smsd 6.1.1

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

SMSD Pro — Molecular Graph Matching for Python

Fast substructure search, maximum common substructure (MCS), circular fingerprints, and molecular similarity — with optional tautomer awareness and GPU acceleration. No RDKit or CDK required.

Install

pip install smsd

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.mcs("c1ccccc1", "c1ccc2ccccc2c1")
print(f"MCS: {len(mcs)} atoms")  # 6

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

# Circular fingerprint (ECFP4, tautomer-aware)
ecfp4 = smsd.circular_fingerprint("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	15 transforms with pKa-informed weights, 6 solvents, pH-sensitive
Tautomer validation	validate_tautomer_consistency() — proton conservation check

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)
Path fingerprint	Graph-aware DFS path enumeration, tautomer-invariant
MCS fingerprint	MCS-aware, uses chemical matching rules for path compatibility
Format conversions	to_hex(), to_binary_string(), from_hex() — for database storage and REST APIs
Tanimoto similarity	Works across all fingerprint types
Subset check	fingerprint_subset() — fast substructure pre-screening

Infrastructure

Feature	Description
Similarity screening	RASCAL O(V+E) upper bound for fast pre-filtering
RDKit interop	from_rdkit(), to_rdkit(), depict_mcs(), export_sdf()
Lenient parser	Best-effort recovery from malformed SMILES
Batch operations	OpenMP-parallel batch_substructure(), batch_mcs()
GPU acceleration	CUDA + Apple Metal for domain init and RASCAL screening

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	51 us	569 ms	Complex ring system
Coronene self-match	6 us	780 us	Symmetric PAH
Caffeine / Theophylline	17 us	564 us	N-methyl difference
PEG-12 / PEG-16	1.6 ms	2.2 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.circular_fingerprint("c1ccccc1", radius=2, fp_size=2048)

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

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

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

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

# Tanimoto similarity (works with any fingerprint type)
sim = smsd.tanimoto(
    smsd.circular_fingerprint("c1ccccc1", radius=2),
    smsd.circular_fingerprint("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
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.mcs("CC(=O)C", "CC(O)=C", chem=opts)

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

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.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.tanimoto(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)

# 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:6.0.0 on Maven Central
• C++: Header-only, zero dependencies — GitHub

Citation

If you use SMSD in your research, please 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.