Announcing RosettaCommons Chaperone Support in pg_bio & Mining the Twilight Zone at 0.81 Å (DUF7861)

Today we are excited to announce a major new milestone in pg_bio: Layer 13 — Native In-Database Support for RosettaCommons Chaperone (RosettaCommons/chaperone) Protein Design Workflows, alongside a new Bell-Curve Twilight-Zone Homology Index inside our Swarm Discovery engine! By compiling Shrake–Rupley Golden-Section Fibonacci-sphere RASA (bio_rasa), 3D Hydrogen-Bond CSR Networks (bio_hbond_network), Buried Unsatisfied Polar Atom (BUNS) auditing (audit_hydrogen_bonds), and PipelineDataHandler lineage hashing (chaperone_lineage_hash) directly into PostgreSQL 18 via Rust (pgrx), pg_bio evaluates 400 multi-chain complexes in 209.04 ms (0.52 ms/complex — 36.9x faster than Python) and verifies all 8 official Chaperone Core Examples in ~4.5 ms. Even better, when we unleashed the combined Twilight-Zone + Chaperone RASA engine on our 226,242-edge Swarm graph, the #1 ranked hit (id = 26734, twilight_score = 0.8640) revealed a stunning 0.81 Å C$\alpha$ RMSD structural lock (TM-score = 0.927, per-residue RASA Pearson $r = 0.9509$) between two uncharacterized DUF7861 (PF25260 / IPR057183) proteins from an inland Xinjiang salt lake archaeon (Natrinema saccharevitans, A0A1S8AT22) and a Korean marine solar saltern archaeon (Halobellus clavatus, A0A1H3IA37).

🚀 Announcing Layer 13 in `pg_bio`: Native RosettaCommons Chaperone Workflows in SQL

Modern de novo protein design pipelines—recently unified by RosettaCommons Chaperone—chain generative backbone, inverse-folding, and structure-prediction blocks together through a PipelineDataHandler:

\[\text{RFdiffusion3 / RFD3NA (Backbone)} \longrightarrow \text{Foundry-MPNN / LigandMPNN (Sequence)} \longrightarrow \text{AlphaFold3 / Boltz-2 (Oracle)} \longrightarrow \text{Evaluation Blocks (RMSD, RASA, H-Bond)}\]

While GPU inference generates tens of thousands of candidate structures, post-oracle biophysical evaluation (rmsd, rasa, hbond) becomes a severe CPU and disk-I/O bottleneck when separate Python processes must parse loose PDB/mmCIF files to catch three classic wet-lab failure modes:

Surface Hydrophobic Aggregation Traps (rasa): Candidates that fold with high pLDDT in silico but aggregate in E. coli expression because too many bulky apolar side-chains (Val, Ile, Leu, Met, Phe, Trp, Tyr) remain exposed on the surface ($\text{RASA}_i \ge 0.25$).
Buried Unsatisfied Polar Atoms (hbond / BUNS): Burying a polar $\text{N}$ or $\text{O}$ atom inside a core or binding interface ($\text{SASA} < 0.5\text{–}1.0\text{ Å}^2$) without a hydrogen-bond partner costs $+3\text{ to }+5\text{ kcal/mol}$ in desolvation energy—silently killing binding affinity and fold stability.
Multi-Cycle Lineage Fragmentation (PipelineDataHandler): Iterative MPNN <-> AF3/Boltz-2 refinement loops require deterministic parent–child lineage hashes so winning constructs trace cleanly back to their seed backbone.

The 5-Stage `pg_bio` Chaperone Engine (36.9x Faster Than Python)

We built the entire Chaperone evaluation suite natively into src/lib.rs:

