* add ability to block atoms/bonds from participating in tautomer zones
* be more structured with the atom flag
* response to review
---------
Co-authored-by: = <=>
sortAndUniquifyToTry previously built a parallel vector of (index, string)
pairs, sorted by string, erased duplicates, then rebuilt the original vector
— O(N log N) with one heap allocation per candidate product.
Replace with an erase-remove over a boost::unordered_flat_set<size_t> keyed
on buildProductHash (boost::hash_combine over synthon IDs + reaction ID).
Dedup is now O(N) average with no string allocations on the hot path.
Also switch SearchResults::d_molNames from std::unordered_set<std::string>
to boost::unordered_flat_set<std::string> for the same open-addressing cache
locality benefit during mergeResults.
Perf (42-rxn / 140B-product Freedom space, maxHits=3000, hitStart=1000,
9 queries; vanilla.log → 2unordered_flat_set.log):
Benzene: 6.92s → 5.64s (−19%)
Tolueneish: 6.19s → 5.07s (−18%)
Acetaminophen: 4.50s → 3.63s (−19%)
Allopurinol: 4.41s → 3.94s (−11%)
Theophylline: 4.39s → 3.90s (−11%)
Nicotine: 4.87s → 3.97s (−18%)
Ciprofloxacin: 6.82s → 6.09s (−11%)
Aspirin: 4.51s → 3.42s (−24%)
Metoprolol: 5.11s → 4.07s (−20%)
Total: 48.40s → 40.33s (−17%)
Hit counts and MaxNumResults unchanged across all queries.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
#9287 tagged these install rules with COMPONENT dev, which routes
rdkitpython-config.cmake / rdkitpython-targets.cmake into the dev
component alongside the python-agnostic C++ cmake config. That's
correct progress over the prior Unspecified default, but `dev` is the
wrong group: these files hardcode a single python version via
configure_file substitution (find_dependency(Python3 3.X ...)
and Boost component python3XX/numpy3XX), so they belong with the
python wrappers — themselves python-version-specific — rather than
with python-agnostic dev artifacts.
Change the two affected COMPONENT tags from dev to python.
In Code/Numerics/Optimizer/BFGSOpt.h, the gradient-convergence check
computed
double term = std::max(funcVal * gradScale, 1.0);
...
test /= term;
if (test < gradTol) return 0;
When funcVal (the current energy) is negative, funcVal * gradScale is
negative and std::max clamps the denominator to 1.0. The convergence
test therefore divides the gradient norm by 1 instead of by the
intended |E| * gradScale, which over-tightens the criterion by a factor
of |funcVal * gradScale| whenever |funcVal * gradScale| > 1.
Negative energies are a normal mid-minimization state for force fields
that include stabilizing terms (MMFF94, UFF with charges, AMBER-style
potentials), so this affects realistic workloads: extra BFGS iterations
or, occasionally, hitting MAXITS and returning the "too many iterations"
status when convergence would otherwise have been reached.
The fix is to use |funcVal| in the denominator, matching the pattern
used three lines below ('std::max(fabs(pos[i]), 1.0)') and matching the
intended interpretation as a magnitude.
A new test case 'testBFGSOptimizationNegativeEnergy' in
testOptimizer.cpp minimizes a 2D quadratic whose value is always
negative along the convergence path and verifies the optimizer reaches
the analytic minimum.
git blame attributes the original line to commit e08e0d16d (Nov 2015),
when the optimizer was restructured; the surrounding code does use
absolute values, so this reads as an oversight rather than an
intentional choice.
* add checked iterators
* support checked atom and bond iterators
the idea here is to allow optional checking that the graph is not being
modified while an iterator is active
* ignore new member functions
---------
Co-authored-by: Ric R <ricrogz@gmail.com>
Headers under Code/RDGeneral/hash and the generated cmake package config
files (rdkit-config*.cmake, rdkit-targets*.cmake, rdkitpython-config*.cmake,
rdkitpython-targets*.cmake) are currently installed without an explicit
COMPONENT, so they default to "Unspecified" and cannot be cleanly separated
from runtime artifacts by packagers that use cmake's per-component install
(e.g. -DCMAKE_INSTALL_COMPONENT=dev).
