ProPrep 1.16.0

Added

• Externally obtained parameters can be imported and used. The import wizard
  browses for the .frcmod, .lib/.off and optional .prep with the same
  file browser the rest of ProPrep uses, shows what each parameter category
  means and where it will be stored, and reads any atom types the frcmod
  declares so they reach tLEaP as addAtomTypes entries. Generated parameters
  cannot introduce new types, which is why the wizard never asked; imported
  ones can.

• A transformer can be saved that only binds a force-field library, with no
  renaming. Renaming and parameter-binding are separate jobs: MCPB output needs
  both, while a cofactor already named as its library names it needs only the
  binding. There was no way to express the second, so an imported cofactor's
  library was unreachable. The library and its redox/spin state are chosen from
  lists of what exists rather than typed as a path.

• Deposits record which force fields their parameters require, derived from the
  atom types in the files. The Topology Generator already enforced declared
  prerequisites, but almost nothing declared any.

• Pure inorganic metal clusters can be given hydrogens before parameterization.
  Hydrogen addition covered protein and organic residues only, and reduce has
  no chemistry for a Mo-S-O or Fe-S core, so a cofactor whose resting state
  carries a hydroxo — Mo(=O)(=S)(OH) in a molybdenum cofactor — reached the QM
  model as a bare oxo with the wrong electron count and charge, and there was no
  way to correct it. Editing the Gaussian input by hand is not an alternative:
  it and the model PDB are matched by index, so an atom added to one shifts the
  Seminario force-constant indices and leaves the deposited residue template
  without the hydrogen its charges were fitted with. Offered for every cluster,
  defaulting to no.

• The structure viewer focuses on a cluster while its hydrogen-addition prompt
  is on screen, and re-reads the file once a hydrogen is added so the new atom
  is visible. ViewerCoordinator.refresh_structure() re-serves the current
  path, which show_structure deliberately will not do — it treats a repeated
  path as a no-op, correct until the file is edited in place. It never starts a
  viewer or opens a tab.

• A hydrogen added to an atom with only one bond is now placed at an angle to
  that bond, rotated to the least crowded side, instead of directly opposite it
  — a hydroxo, thiol or amine hydrogen is bent, never linear. The opening angle
  can be set at the prompt.

Changed

• The cofactor prerequisites panel reports what the selected parameter sets
  declare instead of describing their chemistry. It asserted that every
  cofactor needs a protein force field and inferred GAFF2 from residue names,
  while explaining the requirement in terms of a ribitol tail and which bond
  would fail — none of which is knowable about an arbitrary parameter set.

• MCPB step results are stored per site. A single shared record meant step_1
  belonged to whichever site ran last, which cross-wired one site's atom-type
  fingerprint and RESP charge constraint onto another.

• The session editor table labels its status column Status instead of St.

• The parameterization prompt selects sites, and each selected site goes to
  the parameterizer its own category implies. It no longer groups residues:
  that was for a modified amino acid covalently bound to a cofactor, which is
  now expressed by defining the pair as a site in the Redox Site Detector, so
  the combine/separate and category-conflict prompts are gone. A mixed
  selection is ordinary rather than a conflict.

• Selecting metal sites now parameterizes those sites and no others. MCPB
  previously re-detected and parameterized every metal site in the structure
  regardless of the selection, so a structure with two equivalent Fe2S2
  clusters could not have one derived and the other served by the reuse
  transformer. Sites left out are named, with a note that every metal site
  needs parameters before the topology build will succeed.

Fixed

• A Gaussian output that describes a different model than the input beside it
  is refused rather than fitted. Re-running the atom-typing step rebuilds the
  models but Gaussian is run by hand, so a leftover log gives the Hessian or
  ESP of a superseded model — and the result still looks like force constants
  or charges. Compared by content: atom count, elements, then coordinates.

• A withheld metal cluster's own bonds get force constants. Fe-S inside an
  Fe2S2, Mo-S/Mo-O inside a Mo cofactor and the O-H of a hydroxo reached
  neither the coordination bond list nor the prmtop, so nothing derived a
  parameter for them and tLEaP reported one missing for each.

• Metal covalent radii are used when perceiving a cluster's bonds. Molybdenum
  fell to a carbon-like default, so a Mo cofactor was deposited with two of its
  four bonds; Fe-S was passing only because it sat just under that accidental
  cutoff. Metal-metal pairs are excluded — a cluster's metals are bridged
  through their ligands.

• An inferred hydrogen is typed for the atom it is bonded to. Amber names a
  hydrogen after its neighbour, so a metal-bound hydroxo proton typed H took
  the amide hydrogen's van der Waals radius where the hydroxyl convention HO
  is zero.

• RESP scaffolding residues are constrained to their own charge rather than to
  zero. Keeping a real ARG as a gap bridge left its +1 nowhere to go but the
  neighbouring coordinating residues, one of which came out positive.

• Charges are found for metal-site atoms again. The lookup is keyed by
  coordinate tuple, and BioPython's float32 does not compare equal to the
  float64 a resumed session restores, so every site atom summed as zero and a
  −2 site was proposed as 0.

