AMYPdb is freely accessible 1. Users can browse web pages to obtain descriptions of each family, visualize protein sequences enriched with links to both UniProtKB and the Protein DataBank (PDB), study multiple sequence alignments using the Jalview editor 32, or access bibliographic references. An identity card for each protein is available from the "protein menu". Links to Wikipedia provide further information on some families. Sequences can be selected and exported in FASTA format for further analysis.

The amino acid sequence patterns are accessible by browsing the pages from the "pattern" menu, or by using the search interface. The search page contains several menus, allowing the user to focus on particular data. For instance, they can interrogate AMYPdb with UniProtKB or PROSITE identifiers or patterns to determine whether a particular protein or pattern is stored in AMYPdb. They can also submit a personal signature to find any matching amyloid proteins, or inversely, they can submit a sequence to find matching AMYPdb patterns. It is also possible to select patterns using thresholds on quality scores. This method is useful for discovering patterns shared between families.

Beyond functioning as a pure repository of knowledge, AMYPdb also provides private workspaces to anyone interested in further analysis. This allows users to manage their own working sets of proteins and patterns, which can be easily manipulated and organized accordingly to their research interests.

Below, we have illustrated several ways that AMYPdb can be useful in pattern research on amyloid proteins.

AMYPdb patterns can be used to highlight residues though to be important to the structure, function and evolution of protein families. One of the objectives of our project was to propose a list of specific amino acid sequence patterns for each family stored in the database. On the web interface of AMYPdb, the 3,621 amino acid sequence patterns are classified by their CF value. About 27% of the patterns have a CF ≥ 0.9 (956 of 3,621) and can be considered as very good descriptors for the corresponding amyloid families (Table 1, column n°5). The best results are shown in Table 2. It is interesting to note that all of the AMYPdb patterns noted in that table are of better quality than those in the PROSITE database.

There are many advantages in describing a protein family using several patterns rather than only one or two, as is done in PROSITE. First, the occurrence of more than one pattern increases confidence that a protein belongs to a specific family. Pattern distribution along sequences can also be used to assess conserved and variable regions in proteins. Indeed, highly specific patterns only describe conserved regions in proteins. Examples of this are the Tau and prolactin protein families. Human Tau protein is characterized by 13 patterns with CF ≥ 0.9, all found in the C-terminal region of the protein and covering barely 14% of the sequence. This suggests that the C-terminal part of Tau is the protein's main domain (indeed it is the microtubule-interacting region). On the other hand, 54 patterns with CF ≥ 0.9 are characteristic of human prolactin, and they are distributed all along the sequence, covering 32% of it. This suggests the presence of numerous important regions, which are likely correlated to the many known biological effects of prolactin.

AMYPdb is a MySQL relational database. The web-based interface was created using PHP and JavaScript, and relies on a modified version of the e107 content management system. The data are stored in 23 tables and occupies nearly 4 GB of disk space (Figure 1). The central table contains general information relevant to protein sequences. This is linked to additional tables containing two kinds of data: information collected from public libraries; and the results of protein sequence analysis. This data organization is suitable for complex queries, such as the extraction of amino-acid signatures matching the regions involved in aggregation.

The data flow is described in Figure 2. Raw data were extracted from three main public databases: protein files from UniProtKB (release 3.2); references from MEDLINE; and known patterns from PROSITE. We used the Sequence Retrieval System (SRS) program implemented at the European Bioinformatics Institute 41. SRS has the advantage of possessing a unique interface for interrogating multiple databases. In addition, results can be saved in eXtended Markup Language (XML) format, thus facilitating the exchange and manipulation of large amounts of data between different programs. UniProtKB and MEDLINE were searched using keywords describing the amyloid protein families. The keywords used were mainly protein and gene names commonly used in the amyloid research field. Several phylogenetically distant sequences of each family were then submitted to ScanProsite 42. This program scans protein sequences for the occurrence of signatures stored in the PROSITE database. The accession numbers of PROSITE patterns were then queried in SRS. All retrieved XML files were imported into AMYPdb. After this first selection step, we obtained a catalog of 31 amyloid families (Table 1), containing 1,284 amyloid proteins, 1,692 references and 38 PROSITE patterns. Data stored in AMYPdb are those of precursor proteins, amyloidogenic peptides and partial sequences. In this first version of AMYPdb, no immunoglobulin light/heavy chains have been stored, due to the high sequence diversity of those proteins.

