Repository navigation
Releases
5.0.0 (2026-11-01)
Release Notes
Pixelfold now scores docked poses from SD files, laid into the reference or a
receptor, and checks each pose's own geometry and its double bonds' E and Z as
PoseBusters does, against the reference ligand. pixelfold rin builds one
network over several files, such as a co-folding model's samples. The version
is major because the viewer library's Pose gains fields and view_poses
takes a receptor; pixelfold-core only gains API, though
ParsedStructure::read no longer prints its warnings, and the poses table
gains a geometry column.
Added
- Docked poses from SD files.
pixelfold posestakes.sdf,.sdand
.molfiles beside structure files, each record a pose, as docking programs
such as DiffDock, Gnina and Vina (through Meeko) write them. A record holds
the ligand alone, so it is laid into a receptor: by default the reference
without its ligand, or the structure--receptornames, which is superposed
on the reference's pocket like any pose and loses its own copies of the
reference ligand, whose definitions give way to the reference's. An SD file
names no atoms, so the ligand's heavy atoms are paired with the reference
ligand's by bond graph, bond orders included, and take their names and
chemical definitions, so a pose turned end for end is not credited with the
reference's hydrogen bonds; a ligand whose bond orders agree in no pairing,
as another tautomer's would not, is paired by its bonds alone with a warning,
and a molecule that does not pair is scored untyped, with a warning.
--model Npicks one record, and--viewdraws docked poses in their
receptor. Only the V2000 layout is read, gzipped (.sdf.gz, as Gnina writes
it) or not. The library gains thesdfmodule andProtein::without. - One network over several files.
pixelfold rintakes several structure
files and folds them into one network, each edge weighted by the share of
models it is present in, as--model alldoes for the models of one file:
co-folding models such as AlphaFold 3 and Boltz write each sample as its own
file. Each file gives the model--modelnames, or every model with
--model all, and--min-frequencyapplies. Residues are matched by chain,
number and insertion code, so a file naming a polymer residue otherwise than
the files before it is refused, as is one naming otherwise a ligand or water
that both networks hold.--mvsjshows the first file. The library gains
rin::Ensemble::try_addandrin::Renamed. - Pose geometry and E/Z double bonds.
pixelfold posesmakes
PoseBusters' checks of each pose's own geometry, with the reference ligand's
geometry in place of RDKit's ideal one: bond lengths and angles more than a
quarter off the reference's, atoms neither bonded nor sharing a neighbour
that come within seven tenths of their lower bound (never further apart than
the reference holds them), and aromatic rings and carbon double bonds more
than 0.25 A from flat where the reference's are flat. A newgeometry
column names the checks a pose fails, with counts (an object of counts in
JSON), and the viewer's pose table names them too.invertednow also names
the reference's E and Z double bonds a pose turns the other way, read from
its chemical definitions; without them, a warning says neither flatness nor
E and Z is checked. The ranking is unchanged. The library gains
rmsd::RmsdReference::assess,rmsd::Assessmentandrmsd::Geometry.
Changed
- The parser prints nothing.
ParsedStructure::readgathers what it
skipped and the identifiers too long for their fields into
ParsedStructure::warnings, and the command line prints them, once per file
where the truncation note used to repeat at every model built. - The viewer library's poses.
pixelfold_tui::Posegainsdockedand
geometry, andview_posestakes the receptor docked poses are laid into:
a breaking change to thepixelfold-tuilibrary. - The pose table's columns.
poseswrites ageometrycolumn between
invertedandlost, andinvertedmay name double bonds beside
stereocentres.
Fixed
--mvsjshowed the first model. A MolViewSpec scene ofrin --model N
drew model 1 with model N's network, since Mol* shows a file's first model
unless told otherwise. The scene now shows the model--modelnames, and
under--model allevery model, overlaid, each with the network's residues
picked out.- The model that failed under
--model all.ss,sasaand
interactionsnow name the model a load error comes from, asrindoes. - Parser warnings over the viewer. A file's warnings, printed while the
viewer held the terminal, landed on its drawing, and the note of a
truncated identifier came again at every model or pose stepped to. The
viewer now shows the shown file's warnings on a header line.