• Atom types are read from a library once. The section test also matched
  atomspertinfo, whose rows repeat every name, so an 84-atom residue was
  reported as 168.

• A supplied library is matched on heavy atoms when the structure has no
  hydrogens, instead of demanding a hand-written mapping for a library that
  fits.

• Caps stay in the chain when the PDB is reordered for tLEaP. Caps bracketing
  an unfilled internal gap were moved to opposite ends of the chain, leaving a
  70 Å peptide bond and the gap they had guarded open.

• The PDB written back from a topology keeps the topology's atom names. They
  were translated to PDB v3 conventions, so a library using older names — O1P
  rather than OP1, say — built once and then failed against the file it had
  just produced.

• Triage categories survive a resume. They lived only on the instance, so a
  resumed run found no organic residues, no clusters and no isolated metals,
  and ticked those steps off as complete.

• Resuming a pending small-molecule or modified-amino-acid parameterization
  works; both dispatched to methods that were never written.

• Force Field Integration names a deposited frcmod for its library entry rather
  than the working directory's site_N.

• Implicit solvation for the large model is now asked for rather than
  inherited. The prompt sat entirely under an anion check, so an anionic large
  model silently adopted the small model's SCRF (announced only after the fact)
  while a neutral or cationic one dropped it with no message at all, leaving the
  two calculations at different levels of theory for no stated reason. It is now
  offered in every case, defaulting to yes whenever the small model was solvated,
  and declining says plainly that the models will differ.

• The Merz-Kollman ESP radius is taken from MCPB.py's own vdwRadiiDict2023
  (Smith et al. JCTC 2023, 19, 2064) rather than from the force-field IOD
  parameters. They are different quantities — the MK value decides how close to
  the nucleus ESP grid points may fall, the IOD value is a Lennard-Jones term —
  and the IOD tables stop at tetravalent, so a Mo(VI) cofactor could not resolve
  one at all and fell back to a generic 1.5 A with no cited source. Force-field
  parameters are unchanged. tools/patch_readradii.py adds the block to
  large_resp.gjf files generated before this, so they need not be rebuilt.

• Checklist state stores numbers as numbers. The serializer recognised only
  Python int/float, and a numpy scalar is neither — np.float32 fails
  isinstance(v, float) and has no __dict__ — so every coordinate read from
  BioPython was written as {"__type__": "str", "value": "-46.078"}. On resume,
  metal reinsertion then handed PDBIO a dict where it wanted a number and the
  step failed with "must be real number, not dict". numpy scalars now serialize
  as plain numbers, and MetalInfo.from_dict unwraps and coerces so a state
  file written before this still resumes.

• Resuming a metal-site run from saved checklist state no longer fails at
  structure recombination with tLEaP error: 'site_id'. Checklist state stores
  objects wrapped as {"__type__", "value"} and uses the dataclass field names,
  while the exporters write a flat dict with three fields spelled differently
  (coordinates for coords, and so on). dict_to_redox_site, the documented
  normalizer for exactly this case, understood only the exported shape; it now
  accepts both.

• A checklist step that reports a failure without raising is recorded as failed
  rather than completed. Structure recombination printed "tLEaP failed" and was
  ticked off, so the following step ran and died on the prepared structure it
  had never produced, hiding the real cause. Handlers signal this by returning
  success: False.

• RESP is constrained to the charge its own ESP was computed with. Per-site
  results are restored from a single shared workspace key, so in a multi-site
  run the charge came from whichever site ran last: a -1 Fe2S2 model was fitted
  against a -3 constraint and RESP spread the missing two electrons over the
  site, reaching +5.6 on a metal. It does not fail on a mismatched total — it
  fits a different molecule. Charge and multiplicity now come from the site's
  own Gaussian log (falling back to its input file), which cannot be another
  site's, and a disagreement with the stored step-1 value is reported.

• The large model's Gaussian input carries the metal van der Waals radii again.
  They were gathered by looking each redox_site.centers entry up in the atom
  assignments, which only works for a lone metal ion, where the center is the
  metal atom. For an organometallic cofactor or a pure cluster the center
  describes the residue instead, so nothing matched and the file was written
  with a bare Pop=MK and no radii block — the Merz-Kollman ESP then sampled
  the metals with Gaussian's defaults. The radii now come from the metal atoms
  themselves, and a site that resolves none says so instead of quietly falling
  back.

• The large model's ESP calculation now defaults to the same level of theory as
  the small model rather than HF/6-31G*. The MCPB.py tutorial states it
  directly: "we used the B3LYP/6-31G* level of theory to perform the
  calculations for both the small and large models". HF/6-31G* is the generic
  RESP convention for organics, and it silently differed from the functional
  just chosen for the small model.

• A hydrogen inside a metal cluster is no longer typed as a metal ligand. Every
  non-metal atom of a pure cluster was given a unique Y* type on the grounds
  that such a residue has no non-core atoms — true until a cluster could carry
  a hydroxo hydrogen, which bonds the oxygen rather than the metal.