Stage 1 — Voxel-Accelerated Shrake–Rupley SASA & Tien (2013) RASA (bio_rasa): Evaluates 64 Golden-Ratio Fibonacci sphere points per heavy atom ($r_{\text{probe}} = 1.4\text{ Å}$) using $5\text{ Å}$ 3D spatial voxel hashing (reducing complexity from $O(N^2)$ to $O(N)$) and normalizes per-residue SASA by the Tien et al. (2013) empirical maximum ASA scale (plus DNA/RNA nucleotides) to return total_sasa, polar_sasa, apolar_sasa, mean_rasa, core_buried_fraction ($\text{RASA} < 0.20$), surface_hydrophobic_fraction, and per_residue_rasa.
Stage 2 — 3D Hydrogen-Bond CSR Network (bio_hbond_network) & BUNS Auditor (audit_hydrogen_bonds): Identifies backbone, side-chain, nucleobase, and HETATM ligand donors/acceptors, enforces donor–acceptor distance ($2.20\text{ Å} \le d_{\text{DA}} \le 3.50\text{ Å}$) and antecedent angle ($\ge 100^\circ$), projects the 3D H-bond graph into a Compressed Sparse Row (SparseAttentionMap) matrix, and counts cross_interface_hbonds and buried_unsatisfied_polars (BUNS).
Stage 3 — Fused Single-Pass Block Evaluator (evaluate_chaperone_block): Shares the precomputed atomic SASA buffer with the Kabsch self-consistency $\text{scRMSD}$ aligner, interface $\Delta\text{SASA}$ calculator, BUNS auditor, and 3D gyration tensor globular gate (is_globular) in a single zero-copy Rust call.
Stage 4 — Content-Addressable Lineage Hashing (chaperone_lineage_hash): Generates deterministic 64-bit MurmurHash3 (fmix64) lineage IDs (chp_<stage>_c<cycle>_<hash16>) across design cycles.
Stage 5 — All 8 Official Chaperone Core Examples Built Into SQL (chaperone_core_examples): Verifies rosettacommons.github.io/chaperone/core-examples/ (nanoparticle-design, peptide-binder, cortisol-binder, dna-binder, enzyme-glycosidase, enzyme-protease, ribozyme-hammerhead, and multi-mpnn) in ~4.5 ms total:

Benchmark Metric (PostgreSQL 18 on Apple Silicon)	Python Pull + Evaluation	`pg_bio` Native Rust Engine	Speedup / Result
400 Multi-Chain Complexes (`144` Heavy Atoms Each)	`7,709.11 ms`	`209.04 ms` (`0.52 ms`/complex)	`36.9x` faster
4-Archetype Failure Mode Discrimination (Pass vs. Drift/Hydro/BUNS)	Partial (no BUNS/RASA)	`100 / 400` exact recall	100% precision
8 Official Chaperone Core Examples Suite (`chaperone_core_examples`)	Manual multi-script setup	`~4.5 ms` total (`8 / 8` verified)	Zero external scripts

📖 Full Documentation: docs/15_chaperone_protein_design_workflows.md
📓 Interactive Benchmark Notebook: experiments/13_chaperone_protein_design_workflows.ipynb

The Mathematical Filter: Bell-Curve Twilight-Zone Scoring + Chaperone Auditing in SQL

When you rank 226,000 protein similarity edges purely by vector distance (<=>), two extremes crowd the top of the leaderboard: near-identical sequence clones (100% identity across sister strains) and short compositional k-mer collisions that share no global 3D fold. Why do traditional BLAST or k-mer pipelines struggle to rank twilight-zone discoveries? Look at the raw sequence numbers for A0A1S8AT22 and A0A1H3IA37:

Across their 81 residues, 32 positions have mutated (49 / 81 positional identity, or 60.5%), and their 3-mer compositional cosine distance is a massive 0.6896.
Under a strict pg_bio local Smith-Waterman alignment (match = +2, mismatch = -1, gap = -2), their raw score is 67 out of a theoretical maximum of 162—a normalized Smith-Waterman similarity of just s = 41.4%.

To make PostgreSQL automatically surface pairs with tight 3D fold lock (<=>) and heavy sequence divergence (s ≈ 35%–65%), we paired our Chaperone evaluator with a Bell-Curve Twilight-Zone Index. Given vector cosine distance $d_{\cos}$, normalized Smith-Waterman similarity $s \in [0, 1]$, length ratio $L_{\text{ratio}} = \min(L_1, L_2)/\max(L_1, L_2)$, and Shannon sequence entropy $H$:

\[\text{Twilight Score} = (1 - d_{\cos})^2 \times \underbrace{\left[4 \cdot s \cdot (1 - s)\right]}_{\text{Bell curve peaking at } s = 0.50} \times L_{\text{ratio}}^2 \times \min\left(\frac{H}{4.0}, 1.0\right) \quad \text{for } s \in [0.10, 0.96]\]

Notice how $4s(1-s)$ acts as a natural biological bandpass filter:

Identical Clones ($s = 1.0$): $4(1)(0) = 0.0$ — trivial strain duplicates are zeroed out.
Unrelated Noise ($s < 0.10$): Zeroed out by the homology floor.
The Evolutionary Sweet Spot ($s = 0.414$ for A0A1S8AT22 vs. A0A1H3IA37): $4(0.414)(0.586) = 0.9704$, propelling edge id = 26734 to twilight_score = 0.8640—the single highest cross-genus Twilight-Zone score in the database!