In conda-forge we split the package into librdkit (runtime) and librdkit-dev
(headers + cmake config), with librdkit installing components
"Unspecified base data runtime" and librdkit-dev installing the "dev"
component. With the current upstream tagging, the hash headers and all
cmake config files end up in librdkit instead of librdkit-dev, which both
ships build-time artifacts in a runtime package and leaves librdkit-dev
without the cmake config needed for find_package(RDKit) to work.
This commit tags all five affected install() calls with COMPONENT dev so
per-component installs work correctly. Default (non-component) installs
are unaffected.
The previous `set(CMAKE_EXE_LINKER_FLAGS ...)` replaced the variable
wholesale, which clobbers any toolchain-supplied linker flags. In
particular, conda-forge's clang_osx-64 / clangxx_osx-64 packages set
`-stdlib=libc++ -L${PREFIX}/lib -Wl,-rpath,${PREFIX}/lib` via
`CMAKE_EXE_LINKER_FLAGS`. Losing those flags causes the postgres
extension link to pick up the wrong libc++ and fail to resolve
ABI-tagged symbols on libc++ 19+:
[ 94%] Linking CXX executable rdkit.dylib
Undefined symbols for architecture x86_64:
"VTT for std::__1::basic_stringstream<...>"
"vtable for std::__1::basic_stringbuf<...>"
"vtable for std::__1::basic_stringstream<...>"
"vtable for std::__1::basic_istringstream<...>"
ld: symbol(s) not found for architecture x86_64
The missing symbols carry the `[abi:ne190107]` ABI tag introduced by
libc++ 19+ — references that only resolve against the conda-forge
libc++, not the system one the link was falling back to.
Append to `CMAKE_EXE_LINKER_FLAGS` instead so the toolchain flags
survive. The other rdkit `.dylib`s in the same build are linked via
the standard cmake toolchain path and were never affected.
Verified by building rdkit-postgresql on osx-64 + osx-arm64 via the
conda-forge feedstock (https://github.com/conda-forge/rdkit-feedstock)
with this fix applied as a downstream patch.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* refactor: clean up AAP similarity logic and add type hints
* Refactor AAP similarity implementation for clarity and maintainability
* Refactor AAP similarity implementation for clarity and maintainability
* ran yapf over the code
* Fix CosineSimilarity doc formula and clarify similarity metric names in BitOps.h
* Fix CosineSimilarity formula in BitOps.h and adjust similarity docs
* CIPLabeler performance: Store vector of bonds
CIPLabelling refers to bonds by index over and over again. This
causes a measurable hit in performance in findConfigs() because
we iterate over a bitset of "allowed" bonds. For very large
molecules with many bonds, this can be a rate-limiting step!
This affects many PDB-sized structures.
2J3N goes from 0.7s to 0.25s with this change.
I had another example for which the findBondWithIdx() call was
taking 500ms of a 700ms call (after the performance update
in #9171 was implemented)
* yikes, XXL reserve
thanks, greg
Co-authored-by: Greg Landrum <greg.landrum@gmail.com>
---------
Co-authored-by: Greg Landrum <greg.landrum@gmail.com>
* CIP labeller: Don't calculate auxiliary descriptors unnecessarily
The first 3 rules (the constitutional rules) are pretty easy
to understand. After rule 3, we need to calculate auxiliary
stereo descriptors to break ties.
However, we _were actually_ calculating auxiliary stereodescriptors
for all centers! We should only need to calculate auxiliary
stereocenters for sites that are needed to break ties.
This cost time - it also caused errors if the auxiliary descriptors
needed a graph expansion, because bonds in the digraph might be
pointed in the wrong direction.
Example case PDB ID 4AXM
Before this commit, errored with "Could not calculate parity! Carrier mismatch"
after 14s. After this commit, completes successfully in 0.036s.
Labelled centers all match (for the centers that had labels in
the failure case).
Includes a test that I can imagine breaking with this optimization.
The reference labels are from before this change
* Ensure all "arms" of stereo bonds and atropisomer bonds are expanded
For tetrahedral centers, ranking using the constitutional rules
always expands as far as is needed (but no further). For SP2bond
and atropisomers, if the first side is not resolvable, the
second side is never visited.
