Rich structure alphabets enable highest accuracy protein search

30-second read

Reseek v3 combines multiple structure alphabets with hierarchy-specific statistical models for family, superfamily, and fold search.

Central question

Can layered filters, richer structural alphabets, and family/superfamily/fold-specific statistical models improve search accuracy at AI-scale?

Intuition

Reseek v3 first finds two-hit diagonals with a tiny κ alphabet, then runs SIMD Smith–Waterman on a one-byte ν alphabet, and only then reranks with Mega′ spanning about 5×10^22 states. High-scoring alignments combine reverse score, LDDT, and DALI signals.

Why it matters

The preprint reports median rank 1 for Reseek-sensitive across collection and categorization metrics on SCOP40/SCOP40c/CATH40, with significant pairwise rank tests. It remains a single-author, unreviewed study using custom summary metrics, and FAST is absent.

Prerequisites

• Understand structural alphabets, spaced seeds, and SIMD Smith–Waterman.
• Know the differences among CVE, PR, ROC, collection, and categorization.
• Recognize the evidence levels of preprints, author-run benchmarks, and independent validation.

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

Method walkthrough

1. 01 · Fast κ spaced-seed filter

   Compress each residue into one of 32 coarse states and keep only pairs with two similar 7-mers on the same diagonal, quickly discarding most implausible comparisons.

   Technical reading: Kappa32=aa4×sec4×pm2; k=7 with spaced pattern 1010011. Reseek builds an inverted index over query k-mers, so RAM scales roughly with query size rather than database size.

   Input: κ sequences for queries and database structures.

   Output: A reduced set of pairs passing the two-hit diagonal test.

   Boundary: Filtering discards some true positives; the fast/sensitive presets explicitly trade speed for sensitivity.

   PDF pp. 2 and 5, Methods: Algorithm workflow; PDF p. 10, Fig. 1

2. 02 · Filter with ν, then refine with Mega′

   The second gate runs fast local alignment with a 256-state one-byte representation; only high scorers pay to construct expensive Mega′ profiles combining many features.

   Technical reading: Nu256=aa4×sec32×pm2 and uses modified Parasail SIMD Smith–Waterman; a forward-plus-reverse ν score filters again. Mega′ is a weighted product of component alphabets with about 5×10^22 states, aligned forward and reverse through on-the-fly profiles.

   Input: κ-passing candidates and ν/Mega′ profiles.

   Output: Information-rich local alignments and forward/reverse scores.

   Boundary: Mega′ construction is expensive; runtime depends on earlier filter hit rates and query/database chunking.

   PDF pp. 2–3 and 5, Algorithm workflow; PDF p. 10, Fig. 1

3. 03 · Rank with task-specific test statistics

   Family, superfamily, and fold are not the same kind of similarity. Version 3 learns separate weights for each and combines alignment, short-range LDDT, and long-range DALI signals into a P-value.

   Technical reading: TS is a weighted combination of forward/reverse Mega′ scores, LDDT, and DALI, tuned with Hooke–Jeeves optimization; P-values use truth-standard-specific log-linear fits. LDDT contributes almost exclusively to family and DALI almost exclusively to fold.

   Input: Alignment signals from fully passing candidates and a selected family/superfamily/fold model.

   Output: Task-specific P-values, rankings, and top hits.

   Boundary: Choosing the wrong statistical model materially changes hit order; users must define the search objective first.

   PDF pp. 3–4, Results: Model weights; PDF p. 13, Fig. 4

Key result

The author reports median rank 1 across metrics, but this remains a single-author preprint about search accuracy.

Evidence-guided deep reading

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

paper-fact

What median rank 1 actually means

Table 3 ranks methods across reference×truth×metric combinations: median rank is 1 for Reseek-sensitive, 2 for Reseek-fast, 3 for TM-align, and 4 for DALI/Foldseek. Metrics include collection Sum3/PR90 and categorization Top3.

Table 4 compares ranks with Wilcoxon signed-rank tests: Reseek-sensitive versus DALI p=.0007144, Foldseek 1.198×10^-5, TM-align .0005855, and Reseek-fast 2.705×10^-6. This supports the author's multi-task ranking, not per-pair geometric superiority.

Source locator: PDF pp. 3 and 21–22, Tables 3–4

project-reading

Summary metrics encode value choices

Sum3=2×SEPQ0.1+1.5×SEPQ1+SEPQ10 explicitly weights low-false-positive regimes more; PR90 is recall at precision=.9; Top3 similarly weights .1%, 1%, and 10% error rates.