Install pixelfold 5.0.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fuyu-myk/pixelfold/releases/download/v5.0.0/pixelfold-installer.sh | shInstall prebuilt binaries via powershell script
powershell -ExecutionPolicy Bypass -c "irm https://github.com/fuyu-myk/pixelfold/releases/download/v5.0.0/pixelfold-installer.ps1 | iex"Install prebuilt binaries via Homebrew
brew install fuyu-myk/tap/pixelfoldDownload pixelfold 5.0.0
| File | Platform | Checksum |
|---|---|---|
| pixelfold-aarch64-apple-darwin.tar.xz | Apple Silicon macOS | checksum |
| pixelfold-x86_64-apple-darwin.tar.xz | Intel macOS | checksum |
| pixelfold-x86_64-pc-windows-msvc.zip | x64 Windows | checksum |
| pixelfold-aarch64-unknown-linux-gnu.tar.xz | ARM64 Linux | checksum |
| pixelfold-x86_64-unknown-linux-gnu.tar.xz | x64 Linux | checksum |
4.0.0 (2026-11-01)
Release Notes
Pixelfold now triages ligand poses: pixelfold poses ranks every model of the
pose files against a reference by the contacts its ligand keeps, its ligand
RMSD, its clashes and the stereocentres it inverts, and --view lays the
ranking out in a table under the structure, each pose drawn on the reference's
pocket. pixelfold rin --model all folds an ensemble into one network weighted
by how often each edge is present, and the viewer shows it. The version is
major only because rin::Edge and rin::Network gain a public models field:
code outside this workspace that builds either with a struct literal must set
it. Nothing is removed, and a single structure's network reports what it did in
3.0.0 plus models and frequency, both 1.
Added
pixelfold poses, ranking ligand poses against a reference by the
contacts the ligand keeps:pixelfold poses --reference crystal.cif --ligand STI pose_*.cif. Every model of each pose file is a pose, so co-folding
samples, docking output and an NMR ensemble all read the same way. A pose's
fingerprint is the set of residues its ligand touches, each with the kind of
interaction; each pose reports its contacts, those it shares with the
reference, its recovery (the share of the reference's contacts it keeps), its
Tanimoto similarity, its ligand RMSD, its clashes, the reference ligand's
stereocentres it inverts and the contacts it lost, best first. The RMSD
follows AlphaFold 3's protocol: the pose is superposed
on the reference pocket's backbone (N, CA and C within 10 A of the ligand, on
the chain holding the most), and the ligand's heavy atoms pair by name or, for
output named otherwise, by bond graph, in the pairing closest in space that
the bonds allow, so swapped equivalent atoms and reordered copies cost
nothing. The same pairing checks chirality, PoseBusters-style but against the
reference: a tetrahedral stereocentre of the reference ligand, one whose
neighbours no symmetry of the bond graph exchanges, is inverted when the
pose's atoms around it clearly turn the other way, a centre left flat not
counting, and a symmetric ligand merely turned about, such as a meso
compound, inverts nothing. A clash is a ligand-protein heavy-atom pair closer
than 0.75 of their summed van der Waals radii, PoseBusters' default; a
covalent ligand's bonds to the protein, read off the reference, are not
clashes. A pose file without its ligand's chemical definitions, as co-folding
output often is, borrows the reference's.
Residues are matched with the reference's by aligning chain sequences, so a
co-folding model numbered from 1 compares with a crystal structure in author
numbering: of identical chains, as in a homodimer, the pairing that keeps the
most reference contacts wins, a cofactor or an ion matches by component, and a
pose of another protein is refused.--reference-selectpicks the reference's
ligand apart from the poses', such as one copy of a ligand the reference holds
several of; a pose holding several copies, as a homodimer's co-folding model
does, is scored by the copy most like the reference's, named in acopy
column. Water-mediated contacts are left out, since predicted poses carry no
water. The library gains thepose,numberingandrmsdmodules and
Dictionary::with_fallback.- Interaction networks of an ensemble.
pixelfold rin --model allfolds
every model of a file, an NMR bundle or the frames of a trajectory, into one
network whose edges are weighted by the fraction of models they are present
in, the probabilistic network of RING 3.0. An edge's count and energy are
summed over the models and its distance is the shortest in any; a residue
takes the secondary structure most models give it.--min-frequency Fkeeps
the edges present in at least that fraction of the models (the exact
fraction, not the rounded one shown), so--analyzeand--path, which read
the network unweighted, can work on its consensus. The library gains
rin::Ensemble(withcontacts, how many models each residue pair
interacts in by any kind) andNetwork::frequency. - The ensemble's network in the viewer. For a multi-model file,
nturns
the network pane to the network of every model, each edge brightened by the
share of models it is present in; pressing it again hides the edges in fewer
than half of the models, then in fewer than nine in ten, then returns to the
shown model's network. The network stays asmsteps through the models,
laid over each one's atoms, and a selected residue lists the residues it
touches and in how many models, by any kind of interaction. - The pose table in the viewer.
pixelfold poses ... --viewopens the
viewer on the ranked poses instead of printing them: a table under the
structure lists each pose's similarity, recovery, RMSD, clashes and inverted
stereocentres,osorts it by RMSD, clashes or inversions in turn,phides
it, andm/Mor a click on a row draws a pose. Every pose is drawn
superposed on the reference's pocket, so the camera stays put and only the
ligand moves; a pose whose pocket cannot be matched is drawn as its file has
it, and the table says so. The ligand is tinted magenta over a C-alpha trace,
and the view first looks at it down the line of sight the trace blocks least,
since a pocket hides it from most directions. The table's bottom border names
what the pose drawn lacks: the contacts it loses and the centres it inverts.
fframes the pose drawn and looks at its ligand again, a pose file gone
since it was scored is stepped past and named, and the network pane follows
the pose drawn. In the library,
rmsd::SuperpositionandRmsdReference::superpositionare public.
