Protein structure alignment by Reseek improves sensitivity to remote homologs

30-second read

Reseek's mega-alphabet jointly encodes amino acids and eight structural features, then reuses mature local sequence-alignment machinery for remote-homology search.

Central question

How can rich local and long-range structural states be turned into a sequence-alignable representation while keeping E-values calibrated to actual errors in large databases?

Intuition

Reseek avoids learning one enormous substitution matrix for 85.9 billion mega-letters. It factorizes amino acid, local conformation, nearest 3D-neighbor distances, and other features, learns separate log-odds matrices, sums them into residue-pair scores, and applies Smith–Waterman.

Why it matters

On the SCOP40 search benchmark, Reseek v2 achieves higher homolog sensitivity than DALI, TM-align, and Foldseek at matched false-positive errors per query, with better-calibrated E-values. This is retrieval evidence, not a FAST RMSD/coverage head-to-head.

Prerequisites

• Understand structural alphabets and substitution log-odds scores.
• Know Smith–Waterman local alignment and affine gap penalties.
• Be able to interpret sensitivity, FPEPQ, CVE curves, and E-value calibration.

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

Method walkthrough

1. 01 · Factor structural context into small alphabets

   Each residue records not only amino-acid type but also local backbone shape and distances/conformations of nearest and reverse 3D neighbors; each feature is discretized separately.

   Technical reading: The feature vector includes AA, DistNEN/DistREN, Conf represented by pairwise distances within ±3 residues, and NEN/REN contexts. Scalar features use equal-frequency bins; vector features use K-means representatives.

   Input: A protein Cα backbone and a training structure collection.

   Output: A discrete feature vector for every residue.

   Boundary: Discretization and feature selection discard coordinate detail; equal mega-letters do not imply identical full 3D environments.

   PDF pp. 2–3, §2 Materials and methods

2. 02 · Sum factorized log-odds and run local alignment

   When comparing two residues, each feature contributes a weighted vote. Summing these votes yields a pair score without observing every enormous state combination.

   Technical reading: Each feature f has log-odds matrix Mf and weight wf, with Sij=Σf wf Mf[Vf(i),Vf(j)]. AA plus eight 16-letter structural features create 20×16^8=85,899,345,920 mega-states. Smith–Waterman/BLAST-like search maximizes substitution scores minus affine gaps.

   Input: Two discrete feature-vector sequences and per-feature scoring matrices.

   Output: A local alignment and raw alignment score.

   Boundary: This is fundamentally local search; homolog-detection ability is not global rigid-superposition quality.

   PDF p. 3, Equations 2–3

3. 03 · Convert alignment into AQ and calibrated E-values

   Combine alignment score, local-distance preservation, and a reversed-sequence control into 0–1 AQ, then calibrate E-values against empirical errors so a threshold predicts false positives.

   Technical reading: The AQ statistic combines LDDT-mu, forward score, query-versus-reversed-query score, and mean length, then maps to [0,1] with a logistic function. Constants are trained with SCOP40 holdout and E-values fitted to an empirical AQ distribution.

   Input: An alignment, coordinate-derived LDDT-mu, reverse score, and lengths.

   Output: AQ, E-value, ranking, and an optional Kabsch superposition.

   Boundary: E-value extrapolation to AFDB scale still depends on the null model and database composition; the number of superfamilies is unknown.

   PDF p. 3, Equation 4 and SCOP40 assessment protocol

Key result

SCOP40 all-v-all results emphasize remote-homology sensitivity and calibrated E-values, not superposition RMSD.

Evidence-guided deep reading

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

paper-fact

Read sensitivity at a fixed error rate

Table 1 reports sensitivity at FPEPQ=.1/1/10: Reseek-sensitive scores .22/.34/.48, whereas Reseek-very-sensitive scores .22/.35/.51; TM-align gives .12/.24/.42, DALI .075/.17/.44, and Foldseek .11/.22/.41. Both the preset and error budget affect the conclusion, so the two Reseek modes must not be conflated.

The same table reports 49 s for Reseek-fast, 78 for Foldseek, 178 for Reseek-sensitive, 840 for Reseek-very-sensitive, 3.7×10^5 for TM-align, and 4.1×10^5 for DALI, on a 32-thread i9-14900K excluding database-build time.

Source locator: PDF pp. 3–5, Fig. 3 and Table 1

paper-fact

Why E-value calibration is a core result

A CVE plot asks how many homologs are recovered when allowing x false positives per query, matching threshold decisions in large databases more directly than ROC or precision alone. Ideally E≈FPEPQ.