If the constitutional rules don't resolve a side, we need to
label the auxiliary centers. It's important to label all
auxiliary centers that _will_ be visited, so we need to know
what centers will be visited.
This commit updates the label() call in SP2 and Atropisomer
bonds to always attempt to label both sides if using the
constitutional rule set.
The constitutional rules are cheap, and if they fail, we
always go on to the full rule set. It is not a savings to skip
the search on the second side if we're going to keep going
anyway!
Includes a test that reproduces Ricardo's example.
This has no measurable effect on performance relative to the
original solution
* If any parts of the center have been seen, label it.
I couldn't make an example hit this, but Ric is totally
theoretically right
* Greg's ranges suggestion #2
Co-authored-by: Greg Landrum <greg.landrum@gmail.com>
* any_of for container search
Co-authored-by: Greg Landrum <greg.landrum@gmail.com>
---------
Co-authored-by: Greg Landrum <greg.landrum@gmail.com>
* fix extended boundary issue (3 mols)
* clang pass
* no change. retrigger CI for failed java test
there's a failing java test that seems to be failing by chance rather than by changes, as it depends on rng. this is just to retrigger the CI pipeline to confirm this
* no change. retrigger the CI (yet again)
* raw strings and removed garbage collector
* SHARED-12256: Add test and change function.
* SHARED-12256: Update to only wrapping changes.
* SHARED-12256: Parameterize tests.
* SHARED-12256: GetPropIfPresent changes.
* Revert "SHARED-12256: GetPropIfPresent changes."
This reverts commit f598f8c161.
* SHARED-12256: Make default the keyword in the boost wrappings.
* SHARED-12256: Overload function instead of using a sentinel.
* SHARED-12256: Extend GetProp changes.
* SHARED-12256: Add entry point for tests and fix tests.
* iterators for random-access MolSuppliers
add optional caching to SDMolSupplier
* add support to SmilesMolSupplier too
There is a lot of duplicate code between the random-access suppliers that would be worth trying to remove
but at the moment it looks like it would require multiple inheritance, and I think we want to avoid that
* add input iterators for ForwardSDMolSupplier()
* throw when calling begin() on a used supplier
* switch to use the spaceship operator
* init() should reset the mol cache
* Make SDMolSupplier and SmilesMolSupplier safe for multi-threaded reads
* add benchmarking
* add TDTMolSupplier support
improved testing
add benchmarks for parallel iteration
optional TBB support
* better const handling, add reverse iterators
doesn't look like const_iterator is possible since getting data from the underlyng supplier object is non-const
* improve docs
more usings
add reverse iterator to TDTMolSupplier
* tests only try execution::par when it is there
* fix typo
* more testing/demo
* remove accidentally added files
* review changes
* add default ctors
* disable a false-positive compiler warning
it is stupid to have to do this
---------
Co-authored-by: = <=>
* Tautomer insensitive hash v2, E/Z and stereocenter-preservation
* Preserve E/Z stereochemistry and stereocenters in TautomerHashv2
Simplify extension logic to better protect stereocenters connected via
single bonds to aromatic systems. Preserve E/Z stereo on exocyclic
double bonds to distinguish geometric isomers (e.g., E/Z hydrazones).
* add helper function to remove duplicated code
* Fix ring info and bond aromaticity handling in MolHash
- Add fastFindRings check in TautomerHashv2 before ring queries
- Set isAromatic consistent with bond type (true for AROMATIC bonds)
- Fix inverted condition in RegioisomerHash
* more consistent hashes regardless of stereo annotation
* check for duplicate atoms/bonds in StereoGroups
* explicit handling of duplicate stereogroup atoms in CTAB and CXSMILES parsers
---------
Co-authored-by: = <=>
Adds a new function MolDraw2D_detail::getSGroupDataLabels() that returns
the text and molecule-coordinate positions of DAT SGroup labels, using
the same placement logic as the drawing code. This allows external
renderers to display SGroup labels consistently with RDKit's placement.
Refactors DrawMol::extractSGroupData() to call getSGroupDataLabels()
internally, eliminating the duplicate FIELDDISP parsing and position
computation logic.
