AGGRESCAN: a server for the prediction and evaluation of "hot spots" of aggregation in polypeptides

Input

The polypeptide sequence(s) can be typed or pasted on screen using FASTA format. Despite supporting up to 100 characters for name entries, use of very long names is discouraged as it disturbs the visualization of the output. Sequence entries may not contain more than 2,000 residues and the letters must correspond to those associated to the 20 natural amino acids. If these two conditions are not satisfied an error message will appear on screen. White-space, enter and tab characters are ignored. Characters may be entered as lower and/or upper case, and so will remain in the output.

Processing

The calculations are based on aggregation-propensity values per amino acid (aaAV, or a3v) derived previously from experimental data [12]. The program calculates the a3v average (a4v) over a sliding window of a given length and assigns it to the central residue in the window. The size of the sliding window ([5, 7, 9], and [11] residues) was trained against a database of 57 amyloidogenic proteins, in which the location of "hot spots" was experimentally known. In general, the predictions of the overall aggregation-prone regions do not depend on the length of the used windows and only slightly affect their limits. There are, however, two remarkable exceptions: 1) The use of long windows on top of very short sequences results on excessive smoothing of the profile and experimentally different "hot spots" become grouped and masked and cannot be individualized in the prediction. 2) The use of short windows on top of very long sequences results in the appearance of a number of short experimentally non-relevant predicted "hot spots" with associated low areas. Thus, the procedure incorporates a ponderation of the window length relative to the size of the analyzed protein. The best predictions were obtained using a window size of 5 for < = 75 residues, 7 for < = 175, 9 for < = 300 and 11 for > 300, respectively, probably reflecting that for longer sequences larger "hot spots" are necessary in order to significantly increase their aggregation propensities, while short-stretches suffice for smaller peptides. To account for charge effects at the polypeptide's termini (NH₃ ⁺ and COO^-) a virtual residue is added to each side of the chain (residue 0 at the N-terminus and residue n+1 at the C-terminus, n being the original sequence length). The a³v of residue 0 equals the average a³v of the basic residues (K, R), while that of residue n+1 equals the average a³v of the acidic residues (D, E). The first window, ranging from residue 0 to residue 4, 6, 8 or 10 (depending on window size), will serve to assign an a⁴v to residue 2, 3, 4 or 5, respectively. Thus, the off-centre residues 1, 1–2, 1–3 or 1–4 may not have an associated a⁴v. This is solved by giving these residues the value corresponding to the first window centre. The same procedure is followed at the C-terminus. The "hot spot" threshold (HST) has been defined as the average of the a³v of the 20 amino acids weighted by their frequencies in the SwissProt database [15]. The aggregation profile (AP) of the polypeptide is defined by the complete sequence of a⁴v. The sum of a⁴v and the average of a³v over the entire sequence (a⁴vSS and a³vSA, respectively) are also calculated. A region in the polypeptide sequence is considered an aggregation "hot spot" (HS) if there are 5 or more sequentially continuous residues with an a⁴v larger than the HST and none of them is a proline (aggregation breaker) [16]. The average a⁴v in each "hot spot" is then calculated (a⁴vAHS). Finally, the area of the AP above the HST (AAT), the total area (TA, HST being the zero axis), and the area above the HST of each profile peak identified as "hot spot" (HSA) are integrated numerically using the trapezoidal rule (see additional file 2).

Output

With current service resources, the delay time between pressing the submit button and receiving the output on screen is of 10 minutes for an input set of 100 sequences of sizes between 40 and 1,000 residues. The output is structured in tables, one per sequence and an additional one with averages over all sequences, an excel-readable document with output values and a list of sequences sorted by normalized a⁴vSS for 100 residues (Na⁴vSS). The first row in the output contains the sequence names. The second row displays links to the three graphics produced per sequence, i.e., Profile graphic: AP (red), a³vSA (green), HST (blue); Area graphic: HSA (same value assigned to all residues in the "hot spot"); Normalized-Area graphic: normalized HSA for a 100-residue "hot spot" (NHSA). In the following rows we find the a³vSA, the number of "hot spots" identified (nHS), the normalized number of "hot spots" for 100 residues (NnHS), the AAT, the THSA, the TA, the AAT and THSA divided by the number of residues (AATr and THSAr, respectively), and Na⁴vSS. Finally, a row per residue is given with columns for the residue number, its one-letter code, a⁴v, HSA, NHSA, and a⁴vAHS (see additional file 3).

