A secondary structure-based position-specific scoring matrix applied to the improvement in protein secondary structure prediction

30-second read

Instead of changing only the model, the paper changes the input by building an SSE-PSSM over predicted secondary-structure states to complement amino-acid PSSMs.

Central question

Can converting homolog alignments into secondary-structure states before building a PSSM yield denser, more generalizable SSP features than an amino-acid PSSM?

Intuition

Twenty amino-acid symbols can produce sparse counts in distant-homolog alignments; mapping residues into three or eight SSE codes reduces zeros and directly encodes homolog structural tendencies at each position.

Why it matters

Rather than adding a deeper model, it changes feature representation. Under the same simple machine-learning setup, SSE-PSSM beats AA-PSSM and adds value when indirectly integrated with existing SSP outputs; this is not native end-to-end integration into state-of-the-art architectures.

Prerequisites

• Understand MSAs, PSSMs, and log odds
• Know DSSP three-state/eight-state secondary structures
• Understand independent tests and the distinction between features and models

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

Method walkthrough

1. 01 · Retrieve homologs and obtain SSE strings

   Use PSI-BLAST/HHBlits to find query homologs; apply DSSP when structures exist, otherwise vote position-wise from secondary homolog SSEs.

   Technical reading: Hit alignments remain amino-acid based, but each aligned residue is replaced by its SSE code; no-hit or uncovered positions use an SSE-substitution-matrix pseudocount.

   Input: A query sequence, sequence database, and structural reference database

   Output: SSE-transformed hit strings aligned to the query

   Boundary: Predicted or voted hit SSEs propagate upstream errors into the feature.

   PDF pp. 18–20, Algorithm of the SSE-PSSM, Fig. 4

2. 02 · From occurrence counts to a log-odds matrix

   At each query position, count aligned SSE codes, weight by hit similarity to form a PPM, then divide by background frequency and take log2.

   Technical reading: Occ(c,p) sums weighted code matches, PPM normalizes occurrences, and PSSM equals log2(PPM/background); main experiments use the eight-state DSSP alphabet.

   Input: SSE-transformed alignments and background SSE frequencies

   Output: Per-residue SSE log-odds features

   Boundary: A smaller code set reduces sparsity but may lose amino-acid-specific signal; hence the paper tests combined features.

   PDF pp. 19–22, Fig. 4, Eqs. 3–5, alphabet experiments

3. 03 · Strict independent tests and indirect integration

   Compare SSE-PSSM with AA-PSSM using training/independent sets below 25% identity, then feed existing SSP probability outputs plus SSE-PSSM into a second-level model.

   Technical reading: QuerySet-T/I and TargetSet-nr25 are stringently homology-reduced, with TS115/CASP12/13 also used. For DeepCNF on CASP12, Q3 rises 82.9 to 84.1 and Q8 rises 68.4 to 71.8.

   Input: SSE-PSSM, AA-PSSM, or existing predictor probabilities

   Output: Q3/Q8, SOV, boundary/internal accuracy, and error rates

   Boundary: Training source was unavailable for compiled programs, so this is stacking-like indirect integration, not native feature fusion.

   PDF pp. 3–8 and 22, Tables 1–3, integration procedure

Key result

Preliminary integration improved three-state accuracy by 2.0% and eight-state accuracy by 5.2% on average, showing room remains in feature representation.

Evidence-guided deep reading

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

paper-fact

Why SSE representation may be more informative

For a novel query with few homologs, a 20-column amino-acid propensity matrix contains many zeros; reducing to three or eight SSE symbols yields a denser PPM from the same hits.

Under the same machine-learning setup in Table 1, SSE-PSSM scores Q3 0.785 versus AA-PSSM 0.654 and Q8 0.663 versus 0.479 on QuerySet-I; SOV and boundary/internal measures agree.

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

paper-fact

Why Table 1 is not a leaderboard

The authors explicitly say the SSE-PSSM model and state-of-the-art programs are not on equal footing: the former uses new features plus simple iMLS, while the latter uses traditional PSSMs plus distinct advanced models.

TargetSet-nr25 contains only about 11.4k sequences, far smaller and stricter than UniRef90-2015 at 38.2M, so absolute scores are lower; the key comparison is the controlled feature contrast.

