BGPT: Paper Review: AMYPdb: A database dedicated to amyloid precursor proteins

Explore by Goal

Quick Explanation Copied

AMYPdb (2008) at a glance

AMYPdb is a free amyloid-precursor knowledge database that centralizes 31 amyloid families, 3,621 short sequence patterns (incl. 38 from PROSITE), and library/PDB links, using pattern discovery + pattern matching to enable family assignment and motif exploration.

Primary evidence:

Long Explanation

Paper Review: AMYPdb (amyloid precursor proteins)

Evidence basis: full-text content you provided for 10.1186/1471-2105-9-273. All quantitative claims below are taken from that text.

1) Visual: dataset + pattern inventory (from the paper)

The paper reports: 31 families, initial 1,284 amyloid proteins, updated 1,705 proteins (split into 1,063 full-length + 642 fragments), 3,621 patterns total (incl. 38 PROSITE patterns), 956 patterns with CF ≥ 0.9, and 267,490 UniProtKB sequences matching AMYPdb patterns (with 267,069 remaining non-amyloid after curation via specificity thresholds).

2) Visual: pattern “high CF” proportion (CF ≥ 0.9)

The paper states that 27% of patterns have CF ≥ 0.9, reported as 956 of 3,621.

3) What AMYPdb claims to do (capabilities)

3.1 Database browsing + analysis workflow

Users can browse family pages, visualize sequences enriched with links to UniProtKB and PDB, and study alignments using Jalview; sequences can be exported in FASTA.
The database supports pattern-based interrogation: searching by UniProtKB/PROSITE identifiers, querying by submitting personal signatures, or submitting sequences to find matching patterns.
Private workspaces and a wiki system are provided for user curation/sharing.

3.2 The core computational idea: flexible sequence-pattern discovery + matching

Raw data were assembled from UniProtKB, MEDLINE, and PROSITE; the paper uses SRS as an interface for interrogation and XML exchange.
Pattern discovery uses PRATT (advanced PRATT) with maximum pattern length limited to 20 amino acids; clusters are chosen to reflect default physico-chemical grouping and two amyloid-relevant cluster sets (secondary-structure propensity; protein-protein interface relevance).
Pattern matching scans UniProtKB using WAPAM, and the paper reports runtime acceleration via Rdisk (15 hours reported vs 2,000 hours estimated without it).

3.3 How pattern “quality” is quantified (and what that does/does not prove)

Pattern performance is evaluated using sensitivity (Sen), specificity (Spe), and a correlation factor (CF). The paper defines TP/FN/SF/TN/FN logic in terms of matches across amyloid-family vs non-amyloid sets.
Skeptical interpretation: high Sen/Spe/CF indicates a pattern discriminates sets in the curated reference sets, not that the pattern is causally responsible for aggregation in vivo. This limitation follows directly from the paper’s own statement that it cannot yet automatically predict aggregation regions due to sparse in vivo data.

4) Evidence examples presented in the paper (what is shown)

The paper presents case studies using AMYPdb patterns and compares to known annotations/motifs, including:

Human transthyretin (TTR): it reports that 97 patterns (CF ≥ 0.9) mapping across the sequence correspond to 31 of 37 single-site amyloidogenic variants described in UniProtKB, and it provides a pattern representation linking underlined residues to pathogenic missense mutation positions (73–92) in human TTR.
Tau isoforms: it reports that using the search interface, it finds 8 patterns (CF ≥ 0.9) matching 1–6 isoforms, including a junction-located pattern (P-G-G-G-[KNS]-V-Q-I-[FIV]-[DHNY]) that matches regions across Tau isoforms and includes tau pathogenicity variants per UniProtKB numbering.
Prion (PrP): it uses previously predicted amyloidogenic determinants (Hamodrakas et al.) and reports that AMYPdb search for pattern signatures overlapping those segments is summarized in a table, noting a high density of mutation/modification sites overlapping early amyloidogenic segments and linking to cysteine/disulfide and glycosylation features in UniProtKB context.
Cross-family conserved motifs: it describes an observation of shared patterns between huntingtin and prolactin, then adds a follow-up using PRATT/PRATT with a larger dataset and a new pattern (N°5) with high specificity claimed by Sen=1 for a new pattern and notes that in a later UniProtKB release it retrieves additional sequences with low false positives reported.

