To gather the data to generate the database, the sequences of all IDRs were downloaded in FASTA format from DisProt (2021_08 release), a manually curated database of disordered proteins (Quaglia et al., 2021). The Waltz algorithm was used to identify potential amyloidogenic segments. The “Best overall performance” of Waltz sets the detection threshold at a value of 92.0 (Maurer-Stroh et al., 2010), which performs very well on classic amyloid regions but is too high to identify CARs. We previously showed that thresholds of 85.0, 80.0 and 73.5 allowed us to identify and rank those regions according to their differential polar and ionizable residues content (Santos et al., 2021). IDRs longer than 20 amino acids were scanned with these parameters. Only CARs longer than six residues were included in the final database. Information about each detected CAR can be accessed through different links that map to its correspondent disordered region (DisProt) or the entire protein (UniProt).

The server site was developed using the Django 3.0 web framework. The front-end is written in HTML/CSS. The database search engine is based on a JavaScript filter table that allows to retrieve CARs specific information upon querying by distinct identifiers. It is platform-independent and does not require previous registration.

The peptides with the sequences Ac-EGVSISVYR-NH₂ and Ac-GAVAVSAA-NH₂ corresponding to the CARs-derived segment of titin and RPLP2 proteins respectively were purchased from Genscript Biotech (New Jersey, United States) with a purity >95%. Peptides were N-terminal and C-terminal protected as a strategy to mimic the protein environment. Peptide stock solutions were prepared by solubilizing the lyophilized peptides at a final concentration of 1 mg/ml in 100% dimethyl sulfoxide, divided into aliquots and stored at −80°C. For aggregation assays, peptides were diluted to 200 µM in phosphate saline buffer (PBS) and incubated at 37°C for 72h-7 days in a 96-well plate (SIGMA-Aldrich, Saint Louis, United States) under continuous agitation (100 rpm). The final concentration of DMSO in the assay was 4 and 1.3% (v/v) for titin and RPLP2 peptides, respectively. Additional details from the aggregation experiments are described on the MIRRAGGE spreadsheet (Martins et al., 2020) of the Supplementary material.

The fluorescence emission spectra of the binding of 40 µM Thioflavin-T (ThT) to the aggregated peptides were recorded using a Spark plate reader (Tecan, Männedorf, Switzerland). The samples were excited at 440 nm with an excitation bandwidth of 5 nm. The emission spectra were recorded from 460 to 600 nm with an emission bandwidth of 5 nm and a 2 nm interval.

Congo red (CR) interaction with the peptide solutions was tested using a Cary 100 UV/Vis Spectrophotometer (Varian, Palo Alto, United States). The absorbance spectra were recorded from 400 to 650 nm using a 1 cm optical length quartz cuvette at room temperature. 10 µl of peptide samples were preincubated with 90 µl of CR at a final concentration of 5 µM for 10 min at room temperature. Solutions of PBS with 5 µM CR and without peptide were used as negative controls.

The secondary structure of incubated peptides was analyzed by attenuated total reflectance Fourier transform infrared (ATR FT-IR) spectroscopy using a Bruker Tensor FT-IR Spectrometer (Bruker, Massachusetts, United States) with a Golden Gate MKII ATR accessory. The aggregated peptide solutions were dried out under N₂ (g) atmosphere. Each spectrum consisted of 32 scans and was measured at a spectral resolution of 4 cm⁻¹ within the 1800–1,500 cm⁻¹ range. All spectral data were acquired and normalized using the OPUS MIR Tensor 27 software and the Peak Fit 4.12 program (Systat Software Inc., San Jose, United States) was used for data deconvolution.

For negative staining, aggregated peptide samples were sonicated at intensity two for 5 min in an ultrasonic bath (VWR ultrasonic cleaner, Leuven, Belgium), placed onto carbon-coated copper grids, and incubated for 1 min. The excess of the sample was removed using ashless filter paper. Then, the grids were washed with distilled water and incubated for 1 min with 2% (w/v) uranyl acetate and the excess of uranyl acetate was removed carefully using ashless filter paper. The micrographs of aggregated peptides were obtained using a TEM JEM-1400 (JEOL, Tokio, Japan) operating at an accelerating voltage of 120 kV. Representative images of each sample were selected.

We believe that the CARs-DB can help identify regions at risk of promoting aberrant interactions that may eventually lead to pathology, since the presence of CARs has been associated with neurodegenerative diseases and cancer (Santos et al., 2021).

