Internal duplication and distant family relationships in DehI

A 3BJX-guided comparison with CMD-like and AhpD proteins

Author

Jason Schmidberger

Published

September 3, 2026

Summary

The configuration-inverting 2-haloacid dehalogenase DehI has two structurally similar regions of approximately 130 amino acids. This analysis tests whether the two regions retain detectable sequence homology consistent with an ancient tandem duplication and fusion.

The N- and C-terminal regions were defined using the experimentally determined DehI structure 3BJX, extracted from a 69-sequence family alignment, converted into separate HH-suite profile hidden Markov models, and compared using HHalign. The profiles align over nearly their entire lengths with 99.54% HHalign probability, an E-value of 2.9 × 10^-23, and 17% direct sequence identity. The combination of low identity, strong profile-profile support, and broad alignment coverage is consistent with remote homology between the two DehI repeats.

Broad CMD-like and AhpD profiles produced no statistically convincing match to either DehI half. The compact CMD proteins were therefore divided into sequence-coherent monomer subfamilies before profile construction. This revealed reproducible but sub-threshold signals: the strongest were DehI-N/CMD_monomer_017 at 87.72% probability and DehI-N/CMD_monomer_015 at 85.88%. CMD_monomer_015 also matched DehI-C at 80.99% over 64 columns. The first three corrected P-values remain below 0.001 after a conservative correction across 54 comparisons, but none reaches the usual ≥95% HH-suite probability expected for a confident assignment.

Principal result

The N- and C-terminal profiles align across 108 columns: 88% of the N-terminal model and 96% of the C-terminal model. This is a domain-scale match rather than a short conserved motif.

External-family result

The current sequence-only profiles strongly support DehI internal duplication. The monomer-subfamily analysis supplies a credible DehI-to-CMD lead, especially CMD_monomer_015, but does not yet establish that relationship. Independent profile-database and structure-based tests are needed before treating it as a positive family assignment.

Biological question

DehI from Pseudomonas putida PP3 is an unusual α-helical dehalogenase with a deep Stevedore knot. Its monomer contains two regions with closely similar structures but low pairwise sequence identity. The original structural study described two approximately 130-residue domains with about 16% sequence identity (PDB 3BJX; DOI: 10.1016/j.jmb.2008.02.035). A later topology analysis interpreted these regions as likely products of tandem sequence duplication and described the proline-rich arc connecting them (King et al., 2010).

The specific question addressed here is:

Does a profile-profile comparison of the DehI N- and C-terminal families recover statistically strong, domain-scale sequence homology after the linker is excluded?

A second, exploratory question asks whether either DehI repeat retains a detectable relationship to intact, compact carboxymuconolactone decarboxylase (CMD)-like proteins or alkyl hydroperoxide reductase D (AhpD) proteins.

Analysis design

flowchart LR
    A[DehI family<br>core MSA] --> B[Map 3BJX<br>structural boundary]
    B --> C[N-terminal MSA<br>residues 1–129]
    B --> D[C-terminal MSA<br>residues 162–296]
    B -.-> E[Exclude linker<br>residues 130–161]
    C --> F[N-terminal HHM]
    D --> G[C-terminal HHM]
    F --> H[HHalign<br>profile–profile comparison]
    G --> H
    I[CMD-like<br>PF02627 sequences] --> L[Cluster intact<br>80–119-aa monomers]
    L --> M[CMD subfamily HHMs]
    M --> J[External<br>comparisons]
    K[AhpD<br>IPR004674 profile] --> J
    F --> J
    G --> J

The workflow deliberately separates computation from presentation. Alignment splitting and HH-suite profile construction are performed by version-controlled scripts. This document reads the resulting files; it does not rerun HH-suite during website rendering.

Family alignment and structural reference

The input was the 69-sequence aligned FASTA file dehi_core_msa.fasta. The first record is the 3BJX chain-A reference. Its aligned sequence contains a 15-residue recombinant N-terminal expression tag followed by native DehI residues 1–296.

Table 1: Input family-alignment summary.

Property	Value
Family sequences	69
Alignment columns	384
3BJX aligned residues	311
3BJX expression-tag residues	15
3BJX native residues	296

Secondary-structure context

The deposited HELIX records from 3BJX were mapped through the gaps in the 3BJX alignment row. Red rounded bars denote helices. The figure shows the first 20 family members for legibility; the domain splitting and HHM construction use all 69 sequences.

Wrapped DehI multiple-sequence alignment with the 3BJX secondary structure displayed as red helix bars above each block. — Figure 1: 3BJX secondary structure mapped onto a representative subset of the DehI family alignment.

Defining and extracting the repeats

The last long helix of the N-terminal repeat ends at 3BJX residue 129. A proline-rich region occupies residues 130–161, and the next repeat begins at residue 162 with a short 3₁₀ helix followed by the first long C-terminal helix. The linker is therefore 32 residues by these structure-defined bounds, corresponding to the approximately 30-residue linker described informally.

Region	Native 3BJX residues	Treatment
N-terminal repeat	1–129	Retained
Proline-rich linker	130–161	Excluded
C-terminal repeat	162–296	Retained

The boundaries were mapped from native 3BJX residue numbers to MSA columns. The same columns were then extracted from every family member. All-gap columns were removed independently from each resulting alignment. The procedure is implemented in split_dehi_msa.py.

Table 2: Structure-guided split of the DehI family alignment.

Region	3BJX residues	Source MSA columns	Output columns	Sequences
N	1–129	16–154	139	69
Linker excluded	130–161	155–217	—	—
C	162–296	218–384	167	69

The N-terminal output has 139 aligned columns and the C-terminal output has 167. The ungapped 3BJX fragments contain 129 and 135 residues, respectively. All 69 input records are represented in both outputs.

HH-suite profile construction

HH-suite was chosen because HHalign directly compares profile HMMs and is designed to detect remote protein homology through alignment-to-alignment or HMM-to-HMM comparison (HH-suite documentation; Steinegger et al., 2019).

Each split FASTA alignment was converted to A3M format. A column was designated as a match state when fewer than 50% of sequences contained a gap (-M 50). Columns not meeting this condition were represented as insert states. HHmake then applied its standard sequence filtering and weighting before constructing each HHM.

reformat.pl fas a3m results/dehi_N_terminal_msa.fasta \
  results/dehi_N_terminal.a3m -M 50
reformat.pl fas a3m results/dehi_C_terminal_msa.fasta \
  results/dehi_C_terminal.a3m -M 50

hhmake -i results/dehi_N_terminal.a3m \
  -o results/dehi_N_terminal.hhm -name DehI_N_terminal
hhmake -i results/dehi_C_terminal.a3m \
  -o results/dehi_C_terminal.hhm -name DehI_C_terminal

Table 3: Properties of the two HH-suite profiles.