4.1 Skeptical appraisal of these examples

Strength: the examples demonstrate pattern matching against curated/annotated sequence features (UniProtKB variant lines and described determinants) within the same data ecosystem.
Blind spot: pattern overlap with annotated disease variants does not directly establish aggregation mechanism (shared motifs may correlate with co-factors, structural constraints, or curated annotation biases). The paper itself flags that it cannot yet automatically predict aggregation regions in vivo because in vivo fibril-forming sequence data are sparse.

5) Methods audit (reproducibility-critical points)

Data provenance & versioning: the paper states raw data were drawn from UniProtKB (release 3.2 early), UniProtKB release 6.1 later for scanning/matching, MEDLINE, and PROSITE, with SRS used to retrieve XML. This is good provenance, but it also means results can drift with database releases.
Pattern discovery parameterization: the paper sets maximum pattern length to 20 and uses specific residue-cluster sets. It also reports a “select patterns matching 100% of the sequences of each set” criterion. This risks overfitting to small curated family sets—especially for families with limited numbers of known sequences—an issue partially addressed by excluding some families (the paper says 9 families could not be submitted due to small number of sequences).
Non-amyloid negative set design: the paper describes multiple non-amyloid datasets used for performance computation depending on pattern match volume (using UniProtKB full set vs the WAPAM-derived non-amyloid pool vs a random pool). This heterogeneity matters because it directly affects Spe/CF values.
Classification and updating procedure: it reports scanning for 3,621 patterns yielding 267,490 matching sequences, then semi-automatically sorting to assign 421 sequences to amyloid families, increasing the amyloid sequence group to 1,705 and leaving 267,069 in a non-amyloid group. The semi-automatic step may inject subjective thresholds/possible inconsistencies across time unless fully automated/reported.

6) Visual: “method pipeline” as a directed graph (from the paper)

This schematic is a faithful condensation of the paper’s stated flow: raw data extraction (UniProtKB/MEDLINE/PROSITE) via SRS; family curation; pattern discovery with PRATT; pattern matching with WAPAM; quality scoring with Sen/Spe/CF; then database/user-facing analysis including exports and search.

7) Critical appraisal (knowns vs uncertainties)

What is well-supported by the paper text

Scale and scope (as reported): 31 families, 3,621 patterns, and updated counts for proteins and matches.
Internal logic of quality evaluation: the paper defines Sen/Spe/CF in terms of matching behavior across amyloid-family vs non-amyloid sets and describes how negative sets are constructed depending on match volume.
Explicit limitation: lack of dense in vivo fibril-forming sequence data prevents automatic aggregation-region prediction.

Uncertainties / possible failure modes

Negative-set construction can change scores: Spe/CF values depend on which non-amyloid proteins were used as negatives for each pattern category.
Pattern discovery selection criterion may overfit: selecting patterns that match 100% of sequences in a discovery set can produce brittle patterns if the family set is incomplete or biased. The paper notes missing families due to small sample sizes and also relies on curated amyloid-family membership.
Database drift with external sources: AMYPdb updates depend on external database releases; the paper itself describes differences between UniProtKB releases during development.
Mechanism vs correlation: even if a pattern correlates with known amyloidogenic regions/variants, causal mechanistic claims are not established by pattern-matching alone; the paper’s own limitation acknowledges inability to automatically predict in vivo aggregation regions.

8) How you can use this paper inside BGPT

Use BGPT to:

Recreate AMYPdb pattern summary charts (counts, CF thresholds) and compare across amyloid families as extracted from AMYPdb pages (if available via BGPT queries).
Generate a “pattern signature” interpretation workflow: map patterns to UniProtKB annotated features and then cross-check with external aggregation predictors (you would supply those later, but AMYPdb’s role is defined in this paper).

Author Review Buttons (bespoke BGPT pages)

Feedback:

Updated: April 27, 2026