Amyloid formation has been experimentally validated for short peptides or motifs belonging to different IDPs linked to disease. That is the case of α-synuclein, an IDP whose deposition in dopaminergic neurons is intimately ligated to the onset of Parkinson’s disease (Luk et al., 2012). A stretch of 12 amino acids in the NAC domain of α-synuclein is essential for filament assembly (Giasson et al., 2001). We sought this short region (71-VTGVTAVAQKTV-82) in our CARs-DB and found it under the 73.5 threshold with a Waltz score of 78.51. Another interesting example is the microtubule-associated protein tau, an IDP that is found as intracellular solid aggregates in Alzheimer’s disease. A short hexapeptide in the C-terminal end of the protein (304-VQIVYK-312) is essential and sufficient for tau fibrillation (von Bergen et al., 2000; Meng et al., 2012). A CAR containing this peptide is present in the database with a Waltz score of 89.37. Aberrant aggregation can also disturb liquid-liquid phase separation (de Oliveira et al., 2019). In amyotrophic lateral sclerosis (ALS), the amyloidogenic core region of TDP-43 (residues 318–343) initiates its aggregation and facilitates its cytoplasmatic inclusion (Jiang et al., 2013). This region is positive at the 73.5 threshold and could not be detected with the default Waltz cutoff.

The amyloidogenic capacity of CARs in cancer-related proteins has also been demonstrated for peptides in p53 (326-EYFTLQIR-333) and retinoblastoma-associated protein (831-ILVSIGESFG-840) (Santos et al., 2021), both present in CARs-DB under the 85 and 80 thresholds, respectively.

Overall, these examples illustrate how amyloidogenic regions in pathogenic proteins may escape detection with conventional algorithms and the need to redefine the properties of amyloidogenic sequences since they can exhibit a marked polar character when they are embedded in a disordered context, as the regions listed in CARs-DB.

The dynamic nature of IDPs makes them highly exposed to the solvent and sensitive to the environment (Uversky, 2009). From an evolutionary point of view, the prevalence of exposed CARs in IDPs makes sense only if they contribute to the protein function since they necessarily endorse IDRs with a moderate but significant propensity to misassemble and malfunction (Santos et al., 2021).

The IDPs’ ability to establish interactions depends largely on the presence of modular interaction units in their sequences. These interaction-favoring regions are known as short linear motifs or linear interacting peptides (LIPs) (Davey et al., 2012; Pancsa and Fuxreiter, 2012; Piovesan et al., 2021). We demonstrated previously that the probability of any CAR residue to map into a LIP is > 2-fold higher than the one expected by chance (Santos et al., 2021), which suggested a role of CARs in facilitating IDPs’ PPIs. This is especially true when the interaction is mediated by a folding-upon-binding mechanism because the acquisition of structure in the bound state requires physicochemical properties that make the sequence more sensitive to aggregation in the unbound conformation (Langenberg et al., 2020).

The above observations suggest that CARs-DB might be used, in combination with dedicated applications, like ANCHOR (Dosztanyi et al., 2009), to detect LIPs susceptible to go on disorder-to-order transitions upon binding. To illustrate this potential, we describe four examples of CARs involved in disorder-to-order transition. These include short sequences from the receptor-binding domain (RBD) of the SARS-CoV-2 spike protein (482- GVEGFNCYFPLQSYGFQPTNGV-503, Waltz score = 80.1), the human microtubule-associated protein tau (350-VQSKIGSLDNITH-362, Waltz score = 82.0), the SUMO-interacting motif (SIM) PIAS2 (466-KVDVIDLTIESSSDEE-481, Waltz score = 87.1) and the p27^Kip1 (p27) protein (70-LEGKYEWQEVEK-81, Waltz score = 75.4).

The RBD of the SARS-CoV-2 spike protein interacts with the angiotensin-converting enzyme-2 (ACE2) receptor found in human cells and is key to establishing cell host-virus interactions (Sternberg and Naujokat, 2020). Five antiparallel β-sheets form the core of this domain. From this core, it protrudes the receptor-binding motif (RBM), a region that is a flexible and disordered loop in the unbound state, and becomes ordered in the ACE2-bound SARS-CoV-2 Spike structure (Lan et al., 2020), while containing most of the interacting residues with the receptor. Interestingly, a 22-aa long CAR is found in the RBM overlapping with the RBD β6 strand in the bound state, indicative of an active role in binding (Figure 4A).

Tau is intrinsically disordered in solution and locally folds upon binding to microtubules (Kadavath et al., 2015). Specific intermolecular interactions stabilize the tau-microtubule complex (Brotzakis et al., 2021). Importantly, a Tau CAR of 12 residues is found to fold into an α-helical CAR within a strongly interacting Tau region that establishes contact with two α-helices of α-tubulin (Figure 4B). When the tau protein is detached from these structures, it becomes more sensitive to aggregation and may lead to the formation of neurofibrillary tangles, a trademark of the development of Alzheimer’s disease (Hardy and Selkoe, 2002).

