ted from the organs of patients [80] and AI-assisted properties, and this should be revisited to fish se- elucidation of thousands of human protein structures quences in this uncharted amyloid territory. In addition, [81]. Third-generation algorithms and databases are just it should not be overlooked that aggregation is not around the corner. uniquely dependent on APRs since flexible segments often located remote from these stretches may act as Funding conformational switches, masking APRs and thus Our lab is funded by the Spanish Ministry of Science modulating the fibrillation propensity [72]. Moreover, and Innovation (PID2019-105017RB-I00) by ICREA, even if we often assume that APRs are exposed to sol- (ICREA-Academia 2020) and by EU (PhasAge/H2020- vent during the process of protein folding and accord- WIDESPREAD-2020-5). ingly that sequence-based predictions are accurate at this stage, these sequences are usually protected by chaperones and self-chaperoning in vivo, not contrib- Conflict of interest statement uting to fibrillation [73]. Finally, most existing programs Nothing declared. do not consider post-translational modifications, even though they can dramatically alter aggregation pro- References pensities [74]. Thus, it is apparent that despite current Papers of particular interest, published within the period of review, progress, we are addressing a complex process with have been highlighted as: rather simplistic approaches. Because we have in our  of special interest hands accurate tools to predict conformational switches,  of outstanding interest posttranslational modifications, and interactions with 1. Kuan SL, Bergamini FRG, Weil T: Functional protein nano- chaperones, co-factors, or other non-proteinaceous structures: a chemical toolbox. Chem Soc Rev 2018, 47: molecules, we should make an effort to integrate these 9069–9105. parameters in present programs to provide accurate 2. Alberti S, Hyman AA: Biomolecular condensates at the nexus predictions in biological relevant scenarios. of cellular stress, protein aggregation disease and ageing. Nat Rev Mol Cell Biol 2021, 22:196–213. Regarding extrinsic factors, viscosity, temperature, pH, 3. Garcia-Seisdedos H, Villegas JA, Levy ED: Infinite assembly of folded proteins in evolution, disease, and engineering. Angew ionic concentration, protein concentration, solvent Chem Int Ed Engl 2019, 58:5514–5531. identity, and the interaction with other molecules are all 4. Dey S, Levy ED: PDB-wide identification of physiological known to influence intrinsic aggregation propensities hetero-oligomeric assemblies based on conserved qua- [75]. Subtle variations in these factors are behind the ternary structure geometry. Structure 2021, 29: 1303– 1311.e3. formation of structurally different amyloid polymorphs [76], leading to distinct pathological phenotypes in 5. Riek R, Eisenberg DS: The activities of amyloids from a structural perspective. Nature 2016, 539:227–235. neurodegenerative disorders [77]. Despite this evi- 6. Dobson CM, Knowles TPJ, Vendruscolo M: The amyloid phe- dence, scarce efforts have been made to include them in  nomenon and its significance in biology and medicine. Cold the predictions, and only recently the effect of the pH Spring Harbor Perspect Biol 2020, 12. has been accurately parametrized [78] and incorporated An overview of the field that addresses the nature of the amyloid state, its conection with protein homoestasis and strategies to target its link into a webserver [79], although only for disordered with disease. proteins. The main limitation to developing methods 7. Lall D, Lorenzini I, Mota TA, Bell S, Mahan TE, Ulrich JD, that can mimic the protein microenvironment in their Davtyan H, Rexach JE, Muhammad AKMG, Shelest O, et al.: predictions is the lack of