Changed
- Every network edge carries
modelsandfrequency: how many of the
network's models it is present in, and that as a fraction (to three places),
both 1 for a single structure. They are the TSV's last two columns and JSON
and GraphML edge attributes, and the JSON gains a top-levelmodelscount (a
graph attribute in GraphML). In the library,rin::Edgeandrin::Network
each gain a publicmodelsfield, so code that builds either with a struct
literal must set it: a breaking change forpixelfold-core.
Install pixelfold 4.0.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fuyu-myk/pixelfold/releases/download/v4.0.0/pixelfold-installer.sh | shInstall prebuilt binaries via powershell script
powershell -ExecutionPolicy Bypass -c "irm https://github.com/fuyu-myk/pixelfold/releases/download/v4.0.0/pixelfold-installer.ps1 | iex"Install prebuilt binaries via Homebrew
brew install fuyu-myk/tap/pixelfoldDownload pixelfold 4.0.0
| File | Platform | Checksum |
|---|---|---|
| pixelfold-aarch64-apple-darwin.tar.xz | Apple Silicon macOS | checksum |
| pixelfold-x86_64-apple-darwin.tar.xz | Intel macOS | checksum |
| pixelfold-x86_64-pc-windows-msvc.zip | x64 Windows | checksum |
| pixelfold-aarch64-unknown-linux-gnu.tar.xz | ARM64 Linux | checksum |
| pixelfold-x86_64-unknown-linux-gnu.tar.xz | x64 Linux | checksum |
3.0.0 (2026-10-30)
Release Notes
Pixelfold now detects pi-hydrogen bonds, and the validation harness scores
them against RING 4.0. The version is major only because InteractionKind
gains a variant: code outside this workspace that matches on it exhaustively
needs a PiHydrogenBond arm or a wildcard. Nothing is removed, and every
existing interaction kind reports what it did in 2.1.0; residue networks gain
the new edges.
Added
- Pi-hydrogen bonds, a new interaction kind (
pi-hydrogen-bond): an N-H or
O-H pointing into the face of an aromatic ring. PLIP has no such type, so the
geometry is RING 4.0's: the donor within 4.3 A of the ring centre and within
25 degrees of its axis, and the D-H...centre angle above 120 degrees. The
donors are the hydrogen bond donors, and a rotatable hydrogen is turned toward
the ring. The kind reaches every surface that lists interactions:
interactionsandrin(with--type pi-hydrogen-bond), their exports, and
the network pane, where it is drawn asπ-H. A donor bonds a residue once,
through the nearer of a tryptophan's two rings where it sits over both. Its
energy weight is 12.5 kJ/mol, after Levitt and Perutz's 3 kcal/mol. Against
RING, now scored on eight types, it finds
95 of RING's 97 pairs at a precision of 0.65 (F1 0.66); most of the extra
pairs come from the rotatable hydrogens, which RING fixes at one torsion. The
mean RING F1 moves from 0.729 to 0.721.InteractionKindgains a
PiHydrogenBondvariant, which breaks an exhaustivematchon it outside
this workspace.
Install pixelfold 3.0.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fuyu-myk/pixelfold/releases/download/v3.0.0/pixelfold-installer.sh | shInstall prebuilt binaries via powershell script
powershell -ExecutionPolicy Bypass -c "irm https://github.com/fuyu-myk/pixelfold/releases/download/v3.0.0/pixelfold-installer.ps1 | iex"Install prebuilt binaries via Homebrew
brew install fuyu-myk/tap/pixelfoldDownload pixelfold 3.0.0
| File | Platform | Checksum |
|---|---|---|
| pixelfold-aarch64-apple-darwin.tar.xz | Apple Silicon macOS | checksum |
| pixelfold-x86_64-apple-darwin.tar.xz | Intel macOS | checksum |
| pixelfold-x86_64-pc-windows-msvc.zip | x64 Windows | checksum |
| pixelfold-aarch64-unknown-linux-gnu.tar.xz | ARM64 Linux | checksum |
| pixelfold-x86_64-unknown-linux-gnu.tar.xz | x64 Linux | checksum |
2.1.0 (2026-10-30)
Release Notes
The Linux release binaries now start on RHEL and Rocky 8, the usual cluster
login node, and pixelfold sasa stops dropping modified residues. A water now
bridges two donors and two acceptors at most, as its lone pairs and hydrogens
allow. The validation harness gains RING 4.0 as a second interaction reference
and scores PLIP only on the ligands PLIP profiled. The library adds one method,
Atom::is_polymer, and removes nothing.