Generation of protein aggregation profiles and prediction of aggregation "hot spots"

One remarkable example in the test set is lung surfactant protein C (SP-C). This protein is expressed as a 197-amino acid proprotein that is processed to the 35-amino acid mature peptide. This fragment is associated with the development of pulmonary alveolar proteinosis (PAP). The bronchoalveolar fluid from PAP patients is rich in insoluble SP-C aggregates which exhibit the characteristic properties of amyloids by Congo red staining and electron microscopy. Moreover, the isolated peptide has been shown to form amyloid fibrils in vitro [25]. In good agreement with this data, AGGRESCAN predicts the SP-C region within the precursor as the HS with the highest aggregation propensity (Figure 1).

Other two interesting molecules are serum amyloid A (SAA) and Tau proteins, involved in systemic amyloidosis and Alzheimer's disease, respectively. AGGRESCAN detects only one HS in SAA and a very dominant one in Tau (Figure 1). In both cases, these sequence stretches correspond to the unique regions in SAA and Tau proved to be relevant for amyloidosis [26, 27]. Importantly, the SAA and Tau sequences display highly negative Na⁴vSS values, -28.2 and -32.5 respectively. Although this suggests an overall low aggregation propensity, the presence of specific HS that can act as nucleation points from which the ordered fibrillar structure can be expanded under certain circumstances, turn these proteins amyloidogenic. Actually, Tau is an usually highly soluble microtubule-associated protein [28] but in Alzheimer's disease it aggregates into fibres with a tendency to form neurofibrillary tangles.

There are several computational approaches for detecting aggregation-prone regions and predicting polypeptide propensities for amyloid fibril formation. Some of them, including AGGRESCAN, rely on experimental or theoretical calculations of individual amino acid aggregation propensities and on the use of these values to scan protein sequences. The main difference between these algorithms is the way aggregation propensities are obtained. Pawar and co-workers proposed an aggregation scale based on phenomenological expressions relating protein intrinsic factors with the aggregation rates of a set point mutants scattered along acylphosphatase sequence and of a few other polypeptides [30]. As the fitting was done considering effects in both aggregation relevant and non-relevant regions, it is possible that the data do not necessarily reflect propensities within nucleating sequences. To address this point, Rojas Quijano and co-workers derived propensities from the analysis of the Tau-related amyloidogenic peptide Ac-VQIVYK-amide and its single site mutants Ac-VQIVXK-amide (X≠Cys) [19]. In AGGRESCAN, we somehow combine both approaches, in the sense that (i) propensities are calculated from the analysis of single mutants in a nucleating sequence (the central hydrophobic cluster of Aβ) which is perhaps the best well characterized aggregation-prone sequence in the literature and one of the few for which a high-resolution structure in the amyloid conformation is available, and (ii) we consider it in the context of the full length polypeptide (in fact fused to GFP, which acts as aggregation reporter) and not in an isolated manner as a short peptide. In addition, to the best of our knowledge our method is the only one in which aggregation propensities have been derived from experiments inside the cell, where the presence of the folding machinery might modulate the aggregation tendencies of polypeptides. Besides these three experimentally calculated propensity scales, Galzitskaya and co-workers have used the mean packing density for natural amino acid residues in protein structures, as a scale to predict amyloidogenic regions in proteins [31]. A comparative analysis of the four different scales shows that, despite these differences, there is a striking correlation between our in vivo obtained amino acid aggregation propensities and the others (Table 3), probably because they reflect a combination of properties characteristic of protein aggregation, such as hydrophobicity, secondary structure propensity or packing density. Importantly, although our method was not aimed at the specific identification of short amyloidogenic peptides, but rather of aggregation-prone sequences within natural proteins, AGGRESCAN identifies the presence of at least one hot spot in more than 80% of the amyloid forming sequences in a set of experimentally characterized peptide fragments of amyloidogenic proteins [32]. Also, using a database of six-residue peptides containing both amyloid formers and non-formers [32, 33] the receiver operator characteristic (ROC) curve for our method compares well with those obtained using structure-based data, such us packing density on protein structures or the 3D profile method, based on the threading of six-residue peptides through the known crystal structure of the cross-β spine formed by the peptide NNQQNY from Sup35 yeast prion [32] (Figure 2).