Profile	Match states	Sequences retained	Effective sequences
DehI N-terminal	123	68/69	5.5
DehI C-terminal	113	68/69	6.0

HHmake retained 68 of 69 sequences in each model after its default redundancy and quality filters. The effective sequence counts (Neff) of 5.5 and 6.0 are much smaller than the raw sequence count because related sequences do not contribute as independent observations.

N-terminal versus C-terminal HHalign result

The two precomputed HHMs were compared directly:

hhalign \
  -i results/dehi_N_terminal.hhm \
  -t results/dehi_C_terminal.hhm \
  -o results/dehi_N_vs_C.hhr \
  -Oa3m results/dehi_N_vs_C.a3m \
  -atab results/dehi_N_vs_C.tsv

Table 4: HHalign profile-profile comparison.

Metric	Value
HHalign probability	99.54%
E-value	2.9 × 10^-23
Profile score	124.39 bits
Aligned columns	108
Direct identity	17%
Profile similarity	0.123
Sum of posterior probabilities	94.6
N-profile coverage	87.8%
C-profile coverage	95.6%

The alignment spans N-profile match states 5–113 and C-profile match states 2–112. It therefore covers most of both models rather than selecting a small highly conserved segment. The 17% direct identity lies in the range where pairwise sequence comparison alone is generally difficult, while the profile score remains strong because residue preferences and gap patterns are aggregated across the family.

Profile-profile alignment

The following is the consensus and 3BJX portion of the HHalign output. Digits in the Confidence lines are posterior alignment confidence on a 0–9 scale.

Q 3BJX       12 QLDVSGEMESTYEDIRLTLRVPWVAFGCRVLATFPGYLPLAWRR.SAEALITRYAEQAADELRERSLLNIGPLPNL-K-E 88 (129)
Q Consensus   5 e~eA~g~~~~vY~DIk~tLrVP~Vn~ifRtlA~yp~fl~~~W~~.~rP~~~T~~fe~aA~~lR~~~~~~~~~~~~~-~-~ 81 (123)
                +.++.+.++++|+|||+++|.|.||..+|+||+||+||..+|+. +||..+|.+|++++++++..+...+...+-+ . .
T Consensus   2 ~~~~~~~~~~l~~dI~~~h~~~~v~SdYR~LA~wP~~L~~~W~~.LkP~i~s~~y~~~~~~l~~~A~~~v~~lP~~v~l~ 80 (113)
T 3BJX        2 ATKASTEVQGLLKRVADLHYHHGPASDFQALANWPKVLQIVTDEvLAPVARTEQYDAKSRELVTRARELVRGLPGSAGVQ 81 (135)
Confidence      56889999999999999999999999999999999999999999 9999999999999999999888877765332 1 1

Q 3BJX       89 --RLYAAGFDDGEIEKVRRVLYAFNYGNPKYLL 119 (129)
Q Consensus  82 ~-~~~~~~~~~~~~~~ir~~l~~f~~vnPklLl 113 (123)
                 . .+.+.+ ++.++++|.+.+..|...-|-++|
T Consensus  81 ~~~L~~~~-~~~~i~~i~~l~~~F~~~lp~Lil 112 (113)
T 3BJX       82 RSELMSML-TPNELAGLTGVLFMYQRFIADITI 113 (135)
Confidence      1 555555 899999999999999988877654

Comparison with CMD-like and AhpD proteins

Defining the reference families

The external comparisons use two UniProt-derived, 60%-identity-reduced source sets already present in the project. These labels need care. Pfam PF02627 is named for CMD but represents the broader AhpD/CMD-like fold; bona fide AhpD proteins can also carry PF02627. The two groups below are therefore operational sequence sets, not mutually exclusive fold classifications. This relationship is independently reflected in the domain annotation of an AhpD example (UniProt P9WQB4), the HAMAP AhpD rule, and the 2OUW structural annotations.

CMD-like: PF02627 records explicitly annotated as carboxymuconolactone decarboxylase, excluding descriptions containing AhpD, and restricted to 80–220 residues to reduce fragments and fusions.
AhpD: records from the more specific IPR004674 source set, restricted to 120–230 residues.

Table 5: Construction of the external-family profiles.

Profile set	Source records	Retained before HHmake	Length range	HHmake retained	Match states	Neff
CMD-like	12,656	7,266	80–220	104	124	9.4
AhpD	178	172	120–229	121	175	4.2

Two non-standard residues in the CMD-like set were converted to X; none were converted in the AhpD set. Selection and normalization are implemented in prepare_cmd_ahpd_comparison.py. The retained sequences were aligned independently with MAFFT FFT-NS-1, converted to A3M with the same -M 50 match-state rule as the DehI halves, and passed to HHmake. The complete pipeline is in build_cmd_ahpd_hhsuite_profiles.sh.

HHalign comparison matrix

Each row is a direct comparison of two precomputed profiles. Because exactly one target was supplied in each run, the E-values are useful for comparing these pre-specified runs but are not calibrated to a large database search.

Table 6: Internal and broad-family HHalign results. Bold marks the established internal match, the strongest broad-family DehI/external lead, and the external-family control.

Query	Template	Probability	E-value	Score	Aligned columns	Identity	Query coverage	Template coverage
DehI-N	DehI-C	99.54%	2.9 × 10^-23	124.39	108	17%	87.8%	95.6%
DehI-N	CMD-like	0.21%	0.28	9.46	10	30%	8.1%	8.1%
DehI-C	CMD-like	0.14%	0.42	8.08	1	0%	0.9%	0.8%
DehI-N	AhpD	0.23%	0.26	12.12	21	24%	17.1%	12.0%
DehI-C	AhpD	4.12%	0.01	19.65	72	15%	63.7%	41.1%
CMD-like	AhpD	97.04%	8.3 × 10^-11	55.41	43	26%	34.7%	24.6%

The DehI-N/C result is qualitatively distinct from all four external comparisons: it combines near-certain probability, an extremely small E-value, and near-domain-scale coverage. DehI-C/AhpD is the only external comparison with appreciable span, aligning C-profile states 34–106 to AhpD states 9–87. Its probability remains very low, however, and the aligned region is rich in generic helical-protein preferences. It should therefore be treated as a hypothesis for structural and database-search follow-up rather than as evidence of common ancestry.