Added
Atom::is_polymerinpixelfold-core: whether an atom belongs to a
polymer chain, decided by its entity, so modified residues written as HETATM
records count. Without entity records it falls back to the record type.- Interaction agreement against RING 4.0, the whole-structure reference
beside PLIP's ligand sites. Its golden files cover all 205 benchmark entries
and come from the public RING 4.0 web server
(benchmarks/tools/generate_ring.py), since RING's standalone binary may not
be passed on. Seven RING types map onto pixelfold's; van der Waals contacts,
pi-hydrogen bonds and generic contacts have no counterpart and are left
unscored. Pixelfold is scored only where RING looks: edges to water, and
amino acids of one chain fewer than three positions apart, are left out.
Per-type F1 runs from 0.32 (pi-stacking, where RING's wider window finds
seven times as many) to 1.00 (disulfide), with a mean of 0.729, and
--checknow fails below 0.65 (min_ring_f1).
Changed
- The Linux binaries need glibc 2.28 instead of 2.34. Both Linux targets
now build inside a pinned manylinux_2_28 container rather than on the bare
Ubuntu 22.04 runner, so the prebuiltpixelfoldstarts on RHEL/Rocky 8 and
later. The macOS and Windows binaries andcargo installare unaffected. - A water bridges two donors and two acceptors at most. A water accepts a
hydrogen bond on each of its two lone pairs and donates one from each of its
two hydrogens, but the geometric gates alone let one water in a crowded site
bridge every donor near it to every acceptor near it. Where a water has more
candidates than it can serve,pixelfold interactionsandpixelfold rinnow
keep the bridges whose angle at the water is nearest tetrahedral. Against
PLIP, water-bridge precision moves from 0.53 to 0.59 and recall from 0.87 to
0.83, so F1 rises from 0.62 to 0.65 and the mean PLIP F1 from 0.828 to 0.832.
RING scores no water bridges, so its numbers do not move.
Fixed
pixelfold sasaleft out modified residues. Surface atoms were chosen by
record type, so a selenomethionine or phosphoserine inside a chain, written as
a HETATM record, was neither reported nor allowed to shield the residues
beside it, which read too exposed: 3G5S's 14 MSE were missing, and--select 'resn MSE'printed an empty table. The surface is now every heavy atom of a
polymer entity (Atom::is_polymer); a file without entity records keeps the
record-type rule. Selenium also had no radius of its own and took carbon's
1.70 A; it now takes Bondi's 1.90 A. The FreeSASA golden files are
regenerated on the same atom set, the 51 entries holding modified residues
change, and 8A71's modified DNA is scored for the first time: SASA mean
absolute error against FreeSASA moves from 4.93 to 4.77 A^2 over 203 entries.- The validation harness charged pixelfold for ligands PLIP never profiled.
It decided what a ligand was by the HETATM record, while PLIP leaves out
modified residues of a chain and sets aside some crystallisation additives,
so pixelfold's contacts to those counted as false positives. PLIP golden
files now list the residues PLIP profiled, and only edges touching one of
them are scored. Hydrogen-bond precision moves from 0.58 to 0.64 and
water-bridge precision from 0.50 to 0.53; no recall changes. A golden file
that fails to parse, or a ligand-scoped one whose interactions name a ligand
it does not list, now stops the harness instead of skipping its entry, which
had left the gates it feeds silently unchecked.
Install pixelfold 2.1.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fuyu-myk/pixelfold/releases/download/v2.1.0/pixelfold-installer.sh | shInstall prebuilt binaries via powershell script
powershell -ExecutionPolicy Bypass -c "irm https://github.com/fuyu-myk/pixelfold/releases/download/v2.1.0/pixelfold-installer.ps1 | iex"Install prebuilt binaries via Homebrew
brew install fuyu-myk/tap/pixelfoldDownload pixelfold 2.1.0
| File | Platform | Checksum |
|---|---|---|
| pixelfold-aarch64-apple-darwin.tar.xz | Apple Silicon macOS | checksum |
| pixelfold-x86_64-apple-darwin.tar.xz | Intel macOS | checksum |
| pixelfold-x86_64-pc-windows-msvc.zip | x64 Windows | checksum |
| pixelfold-aarch64-unknown-linux-gnu.tar.xz | ARM64 Linux | checksum |
| pixelfold-x86_64-unknown-linux-gnu.tar.xz | x64 Linux | checksum |
2.0.0 (2026-10-22)
Release Notes
The structure-IO milestone. PixelFold now reads structures through
mmcif-model and mmcif-cif instead of pdbtbx, and comparing the two readers
across the benchmark corpus turned up a string of numbering and conformer bugs,
fixed here. Multi-model files and biological assemblies become first-class:
pick a model or report on all of them, and build an assembly with each copy of
a chain named. Loading a large structure is about 30 times faster.