-- 1. Audit all 8 Official RosettaCommons Chaperone Core Examples in ~4.5 ms
SELECT
    ex.example_name,
    ex.category,
    ROUND(audit.sc_rmsd::numeric, 3) AS sc_rmsd,
    ROUND(audit.interface_delta_sasa::numeric, 1) AS delta_sasa,
    audit.cross_interface_hbonds AS iface_hb,
    audit.buried_unsatisfied_polars AS buns,
    ROUND(audit.active_site_or_motif_rmsd::numeric, 3) AS motif_rmsd,
    audit.passes_example_gate AS pass_gate
FROM chaperone_core_examples() AS ex
CROSS JOIN LATERAL evaluate_chaperone_core_example_block(
    ex.example_name, ex.candidate_pdb, ex.reference_pdb, ex.ligand_smiles
) AS audit;

-- 2. Execute the full Twilight-Zone + RosettaCommons Chaperone RASA/H-Bond correlation on Edge #26734
SELECT
    od.id AS edge_id,
    p1.uniprot_id AS bait_id,
    p2.uniprot_id AS discovery_id,
    ROUND((p1.embedding <=> p2.embedding)::numeric, 4) AS esm2_cosine_dist,
    ROUND(
        (ROW(p1.embedding::vector, p1.sequence)::bio_feature
         <~> ROW(p2.embedding::vector, p2.sequence)::bio_feature)::numeric, 4
    ) AS hybrid_bio_dist,
    smith_waterman_score(p1.sequence, p2.sequence, 2, -1, -2) AS sw_score,
    od.seq_identity_pct AS sw_sim_pct,
    od.twilight_score,
    ROUND(ABS(isoelectric_point(p1.sequence) - isoelectric_point(p2.sequence))::numeric, 2) AS delta_pi,
    ROUND(ABS(gravy_score(p1.sequence) - gravy_score(p2.sequence))::numeric, 2) AS delta_gravy,
    ROUND(c1.mean_rasa::numeric, 3) AS bait_rasa,
    ROUND(c2.mean_rasa::numeric, 3) AS disc_rasa,
    ROUND((c2.core_buried_fraction * 100.0)::numeric, 1) AS buried_core_pct,
    c2.total_hbonds,
    c2.buried_unsatisfied_polars AS buns_count,
    chaperone_lineage_hash('SwarmCompare', p1.uniprot_id, p2.sequence, 1) AS lineage_id
FROM orphan_discoveries od
JOIN proteins p1 ON p1.uniprot_id = od.bait_id
JOIN proteins p2 ON p2.uniprot_id = od.orphan_id
CROSS JOIN LATERAL evaluate_chaperone_block(
    pg_read_file('/Users/jonatas/code/ideia.me/assets/models/AF-A0A1S8AT22.pdb'),
    pg_read_file('/Users/jonatas/code/ideia.me/assets/models/AF-A0A1S8AT22.pdb'),
    'A', 'B'
) c1
CROSS JOIN LATERAL evaluate_chaperone_block(
    pg_read_file('/Users/jonatas/code/ideia.me/assets/models/AF-A0A1H3IA37.pdb'),
    pg_read_file('/Users/jonatas/code/ideia.me/assets/models/AF-A0A1S8AT22.pdb'),
    'A', 'B'
) c2
WHERE od.id = 26734;