systematic experimental data C9orf72 deficiency promotes microglial-mediated synaptic loss in aging and amyloid accumulation. Neuron 2021, 109: covering all possible variables combinations for a set of 2275–2279. structurally and sequentially unrelated proteins. Gath- 8. Hamrang Z, Rattray NJW, Pluen A: Proteins behaving badly: ering these data is often seen as a “low regard” objective, emerging technologies in profiling biopharmaceutical ag- but we have all witnessed how artificial intelligence (AI) gregation. Trends Biotechnol 2013, 31:448–458. has revolutionized structural biology, which would not 9. Shanmugam N, Baker MODG, Ball SR, Steain M, Pham CLL,  Sunde M: Microbial functional amyloids serve diverse pur- have been possible without the previous existence of an poses for structure, adhesion and defence. Biophys Rev 2019, extensive collection of protein structures. 11:287–302. 2019 113. An extensive and detailed review on microbial functional amyloids and their roles. Our opinion is that building up algorithms that can 10. Levkovich SA, Gazit E, Laor Bar-Yosef D: Two decades of effectively predict aggregation in the specific conditions studying functional amyloids in microorganisms. Trends occurring at the neuronal synapsis or allow stable for- Microbiol 2021, 29:251–265. mulations of antibodies for immunotherapies, just to 11. Rubel MS, Fedotov SA, Grizel AV, Sopova JV, Malikova OA, mention a couple of applications, is indeed a “high Chernoff YO, Rubel AA: Functional mammalian amyloids and amyloid-like proteins. Life 2020, 10:1–32. reward” objective. 2021 has witnessed the development of impressive advances that should help in attaining 12. Santos J, Ventura S: Functional amyloids germinate in plants. Trends Plant Sci 2021, 26:7–10. these challenging objectives, like the obtention of high- Current Opinion in Structural Biology 2022, 73:102343 www.sciencedirect.com Protein aggregation prediction Navarro and Ventura 9 13. Zozulia O, Dolan MA, Korendovych IV: Catalytic peptide as- database containing structural information of experimentally semblies. Chem Soc Rev 2018, 47:3621–3639. determined amyloid-forming peptides. Nucleic Acids Res 2020, 48:D389–D393. 14. Shen Y, Levin A, Kamada A, Toprakcioglu Z, Rodriguez- An updated open-access database storing >1400 amyloid-forming Garcia M, Xu Y, Knowles TPJ: From protein building blocks to hexapeptide entries. functional materials. ACS Nano 2021, 15:5819–5837. 32. Tsiolaki PL, Nastou KC, Hamodrakas SJ, Iconomidou VA: Mining 15. Zeng R, Lv C, Wang C, Zhao G: Bionanomaterials based on databases for protein aggregation: a review. Amyloid 2017, 24: protein self-assembly: design and applications in biotech- 143–152. nology. Biotechnol Adv 2021, 52:107835. 33. Szulc N, Burdukiewicz M, Gasior-Głogowska M, 16. Santos J, Pujols J, Pallarès I, Iglesias V, Ventura S: Computa- Î Wojciechowski JW, Chilimoniuk J, Mackiewicz P, tional prediction of protein aggregation: advances in prote-  Sneideris T, Smirnovas V, Kotulska M: Bioinformatics omics, conformation-specific algorithms and methods for identification of amyloidogenic peptides biotechnological applications. Comput Struct Biotechnol J show robustness to misannotated training data. Sci Rep 2020, 18:1403–1413. 2021, 11:8934. 17. Langenberg T, Gallardo R, van der Kant R, Louros N, Michiels E, 34. Prabakaran R, Rawat P, Kumar S, Michael Gromiha M: ANuPP: a  Duran-Romaña R, Houben B, Cassio R, Wilkinson H, Garcia T,  versatile tool to predict aggregation nucleating regions in et al.: Thermodynamic and evolutionary coupling between the peptides and proteins. J Mol Biol 2021, 433:166707. native and amyloid state of globular proteins. Cell Rep 2020, An ensemble-classifier trained to identify amyloid-fibril forming pep- 31. tides and regions in protein sequences. Describes how positive selection of amyloid propensity favors globular proteins stability. 