This is a major release because the library API changes. Atom::residue_seq
is signed and Atom::chain_id is an eight-byte ChainId; loading goes through
parser::Load; and pixelfold-tui and pixelfold-core now expose only what
the other crates use (see Removed). The command line keeps every option it had;
single-model output keeps its shape, with residues listed in label-chain order.
Added
- Biological assemblies, built on request with
--assembly(the primary
assembly) or--assembly=ID, on every subcommand and in the viewer, whereu
switches between the deposited coordinates and the primary assembly. Copies
of a chain are namedA,A-2,A-3, with a ligand or water following its
protein copy, soselect chain A-2picks one copy and the residue network
shows the interfaces between copies (6ZPA's hexamer, 7AEW's two copies of
its asymmetric unit). Where the deposited coordinates hold more than one
biological unit, the assembly keeps only its own chains. An unknown id is an
error naming those the file declares; a legacy PDB file builds the assemblies
itsREMARK 350declares; andrin --mvsjrefuses--assembly, since Mol*
names copies by operator. Library users name what to build with
parser::Load(model, assembly, altloc policy), which
ParsedStructure::protein,App::load_proteinandpixelfold_tui::view
now take. --model N|all, a global option choosing which model of a multi-model
file (an NMR ensemble, say) every subcommand and the viewer load; the first
model remains the default, and an absent one is an error naming the models
the file has.select model Nnow has something to match.--model all
makesss,sasaandinteractionsreport every model, with a leading
modelcolumn in the table and TSV and amodelfield in the JSON; output
for a single model keeps its shape. A selection is then an error only when it
matches nothing in any model.rin,renderand the viewer work on one model
at a time and refuseall.- Model stepping in the viewer:
mandMshow the next and previous
model, wrapping at either end, without re-reading the file. The camera stays
put, the info bar names the model, and the selection and network pane start
over for the new model's atoms. Library users read a file once with
parser::ParsedStructureand build any model from it;App::load_protein
andpixelfold_tui::viewtake the model to open.
Changed
-
Chain ids hold eight bytes, as
ChainId, so a four-character author id
keeps room for a copy number (AAA-2); atom names stay four. -
Loading a large structure is about 30 times faster. Secondary-structure
assignment tried every pair of residues for a beta bridge; it now tries only
the pairs within a step of some hydrogen bond's two ends, the only pairs a
bridge pattern can match. On 1AON (58,870 atoms) the pass fell from 1.7 s to
11 ms and the whole load from about 1.8 s to 60 ms, with every assignment
across the benchmark corpus unchanged. -
Element symbols are matched without allocating. Radii, colours, metal
tests and ligand atom typing upper-cased each symbol into a new string, on
every atom and every frame; building a frame's scene for 1AON fell from
2.4 ms to 0.6 ms. -
Structures load through
mmcif-modelandmmcif-cifin place of pdbtbx.
One parse now serves the coordinates, the component definitions and the
assembly records, so the hand-written CIF reader and the space-group retry are
gone. Reports list residues in label-chain order (polymer chains, then
ligands, then waters, as an mmCIF file writes them) rather than grouped by
author chain, and an atom group lists its atoms in file order. Gzipped files
and RCSB's numbered assembly files (.pdb1) still load, BinaryCIF (.bcif)
loads too, and a file whose assembly or entity records cannot be read loads
its coordinates with a warning instead of failing; a file whose extension
names no structure format is refused. A missing B-factor is missing, shown as
n/aand coloured grey, rather than read as 0.
Removed
- Most of
pixelfold-tui's public API. It exposesview,searchand
ViewProtocol;App, its fields and methods,RenderedFrame, the picking
helpers, and thesearchmodule are internal, andDisplayModemoves here
frompixelfold-core. pixelfold-coreinternals from the public API: theassembly,dssp
andspatialmodules (HBondstays, re-exported), the interaction
submodules (aromatic,classify,connectivity,hydrogens,params,
topology) andCharge, the selection syntax tree (Selection,Cmp,
ResiRange,AtomSet::of),filter_altlocs, and
Dictionary::from_block, which took anmmcif-ciftype. Nommcif-*type
is part of the public API, so a future mmcif upgrade is not a breaking change.- Unused code:
rin::HBondGraphwith its motifs and analysis,
load_protein_backbone,load_protein_ca_only,
SecondaryStructure::color_256,params::GRID_CELL_SIZE, and
Grid::within. App::residue_colorsinpixelfold-tui. It mapped a residue number to a
secondary structure with no chain, and every atom already carries its own
residue's secondary structure.