Closes#7829
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* initial ranges support for Atom/Bond iterators.
needs more testing
* support random access
test sort
more testing please
* compiles on windows
* fix size()
more testing
add some benchmarking
* disable benchmarking code by default
* do not allow modifying the graph through the iterators
---------
Co-authored-by: = <=>
MolToInchi has called MolOps::Kekulize(*m, false) for years, but PR #9125 changed the default traversal to canonical=true. That pulls rankFragmentAtoms() and the canonicalization path into the InChI conversion even though the tested InChI outputs stay the same.
Validation:
- rdkit.Chem.UnitTestInchi passed before and after this change on upstream/master (18 tests, OK in both runs).
- No InChI output drift was observed between stock and patched builds on Regress/Data/mols.1000.sdf.gz, rdkit/Chem/test_data/pubchem-hard-set.sdf.gz, or the atom-order regression molecules added in Code/GraphMol/catch_graphmol.cpp.
Performance:
- Release_2026_03_1 Python MolToInchi on Regress/Data/mols.1000.sdf.gz improved from 0.40712s to 0.38871s median (-4.52%).
- Release_2026_03_1 rdinchi MolToInchi on the same dataset improved from 0.39755s to 0.37814s median (-4.88%).
- Release_2026_03_1 standalone C++ MolToInchi on /tmp/mols.1000.sdf improved from 7.66775s to 7.03474s wall time (-8.26%), from 20.57B to 19.04B cycles (-7.46%), and from 121.78M to 114.05M cache misses (-6.35%).
* Configurable legend position (Top/Left/Right/Bottom) and vertical text (GitHub #9023)
- Add LegendPosition enum and legendPosition, legendVerticalText to MolDrawOptions
- Support legend at Top, Left, Right, Bottom; vertical text for Left/Right
- Python: MolDrawOptions.legendPosition, .legendVerticalText; LegendPosition enum
- Python: MolToSVG() wrapper with legend/drawOptions; doc updates for MolToImage
- JSON: legendPosition (string), legendVerticalText (bool) in draw options
- C++ and Python tests; release note and Cartridge.md docs
* MolDraw2D: legend gutter for horizontal side legends; vertical side height fit
- Reserve horizontal gap between molecule and left/right horizontal legends
(scale mol to molWidth-gutter, align toward legend strip).
- Position horizontal side legend by measured text width from partition edge.
- Vertical side legends: iterative scale so n*max_h+(n-1)*gap fits panel.
- Catch: long vertical side legend section.
* Update legend-position tests and review-driven cleanup
Use enum/default wording for legendPosition docs, move the lightweight Python test to Wrap, add regex-based placement checks (including horizontal side and vertical stacking), and refactor extractLegend helpers per style guidance.
* Fix MolDraw2D legend edge cases
* MolDraw2D: review follow-up (legend tests, bounds, DRY Top/Bottom)
* Update no-FT legend test coords
* Address PR review: document constants, remove release-note text, and simplify extra-padding logic
* Speed up tautomer canonicalization by deferring on SSSR calc
* Lazy kekulization for tautomer enumeration
Defer kekulization of tautomers until they are actually needed for
transform matching. This avoids creating kekulized copies for:
1. The initial tautomer (until first iteration)
2. New tautomers that may never be processed (if enumeration ends early)
The Tautomer class now supports lazy initialization of the kekulized
form via getKekulized() method.
Performance improvement: ~7% additional speedup (total ~22-24% from baseline)
* Use count-only substructure matching in tautomer scoring
* Add SubstructMatchCount regression test
* MolStandardize: reduce enumerate overhead
* MolStandardize: avoid per-tautomer ring recomputation
* Atom: cache PeriodicTable pointer in valence calcs
* Atom: reuse PeriodicTable in getEffectiveAtomicNum
* PeriodicTable: add atomic fast path for getTable
* GraphMol: reduce ROMol copy reallocations
* MolStandardize: use quickCopy for per-match product copies
Use RWMol(*kmol, true) in tautomer enumeration to avoid copying properties/bookmarks/conformers for each candidate. This reduces deep-copy overhead without changing chemistry.
* MolStandardize: pre-filter scoring patterns by element/connectivity
For tautomer scoring, pre-compute which SubstructTerms are relevant for
a given input molecule. Since tautomerization only moves H atoms and
changes bond orders (never creates/destroys heavy-atom bonds), patterns
requiring missing elements or connectivity can be skipped for all
tautomers of that molecule.