• Session replay is strict again: a recorded answer is used only when it is the
  next unconsumed interaction and matches the prompt exactly. The forward scan
  that replaced it could not tell a recorded prompt the run skips apart from the
  same question asked at a different point in the workflow, and leapt for the
  latter — a hydrogen-editor prompt now asked during step 8 matched its
  recording from step 12 sixty-five interactions later, consuming three
  checklist decisions on the way, after which replay ran step 13 while the
  checklist sat at step 9. A prompt that does not match now announces the
  divergence once, falls through to live input, and leaves the position alone,
  so replay resynchronises as soon as the recorded question comes round again.

• Integer and decimal prompts are replayed by question rather than by position.
  IntPrompt and FloatPrompt are not subclasses of Prompt, so patching
  Prompt.ask never covered them: they fell through to the built-in input
  interception, where Rich has already printed the question itself, and every
  numeric answer was recorded with an empty prompt string. Any numeric question
  would then take the next numeric answer in the file, so adding one silently
  shifted the rest — a newly added atom-type offset prompt was answered with a
  Gaussian memory value recorded for something else. Existing logs record these
  answers anonymously and cannot be matched; those prompts now ask rather than
  supply a number from an unrelated question.

• A leftover workflow_state.json no longer blocks replay of a session log.

  The checklist asked "Resume from saved state?" whenever the file was present,
  a question the recorded run never faced and the log therefore cannot answer.
  Replay now starts such a workflow fresh instead of asking. Underneath, an
  unmatched prompt no longer consumes the rest of the log: the replayer scans
  forward to tolerate a recorded prompt the current run does not ask, but it
  now rewinds when nothing matches, so a single unexpected question costs only
  itself rather than ending the replay and dropping every later prompt to live
  input.

• Force Field Integration (checklist step 15) deposits one library entry per
  metal site instead of merging every site into a single entry. A merged entry
  was keyed by one site type, so it could not be reused on a structure carrying
  only one of the sites, and the reuse transformer it emitted matched neither
  site: transformer matching requires every residue in the rename table to be
  present in a single site, which a table spanning two sites never satisfies.
  Site type, redox state and spin state are now asked per site (the previous
  site's answers carry forward as defaults), and reusing one identity for two
  sites is refused rather than silently overwriting the first. Residue naming
  is unchanged: it still runs across all sites so names cannot collide.

• Metal-site parameterization now asks for the formal charge of a withheld
  cluster's non-metal core atoms (the bridging sulfides of an Fe-S cluster, the
  S/O core of a Mo cofactor), not only of its metals. A pure inorganic cluster
  is withheld from the force-field pass as a whole residue, so none of its atoms
  arrive with a charge; the core atoms were left unset and counted as zero, so
  the suggested QM charge came out short by their formal charge. An Fe2S2 site
  with four cysteinates proposed +2 where [Fe2S2(SCys)4]2- is -2. Correcting
  that by inflating a metal's charge was worse than it appeared, because the
  metal's formal charge is also the van der Waals radius key and is stored in
  the deposited library; a charge the radius database cannot resolve is now
  flagged as it is entered.

• The M*/Y* atom-type numbering no longer restarts at M1/Y1 when metal sites
  are parameterized in more than one pass, which would collide with types
  already deposited once every site's files load into one tLEaP session. The
  starting point comes from the workspace within a session and from earlier
  runs' fingerprint files in a fresh one; what was found is shown and can be
  overridden. Re-running a site offers to reuse the names it had before, so
  correcting one site replaces its entry instead of stranding its old types.

• Metal-site nuclearity counts metal atoms rather than metal-bearing residues,
  so an Fe2S2 cluster reads as binuclear instead of mononuclear.

• Metal-site and small-molecule counts in the menu suggestion and the status
  view count sites rather than residues; a site's coordinating ligand no longer
  inflates the total. The option numbers those messages cite now match the
  menu, which had gained an entry without them being updated.

Installers (attached below)

Self-contained installers that include ProPrep, AmberTools 26, MODELLER and all other dependencies. No conda installation is needed.

• ProPrep-1.16.0-MacOSX-arm64.sh (Macs with an Apple chip: M1, M2, M3, M4)
• ProPrep-1.16.0-MacOSX-x86_64.sh (Macs with an Intel chip; built but not yet verified on Intel hardware)
• ProPrep-1.16.0-Linux-x86_64.sh (Linux, including Windows via WSL2)

bash ProPrep-1.16.0-<platform>.sh

MODELLER (structure repair) requires your own free academic license key from https://salilab.org/modeller/registration.html. For this release supply it at install time:

KEY_MODELLER='your_key' bash ProPrep-1.16.0-<platform>.sh

SHA-256:

5d20369e048e3435199f1693bbfbbe00e1a5860e5a303f0655271fbd29a3a863  ProPrep-1.16.0-MacOSX-arm64.sh
63433d26464f13ca2abab7abcd47aef8dc278d68c6d49e5ee7b5f2b4846075d8  ProPrep-1.16.0-MacOSX-x86_64.sh
2ad5ed4313e194809f12709e1dbcf5cbfb952afe7cff55b025292d3f39c6b783  ProPrep-1.16.0-Linux-x86_64.sh

Archived on Zenodo: https://doi.org/10.5281/zenodo.22012972