Full Biophysical & Chaperone Comparison: `A0A1S8AT22` vs. `A0A1H3IA37`

Biophysical & Structural Metric	Bait: `A0A1S8AT22`	Discovery: `A0A1H3IA37`	Biological & Structural Interpretation
Organism & Taxon	Natrinema saccharevitans (`OX=301967`, Natrialbaceae)	Halobellus clavatus (`OX=660517`, Haloferacaceae)	Cross-family divergence across Halobacteriales
Native Habitat	Inland hypersaline lake (Xinjiang, China; sugar-utilizing)	Marine solar saltern (Korea; coastal evaporative brine)	Continental sulfate/chloride lake vs. thalassohaline marine saltern
Gene / Locus Tag	`A6E15_02450` (Uncharacterized protein)	`SAMN04487946_10985` (Uncharacterized protein)	Both annotated as uncharacterized dark proteome
Domain Architecture	`IPR057183` / `PF25260` (`DUF7861`)	`IPR057183` / `PF25260` (`DUF7861`)	Domain of Unknown Function `DUF7861` (full chain)
Sequence Length & Mass	`81 aa` (`8.90 kDa`)	`81 aa` (`9.09 kDa`)	Zero indels; exact 81-residue chain length conservation
Pairwise Sequence Divergence	`49 / 81` identical (`60.5%`), `SW(2,-1,-2) = 67` (`41.4%`)	`32` amino acid substitutions (`3-mer dist = 0.6896`)	Heavy primary sequence remodeling across 32 positions
ESM-2 Vector Distance (`<=>`)	`0.0000` (Reference)	`0.0563` (`<~> = 0.1770`)	Deep transformer embedding recognizes identical 3D topology
Twilight-Zone Homology Index	Reference Bait	`0.8640` (#1 Rank in Swarm DB)	Highest bell-curve convergence-to-divergence ratio in 226k edges
3D Kabsch Core RMSD & TM-Score	Reference (`81 CA`)	`0.81 Å` across `79 / 81` C$\alpha$ atoms (`TM = 0.927`)	Sub-angstrom full-backbone lock across 97.5% of the chain
AlphaFold Mean pLDDT	`90.2` (Very High)	`91.3` (Very High)	Exceptionally rigid, high-confidence 3D fold in both species
Secondary Structure (`81 aa`)	`44 aa` $\beta$-sheet (`54.3%`), `15 aa` $\alpha$-helix (`18.5%`), `22 aa` loop	`45 aa` $\beta$-sheet (`55.6%`), `15 aa` $\alpha$-helix (`18.5%`), `21 aa` loop	5-stranded antiparallel $\beta$-cradle capped by $\alpha$-helix `50–58`
Isoelectric Point (`pI`) & Basic %	`pI = 6.95` (`18.5%` Basic: `9 R, 1 K, 5 H`; `14.8%` Acidic)	`pI = 7.56` (`21.0%` Basic: `9 R, 3 K, 5 H`; `13.6%` Acidic)	$\Delta\text{pI} = +0.61$: Rare neutral-to-basic halophilic electrostatics
Hydropathy (`GRAVY` & Apolar Surface)	`GRAVY = -0.206` (`40.7%` hydrophobic, `18.2%` apolar surface)	`GRAVY = -0.479` (`34.6%` hydrophobic, `15.1%` apolar surface)	$\Delta\text{GRAVY} = 0.273$: H. clavatus sheds `3.1%` surface hydrophobicity
Chaperone RASA & Core Burial	`Mean RASA = 0.365` (`28.4%` buried core, `23` residues)	`Mean RASA = 0.367` (`28.4%` buried core, `23` residues)	Per-residue RASA Pearson $r = 0.9509$ across all 81 residues
H-Bond CSR Network & BUNS	`63` H-bonds (`9` BUNS, $R_g = 11.91\text{ Å}$)	`64` H-bonds (`7` BUNS, $R_g = 11.81\text{ Å}$)	Passes RosettaCommons Chaperone Gate (`chp_swarmcompare_c1_ef19a0dd181befda`)
Key Active & Redox Motifs	`His3–Asp4–His32` dyad (`4.29 Å`), `0 Cys` (`Ala33`, `Val58`)	`His3–Asp4–His32` dyad (`4.99 Å`), `2 Cys` (`Cys33`–`Cys58` at `7.11 Å`)	Marine saltern ortholog adds a core-flanking `Cys33/Cys58` dithiol pair

Defying the Halophile Rule: Why Is a Salt-Loving Archaeal Protein Basic (`pI = 7.56`)?

