SARST2 high-throughput and resource-efficient protein structure alignment against massive databases

30-second read

SARST2 combines primary, secondary, tertiary, and evolutionary features in a multi-stage filter-and-refine pipeline for searching hundreds of millions of AlphaFold structures on ordinary hardware.

Central question

How can hundreds of millions of protein structures be searched on ordinary computers while preserving high retrieval precision, refined alignments, and manageable memory/disk cost?

Intuition

Reserve expensive work for very few candidates: first eliminate most subjects with reduced alphabets, diagonal word matching, group heads, and ML; then refine using synthesized DP over AA/SSE/SARST/WCN, entropy-guided gaps, and TM-score.

Why it matters

SARST2 reaches 96.3% average precision versus FAST at 95.3% in the same SCOP retrieval experiment and is over 3,300 times faster in database search. This is clear direct evidence of a FAST win for retrieval, but FAST is absent from the geometric-quality comparison, so lower RMSD plus higher coverage is not established.

Prerequisites

• Understand filter-and-refine and database retrieval
• Know structural alphabets, dynamic programming, and TM-score
• Distinguish precision/recall, alignment geometry, and speed/memory

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

Method walkthrough

1. 01 · Multilayer encoding and four filters

   Encode structures as AA, five-symbol AAT, four-symbol SSE, 23-symbol SARST, and WCN; search only group heads through word matching, SARST DP, SSE+AAT DP, and quick TM-score filters.

   Technical reading: A diagonal shortcut computes word matches more likely to form homologous alignments; DT gates quickly reject nonhomologs, while ANN estimates a homology p-score.

   Input: A query structure and grouped/formatted subject database

   Output: A small candidate-subject pool

   Boundary: ML labels come from SCOP families; other classification schemes and novel folds require independent validation.

   PDF pp. 2–6 and 10–12, Box 1, Fig. 1, Methods

2. 02 · Refine with local and long-range features

   Candidates undergo DP with residue scores synthesized from AA/AAT/SSE/SARST substitution matrices and WCN packing-density differences.

   Technical reading: WCN equals a weighted sum of inverse-square distances and encodes long-range packing; PSSM Shannon entropy becomes conservation C(i), which combines with SSE segments to modulate gap-open/extension penalties.

   Input: Candidate structural strings, coordinates, and a query PSSM

   Output: High-quality residue correspondence and a synthesized score

   Boundary: Ablation precision gains are component effects within this pipeline, not per-case RMSD gains.

   PDF pp. 4–5 and 10–11, Eq. 1–9, Table 1

3. 03 · Superpose, rank, and scale

   Final structural superposition yields TM-score, combined with normalized DP score and ANN p-score into confidence/pC-value for ranked interactive hits.

   Technical reading: Go parallelization, 24-bit C-alpha coordinate encoding, grouped databases, and an optional one-decimal representation reduce runtime/storage; searching 215M AlphaFoldDB structures takes 3.4 minutes and 9.4 GiB with 32 i9 CPUs.

   Input: Refined candidates and superposition coordinates

   Output: A hit list ranked by structural-homology confidence

   Boundary: pC-value is minus log2 confidence and lacks the chance-expectation interpretation of a BLAST E-value.

   PDF pp. 6–8 and 11–13, Figs. 3–4, Eqs. 10–15

Key result

On SCOP-2.07 Qry400 retrieval, mean precision is 96.3% versus FAST's 95.3%, with reported speedups of roughly 12× pairwise and over 3300× for database search.

Evidence-guided deep reading

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

paper-fact

What can and cannot be said about FAST

Qry400-to-SCOP-2.07 family-level retrieval uses one known-answer framework; 11-point average precision is SARST2 96.3%, Foldseek 95.9%, FAST 95.3%, and TM-align 94.1%.

At 100% recall in one-CPU database search, SARST2 is over 3,300 times faster than FAST/TM-align and about 12 times faster pairwise. This supports retrieval accuracy and speed; the paper's TM-score/identity alignment-quality plot compares SARST2, Foldseek, TM-align, and BLAST, not FAST.

Source locator: PDF pp. 2–3 and 5–7, Figs. 2–3; PDF p. 9, Fig. 5

paper-fact

Precision and time contribution of each component

Disabling synthesized DP costs 1.91 precision points; disabling WCN costs 1.27, replacing the SSE scoring matrix with binary scores costs 2.91, and disabling VGP costs 0.59. This supports substantive contributions from WCN and learned substitution matrices.

Disabling diagonal matching raises time from 150 to 470 ms, grouped search to 325 ms, and ML to 822 ms. Filters alter not only time but candidate distributions, so some ablations also affect precision.

