smsd 7.0.0

Fast substructure search, MCS, circular fingerprints, and molecular similarity with tautomer awareness and GPU acceleration

SMSD Pro for Python

Python bindings for SMSD native graph matching, including substructure search, maximum common substructure (MCS), fingerprints, and molecular similarity. RDKit and CDK are not required for the core SMSD path.

Benchmark (Dalke NN, 1,000 pairs): 5x faster than RDKit FindMCS, finds larger MCS on 21% of pairs, zero timeouts. See full results.

Install

pip install smsd

Supported CPython versions: 3.9 through the latest stable release series. Current default test target: Python 3.12. The native SMSD path is CPU-first with optional GPU acceleration. RDKit remains optional for interop and depiction rather than a core dependency.

Quick Start

import smsd

# Substructure search
assert smsd.is_substructure(
    smsd.parse_smiles("c1ccccc1"),     # benzene
    smsd.parse_smiles("c1ccc(O)cc1"))  # phenol

# Maximum Common Substructure
mcs = smsd.find_mcs("c1ccccc1", "c1ccc2ccccc2c1")
print(f"MCS: {len(mcs)} atoms")  # 6

# Tautomer-aware MCS
mcs = smsd.find_mcs("CC(=O)C", "CC(O)=C", tautomer_aware=True)

# Circular fingerprint (ECFP4)
ecfp4 = smsd.fingerprint_from_smiles("c1ccccc1", radius=2, fp_size=2048)

# Similarity
sim = smsd.similarity("c1ccccc1", "c1ccc(O)cc1")

Features

Search & Matching

Feature	Description
Substructure search	VF2++ with 3-level NLF pruning, GPU-accelerated domain init
MCS	11-level funnel: chain/tree DP → greedy → McSplit → BK → McGregor
SMARTS matching	Full SMARTS support including X (total connectivity), D (degree), v (valence), R (ring count), r (ring size), x (ring connectivity), / \ (E/Z stereo), $() (recursive), logical AND/OR/NOT
Tautomer matching	30 transforms with pKa-informed weights, 6 solvents, pH-sensitive
Tautomer validation	validate_tautomer_consistency() — proton conservation check
CIP R/S/E/Z	assign_rs(), assign_ez() — full digraph-based stereo descriptors (IUPAC 2013)
MCS SMILES	find_mcs_smiles() — extract MCS as canonical SMILES string
findAllMCS	find_all_mcs() — top-N MCS enumeration with canonical dedup
SMARTS MCS	find_mcs_smarts() — largest substructure matching a SMARTS pattern
R-group decomposition	decompose_r_groups() — scaffold + R-group extraction

Fingerprints

Type	Description
Circular ECFP	Tautomer-aware structural invariants, configurable radius (2=ECFP4, 3=ECFP6, -1=whole molecule)
Circular FCFP	Pharmacophoric invariants (H-bond donor/acceptor, ionisable, aromatic, hydrophobic)
Count-based ECFP/FCFP	ecfp_counts() / fcfp_counts() — superior to binary for ML
Topological Torsion	topological_torsion() — 4-atom path fingerprint (SOTA on peptides)
Path fingerprint	Graph-aware DFS path enumeration, tautomer-invariant
MCS fingerprint	MCS-aware, uses chemical matching rules for path compatibility
Similarity metrics	overlapCoefficient(), dice(), cosine(), soergel() — binary + count-vector
Format conversions	to_hex(), to_binary_string(), from_hex() — for database storage and REST APIs
Subset check	fingerprint_subset() — fast substructure pre-screening

Infrastructure

Feature	Description
MatchResult	mcs_result() — structured result: size, mapping, overlapCoefficient
RDKit interop	mcs_rdkit_native(), batch_mcs_rdkit(), from_rdkit() with correct indices
Similarity screening	RASCAL O(V+E) upper bound for fast pre-filtering
Lenient parser	Best-effort recovery from malformed SMILES
Batch operations	OpenMP-parallel batch_substructure(), batch_mcs(), batch_mcs_rdkit()
Adaptive timeout	min(30s, 500+n1*n2*2) based on molecule size
GPU acceleration	CUDA + Apple Metal for domain init and RASCAL screening
Force-directed layout	force_directed_layout() for bond-crossing minimisation
SMACOF stress majorisation	stress_majorisation() for optimal 2D embedding
Scaffold templates	match_template() for 10 pre-computed common scaffolds
Reaction-aware MCS	find_mcs(..., reaction_aware=True) or map_reaction_aware() post-filter
Ring perception	compute_sssr(), layout_sssr() — clean SSSR APIs

Performance

Benchmarked alongside RDKit 2025.09.2 on the same machine, same Python process. Both toolkits excel at different tasks — use whichever fits your workflow, or both together.

