CLaM-NP Score

Chemical Language Models for Natural Product-Likeness Score.

CLaM-NP scores how "natural-product-like" a molecule is by comparing the likelihood of its SMILES string under decoder-only language models (GPT-2) trained on different chemical spaces. A molecule that a natural-product model finds likely but a synthetic model finds unlikely gets a high score.

from clamnp import CLaMNPScorer

scorer = CLaMNPScorer.from_pretrained()          # downloads models from the HF Hub
scorer.score(["COc1cc2ncnc(Nc3ccc(F)c(Cl)c3)c2cc1OCCCN1CCOCC1",            # gefitinib
              "CN1CC[C@]23c4c5ccc(O)c4O[C@H]2[C@@H](O)C=C[C@H]3[C@H]1C5"])  # morphine
# -> [0.14, 0.67]   (synthetic-like, natural-like)

Installation

Inference only (lightweight):

pip install clamnp
pip install git+https://github.com/ohuelab/clamnp

With uv, no install needed:

uvx --from clamnp clamnp "CC(=O)Oc1ccccc1C(=O)O"

Extras:

Extra	For	Install
train	Training models, dataset preparation	pip install clamnp[train]
viz	Notebooks & paper figures	pip install clamnp[viz]

For local development on this repo, uv sync installs everything (the dev dependency group bundles all extras).

Usage

Command line

# Single molecule (models auto-downloaded and cached on first run)
clamnp "CC(=O)Oc1ccccc1C(=O)O"

# Several molecules
clamnp "CN1C=NC2=C1C(=O)N(C(=O)N2C)C" "O=C(O)c1ccccc1O"

# A file (one SMILES per line) -> CSV
clamnp --input molecules.smi --output scores.csv

# Raw log-likelihood ratio instead of the normalized [0, 1] score
clamnp --raw-score "CC(=O)Oc1ccccc1C(=O)O"

# Unstabilized (2-model) score, on CPU
clamnp --mode unstabilized --device cpu "CC(=O)Oc1ccccc1C(=O)O"

# Average over 32 randomized SMILES instead of scoring the canonical form once
clamnp --n-enum 32 "CC(=O)Oc1ccccc1C(=O)O"

Python

from clamnp import CLaMNPScorer

scorer = CLaMNPScorer.from_pretrained(scoring_mode="stabilized")

scorer.score(["CCO", "c1ccccc1"])                # batch -> list of normalized [0, 1] scores
scorer.score("CN1CCC[C@H]1c1cccnc1")             # single SMILES -> one score
scorer.score(["CCO", "c1ccccc1"], raw=True)      # raw log-likelihood ratios
scorer.score(["CCO", "c1ccccc1"], n_enum=32)     # mean over 32 randomized SMILES
scorer.score_with_details("CCO")                 # full breakdown (log P, perplexity, α, ...)

Models

Models are hosted on the Hugging Face Hub at kohbanye/clamnp, with one subfolder per model (natural, synthetic, general). The shared SMILES tokenizer is kohbanye/SmilesTokenizer_PubChem_1M.

License

MIT — see LICENSE.