If you have followed our previous discoveries in Haloferax, Halogeometricum, and Haloprofundus, you know the golden rule of halophilic protein evolution: to survive in $3\text{–}5\text{ M }\text{NaCl}$ or $\text{KCl}$ without salting out and aggregating, halophilic proteins coat their solvent-accessible surfaces with negatively charged Aspartate (Asp) and Glutamate (Glu) residues, driving their isoelectric points down to pI ≈ 4.0–4.5.

Why do both A0A1S8AT22 (pI = 6.95) and A0A1H3IA37 (pI = 7.56) break that rule?

An Electropositive Arginine–Histidine Crown (18.5% to 21.0% Basic Residues): Both 81-residue chains pack 9 conserved Arginines (Arg5, Arg9, Arg17, Arg29, Arg54, Arg59, Arg60, Arg73, Arg77) and 5 Histidines (His3, His7, His13, His32 plus His79 in H. clavatus) onto a compact $\beta$-grasp/SH3-like cradle (54–56% $\beta$-sheet). In high-salt cytoplasm ($4\text{ M }\text{KCl}$), simple electrostatic attraction from Lysine ($\text{NH}_3^+$) to DNA/RNA is heavily screened by counter-ions. Arginine’s planar guanidinium group, however, forms directional, bidentate hydrogen bonds and cation–$\pi$ stacks with nucleic acid bases and phosphate backbones that remain stable even in saturated brine.
The Buried Trp18–Trp75–Tyr76 Aromatic Staple (0.00–0.18 RASA): When we ran pg_bio’s Fibonacci-sphere Shrake–Rupley bio_rasa engine on both structures, their per-residue solvent accessibility profiles correlated at Pearson $r = 0.9509$ across all 81 residues. Both proteins lock their N-terminal and C-terminal $\beta$-strands together using an ultra-conserved C-terminal 73-RVWYR-77 clamp that packs against Trp18 (Trp18–Trp75 distance 9.87–10.71 Å, Trp18–Tyr76 10.96–10.99 Å), while Val37, Leu47, Val49, and Val74 exhibit 0.00 RASA (100% solvent-excluded core).
The His3–Asp4–His32 Pocket (4.29–4.99 Å) & Halobellus Cys33/Cys58 Dithiol Switch: Although His3 (1-MAHDRIH-7) and His32 (29-RDGH-32 / 29-RNGH-32) are separated by 29 residues in sequence, the antiparallel $\beta$-sheet folds them directly face-to-face (4.29 Å in Natrinema, 4.99 Å in Halobellus) next to Asp4. Even more intriguingly, while the inland desert-lake Natrinema protein contains zero cysteines (Ala33 and Val58), the marine solar-saltern Halobellus clavatus ortholog substitutes two cysteines (Cys33 on $\beta$-strand 3 and Cys58 at the C-terminus of the 50–58 $\alpha$-helix) whose sulfur atoms ($\text{S}\gamma-\text{S}\gamma$) sit 7.11 Å apart—a classic poised dithiol/metal-sensing switch for shallow sunlit salterns experiencing diurnal photo-oxidative stress.

Interactive 3Dmol.js Structural Validation (`0.81 Å` Core RMSD)

Both AlphaFold structures below are loaded locally from pre-aligned coordinates (/assets/models/AF-A0A1S8AT22.pdb and /assets/models/AF-A0A1H3IA37.pdb). Thanks to our built-in pg_bio_sync.js controller:

Double-click either 3D canvas to lock both cameras into synchronized 3D rotation and zoom.
Click any residue or atom on one protein to automatically highlight the corresponding structural position on its partner.
Use the interactive chemistry buttons below to inspect the 0.81 Å superposition, the His3–Asp4–His32 dyad and Cys33/Cys58 switch, the Arginine–Aromatic nucleic-acid clamp, and the r = 0.951 buried RASA core.

Bait: A0A1S8AT22 (Natrinema saccharevitans — 81 aa, pLDDT 90.2, pI 6.95)

Discovery: A0A1H3IA37 (Halobellus clavatus — 81 aa, pLDDT 91.3, pI 7.56)

Interactive Structure Explorer: Click any button above to inspect how A0A1S8AT22 (Natrinema saccharevitans, cyan) and A0A1H3IA37 (Halobellus clavatus, magenta) preserve a 0.81 Å Cα RMSD DUF7861 β-cradle across 32 amino acid substitutions, a His3–Asp4–His32 catalytic/metal dyad, and a 9-Arginine electropositive crown built for high-salt nucleic acid binding.

