Fingerprints

COSMolKit exposes fixed-length bit vectors plus the source-defined sparse bit and count forms used by AtomPair. The exposed Morgan and MACCS branches are covered by strict RDKit bit-identical parity tests. The source-backed topological, Avalon, Pattern, Layered, AtomPair, and Topological Torsion implementations follow their pinned upstream algorithms and exact maintained-corpus validation. Similarity-shape correlation or structurally similar hashing is not a compatibility claim. The Python Fingerprint object is a sparse view over the binary vector: on_bits() returns the bit indexes whose value is 1. It is not a dense floating-point neural embedding.

Topological Torsion fingerprints

Topological Torsion is a separate fingerprint family from Molecule.topological_fingerprint(). The latter is RDKit’s path/subgraph RDKFingerprintMol algorithm; Topological Torsion enumerates ordered atom paths of a configured length and encodes or hashes their atom invariants. The two names, parameters, and outputs are intentionally not interchangeable.

The modern generator exposes all four RDKit vector forms. Sparse count output retains unfolded 64-bit torsion ids, sparse bit output applies the generator’s bit-domain and count-simulation rules, count output folds counts into fp_size, and bit output returns an explicit Fingerprint of that size:

import cosmolkit

mol = cosmolkit.Molecule.from_smiles("CCCCO")
generator = cosmolkit.get_topological_torsion_generator(
    include_chirality=False,
    torsion_atom_count=4,
    count_simulation=True,
    count_bounds=[1, 2, 4, 8],
    fp_size=2048,
)

sparse_count = generator.get_sparse_count_fingerprint(mol)
sparse_bit = generator.get_sparse_fingerprint(mol)
count = generator.get_count_fingerprint(mol)
bit = generator.get_fingerprint(mol)

print(sparse_count.nonzero_elements())
print(sparse_bit.on_bits())
print(count.nonzero_elements())
print(bit.on_bits())

The live object returned by generator.get_options() controls include_chirality, torsion_atom_count, count_simulation, count_bounds, fp_size, num_bits_per_feature, and only_shortest_paths. Scalar calls additionally accept from_atoms, ignore_atoms, and custom_atom_invariants. from_atoms selects paths by endpoint, while ignore_atoms excludes paths containing any selected atom. only_shortest_paths keeps only source-defined shortest atom paths. Custom invariants replace the default atom-code generator for that call.

The default atom-code generator reads the molecule’s explicit-valence cache, as does RDKit. Molecules returned by the default Molecule.from_smiles() path and by with_hydrogens() are ready for fingerprinting because both operations commit the source-defined explicit-valence state. A molecule parsed with sanitize=False or produced by another operation that invalidates valence state must be sanitized first, for example with mol = mol.sanitize(), unless the call supplies custom_atom_invariants. Missing cache state is reported as a typed exception rather than being recomputed implicitly inside the fingerprint call.

Topological Torsion provenance uses the shared AdditionalOutput container. Allocate only the outputs needed before the call. This family populates atom_to_bits, atom_counts, bit_paths, and atoms_per_bit; bit_info_map belongs to other generator families and remains empty here, matching RDKit. Parity tests still request and compare that empty container so the absence of entries is verified rather than assumed.

output = cosmolkit.AdditionalOutput()
output.allocate_atom_to_bits()
output.allocate_atom_counts()
output.allocate_bit_paths()
output.allocate_atoms_per_bit()

generator.get_fingerprint(mol, additional_output=output)
print(output.atom_to_bits())
print(output.atom_counts())
print(output.bit_paths())
print(output.atoms_per_bit())

Bulk generator methods preserve input order and accept num_threads:

molecules = [
    cosmolkit.Molecule.from_smiles("CCCC"),
    cosmolkit.Molecule.from_smiles("CCCCC"),
]
bits = generator.get_fingerprints(molecules, num_threads=2)

assert bits[0].on_bits() == generator.get_fingerprint(molecules[0]).on_bits()
assert bits[1].on_bits() == generator.get_fingerprint(molecules[1]).on_bits()

The legacy functions are compatibility adapters, not aliases for every modern option:

• get_topological_torsion_fingerprint() returns the historical unfolded sparse-count vector and preserves RDKit’s 2**width - 1 reported-size compatibility behavior.