Best broad-family exploratory alignment: DehI-C versus AhpD

Q 3BJX       34 NWPKVLQIVTDEVLAPVARTEQYDAKSRELVTRARELVRGLP-GSAGVQRSELMSMLTPNELA------GLTGVLFMYQRF 107
Q Consensus  34 ~wP~~L~~~W~~~LkP~i~s~~y~~~~~~l~~~A~~~v~~lP-~~v~l~~~~L~~~~~~~~i~------~i~~l~~~F~~~ 106
                .-|+|...+.-. |+..++++.......-....|...+.+-| .-..+..+. +...++++++      .+|+.-..|.| |
T Consensus   9 ~lP~yAKDlklN~l~~vl~~~~Lt~~q~~g~aLA~A~a~r~~~l~~ai~~eA-~~~ls~~~~~aA~aaAslMaMNNVyYRf  87

This 72-column match has 15% identity and a sum of posterior probabilities of 37.5. It is visually plausible, but HHalign assigns only 4.12% probability; the statistical assessment, rather than visual resemblance alone, governs the interpretation.

CMD-like versus AhpD control

CMD-like and AhpD align at 97.04% probability over a 43-state core. This is consistent with their shared broader classification and shows that the external profiles retain a detectable common signal. It is not a fully independent positive control, because the CMD-like PF02627 umbrella overlaps conceptually with AhpD/CMD-like proteins even after annotation-based separation.

Monomer-focused CMD analysis

Why a subfamily analysis was needed

The broad PF02627 alignment combines many divergent CMD-like lineages. A profile averaged across that entire set can lose a relationship retained in only one lineage. The 1,316 intact, explicitly CMD-annotated proteins of 80–119 residues were therefore clustered with MMseqs2 at 30% sequence identity and 70% bidirectional coverage. This produced 160 clusters. The 27 clusters containing at least five proteins were retained for profile analysis; the smaller clusters do not contain enough sequences to support a useful family HMM.

Each retained cluster was aligned independently with MAFFT and converted to an HH-suite profile. Both DehI halves were compared with every cluster, giving 54 pairwise tests. The complete ranking, including cluster sizes, effective sequence counts, model lengths and coverage, is available as the monomer-cluster result table, with cluster composition in the cluster manifest. The workflow is implemented in build_cmd_monomer_cluster_profiles.sh.

Lead monomer-subfamily matches

Table 7: Highest-scoring DehI comparisons among the 27 intact CMD monomer profiles. The P × 54 column is a conservative Bonferroni correction across both halves and all retained clusters.

DehI query	CMD cluster (representative)	Sequences	Profile Neff	Probability	Pairwise P-value	P × 54	Aligned columns	Query / CMD coverage	Identity
N-terminal	017 (A0ABY7GUB4)	9	2.9	87.72%	3.4 × 10^-6	1.84 × 10^-4	42	34% / 52%	12%
N-terminal	015 (A0ABP8DVF7)	10	3.4	85.88%	5.3 × 10^-6	2.86 × 10^-4	48	39% / 45%	17%
C-terminal	015 (A0ABP8DVF7)	10	3.4	80.99%	1.2 × 10^-5	6.48 × 10^-4	64	57% / 60%	23%
C-terminal	003 (A0A8A7KB27)	117	5.3	74.52%	2.7 × 10^-5	1.46 × 10^-3	78	69% / 70%	17%

These results are much stronger than the pooled CMD profile, showing that subfamily stratification was important. CMD_monomer_015 is especially interesting because the same external profile gives the leading match to both homologous DehI halves. Its C-terminal comparison also spans more than half of both models. Cluster 003 gives the broadest alignment, but at 74.52% probability it is the least secure of the four leads.

The three lead CMD clusters are themselves homologous at profile level:

Table 8: Post-hoc profile comparisons among the three CMD subfamilies that produced the leading DehI matches.

CMD profile comparison	Probability	E-value	Aligned columns	Identity
015 vs 017	98.85%	3.8 × 10^-17	79	28%
015 vs 003	98.60%	8.7 × 10^-16	77	22%
017 vs 003	96.00%	4.3 × 10^-9	40	20%

Their mutual profile similarity argues that the signals arise from a coherent region of CMD sequence space rather than three unrelated annotations. However, the lead clusters were selected after inspecting the DehI comparisons, so this cross-comparison is a consistency check rather than independent confirmation.

Interpretation of the monomer signal

The corrected P-values show that these alignments are unlikely to be ordinary pairwise chance scores within this 54-test experiment. HH-suite probability also incorporates whether a match resembles true homologues rather than common background alignments, and the maximum here is 87.72%, below the conventional high-confidence range. The short 42–48-column N-terminal matches and modest profile diversity (Neff 2.9–3.4) are additional reasons to describe the result as suggestive evidence, not a family assignment.

Interpretation

Three features make this result supportive of internal homology:

High profile probability. HHalign assigns a 99.54% probability to the N/C relationship.
Domain-scale coverage. The 108 aligned columns cover 88% of the N model and 96% of the C model.
Remote rather than trivial similarity. Direct identity is only 17%, yet conserved residue patterns remain detectable when information is combined across 69 family members.

Together with the near-identical structural folds reported for 3BJX, this is consistent with the proposal that an ancestral domain duplicated in tandem and the two copies subsequently diverged while remaining fused in DehI.

The external comparisons do not yet identify the ancestral domain definitively. However, separating intact CMD monomers into coherent subfamilies changes the picture: three related CMD profiles give probabilities of 74.52–87.72%, and CMD_monomer_015 matches both DehI halves. This makes a monomeric CMD relationship the strongest current external hypothesis. Because the probabilities remain below 95%, the most defensible conclusion is: strong evidence for internal DehI N/C homology and suggestive, but not yet conclusive, evidence connecting those repeats to a subset of monomeric CMD proteins. No comparable sequence-only support was recovered for AhpD.

What the result does not prove

HHalign measures compatibility between two profile models. It does not by itself date the duplication, identify the ancestral free-standing family, or exclude every alternative evolutionary scenario. Structural correspondence, taxonomic distribution, conserved catalytic residues, and comparisons with external families provide independent evidence.

Statistical qualification

Only one template HMM was searched (Searched_HMMs 1). Consequently, the reported E-value and P-value are identical (2.9 × 10^-23^); this is not an E-value corrected for a large profile database. The probability, score, and especially the broad alignment coverage remain directly informative for this pre-specified N-versus-C comparison.