The Horizon: Future Research Ideas

The Bait (A0A1S8AT22 from continental salt-lake *Natrinema saccharevitans) can fold into a compact 81-residue DUF7861 $\beta$-cradle with a 9-Arginine nucleic-acid-binding crown and a His3–Asp4–His32 pocket in steady inland sulfate/chloride brines, but lacks a covalent or metal-bridging dithiol switch when exposed to fluctuating redox potentials and UV-driven oxidative stress, whereas our new Discovery (A0A1H3IA37 from coastal solar-saltern Halobellus clavatus) executes the exact same 3D fold (0.81 Å RMSD, TM = 0.927) while installing a Cys33–Cys58 dithiol pair (7.11 Å), shifting its isoelectric point to pI = 7.56, and reducing surface apolar area to 15.1%, making it a prime scaffold for high-salt synthetic transcription factors, redox-gated RNA chaperones, and cell-free halophilic biosensors.*

Here are three concrete experimental and computational pipelines where A0A1S8AT22 and A0A1H3IA37 can be deployed immediately:

Salt-Proof Nucleic Acid Chaperones for High-Ionic-Strength Cell-Free Systems & Nanopore Buffers: Standard mesophilic RNA/DNA-binding modules lose affinity above 0.3 M NaCl because counter-ion condensation screens lysine–phosphate interactions. Because A0A1H3IA37 maintains 9 invariant Arginines (Arg5, 9, 17, 29, 54, 59, 60, 73, 77) and a Trp18–Trp75–Tyr76 cation–$\pi$ clamp on a hyper-compact 9.09 kDa scaffold (RASA buried core = 28.4%) that is natively soluble in 3–4 M NaCl, it can serve as an ultra-small recruitment domain for tethering polymerases or CRISPR effectors to nucleic acids inside high-salt nanopore sequencing buffers (1 M KCl).
Redox- & Metal-Gated Allosteric Switches via the Cys33–Cys58 Bridge: Comparing the natural Ala33/Val58 pair in A0A1S8AT22 against the Cys33/Cys58 pair (7.11 Å) in A0A1H3IA37 gives protein engineers a nature-validated blueprint for introducing a reversible disulfide or $\text{Zn}^{2+}/\text{Cd}^{2+}$ bridge between $\beta$-strand 3 and the 50–58 $\alpha$-helix without perturbing the 0.81 Å backbone (RASA Pearson r = 0.9509).
Scaling Twilight-Zone + Chaperone RASA Screening Across the Remaining 124,000 Dark Proteins: By executing smith_waterman_score(), sequence_entropy(), bio_rasa(), and evaluate_chaperone_block() directly inside PostgreSQL’s shared memory buffers in Rust, pg_bio eliminates the Python serialization bottleneck—turning a 2.34-million-protein sequence database into a real-time structural discovery engine.

Announcing RosettaCommons Chaperone Support in pg_bio & Mining the Twilight Zone at 0.81 Å (DUF7861)

🚀 Announcing Layer 13 in pg_bio: Native RosettaCommons Chaperone Workflows in SQL

The 5-Stage pg_bio Chaperone Engine (36.9x Faster Than Python)

The Mathematical Filter: Bell-Curve Twilight-Zone Scoring + Chaperone Auditing in SQL

Full Biophysical & Chaperone Comparison: A0A1S8AT22 vs. A0A1H3IA37

Defying the Halophile Rule: Why Is a Salt-Loving Archaeal Protein Basic (pI = 7.56)?

Interactive 3Dmol.js Structural Validation (0.81 Å Core RMSD)

The Horizon: Future Research Ideas

Share this article

Jônatas Davi Paganini

Related Articles

🚀 Announcing Layer 13 in `pg_bio`: Native RosettaCommons Chaperone Workflows in SQL

The 5-Stage `pg_bio` Chaperone Engine (36.9x Faster Than Python)

Full Biophysical & Chaperone Comparison: `A0A1S8AT22` vs. `A0A1H3IA37`

Defying the Halophile Rule: Why Is a Salt-Loving Archaeal Protein Basic (`pI = 7.56`)?

Interactive 3Dmol.js Structural Validation (`0.81 Å` Core RMSD)