35. Wojciechowski JW, Kotulska M: PATH - prediction of amyloi- dogenicity by threading and machine learning. Sci Rep 2020, 18. Yagi-Utsumi M, Yanaka S, Song C, Satoh T, Yamazaki C, 10:1–9. 2020 101. Kasahara H, Shimazu T, Murata K, Kato K: Characterization of amyloid b fibril formation under microgravity conditions. NPJ 36. Kim C, Choi J, Lee SJ, Welsh WJ, Yoon S: NetCSSP: web Microgravity 2020, 6. application for predicting chameleon sequences and amyloid fibril formation. Nucleic Acids Res 2009, 37:W469. 19. Kollmer M, Close W, Funk L, Rasmussen J, Bsoul A,  Schierhorn A, Schmidt M, Sigurdson CJ, Jucker M, Fändrich M: 37. Gasior P, Kotulska M: FISH Amyloid - a new method for finding Cryo-EM structure and polymorphism of Ab amyloid fibrils amyloidogenic segments in proteins based on site specific purified from Alzheimer’s brain tissue. Nat Commun 2019, co-occurrence of aminoacids. BMC Bioinf 2014, 15:54. 10. Evidence for the in vivo structural polymorphism of Ab amyloid fibrils. 38. Niu M, Li Y, Wang C, Han K: RFAmyloid: a web server for predicting amyloid proteins. Int J Mol Sci 2018, 19. 20. Hoppe SO, Uzunog lu G, Nussbaum-Krammer C: a-Synuclein strains: does amyloid conformation explain the heterogene- 39. Keresztes L, Szögi E, Varga B, Farkas V, Perczel A, Grolmusz V: ity of synucleinopathies? Biomolecules 2021, 11:931. The budapest amyloid predictor and its applications. Bio- molecules 2021, 11. 21. Conchillo-Solé O, de Groot NS, Avilés FX, Vendrell J, Daura X, Ventura S: AGGRESCAN: a server for the prediction and 40. Louros N, Orlando G, De Vleeschouwer M, Rousseau F, evaluation of "hot spots" of aggregation in polypeptides.  Schymkowitz J: Structure-based machine-guided mapping of BMC Bioinf 2007, 8:65. amyloid sequence space reveals uncharted sequence clus- ters with higher solubilities. Nat Commun 2020, 11. 22. Tartaglia GG, Vendruscolo M: The Zyggregator method for Demonstration that the amyloid space is significantly larger than pre- predicting protein aggregation propensities. Chem Soc Rev viously thought, including soluble protein sequences. 2008, 37:1395. 41. Sawaya MR, Sambashivan S, Nelson R, Ivanova MI, Sievers SA, 23. Maurer-Stroh S, Debulpaep M, Kuemmerer N, de la Paz ML, Apostol MI, Thompson MJ, Balbirnie M, Wiltzius JJW, McFarlane HT, Martins IC, Reumers J, Morris KL, Copland A, Serpell L, et al.: Atomic structures of amyloid cross-b spines reveal varied Serrano L, et al.: Exploring the sequence determinants of steric zippers. Nature 2007, 447:453–457. 2006 4477143. amyloid structure using position-specific scoring matrices. Nat Methods 2010, 7:237–242. 42. Xu Y, Verma D, Sheridan RP, Liaw A, Ma J, Marshall NM, McIntosh J, Sherer EC, Svetnik V, Johnston JM: Deep dive into 24. Zibaee S, Makin OS, Goedert M, Serpell LC: A simple algorithm machine learning models for protein engineering. J Chem Inf locates beta-strands in the amyloid fibril core of alpha- Model 2020, 60:2773–2790. synuclein, Abeta, and tau using the amino acid sequence alone. Protein Sci 2007, 16:906–918. 43. Raimondi D, Orlando G, Fariselli P, Moreau Y: Insight into the protein solubility driving forces with neural attention. PLoS 25. Tartaglia GG, Cavalli A, Pellarin R, Caflisch A: Prediction of Comput Biol 2020, 16. aggregation rate and aggregation-prone segments in poly- peptide sequences. Protein Sci 2005, 14:2723–2734. 44. Yang W, Tan P, Fu X, Hong L: Prediction of amyloid aggre-  gation rates by machine learning and feature selection. 26. Fernandez-Escamilla A-M, Rousseau F, Schymkowitz J, J Chem Phys 2019, 151. Serrano L: Prediction of sequence-dependent and mutational A neuronal network approach to predict amyloid aggregation rates from effects on the aggregation of peptides and proteins. Nat intrinsic and extrinsic factors. Biotechnol 2004, 22:1302–1306. 