Secondary structure was not included in the HHalign score (SS = 0, reported as missing dssp). This makes the profile match a sequence-derived result rather than one driven by the known structural similarity.

Robustness checks and next analyses

The following tests would strengthen the evolutionary interpretation:

Rebuild both profiles after removing 3BJX and repeat the comparison, ensuring that the structural reference is not driving the result.
Repeat profile construction with match-state thresholds of 30%, 50%, and 70% gaps and compare the aligned cores.
Search both DehI profiles against the same broad, curated HH-suite database and test whether their strongest domain-level matches overlap. A database search supplies the calibration and competing alternatives absent from these one-template HHalign runs.
Build structure-informed profiles for the lead CMD monomer subfamilies and compare representative structures with Foldseek/DALI. This is the most direct follow-up to the suggestive sequence-profile matches.
Test the CMD_monomer_015 relationship against independently collected CMD and DehI sequence sets so that subfamily selection and validation use separate data.
Map conserved aligned profile positions back onto the 3BJX structure and determine whether they occupy corresponding structural environments.

Reproducibility and downloads

The report uses precomputed HH-suite output so that the website can be rendered without installing native bioinformatics software. The Python dependencies are managed by uv; HH-suite 3 is required only when regenerating the profiles.

Scripts

split_dehi_msa.py — map the 3BJX boundaries through the family MSA and extract both repeats.
build_dehi_hhsuite_profiles.sh — convert the alignments to A3M, construct both HHMs, and run HHalign.
prepare_cmd_ahpd_comparison.py — define, filter, and normalize the operational CMD-like and AhpD sequence sets.
build_cmd_ahpd_hhsuite_profiles.sh — align both external families, construct their HHMs, and run all five external profile comparisons.
analyse_cmd_monomer_clusters.py — split the compact CMD set into sequence-coherent clusters and summarize all DehI comparisons with multiple-testing correction.
build_cmd_monomer_cluster_profiles.sh — cluster intact CMD monomers, construct subfamily HHMs, and compare each profile with both DehI halves.

Alignments and sequences

Profiles and comparison output

To regenerate the derived files locally:

uv run python split_dehi_msa.py
./build_dehi_hhsuite_profiles.sh
uv run python prepare_cmd_ahpd_comparison.py
./build_cmd_ahpd_hhsuite_profiles.sh
./build_cmd_monomer_cluster_profiles.sh
quarto render report.qmd