	AGGRESCAN	AMYLOID1a	AMILOYD2b	AMYLOID3c
AGGRESCAN	*	0.849	0.794	0.867
AMYLOID1a	0.849	*	0.764	0.837
AMILOYD2b	0.794	0.764	*	0.807
AMYLOID3c	0.867	0.837	0.807	*

^a AMYLOID1 corresponds to the method described in Ref. [22]

^bAMYLOID2 corresponds to the method described in Ref. [19]

^cAMYLOID3 corresponds to the method described in Ref. [31]

Overall, the success of different computational approaches in predicting aggregation-prone regions allows to propose that aggregation propensity in polypeptide chains is ultimately dictated by the sequence. As it happens with the native conformation of proteins, the sequence contains intrinsic information that is relevant for the regular structural arrangement within β-aggregates, implying that the mechanism of amyloid fibril formation is similar for different peptides and proteins.

Prediction of the effects of protein mutation on the aggregation propensity

Aggregation propensity varies sensibly with the composition and especially the sequence of the polypeptide, in such a way that single amino acid substitutions in proteins associated to depositional diseases result in many cases in changes in the global protein aggregation propensity and sometimes lead to premature or acute pathological symptoms. Predicting the effect of a mutation on the aggregation tendency of a protein could help to anticipate the implications of that mutation in disease development or assist the design, production and storage of more soluble variants of biotechnologically relevant proteins and peptides [34].

The algorithm predicts accurately a large set of natural and designed mutations of Aβ42 (the central hydrophobic region of this peptide was used for the derivation of the current a³v parameter set of AGGRESCAN) (Table 4). As an example, Figure 3 shows how the F19T mutation, which strongly decreases the deposition of Aβ42 [35], results in the loss of the central HS in this peptide. Interestingly, it also anticipates the lower aggregation propensity of Aβ40 and the recent observation that longer Aβisoforms possess increased aggregation propensities [36]. Several natural occurring mutations have also been shown to affect the aggregation rate of Tau [37–40]. The predicted changes in the respective Na⁴vSS correlate well with the experimental changes observed in these Tau variants (Table 4). Figure 3 shows the Area plot of wild type Tau and two of its mutants with highest, experimentally tested, aggregation propensities. The P301L substitution increases by 1,4 fold the area associated to the main "hot spot" in Tau. In addition, AGGRESCAN predicts the presence of a new HS in the S320F mutant, absent in the wild type form. This mutant is linked to tauthopaty, in which Tau accumulates in inclusion bodies [40].

Other disease-related protein mutants studied here are the recently described G4R and R68Stop of human Stefin B protein. These mutants have been related with the development of Myoclonus epilepsy of type 1. It has been described that R68Stop is more prone to aggregate than wild type Stefin, while the G4R mutant shows an opposite behavior, with a slower fibril formation rate [41]. In agreement with these experimental observations the algorithm predicts an increase in the Na⁴vSS associated to the R68Stop mutation and a decrease for the Gly4Arg change (Table 4).

Type 1 serum amyloid A protein (SAA1) is associated with Familial Mediterranean fever (FMF). FMF patients' genotypes are thought to correlate with the different phenotypes of the disease. A recent study [42] concludes that the gamma SAA1 allele is more frequently observed in the population devoid of amyloidosis, thus suggesting a protective effect of this allele on the development of the illness. In agreement with these results the AGGRESCAN analysis of amyloid A sequence variants predicts that the gamma variant misses a HS and has a lower Na⁴vSS than other alleles.

The Src homology 3 (SH3) domain of the p58 subunit of phosphatidyl-inositol-3 -kinase (PI3-SH3) is one of the best-characterized examples of a small globular protein unrelated to any known pathological condition that can form amyloid fibrils in vitro [43]. Aggregated species obtained from this protein have been found to be cytotoxic when added to cell cultures [44]. We have previously shown that the α-spectrin-SH3 (SPC-SH3) domain, which shares the same fold and 24% sequence identity with PI3-SH3, is a soluble protein that does not form amyloid fibrils under any conditions tested [45]. Nevertheless, a recent work found that the N47A mutation at the distal loop induces the formation of amyloid fibrils [46]. In contrast, the mutation of residue 47 to Gly does not promote aggregation (Ventura, S., unpublished results). According to AGGRESCAN a new HS occurs in the amyloidogenic mutant relative to both the wild type and N47G species, which could be responsible for its increased aggregation abilities (Figure 3).