Pair	SMSD	RDKit	Notes
Morphine / Codeine	79 us	579 ms	Complex ring system
Coronene self-match	6 us	712 us	Symmetric PAH
Caffeine / Theophylline	17 us	373 us	N-methyl difference
PEG-12 / PEG-16	39 us	2.1 ms	Linear polymer

Full data: benchmarks/results_python.tsv

Circular Fingerprint (Novel)

Tautomer-aware Morgan/ECFP — includes tautomer class in the atom invariant, so tautomeric forms of the same molecule produce more similar fingerprints.

ECFP vs FCFP: SMSD supports both fingerprint types (Rogers & Hahn 2010):

• ECFP (Extended Connectivity): atom invariant = atomic number, degree, charge, ring, aromaticity, tautomer class. Best for structural similarity.
• FCFP (Functional Class): atom invariant = pharmacophoric features (H-bond donor/acceptor, positive/negative ionisable, aromatic, hydrophobic). Best for activity-based similarity and SAR.

Name	Radius	Type	SMSD call
ECFP2	1	Structural	circular_fingerprint(mol, radius=1)
ECFP4	2	Structural	circular_fingerprint(mol, radius=2)
ECFP6	3	Structural	circular_fingerprint(mol, radius=3)
FCFP2	1	Pharmacophoric	circular_fingerprint(mol, radius=1, mode="fcfp")
FCFP4	2	Pharmacophoric	circular_fingerprint(mol, radius=2, mode="fcfp")
FCFP6	3	Pharmacophoric	circular_fingerprint(mol, radius=3, mode="fcfp")
Whole	-1	Either	circular_fingerprint(mol, radius=-1)

import smsd

# ECFP4 (structural, recommended default)
ecfp4 = smsd.fingerprint_from_smiles("c1ccccc1", radius=2, fp_size=2048)

# FCFP4 (pharmacophoric — H-bond donors/acceptors, ionisable, aromatic, hydrophobic)
fcfp4 = smsd.fingerprint_from_smiles("c1ccccc1", radius=2, fp_size=2048, mode="fcfp")

# ECFP6 (radius 3, captures larger environments)
ecfp6 = smsd.fingerprint_from_smiles("c1ccccc1", radius=3, fp_size=2048)

# ECFP2 (radius 1, fastest, less discriminating)
ecfp2 = smsd.fingerprint_from_smiles("c1ccccc1", radius=1, fp_size=2048)

# Whole molecule (radius -1 = expand until convergence)
whole = smsd.fingerprint_from_smiles("c1ccccc1", radius=-1, fp_size=2048)

# Tanimoto similarity (works with any fingerprint type)
sim = smsd.overlap_coefficient(
    smsd.fingerprint_from_smiles("c1ccccc1", radius=2),
    smsd.fingerprint_from_smiles("c1ccc(O)cc1", radius=2))

Using with RDKit

SMSD works standalone or alongside RDKit. Use RDKit for parsing and drawing, SMSD for fast matching:

from rdkit import Chem
import smsd

mol1 = Chem.MolFromSmiles("c1ccccc1")
mol2 = Chem.MolFromSmiles("c1ccc(O)cc1")

# MCS with RDKit molecules
result = smsd.mcs_rdkit(mol1, mol2)

# Depict MCS with highlighted atoms (works in Jupyter)
img = smsd.depict_mcs("c1ccccc1", "c1ccc(O)cc1")

# Export to SDF with the native writer
smsd.export_sdf([smsd.parse_smiles(s) for s in ["CCO", "c1ccccc1"]], "output.sdf")

# Convert between SMSD and RDKit
g = smsd.from_rdkit(mol1)
rdmol = smsd.to_rdkit(g)

RDKit is optional — pip install smsd works without it.

Solvent-Aware Tautomer Matching

import smsd

opts = smsd.ChemOptions.tautomer_profile()
opts.solvent = smsd.Solvent.DMSO       # adjust for DMSO
opts.with_ph(5.0)                       # adjust for pH 5.0

mcs = smsd.find_mcs("CC(=O)C", "CC(O)=C", chem=opts)

Supported solvents: AQUEOUS, DMSO, METHANOL, CHLOROFORM, ACETONITRILE, DIETHYL_ETHER

Native MOL/SDF I/O

import smsd

g = smsd.read_molfile("input.mol")
mol_block = smsd.write_mol_block(g)
mol_block_v3000 = smsd.write_mol_block_v3000(g)
sdf_record = smsd.write_sdf_record(g)

smsd.write_molfile(g, "out_v2000.mol")
smsd.write_molfile(g, "out_v3000.mol", v3000=True)
smsd.write_molfile(g, "out.sdf", sdf=True)

The native writer preserves practical chemistry metadata in 7.0.0:

• names, comments, and SDF properties
• charges, isotopes, atom classes, and atom maps
• R#/R<n> plus M RGP
• practical V2000/V3000 stereo round-trip

Platform & GPU Support

SMSD automatically dispatches to the best available compute backend:

Platform	CPU	GPU
macOS (Apple Silicon)	OpenMP	Metal (zero-copy unified memory)
macOS (Intel)	OpenMP	CPU fallback
Linux	OpenMP	CUDA (if available)
Windows	OpenMP	CUDA (if available)

import smsd

# Check GPU availability
if smsd.gpu_is_available():
    print(smsd.gpu_device_info())
    # e.g. "Metal GPU: Apple M2 Pro [OpenMP 5.0, 10 threads]"
    # e.g. "GPU: Tesla T4 [OpenMP 4.5, 8 threads]"

