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:

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:

1. 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 ($\text{RASA}_i \ge 0.25$).
2. 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.
3. 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:

• 📖 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 $H$:

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

1. Identical Clones ($s = 1.0$): $4(1)(0) = 0.0$ — trivial strain duplicates are zeroed out.
2. Unrelated Noise ($s < 0.10$): Zeroed out by the homology floor.
3. 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

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?

1. 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.
2. 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).
3. 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.

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.