chematic vs RDKit — Detailed Comparison
This page gives a direct comparison between chematic and RDKit for teams evaluating which library to use.
TL;DR
| chematic | RDKit | |
|---|---|---|
| Install | pip install chematic |
conda / cmake required |
| Browser / WASM | Yes — 719 KB | No |
| C++ dependency | None (default) | Required |
| Batch fingerprint speed | ~78 µs/mol (2–3× faster, diverse corpus) | ~160–235 µs/mol |
| AI agent integration | MCP server built-in | None |
| Ecosystem maturity | Growing (2024–) | Established (2006–) |
Descriptor accuracy caveat: Mol.descriptors() returns 194 values, but bulk (4,999-molecule) RDKit-agreement testing currently covers a subset — the physicochemical core (MW/HBA/HBD/TPSA/LogP/MR/Fsp3/ring & stereocenter counts/etc.) is a verified 100% (or near-100%) match. A larger set of descriptors that are named to match RDKit 1:1 (Kappa/HallKierAlpha/BertzCT/BalabanJ/BCUT2D/VSA descriptor families/MQN/SA Score) were found to diverge substantially once measured at corpus scale. See tasks/descriptor_validation_coverage.md for the full per-descriptor breakdown.
1. Infrastructure
Installation
| chematic | RDKit | |
|---|---|---|
| Python | pip install chematic |
conda install -c conda-forge rdkit |
| C++ compiler | Not required | Required (Boost, CMake) |
| Docker image delta | ~4 MB | ~200 MB+ |
| GitHub Actions | pip install chematic |
Separate conda setup step |
| Cloudflare Workers | Yes | No |
| AWS Lambda | Yes | Difficult (binary size) |
| Embedded / no-std | Partial | No |
WASM deployment
| Library | Bundle size | Build toolchain |
|---|---|---|
| chematic | 719 KB gzip | wasm-pack build only |
| RDKit.js | ~30 MB | Emscripten SDK + cmake |
| Indigo WASM | ~40 MB | Emscripten SDK + cmake |
chematic compiles to wasm32-unknown-unknown natively — no Emscripten, no cmake, no clang.
2. Performance
All measurements: Python 3.13.6, Apple M4, chematic v0.4.29, RDKit 2026.03.3 (2026-07-17;
see benchmarks/2026-07-17.md for full methodology). Import
time and SMILES parse throughput below were not remeasured this cycle.
Import time (cold process)
| chematic | RDKit | |
|---|---|---|
import only |
~35 ms | ~400 ms |
import + first parse |
~38 ms | ~430 ms |
| Speedup | ~11× | baseline |
Measured by spawning a fresh subprocess per sample (module cache excluded).
Reproduce:
SMILES parse throughput
| N molecules | chematic | RDKit | Speedup |
|---|---|---|---|
| 1,000 | ~1 ms | ~10 ms | ~10× |
| 5,000 | ~5 ms | ~50 ms | ~10× |
| 10,000 | ~10 ms | ~100 ms | ~10× |
Per-molecule: ~1 µs/mol (chematic) vs ~10 µs/mol (RDKit).
Reproduce:
ECFP4 fingerprint generation (batch)
Small repeated fixture (same 20-molecule set scripts/benchmark_vs_rdkit.py has always used):
| N molecules | chematic (bulk.ecfp4) |
RDKit (Python loop) | Speedup |
|---|---|---|---|
| 100 | 1.4 ms | 8 ms | 6.1× |
| 1,000 | 10 ms | 83 ms | 8.0× |
| 10,000 | 126 ms | 839 ms | 6.7× |
On a large, structurally diverse 5,000-molecule ChEMBL corpus the margin is narrower:
~78 µs/mol (chematic) vs ~160–235 µs/mol (RDKit), ~2–3×. chematic uses Rayon for
parallel batch processing across all CPU cores. Neither number reproduces the
previously-reported 3.6 µs/mol / 5–14× figures, even on the identical small-fixture
script — see benchmarks/2026-07-17.md for what was
measured and what remains an open question.