These choices are reasonable but not natural constants. A user prioritizing exhaustive collection, another error budget, or geometry could obtain different median ranks, so raw curves and metric definitions must remain visible.

Source locator: PDF pp. 6–7, Methods: Collection accuracy and Categorization accuracy

paper-fact

AI-scale feasibility is not a rigorous speed championship

On the roughly 54M-structure AFDB50 demonstration, 100 queries take 9m for Foldseek, 6m for Reseek-fast, and 23m for sensitive; 1,000 queries take 19m, 61m, and 4h. Memory is 227/284GB for Foldseek versus about 4.4–6.7GB for Reseek.

The author explicitly states that opposite index directions and scaling behavior make these anecdotal examples rather than a commensurate rigorous comparison. With a single-author preprint and FAST absent, the evidence label remains indirect-modern/preprint.

Source locator: PDF p. 4, Scaling to large databases; PDF p. 19, Table 1

Study design and evaluation

Data and samples

SCOP40, curated SCOP40c, and CATH40 across family/superfamily/fold truth standards, plus an approximately 54M-structure AFDB50 scaling demonstration.

Baselines

• DALI, Foldseek, and TM-align.
• Reseek-fast and Reseek-sensitive as two operating points.

Metrics

Sum3 / PR90

Collection-search sensitivity/recall across FPEPQ levels or at 90% precision.
Boundary: Sum3 weights are author-designed, and filtered tools do not output complete score curves.

Top3

Weighted sensitivity for leave-category-out top-hit categorization across error rates.
Boundary: It examines only the top hit and does not represent exhaustive homolog collection.

Median method rank

Median rank across references, truth standards, and summary metrics.
Boundary: Ignores effect size and depends on the chosen metric set and weights.

Reported result

Across the author's retrieval metrics, Reseek-sensitive has median rank 1 and significantly outranks baselines; results are unreviewed, not independently replicated, and do not include FAST.

PDF pp. 2–7, Results and Methods; PDF pp. 10–14, Figs. 1–5; PDF pp. 21–22, Tables 3–4

teaching-model · not a reported experiment

Teaching example (not a reported experiment)

Why a family model should not be used for fold search

A query and target are sequence-diverged but share a fold, with similar local secondary-structure packing and low amino-acid identity.

1. The family model assigns about 61% weight to aa20, pushing the low-sequence-similarity pair downward.
2. The fold model reduces primary-sequence weight, raises nensec32 to 27%, and relies chiefly on long-range DALI signal.
3. The same raw alignment ranks differently under a different truth objective; model selection is part of the scientific question.

Takeaway: Protein similarity is not one relation; family, superfamily, and fold require different evidence weighting.

indirect-modern

Evidence boundary versus FAST

FAST is absent, and search accuracy is not equivalent to lower RMSD plus greater coverage.

Lawful source and access

22 pages · SHA-256 5a45b5d76cac3ca54fd5425d166a043f8b90de89258cd58d8ad68bb8c3f8c9f5

Version 1 preprint posted 2026-07-30; not peer reviewed.

bioRxiv preprint PDF

Limits and misreadings

• Not peer reviewed and not independently replicated.

Source locator map

1. PDF pp. 2–5, Results and Methods overview
2. PDF pp. 5–7, Methods: Algorithm workflow and accuracy definitions
3. PDF pp. 19–22, Tables 1–4
4. PDF pp. 10–18, Figs. 1–9

Check understanding

1. Does median rank 1 mean first on every metric?

   Answer: No; Table 3 still contains metrics where Reseek-sensitive ranks second or third.

   The median summarizes overall position without erasing task-specific exceptions.

2. Why can the speed ranking reverse between 100 and 1,000 AFDB50 queries?

   Answer: Reseek indexes queries while Foldseek indexes the database, giving different time/memory scaling with query and database size.

   A speed ratio at one scale cannot be generalized universally.

3. Can this paper prove Reseek v3 beats FAST on RMSD plus coverage?

   Answer: No; FAST is absent and the main outcomes are retrieval ranks.

   The geometry question requires a separate same-pair alignment benchmark with common coverage rules.

Completion task: Find three Table 3 rows where Reseek-sensitive is not first, explain what changes in truth/reference/metric, and assess whether the median hides important exceptions.

Paper-specific glossary

spaced seed

A k-mer seed comparing only specified pattern positions, allowing some differences.

collection search

A task seeking many related structures under an acceptable error budget.

categorization search

A task assigning a query to a family/superfamily/fold category using its top hit.

Mega′

A high-capacity residue representation formed by a weighted product of sequence/structure alphabets.