• get_hashed_topological_torsion_fingerprint() returns folded sparse counts.

• get_hashed_topological_torsion_fingerprint_as_bit_vect() preserves the legacy n_bits_per_entry block sizing and its distinct four-slot [1, 2, 4, 8] versus non-four-slot threshold rule.

All modern and legacy adapters terminate in the same Rust atom-code, path, torsion-code/hash, and shared fingerprint-vector core. They do not call RDKit at runtime. Invalid sizes, unsupported source-undefined shifts, malformed JSON, bad selections, and incompatible custom invariants fail visibly as typed Python exceptions instead of selecting another fingerprint algorithm. Rust exposes the corresponding FingerprintError; Rust MoleculeBatch conveniences additionally report computation failures as indexed BatchValidationError records while retaining existing invalid inputs as None in their original positions.

The exact parity boundary is RDKit 2026.03.1 at revision 351f8f378f8ad6bbd517980c38896e66bf907af8. Focused fixtures cover modern and legacy branches, provenance, JSON, selections, chirality, paths, counts, collisions, and errors. The maintained CI matrix compares all 5,000 molecules across nine source-meaningful profiles. The complete ChEMBL 37 audit compares all 2,897,804 mutually parseable records across 36 vector and eight provenance outputs, totaling 127,503,376 exact comparisons. Neither boundary uses sampling or tolerance. This claim applies to the documented, modeled input state; it is not a claim that unrelated Atom Pair fingerprints or all RDKit fingerprint families are implemented.

Source-backed topological and Avalon fingerprints

Molecule.topological_fingerprint() and Molecule.avalon_fingerprint() execute source-backed Rust implementations. The maintained topological matrix is exact across 5,000 rows and 14 profiles; the Avalon matrix is exact across 5,000 rows and 23 profiles.

The two APIs are source-backed and return a fresh explicit bit vector:

• topological_fingerprint() requires the complete RDKit RDKFingerprintMol/RDKitFP generator behavior, including branched-path enumeration, source random-bit generation, density folding, atom invariants, and the exposed path and atom-selection parameters.

• avalon_fingerprint() follows the complete Avalon/reaccs bit-vector path, including bitFlags, isQuery, hydrogen handling, aromaticity passes, and byte-rounded vector semantics. resetVect is an internal adapter detail and is not exposed on COSMolKit’s value-returning API.

The focused option fixtures and maintained corpus are exact source comparisons. Similarity correlation, partial bit agreement, or a heuristic replacement is not an acceptance condition.

mol = Molecule.from_smiles("c1ccccc1O")

topological = mol.topological_fingerprint(
    min_path=1,
    max_path=7,
    fp_size=2048,
    num_bits_per_feature=2,
)
avalon = mol.avalon_fingerprint(
    n_bits=512,
    is_query=False,
    bit_flags=0xF07FFF,
)

print(topological.on_bits())
print(avalon.on_bits())

Layered fingerprints

Molecule.fingerprint_layered() implements RDKit’s Layered fingerprint algorithm version 0.7.0. RDKit labels this algorithm experimental; COSMolKit preserves that upstream metadata even though the exposed parameter and result types are ordinary project-native APIs. The six active source layer bits are:

Flag	Layer	Encoded state
0x01	topology	path-local atom degree and bond-neighbor topology
0x02	bond order	normalized bond type plus topology
0x04	atom type	atomic numbers plus topology
0x08	ring presence	sparse ring-bond membership
0x10	ring size	minimum source SSSR ring size
0x20	aromaticity	query-aware endpoint aromaticity plus topology

0x07 is the source substructure prefix and 0x3f selects all active layers. The default 0xffffffff retains all ten source flag slots; high bits have no encoder in RDKit and therefore add no components. They are accepted, not repurposed as COSMolKit extensions.