Fixed
- Backbone connections took another chain's colour. In backbone mode a
connection was coloured by the first residue anywhere in the structure with
the same number, so in a multi-chain structure chain B's residue 10 showed
chain A's residue 10 secondary structure. Each connection now takes its own
residue's. - Modified residues were selected as ligands.
proteinmatched only the 20
standard amino acid names andligandany non-water, non-metal HETATM, so a
selenomethionine or phosphotyrosine inside a chain was a ligand and not
protein (3G5S's 14 MSE, 9I24's PTR, 8OW6's HYP).protein,nucleic,
ligandandwaternow follow the file's entities, recorded on each atom as
Atom::molecule; a free amino acid declared as a non-polymer stays a ligand.
A file without entity records keeps the old name and record-type rules. - Hydrophobic contacts depended on atom order. The reduction that keeps
the closest contact per atom and residue ran from whichever side came first
in the file, and the two directions can keep different numbers of contacts,
so a residue network's summed energies followed atom order. It now runs from
the ligand side, as PLIP's does, then from the earlier residue: in 1IEP,
imatinib's contacts with Lys 271 count 2 instead of 1. Hydrophobic contacts,
pi-stacking and disulfides list their sides in that order too. No residue
pair gains or loses an interaction across the benchmark corpus, and the
validation report is unchanged. - Residues differing only by insertion code were merged in two places.
Selecting an atom highlighted every residue with its chain and number, so
clicking thrombin's1Alit up1C,1B,1Aand1together; and
HBondGraphmade one node of them all, pooling their atoms and hydrogen
bonds. Both now key a residue by chain, number and insertion code;
ResidueNodecarriesinsertion_codeandHBondGraph::get_nodetakes it. - The default altloc policy kept the minor conformer wherever a side chain
was split and the backbone was not. pdbtbx copied the shared atoms into every
conformer at a fraction of their occupancy, so the backbone vote always tied
and the first label won: 76 residues across 18 of the 206 benchmark entries
kept the less occupied side chain (1GCI SER 3 kept A at 0.47 over B at 0.53).
Shared atoms now stay unlabelled at their own occupancy,--altloc bno
longer drops the backbone of a residue modelled only in A, and--altloc all
no longer reports shared atoms twice. - Negative residue numbers wrapped to 4294967295 and up, so those residues
printed wrong, could not be selected, and never matched a reference in
pixelfold-validate. Residue numbers are signed, andresitakes negative
numbers and ranges (resi -5--1). - Unknown atoms (UNX) were typed as uranium, the element inferred from the
nameUNKbecause theXelement could not be read. They keep elementX. - A numeric-looking component id was rewritten (
4E2read as400), so the
ligand printed under the wrong name. Component ids are kept verbatim. - A legacy PDB file's insertion-code residues were re-sorted, so a chain
numbered1C, 1B, 1A, 1(thrombin's light chain) was read as1, 1A, 1B, 1C,
and secondary structure and the bond graph saw the wrong neighbours. The PDB
and mmCIF files of one entry now give the same result (3PTB, 5JZY, 5LCE,
6T3Q, 6T4A did not). - Two kinds of partial biological assembly went unflagged: an asymmetric
unit copied whole across separate generation records (7AEW) and a chain group
placed by a crystal operator (3BOG). An assembly is now partial when it places
a chain more than once or places its chains by different operators.
Install pixelfold 2.0.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fuyu-myk/pixelfold/releases/download/v2.0.0/pixelfold-installer.sh | shInstall prebuilt binaries via powershell script
powershell -ExecutionPolicy Bypass -c "irm https://github.com/fuyu-myk/pixelfold/releases/download/v2.0.0/pixelfold-installer.ps1 | iex"Install prebuilt b...
1.0.0 (2026-08-02)
Release Notes
The rendering and network milestone. The viewer stops drawing an approximation of
the structure and draws real pixels, through the same framebuffer path the new
headless render subcommand writes, and the residue interaction network becomes
a first-class linked view with its own analytics, energies, and filters. The
interaction geometry is untouched by this release, so the agreement figures
published for 0.6.0 stand as measured.