--- title: "Internal duplication and distant family relationships in DehI" subtitle: "A 3BJX-guided comparison with CMD-like and AhpD proteins" author: "Jason Schmidberger" date: last-modified format: html: theme: cosmo toc: true toc-depth: 3 code-fold: true code-summary: "Show analysis code" code-tools: true code-copy: true fig-responsive: true link-external-newwindow: true smooth-scroll: true execute: enabled: false resources: - dehi_3BJX_secondary_structure.png - split_dehi_msa.py - build_dehi_hhsuite_profiles.sh - prepare_cmd_ahpd_comparison.py - build_cmd_ahpd_hhsuite_profiles.sh - analyse_cmd_monomer_clusters.py - build_cmd_monomer_cluster_profiles.sh - results/dehi_core_msa.fasta - results/dehi_N_terminal_msa.fasta - results/dehi_C_terminal_msa.fasta - results/dehi_N_terminal.fasta - results/dehi_C_terminal.fasta - results/dehi_N_terminal.a3m - results/dehi_C_terminal.a3m - results/dehi_N_terminal.hhm - results/dehi_C_terminal.hhm - results/dehi_N_vs_C.hhr - results/dehi_N_vs_C.a3m - results/dehi_N_vs_C.tsv - results/dehi_terminal_split_report.tsv - results/cmd_ahpd_comparison_sets.tsv - results/uniprot_PF02627_nr60.fasta - results/uniprot_ahpd_IPR004674_nr60.fasta - results/cmd_PF02627_nr60.a3m - results/ahpd_IPR004674_nr60.a3m - results/cmd_PF02627_nr60.hhm - results/ahpd_IPR004674_nr60.hhm - results/dehi_N_vs_CMD.hhr - results/dehi_C_vs_CMD.hhr - results/dehi_N_vs_AhpD.hhr - results/dehi_C_vs_AhpD.hhr - results/CMD_vs_AhpD.hhr - results/cmd_monomer_cluster_manifest.tsv - results/dehi_vs_cmd_monomer_clusters.tsv - results/cmd_monomer_clusters/dehi_N_vs_CMD_monomer_017.hhr - results/cmd_monomer_clusters/dehi_N_vs_CMD_monomer_015.hhr - results/cmd_monomer_clusters/dehi_C_vs_CMD_monomer_015.hhr - results/cmd_monomer_clusters/dehi_C_vs_CMD_monomer_003.hhr - results/cmd_monomer_clusters/CMD_monomer_015_vs_017.hhr - results/cmd_monomer_clusters/CMD_monomer_015_vs_003.hhr - results/cmd_monomer_clusters/CMD_monomer_017_vs_003.hhr --- ## Summary The configuration-inverting 2-haloacid dehalogenase DehI has two structurally similar regions of approximately 130 amino acids. This analysis tests whether the two regions retain detectable sequence homology consistent with an ancient tandem duplication and fusion. The N- and C-terminal regions were defined using the experimentally determined DehI structure 3BJX, extracted from a 69-sequence family alignment, converted into separate HH-suite profile hidden Markov models, and compared using HHalign. The profiles align over nearly their entire lengths with **99.54% HHalign probability**, an **E-value of 2.9 × 10^-23^**, and **17% direct sequence identity**. The combination of low identity, strong profile-profile support, and broad alignment coverage is consistent with remote homology between the two DehI repeats. Broad CMD-like and AhpD profiles produced no statistically convincing match to either DehI half. The compact CMD proteins were therefore divided into sequence-coherent monomer subfamilies before profile construction. This revealed reproducible but sub-threshold signals: the strongest were **DehI-N/CMD_monomer_017 at 87.72% probability** and **DehI-N/CMD_monomer_015 at 85.88%**. CMD_monomer_015 also matched DehI-C at **80.99% over 64 columns**. The first three corrected P-values remain below 0.001 after a conservative correction across 54 comparisons, but none reaches the usual ≥95% HH-suite probability expected for a confident assignment. ::: {.callout-important title="Principal result"} The N- and C-terminal profiles align across 108 columns: 88% of the N-terminal model and 96% of the C-terminal model. This is a domain-scale match rather than a short conserved motif. ::: ::: {.callout-warning title="External-family result"} The current sequence-only profiles strongly support DehI internal duplication. The monomer-subfamily analysis supplies a credible DehI-to-CMD lead, especially CMD_monomer_015, but does **not yet establish** that relationship. Independent profile-database and structure-based tests are needed before treating it as a positive family assignment. ::: ## Biological question DehI from *Pseudomonas putida* PP3 is an unusual α-helical dehalogenase with a deep Stevedore knot. Its monomer contains two regions with closely similar structures but low pairwise sequence identity. The original structural study described two approximately 130-residue domains with about 16% sequence identity ([PDB 3BJX](https://www.rcsb.org/structure/3BJX); [DOI: 10.1016/j.jmb.2008.02.035](https://doi.org/10.1016/j.jmb.2008.02.035)). A later topology analysis interpreted these regions as likely products of tandem sequence duplication and described the proline-rich arc connecting them ([King *et al.*, 2010](https://doi.org/10.1371/journal.pcbi.1000731)). The specific question addressed here is: > Does a profile-profile comparison of the DehI N- and C-terminal families > recover statistically strong, domain-scale sequence homology after the linker > is excluded? A second, exploratory question asks whether either DehI repeat retains a detectable relationship to intact, compact carboxymuconolactone decarboxylase (CMD)-like proteins or alkyl hydroperoxide reductase D (AhpD) proteins. ## Analysis design ```{mermaid} flowchart LR A[DehI family core MSA] --> B[Map 3BJX structural boundary] B --> C[N-terminal MSA residues 1–129] B --> D[C-terminal MSA residues 162–296] B -.-> E[Exclude linker residues 130–161] C --> F[N-terminal HHM] D --> G[C-terminal HHM] F --> H[HHalign profile–profile comparison] G --> H I[CMD-like PF02627 sequences] --> L[Cluster intact 80–119-aa monomers] L --> M[CMD subfamily HHMs] M --> J[External comparisons] K[AhpD IPR004674 profile] --> J F --> J G --> J ``` The workflow deliberately separates computation from presentation. Alignment splitting and HH-suite profile construction are performed by version-controlled scripts. This document reads the resulting files; it does not rerun HH-suite during website rendering. ## Family alignment and structural reference The input was the 69-sequence aligned FASTA file [`dehi_core_msa.fasta`](results/dehi_core_msa.fasta). The first record is the 3BJX chain-A reference. Its aligned sequence contains a 15-residue recombinant N-terminal expression tag followed by native DehI residues 1–296. | Property | Value | |---|---:| | Family sequences | 69 | | Alignment columns | 384 | | 3BJX aligned residues | 311 | | 3BJX expression-tag residues | 15 | | 3BJX native residues | 296 | : Input family-alignment summary. {#tbl-input-summary} ### Secondary-structure context The deposited `HELIX` records from 3BJX were mapped through the gaps in the 3BJX alignment row. Red rounded bars denote helices. The figure shows the first 20 family members for legibility; the domain splitting and HHM construction use all 69 sequences. ![3BJX secondary structure mapped onto a representative subset of the DehI family alignment.](dehi_3BJX_secondary_structure.png){#fig-msa-ss