Paths are bond paths with inclusive min_path and max_path bounds. The defaults are 1 through 7, 2,048 output bits, and branched enumeration. branched_paths=False selects linear bond paths. from_atoms=None means no root restriction. An explicit empty list is different: it is a present empty root selection and returns an empty fingerprint. Duplicate and reordered roots retain the source aggregation order.

import cosmolkit

molecule = cosmolkit.Molecule.from_smiles("c1ccccc1O")
substructure = molecule.fingerprint_layered(
    layers=0x07,
    min_path=1,
    max_path=7,
    fp_size=2048,
)

even_mask = cosmolkit.Fingerprint.from_on_bits(257, range(0, 257, 2))
counted = molecule.fingerprint_layered_with_output(
    layers=0x3F,
    min_path=2,
    max_path=4,
    fp_size=257,
    atom_counts=[10] * molecule.num_atoms(),
    set_only_bits=even_mask,
    branched_paths=True,
    from_atoms=[0],
)

print(substructure.on_bits())
print(counted.fingerprint().on_bits())
print(counted.atom_counts())

set_only_bits must have exactly fp_size bits. It gates projected bits before insertion and before counting. A supplied atom_counts vector is a seed, not a zeroed output buffer; its length must be at least the molecule atom count. Every atom in a path is incremented once when any layer projection for that path passes the mask, even when several layers set bits or bit collisions occur. Omitting counts makes atom_counts() return None.

Invalid zero/reversed path bounds, zero width, short count vectors, width-mismatched masks, out-of-range roots, and ring-preparation failures are reported as typed exceptions. The operation is read-only: repeated, interleaved, batch, and concurrent calls share the same Rust scalar core and do not mutate molecule topology, properties, coordinates, or derived caches.

Pinned RDKit contains one upstream defect in the unrooted linear call: it requests atom-index paths and later consumes them as bond indices, which can terminate the process for ordinary acyclic molecules. COSMolKit does not copy that crash. It applies the source header’s documented bond-path semantics, the same semantics used by the valid rooted linear branch. This is an explicit process-safety compatibility difference, not a chemistry fallback.

Pattern fingerprints

Molecule.pattern_fingerprint() reproduces RDKit’s ordinary-molecule Pattern fingerprint using the fixed, source-ordered table of 13 SMARTS queries. The default result has 2,048 bits. tautomeric=True enables the source’s tautomer-aware structural hash for single, double, and aromatic bond states:

mol = Molecule.from_smiles("c1ccccc1O")

pattern = mol.pattern_fingerprint()
tautomeric_pattern = mol.pattern_fingerprint(
    n_bits=2048,
    tautomeric=True,
)

print(pattern.on_bits())
print(tautomeric_pattern.on_bits())

The call is value-style and does not mutate the molecule. Query-bearing molecules are supported and follow the source’s Pattern-specific atom and bond suppression rules. The compiled SMARTS table is shared across repeated, concurrent, scalar, and batch calls; Pattern calls can be interleaved with the other fingerprint families without shared option state.

n_bits must be greater than zero. SMARTS-compilation and substructure matching failures remain typed exceptions rather than selecting another fingerprint algorithm. RDKit labels Pattern fingerprint version 1.0.0 experimental; COSMolKit preserves that upstream metadata while its documented ordinary-molecule boundary is tested exactly against the pinned source.

The pinned RDKit ordinary overload accepts atomCounts and setOnlyBits but only validates their sizes; their values are otherwise unused and unchanged. COSMolKit intentionally omits those inert arguments instead of presenting them as functional options. RDKit’s separate MolBundle overload intersects member fingerprints. COSMolKit does not expose it because an ordered MoleculeBatch returns one result per record and is not a bundle.

The committed focused, small, and 5,000-row matrices cover query inputs, tautomer mode, boundary and collision-prone widths, inert-argument behavior, and exact validation errors. The complete ChEMBL 37 audit compares ten full profiles across all 2,897,804 mutually parseable records: 28,978,040 exact vectors with zero mismatch.

AtomPair fingerprints

Molecule exposes the four source generator result forms through one project-native implementation:

• fingerprint_atom_pair_sparse_count() returns the source-width count map;

• fingerprint_atom_pair_count() returns the folded count map;

• fingerprint_atom_pair_sparse_bits() returns the source-width bit set;

• fingerprint_atom_pair() returns the fixed-width explicit bit vector.