Added
- Z-buffered sphere-impostor rasteriser (
render/src/raster.rs). Each atom is
a screen-space disc whose front-hemisphere depth is solved per pixel and depth
tested, so there is no sort and no painter's-algorithm bleeding, and the surface
normal falls out of the same solve for Lambert shading. It writes an RGBA8
framebuffer carrying parallel depth and atom-id buffers, with depth-cue fog over
distance. Everything renders through that one buffer: the interactive viewer,
PNG export, and the tests. - Terminal graphics sinks (
render/src/sink/): the Kitty graphics protocol (a
base64 PNG payload chunked at 4096 bytes), iTerm2's OSC 1337 inline image, and a
truecolor upper-half-block fallback packing two vertical pixels per cell.
detect.rsreads terminal capability and SSH from the environment and
to_terminaldispatches on what it finds. The payload is a compressed PNG, so a
remote session does not pay for raw pixels. Sixel is reachable in the
interactive viewer throughratatui-imagebut not yet throughto_terminal. rendersubcommand, space-filling spheres with no terminal UI involved.
With-oit writes a PNG (1200x900 unless told otherwise); without one it emits
a terminal graphics escape when standard output is a terminal, sized to the cell
grid so half-block output cannot overflow it, or pipes a PNG when redirected. It
takes--width/--height,--color element|bfactor|ss,--backbone,
--select,--radius-scale,--no-depth-cue,--slab FAR NEAR, and
--protocol auto|kitty|iterm2|half-block, and it errors when the selection
leaves nothing to draw rather than writing an empty image.- Slab (clipping planes), a depth band that slices a dense structure open:
--slab FAR NEARheadless, and in the viewerkto toggle,[and]to move
the band through depth,,and.to narrow or widen it. - ID-buffer mouse picking. The rasteriser records the source atom index at
every pixel it wins, so a click reads the atom under it directly instead of
hunting for the nearest projected centre. The last frame's buffer and image area
are kept for exactly this. - Depth-aware overlay lines (
render/src/overlay.rs). Backbone connections and
hydrogen bonds interpolate their endpoint depths against the rasteriser's own
z-buffer, so a bond that passes behind an atom is hidden by it. - Golden-image regression test over a fixed synthetic scene at a fixed camera,
which needs no structure file and so runs in CI. It guards projection,
occlusion, shading, depth cue, and the PNG encode together, with a per-channel
tolerance and a changed-pixel budget so a rounding difference at a sphere rim
does not fail the run. - Linked network pane (
g), the residue interaction network drawn beside the
3D view. Selection is shared both ways, so clicking a node highlights that
residue in the structure and picking an atom locates its node. Two layouts (l):
a deterministic Fruchterman-Reingold force-directed layout
(core/src/rin/layout.rs), which reads topology independent of orientation, and
a spatial layout that projects every residue through the 3D camera, so the graph
turns with the structure. Edges are coloured by interaction kind with a legend on
the border; nodes are haloed by betweenness centrality so hubs stand out. - Network analytics (
core/src/rin/analysis.rs): connected components,
betweenness centrality by Brandes, articulation points, and a breadth-first
shortest path between two residues.pixelfold rin --analyzeprints them as a
structural report (components, hub residues by betweenness, cut residues)
instead of exporting the graph, and--path FROM TOprints the shortest chain of
interactions between two residues, or says they lie in different components. - Per-edge interaction energies (
core/src/interactions/energy.rs), one
literature-sourced magnitude in kJ/mol per interaction kind with a primary
citation for each, on the intrinsic single-contact basis RING's own weights use
rather than gas-phase extremes or net folding free energy. Every network edge
sums the contacts folded into it (weight times count) and carries that sum into
node-link JSON, GraphML, and the TSV edge list, and the pane brightens edges by
it on a log scale. The four kinds RING scores land inside these independently
sourced ranges. Two values carry caveats stated at the table: the hydrophobic
weight is per atom-atom contact, the scale detection works at, so an edge's many
contacts sum to the per-residue-pair magnitude; and the disulfide is the covalent
S-S bond energy, an order of magnitude above every non-covalent contact, so it is
compressed at the display layer rather than distorted in the table. - Relative solvent accessibility.
relative_sasadivides per-residue area by
Tien 2013's theoretical maximum.pixelfold sasareports it as anrsacolumn
beside the absolute area, and in the network paneetoggles node colouring
between secondary structure and burial (per-node RSA, summed once on open from
per-atom SASA), so the structural core and the solvent-facing surface separate at
a glance. - Network filters, both reflected in the pane title with a live count of
visible edges:tsteps the interaction-kind filter (all kinds, then each kind
present in turn) andythe chain filter (all, intra-chain, or the inter-chain
interface only). The energy intensity scale rescales to whatever remains visible,
so a filtered view still spans its own range. - MolViewSpec export:
pixelfold rin --mvsjwrites a.mvsjscene for Mol*,
the structure with the network's residues highlighted. A 4-character id resolves
to its canonical RCSB mmCIF URL, which any Mol* can fetch; a local path becomes
afile://URL for a desktop or locally served viewer. - The
rinTSV edge list prints aligned fixed-width columns when it lands on a
terminal and genuine tab-separated values when written to a file or a pipe, so
reading it by eye and piping it into a script no longer trade against each other. rust-toolchain.tomlandrustfmt.tomlpin the toolchain and the 2024 style
edition, so local formatting and CI cannot diverge by version.