45. Michelitsch MD, Weissman JS: A census of glutamine/ 27. Garbuzynskiy SO, Lobanov MY, Galzitskaya OV: FoldAmyloid: a asparagine-rich regions: implications for their conserved method of prediction of amyloidogenic regions from protein function and the prediction of novel prions. Proc Natl Acad Sci sequence. Bioinformatics 2010, 26:326–332. U S A 2000, 97:11910–11915. 28. Walsh I, Seno F, Tosatto SCE, Trovato A: PASTA 2.0: an 46. Harrison PM, Gerstein M: A method to assess compositional improved server for protein aggregation prediction. Nucleic bias in biological sequences and its application to prion-like Acids Res 2014, 42:W301–W307. glutamine/asparagine-rich domains in eukaryotic proteomes. Genome Biol 2003, 4:R40. 29. Tsolis AC, Papandreou NC, Iconomidou VA, Hamodrakas SJ: A consensus method for the prediction of “aggregation- 47. Toombs Ja, Petri M, Paul KR, Kan GY, Ben-Hur A, Ross ED: De prone” peptides in globular proteins. PLoS One 2013, 8. novo design of synthetic prion domains. Proc Natl Acad Sci U S A 2012, 109:6519–6524. 30. Emily M, Talvas A, Delamarche C: MetAmyl: a METa-predictor for AMYLoid proteins. PLoS One 2013, 8, e79722. 48. Lancaster AK, Nutter-Upham A, Lindquist S, King OD: PLAAC: a web and command-line application to identify proteins with 31. Louros N, Konstantoulea K, De Vleeschouwer M, Ramakers M, Prion-Like Amino Acid Composition. Bioinformatics 2014, 30:  Schymkowitz J, Rousseau F: WALTZ-DB 2.0: an updated 2–3. www.sciencedirect.com Current Opinion in Structural Biology 2022, 73:102343 10 Folding and Binding 49. Espinosa Angarica V, Angulo A, Giner A, Losilla G, Ventura S, Structures of pathological and functional amyloids and Sancho J: PrionScan: an online database of predicted prion prions, a solid-state NMR perspective. Front Mol Neurosci domains in complete proteomes. BMC Genom 2014, 15:102. 2021, Jul 1, 14:670513. 50. Afsar Minhas F ul A, Ross ED, Ben-Hur A: Amino acid compo- 67. Ragonis-Bachar P, Landau M: Functional and pathological sition predicts prion activity. PLoS Comput Biol 2017, 13. amyloid structures in the eyes of 2020 cryo-EM. Curr Opin Struct Biol 2021, 68:184–193. 51. Sabate R, Rousseau F, Schymkowitz J, Ventura S: What makes a protein sequence a prion? PLoS Comput Biol 2015, 11, 68. Tompa DR, Kadhirvel S: Changes in hydrophobicity mainly e1004013. promotes the aggregation tendency of ALS associated SOD1 mutants. Int J Biol Macromol 2020, 145:904–913. 52. Zambrano R, Conchillo-Sole O, Iglesias V, Illa R, Rousseau F, Schymkowitz J, Sabate R, Daura X, Ventura S: PrionW: a server 69. Habibi N, Mohd Hashim SZ, Norouzi A, Samian MR: A review of to identify proteins containing glutamine/asparagine rich machine learning methods to predict the solubility of over- prion-like domains and their amyloid cores. Nucleic Acids Res expressed recombinant proteins in Escherichia coli. BMC 2015, Jul 1, 43:W331–W337. Bioinf 2014, 15. 53. Gil-Garcia M, Iglesias V, Pallarès S, Ventura S: Prion-like pro- 70. Navarro S, Ventura S: Computational re-design of protein  teins: from computational approaches to proteome-wide structures to improve solubility. Expet Opin Drug Discov 2019, analysis. FEBS Open Bio 2021, Sep, 11:2400–2417. 14:1077–1088. State-of-the art computational approaches for the identification of prion-like proteins and their function in proteomes. 