fig-alt="Wrapped DehI multiple-sequence alignment with the 3BJX secondary structure displayed as red helix bars above each block." width=100%} ## Defining and extracting the repeats The last long helix of the N-terminal repeat ends at 3BJX residue 129. A proline-rich region occupies residues 130–161, and the next repeat begins at residue 162 with a short 3~10~ helix followed by the first long C-terminal helix. The linker is therefore 32 residues by these structure-defined bounds, corresponding to the approximately 30-residue linker described informally. | Region | Native 3BJX residues | Treatment | |---|---:|---| | N-terminal repeat | 1–129 | Retained | | Proline-rich linker | 130–161 | Excluded | | C-terminal repeat | 162–296 | Retained | The boundaries were mapped from native 3BJX residue numbers to MSA columns. The same columns were then extracted from every family member. All-gap columns were removed independently from each resulting alignment. The procedure is implemented in [`split_dehi_msa.py`](split_dehi_msa.py). | Region | 3BJX residues | Source MSA columns | Output columns | Sequences | |---|---:|---:|---:|---:| | N | 1–129 | 16–154 | 139 | 69 | | Linker excluded | 130–161 | 155–217 | — | — | | C | 162–296 | 218–384 | 167 | 69 | : Structure-guided split of the DehI family alignment. {#tbl-split-summary} The N-terminal output has 139 aligned columns and the C-terminal output has 167. The ungapped 3BJX fragments contain 129 and 135 residues, respectively. All 69 input records are represented in both outputs. ## HH-suite profile construction HH-suite was chosen because HHalign directly compares profile HMMs and is designed to detect remote protein homology through alignment-to-alignment or HMM-to-HMM comparison ([HH-suite documentation](https://github.com/soedinglab/hh-suite); [Steinegger *et al.*, 2019](https://doi.org/10.1186/s12859-019-3019-7)). Each split FASTA alignment was converted to A3M format. A column was designated as a match state when fewer than 50% of sequences contained a gap (`-M 50`). Columns not meeting this condition were represented as insert states. HHmake then applied its standard sequence filtering and weighting before constructing each HHM. ```bash reformat.pl fas a3m results/dehi_N_terminal_msa.fasta \ results/dehi_N_terminal.a3m -M 50 reformat.pl fas a3m results/dehi_C_terminal_msa.fasta \ results/dehi_C_terminal.a3m -M 50 hhmake -i results/dehi_N_terminal.a3m \ -o results/dehi_N_terminal.hhm -name DehI_N_terminal hhmake -i results/dehi_C_terminal.a3m \ -o results/dehi_C_terminal.hhm -name DehI_C_terminal ``` | Profile | Match states | Sequences retained | Effective sequences | |---|---:|---:|---:| | DehI N-terminal | 123 | 68/69 | 5.5 | | DehI C-terminal | 113 | 68/69 | 6.0 | : Properties of the two HH-suite profiles. {#tbl-profile-summary} HHmake retained 68 of 69 sequences in each model after its default redundancy and quality filters. The effective sequence counts (`Neff`) of 5.5 and 6.0 are much smaller than the raw sequence count because related sequences do not contribute as independent observations. ## N-terminal versus C-terminal HHalign result The two precomputed HHMs were compared directly: ```bash hhalign \ -i results/dehi_N_terminal.hhm \ -t results/dehi_C_terminal.hhm \ -o results/dehi_N_vs_C.hhr \ -Oa3m results/dehi_N_vs_C.a3m \ -atab results/dehi_N_vs_C.tsv ``` | Metric | Value | |---|---:| | HHalign probability | 99.54% | | E-value | 2.9 × 10^-23^ | | Profile score | 124.39 bits | | Aligned columns | 108 | | Direct identity | 17% | | Profile similarity | 0.123 | | Sum of posterior probabilities | 94.6 | | N-profile coverage | 87.8% | | C-profile coverage | 95.6% | : HHalign profile-profile comparison. {#tbl-hhalign-summary} The alignment spans N-profile match states 5–113 and C-profile match states 2–112. It therefore covers most of both models rather than selecting a small highly conserved segment. The 17% direct identity lies in the range where pairwise sequence comparison alone is generally difficult, while the profile score remains strong because residue preferences and gap patterns are aggregated across the family. ### Profile-profile alignment The following is the consensus and 3BJX portion of the HHalign output. Digits in the `Confidence` lines are posterior alignment confidence on a 0–9 scale. ```text Q 3BJX 12 QLDVSGEMESTYEDIRLTLRVPWVAFGCRVLATFPGYLPLAWRR.SAEALITRYAEQAADELRERSLLNIGPLPNL-K-E 88 (129) Q Consensus 5 e~eA~g~~~~vY~DIk~tLrVP~Vn~ifRtlA~yp~fl~~~W~~.~rP~~~T~~fe~aA~~lR~~~~~~~~~~~~~-~-~ 81 (123) +.++.+.++++|+|||+++|.|.||..+|+||+||+||..+|+. +||..+|.+|++++++++..+...+...+-+ . . T Consensus 2 ~~~~~~~~~~l~~dI~~~h~~~~v~SdYR~LA~wP~~L~~~W~~.LkP~i~s~~y~~~~~~l~~~A~~~v~~lP~~v~l~ 80 (113) T 3BJX 2 ATKASTEVQGLLKRVADLHYHHGPASDFQALANWPKVLQIVTDEvLAPVARTEQYDAKSRELVTRARELVRGLPGSAGVQ 81 (135) Confidence 56889999999999999999999999999999999999999999 9999999999999999999888877765332 1 1 Q 3BJX 89 --RLYAAGFDDGEIEKVRRVLYAFNYGNPKYLL 119 (129) Q Consensus 82 ~-~~~~~~~~~~~~~~ir~~l~~f~~vnPklLl 113 (123) . .+.+.+ ++.++++|.+.+..|...-|-++| T Consensus 81 ~~~L~~~~-~~~~i~~i~~l~~~F~~~lp~Lil 112 (113) T 3BJX 82 RSELMSML-TPNELAGLTGVLFMYQRFIADITI 113 (135) Confidence 1 555555 899999999999999988877654 ``` ## Comparison with CMD-like and AhpD proteins ### Defining the reference families The external comparisons use two UniProt-derived, 60%-identity-reduced source sets already present in the project. These labels need care. Pfam PF02627 is named for CMD but represents the broader AhpD/CMD-like fold; bona fide AhpD proteins can also carry PF02627. The two groups below are therefore **operational sequence sets**, not mutually exclusive fold classifications. This relationship is independently reflected in the domain annotation of an [AhpD example (UniProt P9WQB4)](https://www.uniprot.org/uniprotkb/P9WQB4/entry), the [HAMAP AhpD rule](https://hamap.expasy.org/rule/MF_01676), and the [2OUW structural annotations](https://www.rcsb.org/annotations/2OUW). - **CMD-like:** PF02627 records explicitly annotated as `carboxymuconolactone decarboxylase`, excluding descriptions containing `AhpD`, and restricted to 80–220 residues to reduce fragments and fusions. - **AhpD:** records from the more specific IPR004674 source set, restricted to 120–230 residues. | Profile set | Source records | Retained before HHmake | Length range | HHmake retained | Match states | Neff | |---|---:|---:|---:|---:|---:|---:| | CMD-like | 12,656 | 7,266 | 80–220 | 104 | 124 | 9.4 | | AhpD | 178 | 172 | 120–229 | 121 | 175 | 4.2 | : Construction of the external-family profiles. {#tbl-external-profiles} Two non-standard residues in the CMD-like set were converted to `X`; none were converted in the AhpD set. Selection and normalization are implemented in [`prepare_cmd_ahpd_comparison.py`](prepare_cmd_ahpd_comparison.py). The retained sequences were aligned independently with MAFFT FFT-NS-1, converted to A3M