SeqCP: A sequence-based algorithm for searching circularly permuted proteins

30-second read

SeqCP aligns normal and duplicated query sequences, using boundary shifts, coverage, gaps, and improvement ratios to detect CP relationships.

Central question

Without known 3D structures, how can normal and duplicated sequence alignments detect circularly permuted proteins and locate CP sites?

Intuition

Duplicate the query into query-plus-query: if the target is a circular permutant, its two fragments can align continuously across the duplicated-query seam. Boundary shifts, coverage, gaps, and improvement over the normal alignment then separate CP patterns from ordinary similarity.

Why it matters

Structure-based CP searches cannot process the many proteins known only by sequence. SeqCP expands discovery through sequence-database search and is about 33 times faster than CPSARST; its classification AUC or threshold is not structural-alignment quality.

Prerequisites

• Understand circular permutation and the duplicated-query trick
• Understand BLAST local alignment, identity, similarity, and gaps
• Know sensitivity, specificity, MCC, and ROC AUC

paper-specific guide · plain → technical → input → output → source

Method walkthrough

1. 01 · Search normal and duplicated queries

   Search the same target database with q and qq-prime; only targets aligned by both queries enter the candidate pool.

   Technical reading: blastp uses low-identity-oriented settings including word size 2, BLOSUM45, gap open 10 and extension 3; CP sites are computed from first/second-copy boundary positions in the qq-prime alignment.

   Input: A query sequence and FASTA target database

   Output: Hits with normal/duplicated alignments and candidate CP sites

   Boundary: Very low identity or short fragments may fail to produce recognizable two-block alignments.

   PDF pp. 2–5, Methods §§2.1–2.4, Figs. 1–2

2. 02 · Apply a nine-rule initial gate

   Use protein-length ratio, identifiable CP sites, normal/duplicate boundary shifts, identity consistency, and duplicated-alignment rate to remove pairs unlike CP.

   Technical reading: Examples include mean length at least 100, short/long ratio at least 0.8, both boundary shifts at least 20%, and duplicated-alignment rate at least 0.8; on CPDB the nine criteria reduce 2,555 marginal pairs to 22 while retaining 472 of 480 high-quality pairs.

   Input: Geometry and quality features from the two alignments

   Output: A small, high-recall CP candidate pool

   Boundary: Thresholds come from CPDB-derived distributions and may inherit its domain/structure bias.

   PDF pp. 5–10, Methods §§2.6–2.7, Table 2 and Fig. 5

3. 03 · Score seven features with a tunable threshold

   Combine duplicate coverage, gap density, similarity, improvements in identity/similarity/coverage over the normal alignment, and a stability factor into a SeqCP score.

   Technical reading: Weights are derived from CPDB training; AUC is 0.89 on the 2007 test, 0.88 on the 2019 test, and 0.81 on a homology-reduced independent test. The recommended threshold is 0.60; 0.75 trades recall for higher precision.

   Input: Seven features for candidates surviving initial filters

   Output: Ranked CPMs, candidate CP sites, CP identities, and confidence

   Boundary: The score is a discriminator under its training distribution; dataset shift requires recalibration.

   PDF pp. 6–11 and 15–16, Eq. 15, Fig. 6, Table 3, Discussion §4.4

Key result

It searches large sequence databases without known 3D structures, extending CPSARST/CPDB knowledge to sequence scale.

Evidence-guided deep reading

Paper facts, project readings, and teaching models are labelled separately.

paper-fact

Where the high-quality label comes from

CPDB contains 4,169 nonredundant CP pairs; the paper defines 480 high-quality pairs through CE-CP structural alignment, length/block/site/gap/coverage/RMSD criteria, and manual inspection, labeling the rest marginal.

SeqCP therefore learns sequence signatures approximating this structural-curation definition; AUC is not an unbiased prevalence or sensitivity estimate for every natural CP.

Source locator: PDF pp. 3–10, Methods §§2.2 and 2.5, Table 1

paper-fact

0.60 and 0.75 serve different needs

A threshold of 0.60 yields the highest MCC on test data; the paper summarizes sensitivity and specificity near 86%, making it suitable for broad discovery.

If positive follow-up is costly, 0.75 reduces candidates by 21% while removing 96% of unlikely cases. The algorithm has not suddenly become more accurate; the operating point changed.