Changed
- The viewer draws the shared RGBA framebuffer, shown through
ratatui-image
(iTerm2, Kitty, or Sixel, with a half-block fallback), instead of the braille
canvas it rendered to before.pixelfold-tuiis split accordingly intoapp
(event loop),inputs(keyboard and mouse),render(framebuffer build), and
ui(per-frame layout), and the viewer's framebuffer is the same onerender
writes to a PNG. - The network is built on a worker thread the first time the pane opens.
Detection, layout, and the accessibility pass no longer freeze the event loop on
a large structure; the pane shows a placeholder until the result arrives. The
loop polls for it inside the 16 ms wait it already performs, so frame timing is
unchanged. - The interactive iTerm2 draw encodes each frame at the fastest compression
(to_png_bytes_fast), roughly halving the per-frame cost. For a local terminal a
larger payload is cheaper than the time spent shrinking it. Files written by
renderkeep the default compression, where the trade runs the other way. - The event loop drains every buffered event before drawing again, so a burst of
key repeats advances the view once rather than once per event. - The framebuffer is capped in its largest dimension and scaled to fill the image
area, rather than always allocated at the full cell grid times font size.
Removed
- Solvent-surface point generation:
SurfacePoint,
SurfaceCalculator::calculate_surface, and the Fibonacci-sphere distribution
behind it. Nothing had drawn the points since the braille viewer, yet every load
paid to build them. Per-atom Shrake-Rupley areas (calculate_atom_sasa) are
untouched: they remain whatpixelfold sasareports, what the validation harness
compares against FreeSASA, and what the network pane sums per residue for burial
colouring. - The
--no-surfaceflag, onpixelfold viewand on the bare-structure viewer
path. It skipped a calculation that no longer happens.
Fixed
- Mouse picking maps a click proportionally onto the framebuffer instead of
through a fixed cell-to-pixel size. The framebuffer can be a lower resolution
than its image area and the terminal scales it to fit, so the fixed mapping
selected the wrong atom, or none, whenever the two disagreed. - The headless CLI tests took their structures from the gitignored benchmark cache
and skipped themselves when it was absent, so on a clean checkout, CI included,
all 24 reported as passing without running the binary once. The fixtures
(crambin, ubiquitin, thrombin) are now committed undertests/fixturesand a
missing one is a hard failure.
Install pixelfold 1.0.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/fuyu-myk/pixelfold/releases/download/v1.0.0/pixelfold-installer.sh | shInstall prebuilt binaries via powershell script
powershell -ExecutionPolicy Bypass -c "irm https://github.com/fuyu-myk/pixelfold/releases/download/v1.0.0/pixelfold-installer.ps1 | iex"Install prebuilt binaries via Homebrew
brew install fuyu-myk/tap/pixelfoldDownload pixelfold 1.0.0
| File | Platform | Ch...
v0.6.0 - the interaction engine
The interaction-engine milestone. PixelFold now computes non-covalent
interactions and a residue interaction network locally, and publishes their
agreement with reference implementations.
Highlights
- Interaction engine over the uniform spatial grid: hydrogen bonds, salt
bridges, hydrophobic contacts, pi-stacking, pi-cation, halogen bonds, water
bridges, metal coordination, and disulfides, with angular terms and thresholds
derived from PLIP's source. - Ligand support from the PDB Chemical Component Dictionary (bond-graph atom
typing, aromatic ring perception, pH 7.4 polar-hydrogen inference). - Selection language (
--select) and headless subcommandsinteractions,
rin,ss,sasa,fetch, with TSV / JSON / GraphML output. - Validation harness with a CI regression gate, comparing against DSSP 4,
FreeSASA, and PLIP over a stratified 205-entry benchmark.
Validation
- DSSP 4: Q3 97.1% (194 entries); backbone H-bond edge F1 0.97 (205).
- FreeSASA: MAE 4.93 A^2, median 3.77, Pearson r 0.99 (202).
- PLIP (60 protein-ligand complexes, per-type F1): halogen 0.98, pi-stacking
0.92, hydrophobic 0.87, pi-cation 0.84, salt-bridge 0.83, metal 0.81,
hydrogen-bond 0.65, water-bridge 0.59.
See CHANGELOG.md for the full list. RING 4.0 (protein-protein
reference) is license-gated and not yet generated; prebuilt binaries and
crates.io publishing arrive with the packaging milestone.