Two-stage filtering:
1. Element check: skip patterns requiring atoms not in the molecule
2. Connectivity check: skip patterns whose bond-order-agnostic structure
doesn't match the input molecule's connectivity
This reduces the number of VF2 substructure calls per tautomer from 12
to typically 3-5, depending on the molecule's composition.
* MolStandardize: preserve molecule properties for canonical tautomer
Copy molecule properties from the original input to the canonical tautomer
result. Since quickCopy during enumeration skips d_props to avoid overhead,
extended SMILES data like link nodes (LN) was lost. This restores them
on the final result.
* TautomerQuery: preserve molecule properties (e.g. link nodes) in tautomers
TautomerQuery::fromMol() uses TautomerEnumerator::enumerate() which uses
quickCopy for performance. This doesn't copy molecule properties like
_molLinkNodes. Without this fix, XQMol output would lose link node
extensions in the SMILES.
Copy properties from the original query molecule to all enumerated
tautomers before constructing the TautomerQuery. This preserves extended
SMILES data without impacting enumeration performance.
* MolStandardize: use parallel iteration and cache bond lookups
Replace O(n) getAtomWithIdx/getBondWithIdx calls with parallel iteration
over atom/bond ranges in canonicalizeInPlace and enumerate. Cache bond
lookups in setTautomerStereoAndIsoHs to avoid repeated O(n) searches.
* perf: add specialized matchers for simple tautomer scoring patterns
Replace VF2 graph matching with O(n) loops for 6 simple patterns:
- countDoubleOrAromaticBonds: C=O, N=O, P=O patterns
- countMethyls: [CX4H3] methyl groups
- countCarbonDoubleHetero: [C]=[/home/dcvuser/rdkit;Code/GraphMol/MolStandardize/Tautomer.h] aliphatic C=hetero
- countAromaticCarbonExocyclicN: [c]=aromatic C=exocyclic N
Complex patterns (benzoquinone, oxim, guanidine, aci-nitro) still use VF2.
Combined with the pre-filtering optimization, this achieves ~3.7x speedup
(~2500ms vs ~9300ms original) for tautomer canonicalization.
* Fix tautomer canonicalize dropping conformers from quickCopy
quickCopy (RWMol(*mol, true)) skips conformers, so tautomer
enumeration products lose 2D/3D coordinates. This causes InChI
generation to omit the /b (double bond E/Z stereo) layer, since
E/Z is derived from atomic coordinates.
Fix: copy conformers from the original molecule onto the canonical
tautomer after pickCanonical in TautomerEnumerator::canonicalize().
Tests: SMILES-based E/Z check in testTautomer.cpp, molblock-based
conformer preservation check in catch_tests.cpp.
* add test on canonicalize losing stereo
* add regression test for exocyclic C=C tautomer canonicalization
The getTautomerStateKey() pre-filter (commit 2595ef748) can falsely
deduplicate distinct tautomers when their atom-index-ordered state
patterns happen to match, leading canonicalize() to pick the wrong
canonical form for molecules with STEREOTRANS-pinned exocyclic C=C
bonds after RemoveHs.
Test verifies that O=C(CC1=CC2=CC=COC2)NC1=O canonicalizes to the
exocyclic form O=C1CC(=CC2=CC=COC2)C(=O)N1, not the endocyclic form
O=C1C=C(C=C2CC=COC2)C(=O)N1.
Currently expected to FAIL until the state key dedup bug is fixed.
* MolStandardize: expand tautomer connectivity SMARTS
* MolStandardize: scope tautomer pattern enum
* MolStandardize: trim tautomer pattern enum
* MolStandardize: use symmetric ring scoring
* parse templates as smarts
* accept ring templates in SMARTS format
* undo CLAUDE mistake
* rename files
* enable templating for macrocycles
* enable macrocycle templating
* Add test for macrocycle templating
Tests that ring system templates are used only for macrocycles (rings
with size > 8). The test verifies the exact threshold by generating
coordinates with and without templates for rings of size 4-14.
Addresses review feedback on PR #9203.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
The non-terminal bond filter checked len(nb2) > 1 for both atoms,
ignoring nb1 entirely. This could include bonds with a terminal
begin-atom when computing dmax for torsion weights.