with the same `-M 50` match-state rule as the DehI halves, and passed to HHmake. The complete pipeline is in [`build_cmd_ahpd_hhsuite_profiles.sh`](build_cmd_ahpd_hhsuite_profiles.sh). ### HHalign comparison matrix Each row is a direct comparison of two precomputed profiles. Because exactly one target was supplied in each run, the E-values are useful for comparing these pre-specified runs but are not calibrated to a large database search. | Query | Template | Probability | E-value | Score | Aligned columns | Identity | Query coverage | Template coverage | |---|---|---:|---:|---:|---:|---:|---:|---:| | DehI-N | DehI-C | **99.54%** | 2.9 × 10^-23^ | 124.39 | 108 | 17% | 87.8% | 95.6% | | DehI-N | CMD-like | 0.21% | 0.28 | 9.46 | 10 | 30% | 8.1% | 8.1% | | DehI-C | CMD-like | 0.14% | 0.42 | 8.08 | 1 | 0% | 0.9% | 0.8% | | DehI-N | AhpD | 0.23% | 0.26 | 12.12 | 21 | 24% | 17.1% | 12.0% | | DehI-C | AhpD | **4.12%** | **0.01** | **19.65** | **72** | 15% | 63.7% | 41.1% | | CMD-like | AhpD | **97.04%** | **8.3 × 10^-11^** | **55.41** | 43 | 26% | 34.7% | 24.6% | : Internal and broad-family HHalign results. Bold marks the established internal match, the strongest broad-family DehI/external lead, and the external-family control. {#tbl-comparison-matrix} The DehI-N/C result is qualitatively distinct from all four external comparisons: it combines near-certain probability, an extremely small E-value, and near-domain-scale coverage. DehI-C/AhpD is the only external comparison with appreciable span, aligning C-profile states 34–106 to AhpD states 9–87. Its probability remains very low, however, and the aligned region is rich in generic helical-protein preferences. It should therefore be treated as a hypothesis for structural and database-search follow-up rather than as evidence of common ancestry. ### Best broad-family exploratory alignment: DehI-C versus AhpD ```text Q 3BJX 34 NWPKVLQIVTDEVLAPVARTEQYDAKSRELVTRARELVRGLP-GSAGVQRSELMSMLTPNELA------GLTGVLFMYQRF 107 Q Consensus 34 ~wP~~L~~~W~~~LkP~i~s~~y~~~~~~l~~~A~~~v~~lP-~~v~l~~~~L~~~~~~~~i~------~i~~l~~~F~~~ 106 .-|+|...+.-. |+..++++.......-....|...+.+-| .-..+..+. +...++++++ .+|+.-..|.| | T Consensus 9 ~lP~yAKDlklN~l~~vl~~~~Lt~~q~~g~aLA~A~a~r~~~l~~ai~~eA-~~~ls~~~~~aA~aaAslMaMNNVyYRf 87 ``` This 72-column match has 15% identity and a sum of posterior probabilities of 37.5. It is visually plausible, but HHalign assigns only 4.12% probability; the statistical assessment, rather than visual resemblance alone, governs the interpretation. ### CMD-like versus AhpD control CMD-like and AhpD align at 97.04% probability over a 43-state core. This is consistent with their shared broader classification and shows that the external profiles retain a detectable common signal. It is not a fully independent positive control, because the CMD-like PF02627 umbrella overlaps conceptually with AhpD/CMD-like proteins even after annotation-based separation. ## Monomer-focused CMD analysis ### Why a subfamily analysis was needed The broad PF02627 alignment combines many divergent CMD-like lineages. A profile averaged across that entire set can lose a relationship retained in only one lineage. The 1,316 intact, explicitly CMD-annotated proteins of 80–119 residues were therefore clustered with MMseqs2 at **30% sequence identity and 70% bidirectional coverage**. This produced 160 clusters. The 27 clusters containing at least five proteins were retained for profile analysis; the smaller clusters do not contain enough sequences to support a useful family HMM. Each retained cluster was aligned independently with MAFFT and converted to an HH-suite profile. Both DehI halves were compared with every cluster, giving 54 pairwise tests. The complete ranking, including cluster sizes, effective sequence counts, model lengths and coverage, is available as the [monomer-cluster result table](results/dehi_vs_cmd_monomer_clusters.tsv), with cluster composition in the [cluster manifest](results/cmd_monomer_cluster_manifest.tsv). The workflow is implemented in [`build_cmd_monomer_cluster_profiles.sh`](build_cmd_monomer_cluster_profiles.sh). ### Lead monomer-subfamily matches | DehI query | CMD cluster (representative) | Sequences | Profile Neff | Probability | Pairwise P-value | P × 54 | Aligned columns | Query / CMD coverage | Identity | |---|---|---:|---:|---:|---:|---:|---:|---:|---:| | N-terminal | [017 (A0ABY7GUB4)](results/cmd_monomer_clusters/dehi_N_vs_CMD_monomer_017.hhr) | 9 | 2.9 | **87.72%** | 3.4 × 10^-6^ | 1.84 × 10^-4^ | 42 | 34% / 52% | 12% | | N-terminal | [015 (A0ABP8DVF7)](results/cmd_monomer_clusters/dehi_N_vs_CMD_monomer_015.hhr) | 10 | 3.4 | **85.88%** | 5.3 × 10^-6^ | 2.86 × 10^-4^ | 48 | 39% / 45% | 17% | | C-terminal | [015 (A0ABP8DVF7)](results/cmd_monomer_clusters/dehi_C_vs_CMD_monomer_015.hhr) | 10 | 3.4 | **80.99%** | 1.2 × 10^-5^ | 6.48 × 10^-4^ | 64 | 57% / 60% | 23% | | C-terminal | [003 (A0A8A7KB27)](results/cmd_monomer_clusters/dehi_C_vs_CMD_monomer_003.hhr) | 117 | 5.3 | **74.52%** | 2.7 × 10^-5^ | 1.46 × 10^-3^ | 78 | 69% / 70% | 17% | : Highest-scoring DehI comparisons among the 27 intact CMD monomer profiles. The P × 54 column is a conservative Bonferroni correction across both halves and all retained clusters. {#tbl-cmd-monomer-leads} These results are much stronger than the pooled CMD profile, showing that subfamily stratification was important. CMD_monomer_015 is especially interesting because the same external profile gives the leading match to both homologous DehI halves. Its C-terminal comparison also spans more than half of both models. Cluster 003 gives the broadest alignment, but at 74.52% probability it is the least secure of the four leads. The three lead CMD clusters are themselves homologous at profile level: | CMD profile comparison | Probability | E-value | Aligned columns | Identity | |---|---:|---:|---:|---:| | [015 vs 017](results/cmd_monomer_clusters/CMD_monomer_015_vs_017.hhr) | 98.85% | 3.8 × 10^-17^ | 79 | 28% | | [015 vs 003](results/cmd_monomer_clusters/CMD_monomer_015_vs_003.hhr) | 98.60% | 8.7 × 10^-16^ | 77 | 22% | | [017 vs 003](results/cmd_monomer_clusters/CMD_monomer_017_vs_003.hhr) | 96.00% | 4.3 × 10^-9^ | 40 | 20% | : Post-hoc profile comparisons among the three CMD subfamilies that produced the leading DehI matches. {#tbl-cmd-lead-relatedness} Their mutual profile similarity argues that the signals arise from a coherent region of CMD sequence space rather than three unrelated annotations. However, the lead clusters were selected after inspecting the DehI comparisons, so this cross-comparison is a consistency check rather than independent confirmation. ::: {.callout-warning title="Interpretation of the monomer signal"} The corrected P-values show that these alignments are unlikely to be ordinary pairwise chance scores within this 54-test experiment. HH-suite probability also incorporates whether a match resembles true homologues rather than common background alignments, and the maximum here is 87.72%, below the conventional high-confidence range. The short 42–48-column N-terminal matches and modest profile diversity (Neff 2.9–3.4) are additional reasons to describe the result as **suggestive evidence**, not a family assignment. ::: ## Interpretation Three features make this result supportive of internal homology: 1. **High profile probability.