Source locator: PDF pp. 10–11 and 16, Fig. 6, Table 3, Discussion §4.4

paper-fact

How CP search improves modeling

Normal-sequence templates for UPI00057E200D give incomplete coverage; after SeqCP suggests CPM71, SWISS-MODEL builds a full model that superposes with an AlphaFold2 model at RMSD 1.20 A over 154 aligned C-alpha atoms.

For A0A388KU10 CPM475, template coverage rises 49 to 90% and global identity about 22 to 41%; the YbeA CPM74 case shows sequence detection can complement structure-only methods when CP changes a domain-swapped conformation.

Source locator: PDF pp. 11–15, Figs. 7–11, Results §3.6

Study design and evaluation

Data and samples

CPDB with 2,238 proteins and 4,169 pairs for training; NrPDB100-2007 with 26,349 proteins for threshold testing; NrPDB100-2019 with 85,725 for large-scale testing; 20,000 simulated CPMs for site testing.

Baselines

• A CE-CP structural standard, CPSARST site localization, existing motif/sequence methods, and normal versus CPM-assisted modeling

Metrics

ROC AUC

Threshold-free discrimination of high-quality versus marginal pairs by the SeqCP score.
Boundary: Depends on label definition, class balance, and dataset relatedness.

MCC

Threshold quality considering TP/TN/FP/FN, used to choose 0.60.
Boundary: The optimum may shift with deployment prevalence.

CP-site accuracy

Fraction of simulated descendants with the exact CP site recovered.
Boundary: In-silico mutations do not cover every natural evolutionary process.

Reported result

SeqCP reaches about 0.9 AUC across tests with a recommended 0.60 threshold, at least 80% exact-site accuracy above 20% identity, and about 1.75M pairs per minute—roughly 33 times CPSARST.

PDF pp. 3–16, Figs. 1–11 and Tables 1–3

teaching-model · not a reported experiment

Teaching example (not a reported experiment)

Thresholds are decisions: choose by follow-up cost

Teaching model: you can experimentally validate 100 candidates, while SeqCP produces 500 hits at or above 0.60.

1. For a discovery catalog, keep 0.60 and validate a stratified sample to estimate calibration.
2. If each experiment is costly, start with the 0.75 high-confidence pool, then apply diversity criteria to avoid selecting only one family.
3. Report candidate counts, precision targets, and miss risk at each threshold rather than claiming one universally correct threshold.

Takeaway: Score ranking, threshold decision, and experimental budget are three layers of one pipeline.

outside-scope

Evidence boundary versus FAST

SeqCP is sequence-based CP retrieval; its classification performance says nothing directly about FAST RMSD or coverage.

Lawful source and access

17 pages · SHA-256 7723ea237f46a1db5071d7d00e60364b751bc44e0638a0a1e1211ac704c6a76d

Published online in 2022 and assigned to the 2023 volume; lawful open full text.

Europe PMC open-access PDF

Limits and misreadings

• Scores and thresholds are trained on CPDB cases, so dataset bias affects generalization.

Source locator map

1. PDF pp. 2–6, Methods §§2.1–2.7 and Figs. 1–2
2. PDF pp. 7–11, Results §§3.1–3.5, Fig. 6 and Tables 1–3
3. PDF pp. 11–16, Results §3.6 and Discussion §§4.1–4.5

Check understanding

1. Why does a duplicated query reveal CP?

   Answer: The target's two rearranged fragments can align continuously across the seam between query copies.

   A normal alignment often captures only one fragment or shows boundary shifts.

2. Is AUC 0.89 an alignment RMSD?

   Answer: No, it is CP-classification discrimination.

   Structural geometry requires separate coverage/RMSD evaluation.

3. What is the tradeoff of 0.75 versus 0.60?

   Answer: Precision/specificity rise while sensitivity falls.

   It suits settings with expensive follow-up.

Completion task: For a hypothetical CP-search project, state the costs of false positives and false negatives, then choose a threshold and explain a validation-sampling plan.

Paper-specific glossary

duplicated query

Two concatenated copies of a query that let a circularly shifted target align across the seam.

boundary shift

A difference in alignment start/end positions between normal and duplicated queries, used as a CP signature.

MCC

A binary metric that combines all four confusion-matrix cells and remains informative under class imbalance.

operating point

The specific sensitivity/specificity/precision combination after choosing a threshold.