Source locator: PDF pp. 5–7, Table 1

paper-fact

Scale benchmark and independence boundary

In the answer-seeded 215M-structure AlphaFoldDB-2022 test, SARST2 uses 3.4 min/9.4 GiB, Foldseek 18.6 min/19.6 GiB, and BLAST 52.5 min/77.3 GiB; the formatted SARST2 database is about 0.5 TiB versus 59.7 TiB raw CIF.

ML trains on Qry200 and tests on family-disjoint Qry400; on the nrCATH40 independent test, Foldseek precision is 0.89 points above SARST2, while SARST2 slightly beats TM-align and remains fastest. This limits any claim of universal best accuracy.

Source locator: PDF pp. 6–8 and 12–13, Fig. 4, Supplementary Table 7 discussion

Study design and evaluation

Data and samples

SCOP-2.07 with 144,879 domains across 4,022 families; Qry200 train, Qry400 test; nrCATH40 independent; AlphaFoldDB-2022 with 214,459,158 structures.

Baselines

• FAST, TM-align, Fr-TM-align, MICAN-SQ, SARST1/iSARST, MADOKA, Foldseek, and BLAST/bl2seq

Metrics

11-point average precision

Mean precision at eleven recall levels from 0 to 100%, measuring whether homologs concentrate near the top of hit lists.
Boundary: A retrieval metric, not superposition RMSD/coverage.

search time/resources

Wall time, memory, and disk at 100% seeded-answer recall.
Boundary: Depends on hardware, parallel scaling, database format, and parameters.

TM-score/identity

Structural similarity and evolutionary-plausibility analysis on sampled homolog pairs.
Boundary: FAST is absent from this direct geometric comparison.

Reported result

In the authors' SCOP retrieval benchmark, SARST2 slightly exceeds FAST/Foldseek precision while greatly improving speed and resources; it is not universally top precision on the CATH independent challenge.

PDF pp. 2–13, Figs. 1–5, Table 1 and Methods

teaching-model · not a reported experiment

Teaching example (not a reported experiment)

Do not rewrite a retrieval win as a geometric win

Teaching model: a report must answer whether SARST2 simultaneously has lower RMSD and higher coverage than FAST.

1. List direct same-benchmark facts first: average precision 96.3 versus 95.3 and database speed over 3,300-fold.
2. Check participants in the geometric comparison: Fig. 5 omits FAST, so no direct RMSD/coverage pair exists.
3. Conclude 'wins retrieval; geometric double-win not established' and propose running FAST and SARST2 on the same pair set next.

Takeaway: The strongest scientific statement is not the largest claim but the exact task, metric, and comparator directly tested.

direct-search

Evidence boundary versus FAST

This is the clearest direct win over FAST on retrieval; FAST is absent from the geometric-quality experiment, so lower RMSD plus greater coverage is not established.

Lawful source and access

15 pages · SHA-256 4be7873eeb4e4edba364cd1e70ae44b007e79d1b79adab45cccc1c4da3b40e61

Lawful open full text; the site separates retrieval accuracy from geometric alignment quality.

Europe PMC open-access PDF

Limits and misreadings

• The main benchmark is author-designed and still benefits from independent replication.
• Mean precision, speed, and geometric alignment quality are distinct questions.

Source locator map

1. PDF pp. 2–5, Figs. 1–2 and algorithm overview
2. PDF pp. 5–8, Figs. 3–4 and Table 1
3. PDF pp. 9–13, Fig. 5 and Methods

Check understanding

1. Where is SARST2's direct win over FAST?

   Answer: SCOP family retrieval average precision at 96.3% versus 95.3%, plus database and pairwise speed.

   They share the same dataset and answer framework.

2. Can we claim lower RMSD and higher coverage?

   Answer: No; FAST is absent from the geometric-quality experiment.

   Retrieval precision cannot substitute for alignment geometry.

3. Is pC-value an E-value?

   Answer: No.

   pC equals minus log2 confidence and lacks the expected-chance-hits interpretation.

Completion task: Design an evidence-closing benchmark using the same pair set and domain definitions, reporting coverage, RMSD, TM-score, runtime, and failure cases for FAST and SARST2.

Paper-specific glossary

WCN

A packing-density measure based on inverse-square weighted distances from one residue to all others.

filter-and-refine

Use cheap high-recall filters to shrink candidates, then expensive accurate methods to rank them.

average precision

In this paper, mean precision at eleven recall levels, measuring concentration of correct homologs near the top.

pC-value

The negative base-two logarithm of confidence, used as a hit-quality cutoff but not an E-value.