71. Santos J, Pallarès I, Iglesias V, Ventura S: Cryptic amyloido-  genic regions in intrinsically disordered proteins: function 54. Chiti F, Dobson CM: Protein misfolding, amyloid formation, and disease association. Comput Struct Biotechnol J 2021, 19: and human disease: a summary of progress over the last 4192–4206. decade. Annu Rev Biochem 2017, 86:27–68. Demonstration that, in contrast to what is usually assumed, intrinsically disordered proteins contain amyloidogenic sequences that are impor- 55. Van Durme J, De Baets G, Van Der Kant R, Ramakers M, tant for their funtion, with implications for the emergence of globular Ganesan A, Wilkinson H, Gallardo R, Rousseau F, Schymkowitz J: folds. Solubis: a webserver to reduce protein aggregation through mutation. Protein Eng Des Sel 2016, 29:285–289. 72. Ulamec SM, Brockwell DJ, Radford SE: Looking beyond the core: the role of flanking regions in the aggregation of 56. Sankar K, Krystek SR, Carl SM, Day T, Maier JKX: AggScore: amyloidogenic peptides and proteins. Front Neurosci 2020, 14. prediction of aggregation-prone regions in proteins based on the distribution of surface patches. Proteins Struct Funct Bio- 73. Goldschmidt L, Teng PK, Riek R, Eisenberg D: Identifying the informa 2018, 86:1147–1156. amylome, proteins capable of forming amyloid-like fibrils. Proc Natl Acad Sci U S A 2010, 107:3487–3492. 57. Chennamsetty N, Voynov V, Kayser V, Helk B, Trout BL: Design of therapeutic proteins with enhanced stability. Proc Natl 74. Moon SP, Balana AT, Pratt MR: Consequences of post- Acad Sci U S A 2009, 106:11937–11942. translational modifications on amyloid proteins as revealed by protein semisynthesis. Curr Opin Chem Biol 2021, 64:76–89. 58. Zambrano R, Jamroz M, Szczasiuk A, Pujols J, Kmiecik S, Ventura S: AGGRESCAN3D (A3D): server for prediction of 75. Brudar S, Hribar-Lee B: Effect of buffer on protein stability in aggregation properties of protein structures. Nucleic Acids aqueous solutions: a simple protein aggregation model. Res 2015, Jul 1, 43:W306–W313. J Phys Chem B 2021, 125:2504–2512. 59. Kuriata A, Iglesias V, Pujols J, Kurcinski M, Kmiecik S, Ventura S: 76. Holec SAM, Woerman AL: Evidence of distinct a-synuclein  Aggrescan3D (A3D) 2.0: prediction and engineering of pro-  strains underlying disease heterogeneity. Acta Neuropathol tein solubility. Nucleic Acids Res 2019, 47:W300–W307. 2021, 142. An updated AGGRESCAN3D algorithm allowing automated mutation Evidence for the rol of a-synuclein amyloid polymorphs in disease analysis and dynamic calculations of large multimeric proteins, such as manifestation. antibodies. 77. Froula JM, Castellana-Cruz M, Anabtawi NM, Camino JD, 60. Sormanni P, Aprile FA, Vendruscolo M: The CamSol method of Chen SW, Thrasher DR, Freire J, Yazdi AA, Fleming S, rational design of protein mutants with enhanced solubility. Dobson CM, et al.: Defining a-synuclein species responsible J Mol Biol 2015, 427:478–490. for Parkinson’s disease phenotypes in mice. J Biol Chem 2019, Jul 5, 294:10392–10406. 61. Kuriata A, Gierut AM, Oleniecki T, Ciemny MP, Kolinski A,M, Kmiecik S: CABS-flex 2.0: a web server for fast simulations of 78. Santos J, Iglesias V, Santos-Suárez J, Mangiagalli M, Brocca S, flexibility of protein structures. Nucleic Acids Res 2018, 46: Pallarès I, Ventura S: pH-dependent aggregation in intrinsi- W338–W343. cally disordered proteins is determined by charge and lipo- philicity. Cells 2020, 9:145. 2020, 9:145. 62. Gil-Garcia M, Bañó-Polo M, Varejão N, Jamroz M, Kuriata A, Díaz-Caballero M, Lascorz J, Morel B, Navarro S, Reverter D, 79. Pintado C, Santos J, Iglesias V, Salvador V: SolupHred: a server et al.: Combining structural aggregation propensity and sta-  to predict the pH-dependent aggregation of intrinsically bility predictions to redesign protein solubility. Mol Pharm disordered proteins. Bioinformatics 2021, 37. 2018, 15:3846–3859. A server that explicitly considers the impact of the environmental pH on the aggregation of intrinsically disordered proteins. 63. Service RF: “The game has changed.” AI triumphs at protein folding. Science 2020, 370:1144–1145. 