Structure analysis

GL1X reads every MolNFT structure in full atoms once, when it indexes it, and keeps what structural biologists compute first: secondary structure, solvent accessibility, chain interfaces, pockets, side-chain positions and, for nucleic acids, base pairs. The results are stored with the record, so a record's page, the API and MCP answer from the index in milliseconds, and the whole archive can be searched by them: druggable empty pockets lined by a cysteine, large protein–protein interfaces of one UniProt accession, every β-α-β unit holding a P-loop motif, every Ser-His-Asp triad in any fold.

Experimental structures and predicted models are analysed alike; the pocket, interface and modification searches can keep one kind or the other (source). The first model of a file is read; alternate locations keep their first conformation.

What is computed · Pockets · Interfaces · Secondary-structure patterns · 3D templates · In the explorer · API and MCP · Validation

What is computed #

The residues of a chain keep the author numbering of the file (residue_number, insertion codes included), so what GL1X reports is what papers cite: Lys16 of KRAS is 16.

Pockets #

A pocket is a cavity of the protein surface able to hold a small molecule. GL1X finds them on a 1 Šgrid around each structure (coarser by steps of 1.25 for a structure too large for 32 million grid points): a grid point is a candidate when it lies at least 3 Šfrom every polymer atom (a probe fits) and is enclosed, with 21 or more of 30 rays of 12 Šmeeting the protein; candidate points touching each other (26 neighbours) make a pocket, and a pocket of more than 1,200 grid points (a groove along a whole domain) is split where it is least buried. Up to 20 pockets of 50 ų or more are kept per structure.

Each pocket has:

• volume (ų, the number of grid points), radius of gyration;

• buriedness: the mean share of rays that meet the protein (0–1);

• lining residues (an atom within 4.5 Å of a pocket point) with their polar and apolar atoms, apolar share, hydropathy (Kyte–Doolittle) and charge;

• ligands inside: the structure's ligands with at least two fifths of their atoms within 2 Å of the pocket's points. A pocket is occupied when one of them is a ligand or a sugar; one holding only crystallization additives (glycerol, sulfate, PEG…), ions or metals has the state additives; an empty one holds nothing. The pocket search's state=additives keeps every pocket without a ligand, empty or with additives only;

• druggability, a logistic score of size, enclosure and apolarity — the properties that set how strongly a pocket can bind a drug-like molecule (Y Cheng et al. Nat Biotechnol 25, 71–75, 2007; T A Halgren. J Chem Inf Model 49, 377–389, 2009; P Schmidtke, X Barril. J Med Chem 53, 5858–5867, 2010):

  z = −9.44 + 1.42 log₁₀ min(V, 1000) + 7.69 buriedness + 0.47 apolar, druggability = 1 / (1 + e^−z)

  The weights were fitted on ligand-bound structures: the pockets drug-like ligands fill against all the others (validation).

Pockets are ranked by druggability. Searching them across the archive filters on every property — volume, buriedness, the apolar share of the lining and its mean hydropathy, druggability, contents — and finds, for instance, empty pockets of predicted models that look like binding sites, or pockets lined by a cysteine (covalent drug design): GET /api/bio/pockets?state=empty&min_druggability=0.7&residue=CYS; or polar pockets for fragments: ?max_hydropathy=-0.5&min_volume=150.

Interfaces #

Every pair of chains burying 20 Ų or more forms an interface. Its area is half the accessible surface the two chains bury together (each chain's surface alone plus the other's, minus the complex's), the convention of PISA and of most papers. Each side's residues are listed, with the hydrogen bonds (N/O pairs within 3.5 Å) and salt bridges (Lys/Arg–Asp/Glu side chains within 4 Å) across. Kinds: protein–protein, protein–nucleic acid, nucleic–nucleic.

Secondary-structure patterns #

A chain's DSSP string read in three states (H for H, G and I; E for E and B; L for the rest) or in all eight is searched with a regular expression: E{4,}L{2,5}E{4,} is a β-hairpin, H{10,} a long helix. Presets name the classic units and check how their strands pair:

A sequence motif can be required inside the stretch (sequence, PROSITE-like: G-x(4)-G-K-[ST] is the Walker A motif), so "β-α-β units holding a P-loop" is one query. Patterns have up to 300 characters, repeat counts up to 500 and at most 2,000 states once compiled (repeats in a row add up); the presets have fewer than 70.