Reproduce:
Where RDKit is faster or better
| Task | RDKit advantage |
|---|---|
| Publication-quality 3D structures | ETKDGv3 with ML torsion corrections (chematic: good for screening) |
| Exotic molecule handling | 20 years of edge-case fixes |
| Large SDF file streaming | Optimized C++ reader |
3. Feature coverage
chematic has, RDKit does not
| Feature | chematic | RDKit |
|---|---|---|
| Native WASM (no Emscripten) | Yes | No |
| MCP server (AI agent API) | 20 tools (stdio only) | None |
| pKa prediction (built-in) | 15 SMARTS rules | External tool required |
| ADMET profile (built-in) | BBB / Caco-2 / hERG / CYP3A4 | External tool required |
| MAP4 fingerprint | Yes (Minervini 2020) | No (external package) |
| UFF force field for metals | Yes (Zn, Fe, Cu, …) | No |
| IUPAC name generation (offline) | 25+ compound classes | No |
| Retrosynthesis (template-based) | 60 retro-SMIRKS built-in | External tool required |
pip install anywhere |
Yes | No (conda/cmake) |
RDKit has, chematic does not (or is weaker)
| Feature | RDKit advantage |
|---|---|
| Publication-quality 3D conformers | ETKDGv3 with ML torsion corrections; chematic uses chair/envelope + MMFF94 (good for virtual screening) |
| File format support | 100+ formats (chematic: ~20) |
| Validated production docking | Years of benchmarking |
| Community plugins | Large ecosystem |
| Exact InChI (default) | C lib bundled by default |
4. API comparison (Python)
Most common operations map directly:
| Operation | chematic | RDKit |
|---|---|---|
| Parse SMILES | chematic.from_smiles(s) |
Chem.MolFromSmiles(s) |
| Molecular weight | mol.mw |
Descriptors.MolWt(mol) |
| LogP | mol.logp |
Descriptors.MolLogP(mol) |
| TPSA | mol.tpsa |
Descriptors.TPSA(mol) |
| ECFP4 | chematic.bulk.ecfp4(smiles) |
AllChem.GetMorganFingerprintAsBitVect(mol, 2) |
| Substructure | mol.has_substructure("[OH]") |
mol.HasSubstructMatch(Chem.MolFromSmarts("[OH]")) |
| Batch descriptors | chematic.descriptors_df(smiles) |
PandasTools + manual loop |
| Drug-likeness | mol.lipinski_passes |
Descriptors.rdMolDescriptors.CalcNumHBD(mol) <= 5 and … |
| Canonical SMILES | mol.smiles |
Chem.MolToSmiles(mol) |
5. When to choose
Choose chematic if:
- You want chemistry in the browser (WASM, 719 KB, no server)
- You need a pure Rust stack with no C++ toolchain
- You deploy to Lambda, Cloudflare Workers, or other constrained environments
- You build AI agents and want native MCP tool integration
- You need fast batch processing (ECFP4: 2–3× faster, Rayon-parallel)
- You want
pip installto just work — anywhere
Choose RDKit if:
- You need maximum ecosystem compatibility and 20+ years of production validation
- You need publication-quality 3D structures with ML-assisted torsion corrections (ETKDGv3)
- You rely on community plugins written against the RDKit Python API
- You need bit-exact standard InChI without an extra feature flag
- You work with exotic file formats or unusual molecule types
6. Migration quick-reference
If you have existing RDKit code, these are the most common substitutions:
# RDKit
from rdkit import Chem
from rdkit.Chem import Descriptors, rdMolDescriptors, AllChem
mol = Chem.MolFromSmiles("CC(=O)Oc1ccccc1C(=O)O")
mw = Descriptors.MolWt(mol)
fp = AllChem.GetMorganFingerprintAsBitVect(mol, 2, 2048)
# chematic
import chematic
mol = chematic.from_smiles("CC(=O)Oc1ccccc1C(=O)O")
mw = mol.mw
fp = chematic.bulk.ecfp4(["CC(=O)Oc1ccccc1C(=O)O"])[0] # numpy uint8 array
For large-scale migration, chematic's Python API is designed to be familiar to RDKit users while adding batch-first and browser-native capabilities.
Benchmark data: Apple M-series, Python 3.12, chematic v0.4.22, RDKit 2026.03.3.
Reproduce all benchmarks: see benchmark details.