** HHalign assigns a 99.54% probability to the N/C relationship. 2. **Domain-scale coverage.** The 108 aligned columns cover 88% of the N model and 96% of the C model. 3. **Remote rather than trivial similarity.** Direct identity is only 17%, yet conserved residue patterns remain detectable when information is combined across 69 family members. Together with the near-identical structural folds reported for 3BJX, this is consistent with the proposal that an ancestral domain duplicated in tandem and the two copies subsequently diverged while remaining fused in DehI. The external comparisons do not yet identify the ancestral domain definitively. However, separating intact CMD monomers into coherent subfamilies changes the picture: three related CMD profiles give probabilities of 74.52–87.72%, and CMD_monomer_015 matches both DehI halves. This makes a monomeric CMD relationship the strongest current external hypothesis. Because the probabilities remain below 95%, the most defensible conclusion is: **strong evidence for internal DehI N/C homology and suggestive, but not yet conclusive, evidence connecting those repeats to a subset of monomeric CMD proteins.** No comparable sequence-only support was recovered for AhpD. ::: {.callout-note title="What the result does not prove"} HHalign measures compatibility between two profile models. It does not by itself date the duplication, identify the ancestral free-standing family, or exclude every alternative evolutionary scenario. Structural correspondence, taxonomic distribution, conserved catalytic residues, and comparisons with external families provide independent evidence. ::: ### Statistical qualification Only one template HMM was searched (`Searched_HMMs 1`). Consequently, the reported E-value and P-value are identical (`2.9 × 10^-23^`); this is not an E-value corrected for a large profile database. The probability, score, and especially the broad alignment coverage remain directly informative for this pre-specified N-versus-C comparison. Secondary structure was not included in the HHalign score (`SS = 0`, reported as `missing dssp`). This makes the profile match a sequence-derived result rather than one driven by the known structural similarity. ## Robustness checks and next analyses The following tests would strengthen the evolutionary interpretation: - Rebuild both profiles after removing 3BJX and repeat the comparison, ensuring that the structural reference is not driving the result. - Repeat profile construction with match-state thresholds of 30%, 50%, and 70% gaps and compare the aligned cores. - Search both DehI profiles against the same broad, curated HH-suite database and test whether their strongest domain-level matches overlap. A database search supplies the calibration and competing alternatives absent from these one-template HHalign runs. - Build structure-informed profiles for the lead CMD monomer subfamilies and compare representative structures with Foldseek/DALI. This is the most direct follow-up to the suggestive sequence-profile matches. - Test the CMD_monomer_015 relationship against independently collected CMD and DehI sequence sets so that subfamily selection and validation use separate data. - Map conserved aligned profile positions back onto the 3BJX structure and determine whether they occupy corresponding structural environments. ## Reproducibility and downloads The report uses precomputed HH-suite output so that the website can be rendered without installing native bioinformatics software. The Python dependencies are managed by `uv`; HH-suite 3 is required only when regenerating the profiles. ### Scripts - [`split_dehi_msa.py`](split_dehi_msa.py) — map the 3BJX boundaries through the family MSA and extract both repeats. - [`build_dehi_hhsuite_profiles.sh`](build_dehi_hhsuite_profiles.sh) — convert the alignments to A3M, construct both HHMs, and run HHalign. - [`prepare_cmd_ahpd_comparison.py`](prepare_cmd_ahpd_comparison.py) — define, filter, and normalize the operational CMD-like and AhpD sequence sets. - [`build_cmd_ahpd_hhsuite_profiles.sh`](build_cmd_ahpd_hhsuite_profiles.sh) — align both external families, construct their HHMs, and run all five external profile comparisons. - [`analyse_cmd_monomer_clusters.py`](analyse_cmd_monomer_clusters.py) — split the compact CMD set into sequence-coherent clusters and summarize all DehI comparisons with multiple-testing correction. - [`build_cmd_monomer_cluster_profiles.sh`](build_cmd_monomer_cluster_profiles.sh) — cluster intact CMD monomers, construct subfamily HHMs, and compare each profile with both DehI halves. ### Alignments and sequences - [Full DehI core MSA](results/dehi_core_msa.fasta) - [N-terminal aligned FASTA](results/dehi_N_terminal_msa.fasta) - [C-terminal aligned FASTA](results/dehi_C_terminal_msa.fasta) - [N-terminal ungapped FASTA](results/dehi_N_terminal.fasta) - [C-terminal ungapped FASTA](results/dehi_C_terminal.fasta) - [N-terminal A3M](results/dehi_N_terminal.a3m) - [C-terminal A3M](results/dehi_C_terminal.a3m) - [CMD-like A3M](results/cmd_PF02627_nr60.a3m) - [AhpD A3M](results/ahpd_IPR004674_nr60.a3m) - [PF02627 source sequences (nr60)](results/uniprot_PF02627_nr60.fasta) - [IPR004674 AhpD source sequences (nr60)](results/uniprot_ahpd_IPR004674_nr60.fasta) ### Profiles and comparison output - [N-terminal HHM](results/dehi_N_terminal.hhm) - [C-terminal HHM](results/dehi_C_terminal.hhm) - [HHalign report](results/dehi_N_vs_C.hhr) - [HHalign pairwise A3M](results/dehi_N_vs_C.a3m) - [HHalign per-column table](results/dehi_N_vs_C.tsv) - [Domain-splitting report](results/dehi_terminal_split_report.tsv) - [CMD-like profile HHM](results/cmd_PF02627_nr60.hhm) - [AhpD profile HHM](results/ahpd_IPR004674_nr60.hhm) - [DehI-N versus CMD-like report](results/dehi_N_vs_CMD.hhr) - [DehI-C versus CMD-like report](results/dehi_C_vs_CMD.hhr) - [DehI-N versus AhpD report](results/dehi_N_vs_AhpD.hhr) - [DehI-C versus AhpD report](results/dehi_C_vs_AhpD.hhr) - [CMD-like versus AhpD report](results/CMD_vs_AhpD.hhr) - [External-family filtering summary](results/cmd_ahpd_comparison_sets.tsv) - [CMD monomer cluster manifest](results/cmd_monomer_cluster_manifest.tsv) - [Ranked DehI-versus-CMD-monomer results](results/dehi_vs_cmd_monomer_clusters.tsv) To regenerate the derived files locally: ```bash uv run python split_dehi_msa.py ./build_dehi_hhsuite_profiles.sh uv run python prepare_cmd_ahpd_comparison.py ./build_cmd_ahpd_hhsuite_profiles.sh ./build_cmd_monomer_cluster_profiles.sh quarto render report.qmd ```