Source locator: PDF pp. 4–6, Table 1 and accompanying cautions

paper-fact

Class, size, and upstream-error failure modes

Methods generally perform poorly on all-beta proteins; SSE-PSSM shows lower normalized variability across structural classes and protein sizes, though all-alpha is one of its relative weak points.

SSE-PSSM requires known or predicted SSEs for hits; systematic bias in secondary-homolog retrieval or SSE prediction will be aggregated into the matrix. This error channel needs separate deployment measurement.

Source locator: PDF pp. 6–7 and 18–20, Tables 2–3 and Algorithm step 2

Study design and evaluation

Data and samples

QuerySet-T/I and TargetSet-nr25 with less than 25% identity across datasets; TS115 and CASP12/13 with UniRef90-2015 or nrPDB90-2015.

Baselines

• SSE-PSSM versus AA-PSSM under the same iMLS; seven PSI-BLAST SSP and four HHBlits-era SSP programs; indirect combined models

Metrics

Q3/Q8

Three/eight-state residue-micro accuracy.
Boundary: Class imbalance and segment continuity are not fully captured.

SOV3/SOV8

Secondary-structure segment-overlap quality.
Boundary: Not the same scale as Q.

boundary/internal accuracy

Accuracy separated for SSE-segment boundaries and interiors.
Boundary: Depends on boundary definitions and DSSP mapping.

Reported result

SSE-PSSM consistently beats AA-PSSM in controlled tests; preliminary indirect integration with existing predictor outputs improves Q3 by 2.0% and Q8 by 5.2% on average.

PDF pp. 3–8 and 15–22, Tables 1–4, Figs. 1 and 4

teaching-model · not a reported experiment

Teaching example (not a reported experiment)

A simplified SSE-PSSM from five homologs

Teaching model: five aligned hit codes at one query position are H, H, H, E, and C, with equal weights.

1. Count occurrences H=3, E=1, C=1, then divide by five to obtain a PPM.
2. Divide each PPM probability by the SSE background frequency and take log2 to obtain enrichment or depletion.
3. The actual algorithm adds hit-similarity weights, eight-state codes, and pseudocounts for uncovered positions; do not treat this toy example as paper data.

Takeaway: SSE-PSSM does not directly vote a final label; it converts homolog structural tendencies into position-specific log-odds features for a model.

outside-scope

Evidence boundary versus FAST

This concerns SSP features rather than FAST alignment, though its representation-first strategy conceptually echoes structural-alphabet methods.

Lawful source and access

28 pages · SHA-256 ce3ebe3b3430c75f4df5ddaadbd13a45be72839921bfe8431735b93453713c01

Lawful open full text.

Europe PMC open-access PDF

Limits and misreadings

• Gains after integration into newer end-to-end models require separate validation.

Source locator map

1. PDF pp. 3–8, Fig. 1 and Tables 1–3
2. PDF pp. 15–20, Methods and Fig. 4
3. PDF p. 22, alphabet and indirect-integration procedures

Check understanding

1. Is SSE-PSSM itself a predictor?

   Answer: No, it is a feature set.

   A machine-learning model must convert its features into final SSP labels.

2. Why does it contain fewer zeros?

   Answer: The SSE alphabet has only three or eight codes, far fewer than 20 amino acids.

   The same homolog count is distributed across fewer bins.

3. Are the 2.0%/5.2% gains from native integration?

   Answer: No, they come from indirect second-level integration of compiled predictor outputs with SSE-PSSM.

   True end-to-end feature fusion requires access to source and training pipelines.

Completion task: Draw the dataflow from query to homolog search, hit SSEs, transformed alignment, PPM, PSSM, and predictor, labeling one possible error source on every arrow.

Paper-specific glossary

SSE-PSSM

A position-specific log-odds matrix built from secondary-structure rather than amino-acid codes.

position propensity matrix

Normalized occurrence probabilities of codes at each query position.

pseudocount

A prior or substitution statistic used to avoid zero or undefined scores when observations are absent.

indirect integration

Training another layer on existing program outputs rather than modifying the original model.