Since the content of UniProtKB was evolving during the various development phases of our project (631,592 entries added from release 3.2 to 6.1), we updated AMYPdb by semi-automatically sorting the 267,490 sequences extracted with WAPAM. From this group, we picked out 421 protein sequences matching highly specific AMYPdb patterns, and we assigned these sequences to the corresponding amyloid families. This updating procedure increased the AMYPdb sequence group to 1,705 members (1,063 full-length sequences and 642 fragments, Figure 2), and leaving 267,069 sequences in the non-amyloid protein group.

To measure pattern performance, we used three fitness scores commonly used in classification problems: sensitivity (Sen); specificity (Spe); and correlation (CF) 4950. The scores of the 3,621 patterns were calculated for each family, using only full-length sequences. For each pattern, true positives (TP) and false negatives (FN) are sequences of a family respectively either matching or not matching the pattern. False positives (FP) are either amyloid sequences not belonging to the considered family but which match the pattern, or non-amyloid sequences matching the pattern. True negatives (TN) are non-amyloid sequences not matching the pattern. Therefore, when a pattern is specific for one amyloid family, it has high Sen, Spe and CF scores for that family and low scores for other amyloid families. When a pattern is conserved in several amyloid families, the Spe of the pattern remains high for each family, but Sen and CF scores can decrease dramatically. Due to calculation limitations, non-amyloid sequences were obtained from 1 of 3 data sets: the 2,032,835 proteins of UniProtKB was used for patterns matching less than 5000 non-amyloid proteins; the 267,069 non-amyloid sequences resulting from the WAPAM search was used for patterns for matching between 5,000 and 10,000 non-amyloid proteins; and a random pool of 50,000 proteins was used for patterns matching more than 10,000 non-amyloid proteins.

For each pattern, sensitivity corresponds to its ability to describe an amyloid family, while specificity corresponds to its ability to discard proteins not belonging to the amyloid family. The Pearson-Mathews correlation coefficient measures the global prediction accuracy by which a pattern describes a protein family. A CF of +1 indicates that the pattern perfectly differentiate the amyloid family from other amyloid families and non-amyloid proteins.

S e n = T P / ( T P + F N ) S p e = T N / ( F P + T N ) C F = ( ( T P × T N ) − ( F P × F N ) ) ( T P + F P ) × ( F P + T N ) × ( T N + F N ) × ( F N + T P ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@84F1@

The accuracy of hypotheses deduced from bioinformatic methods strongly depends on data sets quality. In the present study, the sequence sets used in the pattern discovery method were those of protein families. In order to facilitate sequence extraction, especially for the discovery of patterns link to misfolding and aggregation, all the proteins were sorted into one of the five following quality categories:

• Amyloid in vivo: the precursor protein, or a specific sub-segment, forms fibrils in human, or animals, or is a yeast prion. Proteins of this class are unambiguously described in literature and are identified by specific keywords in UniProtKB ("Amyloid" or "Prion").

• Amyloid in vitro: the polypeptide forms fibrils under experimental conditions.

• Amyloid in silico: the polypeptide forms fibrils using computational techniques, including protein threading and molecular dynamics simulations.

• Putative amyloid protein: the protein is a member of an amyloid family, but the amyloid properties of that specific member were not assessed.

• Unclassified protein: the protein family does not fulfill the definition of amyloid 4, but sparse data show that at least one member of the family shares some amyloid properties.

At the present time, the classes 2 and 3 are empty. Experts are welcome to contribute to relevance of biological information by changing the status of the proteins using the Wiki system.