The shared parameter surface covers 2D topological or 3D conformer distances, minimum and maximum distance, chirality, count simulation and bounds, num_bits_per_feature, from_atoms, ignore_atoms, conformer selection, and custom atom invariants. A 3D call requires a conformer and sets use_2d=False explicitly.

mol = Molecule.from_smiles("CCCO")

explicit = mol.fingerprint_atom_pair(n_bits=2048)
sparse_count = mol.fingerprint_atom_pair_sparse_count()
folded_count = mol.fingerprint_atom_pair_count(n_bits=2048)
sparse_bits = mol.fingerprint_atom_pair_sparse_bits()

print(explicit.on_bits())
print(sparse_count.nonzero_elements())
print(folded_count.nonzero_elements())
print(sparse_bits.on_bits())

fingerprint_atom_pair_with_output() additionally returns exact atom counts, atom-to-bit rows, endpoint-pair bit information, and atoms-per-bit provenance:

result = mol.fingerprint_atom_pair_with_output()
output = result.additional_output()

print(result.fingerprint().on_bits())
print(output.atom_counts())
print(output.atom_to_bits())
print(output.bit_info_map())
print(output.atoms_per_bit())

The complete ChEMBL 37 validation compares 40 vectors and one full provenance output for every one of the 2,897,804 mutually parseable records: 118,809,964 exact comparisons with zero mismatch. The maintained 5,000-row, ten-profile matrix remains the committed continuous regression gate.

Topological provenance

topological_fingerprint_with_output() returns a TopologicalFingerprintResult. Request atom_bits and/or bit_info to receive the matching source provenance outputs:

result = mol.topological_fingerprint_with_output(
    fp_size=2048,
    atom_bits=True,
    bit_info=True,
)

print(result.fingerprint().on_bits())
print(result.atom_bits())
print(result.bit_info())

Single Molecules

from cosmolkit import Molecule

mol = Molecule.from_smiles("c1ccccc1O")
fp = mol.fingerprint_morgan(radius=2, n_bits=2048)

print(fp.n_bits())
print(fp.on_bits())

Tanimoto similarity is computed directly on Fingerprint values:

phenol = Molecule.from_smiles("c1ccccc1O").fingerprint_morgan()
benzene = Molecule.from_smiles("c1ccccc1").fingerprint_morgan()

print(phenol.tanimoto(benzene))

Additional Output

fingerprint_morgan_with_output() returns a MorganFingerprintResult with the fingerprint and experimental provenance data:

result = mol.fingerprint_morgan_with_output(radius=2, n_bits=2048)
output = result.additional_output()

print(result.fingerprint().on_bits())
print(output.atom_counts())
print(output.atom_to_bits())
print(output.bit_info_map())
print(output.atoms_per_bit())

Supported Parameters

The Python binding exposes the source-backed Morgan generator branches covered by exact RDKit bit parity:

• radius and n_bits

• include_chirality and use_bond_types

• count_simulation and count_bounds

• only_nonzero_invariants

• include_redundant_environments

• from_atoms and ignore_atoms

• custom_atom_invariants and custom_bond_invariants

• atom_invariants_generator="connectivity" | "morgan" | "feature" | "fcfp"

• atom_invariants_include_ring_membership

• bond_invariants_generator="morgan" | "default" | "bond"

• bond_invariants_use_bond_types

• bond_invariants_use_chirality

• num_bits_per_feature

The list is the tested boundary, not a claim that every RDKit fingerprint generator or every input-state preparation branch is implemented. Unsupported branches propagate an error; they do not silently fall back to another fingerprint algorithm.

Batch Fingerprints

MoleculeBatch exposes matching batch APIs. Invalid records kept with errors="keep" produce None in the corresponding output position.

from cosmolkit import MoleculeBatch

batch = MoleculeBatch.from_smiles_list(
    ["CCO", "not-smiles", "CCCO"],
    errors="keep",
).with_parallel_jobs(8)
fps = batch.fingerprint_morgan_list(n_bits=2048)
atom_pairs = batch.fingerprint_atom_pair_list(
    n_bits=2048,
    n_jobs=8,
    progress_bar=False,
)
layered = batch.fingerprint_layered_list(
    layers=0x3F,
    min_path=1,
    max_path=7,
    fp_size=2048,
    n_jobs=8,
    progress_bar=False,
)

print([fp.on_bits() if fp is not None else None for fp in fps])
print([fp.on_bits() if fp is not None else None for fp in atom_pairs])
print([fp.on_bits() if fp is not None else None for fp in layered])