Even protein families that are characterized by their high solubility might eventually self-assemble into β-sheet enriched aggregates (Sabate et al., 2012). It is the case of the ubiquitin-like SUMO proteins, which expose an aggregation-prone β-sheet needed for their binding to target proteins through their SUMO Interacting Motifs (SIMs). SIMs are disordered when unbound and form an intermolecular β-sheet with SUMO in the bound state. In the case of the PIAS2 SUMO substrate, its SIM fairly coincides with a 16 residues CAR (Figure 4C).

A canonical example of disorder-to-order transitions is p27, an IDP that folds upon binding to its Cdk/cyclin partners (Sivakolundu et al., 2005). The interaction regions encompass a 12 residues CAR that in the complex folds to form a short intermolecular β-sheet with the Cdk2 domain (Figure 4D).

Overall, in the four cases, the disordered nature of CARs turns into a partially folded structure upon binding to form the protein complex. This illustrates how CARs-DB may help to uncover LIPs displaying structural transitions upon PPIs which would have remained unnoticed with most conventional aggregation prediction approaches.

Providing additional experimental evidence on the amyloid potential of CARs would allow validation of the concept and justify the development of the CARs-DB. With this aim, here we tested the amyloidogenicity of two identified CARs involving the segment 10429-EGVSISVYR-10437 (Waltz score = 85.9) of titin protein (Q8WZ42) and the stretch 69-GAVAVSAA-76 (Waltz score = 75.3) region of S60 acidic ribosomal protein P2 (RPLP2) (P05387). None of the two proteins have been previously demonstrated to contain amyloid-forming regions.

Titin, the largest known polypeptide, constitutes the elastic filament of the sarcomere, playing a pivotal role in passive muscle force. The modular organization of titin comprises four main regions: the N-terminal domain, the I-band and A-band, and the C-terminal domain (Eldemire et al., 2021). Titin’s elastic capacity is mainly attributed to the I-band, and more specifically to the PEVK segment, an intrinsically disordered region where ∼75% of the amino acids are either proline (P), glutamate (E), valine (V), or lysine (K) (Nagy et al., 2005; Sudarshi Premawardhana et al., 2020). Nevertheless, the molecular interactions that contribute to PEVK’s elastic nature remain poorly understood. We predicted a 9-residue CAR comprising residues 10,429 to 10,437, belonging to the 9,880 to 12,031 disordered sequence that maps at the PEVK region.

As a second candidate, we selected RPLP2, an acidic phosphoprotein that is part of the ribosomal stalk and acts during translation elongation (Campos et al., 2020). The C-terminal domain of RPLP2 is intrinsically disordered and highly conserved as it is determinant in the recruitment of the factor eEF2 (Mishra et al., 2014; Campos et al., 2020), which is essential in the translocation step during translation elongation. We selected an 8-residue CAR corresponding to residues 69–76 mapping at the IDR comprising residues 63–115 of the C-terminal domain.

We assessed the in vitro self-assembly properties of the two selected CARs. The peptides were prepared at 200 µM in PBS and incubated at 37°C and 100 rpm for 72 h or 7 days. First, aggregation into amyloid-like structures from its initial soluble state was evaluated using the amyloid-specific dyes Thioflavin-T (Th-T) and Congo Red (CR). Both peptides promoted a high increase in the Th-T fluorescence emission that evolves significantly with time, reaching a maximum after 7 days (Figure 5A). Accordingly, subsequent experiments were all performed at 7 days of incubation time. As expected, the incubated peptides also displayed increased CR absorbance and the characteristic amyloid redshift of the spectra maximum compared to the free dye (Figure 5B). These results suggest that these peptides have the potential to aggregate into amyloid-like structures. Next, to inspect the secondary structure content of the assemblies, we used FT-IR spectroscopy and recorded the amide I region of the spectra (1,700–1,600 cm⁻¹), corresponding to the absorption of the carbonyl bond group of the protein backbone. The deconvolution of the spectra allowed us to identify the main peaks in the 1,620–1,630 cm⁻¹ region, accounting for approximately 50% of the total area in both peptides, indicating that they have acquired a predominant intermolecular β-sheet structure (Figure 5C). Finally, we performed morphological analysis using transmission electron microscopy to confirm peptides’ internal fibrillar structure. In agreement with the previous results, both solutions contained canonical amyloid fibrils that are twisted in the titin peptide, whereas in the PRLP2 peptide, they are straight and tend to associate laterally (Figure 5D).

Together, the above-described experimental data provide compelling evidence for the amyloidogenicity of these CARs-DB derived peptides.