80. Shi Y, Zhang W, Yang Y, Murzin AG, Falcon B, Kotecha A, van  Beers M, Tarutani A, Kametani F, Garringer HJ, et al.: Structure- 64. Baek M, DiMaio F, Anishchenko I, Dauparas J, Ovchinnikov S, based classification of tauopathies. Nature 2021, 598: Lee GR, Wang J, Cong Q, Kinch LN, Schaeffer RD, et al.: Ac- 359–363. curate prediction of protein structures and interactions using Cryo-electron microscopy structures of tau filaments show that the fi- a three-track neural network. Science 2021, Aug 20, 373: brils folds differ between tauopathies and allow to classify diferent 871–876. clinical manifestations. 65. Pinheiro F, Santos J, Ventura S: AlphaFold and the amyloid 81. Jumper J, Evans R, Pritzel A, Green T, Figurnov M,  landscape. J Mol Biol 2021, https://doi.org/10.1016/  Ronneberger O, Tunyasuvunakool T, Bates R, Zídek A, J.JMB.2021.167059. Potapenko A, et al.: Highly accurate protein structure predic- The authors debate how the emergence of artificial intelligence appli- tion with AlphaFold. Nature 2021, 596:583–589. cations like AlphaFold would impact the study of amyloids. 82. Hamodrakas SJ, Liappa C, Iconomidou VA: Consensus pre- 66. Daskalov A, Mammeri N El, Lends A, Shenoy J, Lamon G, diction of amyloidogenic determinants in amyloid fibril- Fichou Y, Saad A, Martinez D, Morvan E, Berbon M, et al.: forming proteins. Int J Biol Macromol 2007, 41:295–300. Current Opinion in Structural Biology 2022, 73:102343 www.sciencedirect.com Protein aggregation prediction Navarro and Ventura 11 83. Ahmed AB, Znassi N, Château M-T, Kajava AV: A structure- 89. Thompson MJ, Sievers SA, Karanicolas J, Ivanova MI, Baker D, based approach to predict predisposition to amyloidosis. Eisenberg D: The 3D profile method for identifying fibril-forming Alzheimer’s Dementia 2015, 11:681–690. segments of proteins. Proc Natl Acad Sci U S A 2006, 103:4074. 84. Bondarev SA, Bondareva OV, Zhouravleva GA, Kajava AV: 90. Orlando G, Silva A, Macedo-Ribeiro S, Raimondi D, Vranken W: BetaSerpentine: a bioinformatics tool for reconstruction of Accurate prediction of protein beta-aggregation with general- amyloid structures. Bioinformatics 2018, 34:599–608. ized statistical potentials. Bioinformatics 2020, 36:2076–2081. 85. Bryan AW, Menke M, Cowen LJ, Lindquist SL, Berger B: 91. Li Y, Zhang Z, Teng Z, Liu X: PredAmyl-MLP: prediction of BETASCAN: probable beta-amyloids identified by pairwise amyloid proteins using multilayer perceptron. Comput Math probabilistic analysis. PLoS Comput Biol 2009, 5, e1000333. Methods Med 2020, Nov 20, 2020:8845133. 86. O’Donnell CW, Waldispühl J, Lis M, Halfmann R, Devadas S, 92. Liaw C, Tung CW, Ho SY: Prediction and analysis of antibody Lindquist S, Berger B: A method for probing the mutational amyloidogenesis from sequences. PLoS One 2013, 8. landscape of amyloid structure. Bioinformatics 2011, 27: i34–i42. 93. Tian J, Wu N, Guo J, Fan Y: Prediction of amyloid fibril-forming segments based on a support vector machine. BMC Bioinf 87. Bryan AW, O’Donnell CW, Menke M, Cowen LJ, Lindquist S, 2009, 10:S45. Berger B: STITCHER: dynamic assembly of likely amyloid and prion b-structures from secondary structure predictions. 94. Família C, Dennison SR, Quintas A, Phoenix DA: Prediction of Proteins Struct Funct Bioinforma 2012, 80:410–420. peptide and protein propensity for amyloid formation. PLoS One 2015, 10, e0134679. 88. Thangakani AM, Kumar S, Nagarajan R, Velmurugan D, Gromiha MM: GAP: towards almost 100 percent prediction for 95. Burdukiewicz M, Sobczyk P, Rödiger S, Duda-Madej A, b-strand-mediated aggregating peptides with distinct mor- Mackiewicz P, Kotulska M: Amyloidogenic motifs revealed by phologies. Bioinformatics 2014, 30:1983–1990. n-gram analysis. Sci Rep 2017, 7:1–10. 2017 71. www.sciencedirect.com Current Opinion in Structural Biology 2022, 73:102343