Analysis of protein datasets

Besides analyzing the theoretical aggregation properties of single molecules and their individual mutants, AGGRESCAN is also able to deal simultaneously with a large number of sequences. This ability can be specially useful to compare the global aggregation properties of different protein sets and may help to delineate general rules underlying the relationship between the primary structure of proteins and peptides and their specific in vivo and in vitro depositional properties. With this aim we studied the correlation between the structural/aggregational features of 5 different groups of proteins and the predictions provided by AGGRESCAN. These datasets were: 1) natively globular proteins (160 proteins) (from SCOP, the ASTRAL40 set); 2) natively intrinsically unstructured proteins (51 proteins); 3) proteins which are soluble when overexpressed in bacteria (38 proteins); 4) proteins forming inclusion bodies when overexpressed in bacteria (121 proteins) and 5) amyloidogenic proteins (57 proteins) (see additional file 4).

It has been recently shown that proteins that form inclusion bodies (IB) upon recombinant overexpression in E. coli and proteins that form amyloids in vivo and/or in vitro share a good number of structural and functional features, including high purity of the aggregates, enrichment in beta-sheet structure, amyloid-tropic dye binding or enhanced proteolytic resistance [3]. Comparison of the two protein sets in search for similar trends in the predictions showed that, unexpectedly, the AGGRESCAN values for amyloid forming proteins are closer to those for IUP than for any other of the analysed datasets. Amyloid proteins have a lower Na⁴vSS and less HS than proteins in the IB or globular SCOP dataset (Figure 4). In contrast, the HSs in amyloid proteins comprise an area similar to those in IB or globular proteins, which is, however, significantly higher than the average HS area in IUP. These results suggests that, globally, the sequences of amyloidogenic proteins, like those of IUPs, have a low aggregation propensity, although the existence of specific aggregation-prone regions, absent or minor in IUPs, in a context in which they can act as specific and obligatory nucleation points from which the fibrillar structure could be expanded, finally results in highly ordered aggregates (Figure 4). This would explain why point mutations in the HSs of amyloidogenic proteins have usually a huge impact in their solubility, as they would modify the properties of one of the few points in the sequence that can promote and/or modulate aggregation. In contrast, the paradoxically higher-ranking global aggregation propensity of IB protein sequences is likely to indicate that here HS would play a less important role, since aggregation can also occur non specifically from many regions in the protein sequence. This would result in less structured deposits, and would also explain the rather moderate role of point mutations in IB aggregate formation. In other words, a given HS would promote specific amyloid formation in a low aggregating background, as its aggregation tendency outstands from the rest of the sequence. Conversely, the same HS needs to compete with the rest of the sequence to nucleate aggregation in a highly aggregating context (Figure 5). For the same reason unstructured aggregation is usually a much faster event than amyloidogenesis. Recent theoretical and experimental data support this view by showing that prevention of aggregation does not necessarily mean that amyloid fibril formation is abolished and vice versa [48]. This indicates that, despite the fact that aggregates and amyloid fibrils share many features, and the protein regions involved in their formation presumably intersect, they probably differ in the number and specificity of intermolecular contacts involved in the nucleation and stabilization of both types of polypeptide associations.

Recombinant protein production is an essential tool for the biotechnological industry and supports expanding areas of basic and biomedical research, including structural genomics and proteomics. The solubility of proteins expressed in bacteria under mass-production conditions is of major concern, since many recombinant polypeptides produced in bacteria accumulate as insoluble, often refractile, aggregates known as inclusion bodies (IBs) [49], excluding many biotechnologically relevant protein species from the market due to economically inconvenient yields. To date, the solubility of a given gene product has not been anticipated before gene expression. The comparison between the AGGRESCAN output values for proteins shown to be soluble under overexpression conditions in E. coli and those forming inclusion bodies shows that they share a similar number of HSs per 100 residues (NnHS), an expected output when considering that most proteins in both datasets are globular. However, IB-forming proteins have higher Na⁴vSS values than soluble proteins, suggesting that soluble proteins have, on average, a lower intrinsic aggregation tendency than IB-forming species, which may determine, at least partially, their relative behaviour upon overexpression. Overall, the predicted aggregation of proteins in the SCOP database is intermediate between that of soluble and insoluble proteins, suggesting that, in agreement with experimental observations, only a part of them would remain in the soluble cell fraction upon recombinant production. Although AGGRESCAN is able to catch the average trends in the aggregation of IBs and soluble protein groups, the individual outputs for proteins from both groups overlap significantly, making the prediction of the recombinant behaviour of a given sequence difficult in its present form. Besides, aggregation during recombinant production is the net result of several extrinsic and intrinsic factors, their relative importance depending on the protein and expression contexts.