In Fig. 3, a Foldseek E-value of 10^-6 corresponds to measured FPEPQ around .1, underestimating errors by five orders; Reseek E-values track measured errors better. This concerns significance calibration, not coordinate-alignment RMSD.

Source locator: PDF pp. 4–5, Figs. 2–3 and Discussion

project-reading

Evidence distance from the FAST question

FAST is not among the tested methods, and Reseek uses SCOP40 superfamily retrieval as ground truth with sensitivity/FPEPQ/E-value outputs. These do not answer FAST's aligned length and RMSD.

Reseek can reasonably be called a modern candidate for AI-scale structure search. Claiming greater overlap and lower error requires a separate same-pair experiment with fixed coverage rules and RMSD/TM-score computed from both outputs.

Source locator: PDF p. 3, SCOP40 protocol; PDF p. 5, Table 1

Study design and evaluation

Data and samples

SCOP40 v1.75 with 11,206 domains in 1,960 superfamilies, evaluated all-versus-all with same-superfamily as homolog truth.

Baselines

• DALILite v5, TM-align 2022/4/12, and Foldseek 8-ef4e960.
• BLASTP as sequence-search context.

Metrics

Sensitivity@FPEPQ

Fraction of homologs recovered at a fixed number of false positives per query.
Boundary: Using SCOP superfamily as truth can label unknown homology as negative.

E-value calibration

Compares reported E-values with measured FPEPQ.
Boundary: Calibration depends on database size, composition, and the null distribution.

Time / memory

Wall time and peak memory for all-versus-all search on common hardware.
Boundary: Database-build time is excluded and tools differ in filtering/output completeness.

Reported result

Reseek variants generally lead SCOP40 sensitivity at fixed FPEPQ and provide better-calibrated E-values, but speed depends on the preset: 49 s for fast, 178 s for sensitive, and 840 s for very-sensitive versus 78 s for Foldseek. FAST is not tested.

PDF pp. 3–5, §3 Results, Figs. 3–4, and Table 1

teaching-model · not a reported experiment

Teaching example (not a reported experiment)

How E-value underestimation grows in a large database

A tool reports E=10^-6 for a query, but a benchmark measures .1 false positives per query.

1. The user expects one error per million searches; in reality there is about one per ten.
2. They differ by 10^5, so even a useful ranking yields a dangerously wrong threshold interpretation.
3. CVE plots sensitivity against measured FPEPQ, separating ranking from calibration.

Takeaway: Large-scale search must not only rank good hits early; it must know how many errors a cutoff implies.

indirect-modern

Evidence boundary versus FAST

FAST is not compared directly, and the evaluation is not an RMSD/coverage benchmark.

Lawful source and access

6 pages · SHA-256 998dff41139e909b3c0dcdacf8681bff76a985c5988dea5b00152b1740ee62fe

CC BY 4.0 open full text.

Europe PMC open-access PDF

Limits and misreadings

• Single-author, author-run evaluation benefits from independent benchmarking.

Source locator map

1. PDF pp. 2–3, §2 Materials and methods and Equations 2–4
2. PDF pp. 3–5, §3 Results and Figs. 2–4
3. PDF p. 5, Table 1
4. PDF pp. 4–5, Discussion on CVE and E-value calibration

Check understanding

1. Does 85,899,345,920 states require an equally large substitution matrix?

   Answer: No; Reseek factorizes the mega-state into small feature alphabets and sums their log-odds scores.

   This is the key to avoiding sparse observations and memory explosion.

2. Can Reseek-very-sensitive Sens(10)=.51 be interpreted as lower RMSD?

   Answer: No; it is the fraction of homologs recovered at FPEPQ=10.

   It is a retrieval operating point, not superposition geometry.

3. Does this paper directly prove Reseek beats FAST?

   Answer: No. FAST is absent.

   Its primary metrics are sensitivity/FPEPQ, not the RMSD-plus-coverage question posed for FAST.

Completion task: Plot the four structural methods at FPEPQ .1, 1, and 10 from Table 1, then state why this plot cannot replace an Nmat/RMSD table.

Paper-specific glossary

mega-alphabet

The Cartesian-product state of multiple discrete features; Reseek does not enumerate a full matrix.

false-positive errors per query

The average number of false-positive hits per query.

coverage-versus-error curve

A search evaluation plotting sensitivity against measured error rate.

alignment quality (AQ)

Reseek's 0–1 test statistic combining alignment, reverse control, and LDDT-mu.