3D templates #

A 3D template is a few residues in a fixed arrangement — an active site, a metal site — found again in any protein, in any order along the sequence and in any fold: the 3D counterpart of a PROSITE pattern. Each template residue is a Cα and a side-chain point with the amino acids allowed there, as SPASM describes residues (G J Kleywegt. J Mol Biol 285, 1887–1897, 1999); a chain matches when residues of those types superpose onto the template with an RMSD over Cα and side-chain points under the limit (1.0 Å by default, 3 Å at most). The Cα and side-chain points of every indexed chain are in memory, so the archive is scanned directly: residues of the allowed types whose pairwise distances agree with the template's within a tolerance (tolerance; by default twice the RMSD limit, at least 1.5 Å: 2 Å for the default 1 Å) are assembled and fitted exactly (Kabsch superposition), best first, up to ten matches per chain; templates have up to eight residues, each side-chain point within 8 Å of its Cα. A chain whose search has examined two million candidate residues or fitted 200,000 arrangements stops there, and its matches may be incomplete (chains_cut counts such chains); a search stops at the structure time limit (GL1X_STRUCTURE_TIMEOUT, two minutes).

Presets: ser-his-asp (the catalytic triad of serine proteases, taken from trypsin, PDB 1GBT; subtilisin and α/β-hydrolases have the same triad in other folds), c2h2-zinc, cys4-zinc, c3h-zinc, and metal-site, a geometric mode for putative metal sites: three or four Cys, His, Asp or Glu (two of them Cys or His) whose side-chain points lie within 3.8 Å of one point. A template can also be taken from any indexed chain by residue numbers (from), e.g. a site from a paper, and its residue types widened (types).

In the explorer #

The menu's Bioinformatics onchain section opens every tool; its overview (/bio) counts what the archive holds.

• The research workbench (/bio/workbench/{contract}/{token}) shows one structure in Mol* across the page with every tool under it (on a screen tall enough the structure stays in view while the tool's panel scrolls), each drawing the structure the way its question needs: secondary structure in DSSP colours, accessibility and hydropathy as surfaces, half-sphere exposure, the sites residues touch, base pairs of nucleic acids with their dot-bracket, a contact map whose contacts open in 3D; pockets as volumes with their lining residues; interfaces with each side in its colour. Its other tools — conservation, variants, modifications, peptides and HDX, crosslinks, annotation, neighbours — are described in Families and Proteomics. Links name what to show: ?tool=pockets&pocket=2, ?tool=interfaces&chains=A,B, ?residues=57,102,195 or ?residues=10-24.
• Pockets (/bio/pockets), Interfaces (/bio/interfaces), SS patterns (/bio/ss-patterns), 3D templates (/bio/templates) and Modifications (/bio/ptms) search the whole archive and show the chosen hit in 3D across the page above the hits, drawn as the workbench draws it, ready to open there. Every hit links to the workbench in that view, to its record and to PubMed by its title and PDB ID (PubMed search).

API and MCP #

Searches page with cursor (or offset) and answer from the index; a record not yet analysed is analysed on its first request.

Validation #

• DSSP against mkdssp 4.2.2 on 113 structures without alternate locations (27,237 residues): secondary structure identical for 99.97% of residues, accessibility identical (integer Ų) for 99.26% and within 1 Ų for 99.69%; hydrogen-bond partners and energies identical. With every residue of files with alternate locations: 99.86% and 95.99% (mkdssp keeps the last conformation's backbone with every side chain).
• Buried surface against FreeSASA 2.1.2 (Shrake–Rupley, DSSP radii): HIV-1 protease dimer 1HVR 3735.3 vs 3736.1 Ų (0.02%); 4E43 3729.2 vs 3725.5 Ų.
• Pockets: on 14 structures holding drug-like ligands, the ligand's pocket ranks first in 11 and in the top three in 12; with each structure left out of the fit, first in 10 and top three in 12–13.
• Base pairs: DNA duplexes fully paired (1LCD, 11 bp); the flipped base of M.HhaI unpaired (3MHT); the riboswitch 4GXY gives 39 pairs (31 G–C, 5 A–U, 3 G–U) with its pseudoknot.
• Speed on two CPUs: 16,600 atoms analysed in 0.33 s; a ribosome (165,000 atoms, 6ZU5) in 2.0 s; pockets of 4JSV in 0.42 s.