# GPU is used automatically for:
# - Domain initialization in VF2++ substructure search
# - RASCAL batch screening
# - NLF histogram computation

# CPU fallback is seamless — no code changes needed
results = smsd.batch_substructure(query, targets)  # uses GPU if available

API Reference

Core Functions

# Parsing
mol = smsd.parse_smiles("c1ccccc1")
smi = smsd.to_smiles(mol)

# Substructure
smsd.is_substructure(query, target)
mapping = smsd.find_substructure(query, target)

# MCS
mapping = smsd.find_mcs(mol1, mol2)
mapping = smsd.find_mcs("SMILES1", "SMILES2", tautomer_aware=True)

# Similarity
sim = smsd.similarity("SMILES1", "SMILES2")
ub = smsd.similarity_upper_bound(mol1, mol2)
hits = smsd.screen_targets(query, library, threshold=0.5)

# Fingerprints
fp = smsd.path_fingerprint(mol, path_length=7, fp_size=2048)
fp = smsd.circular_fingerprint(mol, radius=2, fp_size=2048)           # ECFP4
fp = smsd.circular_fingerprint(mol, radius=2, fp_size=2048, mode="fcfp")  # FCFP4
sim = smsd.overlap_coefficient(fp1, fp2)
ok = smsd.fingerprint_subset(query_fp, target_fp)

# Format conversions (database storage, REST APIs)
hex_str = smsd.to_hex(fp, fp_size=2048)
fp_back = smsd.from_hex(hex_str)
bits    = smsd.to_binary_string(fp, fp_size=2048)

# Batch (OpenMP parallel)
results = smsd.batch_substructure(query, targets)
results = smsd.batch_mcs(query, targets)
fps     = smsd.batch_fingerprint(mols, path_length=7, fp_size=2048)
hits    = smsd.batch_fingerprint_screen(query_fp, target_fps)

# RASCAL pre-screen + exact MCS in one call
matches = smsd.screen_and_match(query, targets, threshold=0.5)

# Batch find substructure with atom-atom mappings (v7.0.0)
mappings = smsd.batch_find_substructure(query, targets)

# TargetCorpus — prewarm once, query many times (v7.0.0)
corpus = smsd.TargetCorpus.from_smiles(["c1ccccc1", "c1ccc(O)cc1", "CCO"])
corpus.prewarm()
hits = corpus.substructure(smsd.parse_smiles("c1ccccc1"))
sizes = corpus.mcs_size(smsd.parse_smiles("c1ccccc1"))
passing = corpus.screen(smsd.parse_smiles("c1ccccc1"), threshold=0.5)

# GPU
smsd.gpu_is_available()
smsd.gpu_device_info()

Configuration

opts = smsd.ChemOptions()
opts.match_atom_type = True
opts.tautomer_aware = True
opts.ring_fusion_mode = smsd.RingFusionMode.STRICT
opts.match_bond_order = smsd.BondOrderMode.LOOSE

# Profiles
opts = smsd.ChemOptions.tautomer_profile()
opts = smsd.ChemOptions.profile("strict")

Complementary Strengths

SMSD is designed to work alongside existing toolkits, not replace them. Each toolkit brings unique strengths to the cheminformatics ecosystem:

Capability	SMSD Pro	RDKit	CDK
Tautomer-aware circular FP	Yes	—	—
pH/solvent-sensitive matching	Yes	—	—
Multi-level MCS pipeline	11 levels	FMCS	MCSPlus
GPU acceleration	CUDA + Metal	—	—
Descriptor calculation	—	Extensive	Extensive
Reaction handling	Basic	Comprehensive	Comprehensive
3D conformers	—	Yes	Yes
Header-only C++	Yes	—	—

Recommended workflow: Use RDKit or CDK for parsing, descriptors, and 3D — use SMSD for MCS and substructure matching.

Also Available

• Java: com.bioinceptionlabs:smsd:7.0.0 on Maven Central
• C++: Header-only, zero dependencies — GitHub

Citation

If you use SMSD Pro in your research, please cite:

Rahman SA. SMSD Pro: Coverage-Driven, Tautomer-Aware Maximum Common Substructure Search. ChemRxiv, 2025. DOI: 10.26434/chemrxiv.15001534

For the original SMSD toolkit, please also cite:

Rahman SA, Bashton M, Holliday GL, Schrader R, Thornton JM. Small Molecule Subgraph Detector (SMSD) toolkit. Journal of Cheminformatics, 1:12, 2009. DOI: 10.1186/1758-2946-1-12

A machine-readable CITATION.cff is available for automated citation tools.

License

Apache 2.0 — Copyright (c) 2018-2026 Syed Asad Rahman, BioInception PVT LTD. See NOTICE for attribution, trademark, and novel algorithm terms.