kohbanye/clamnp

CLaM-NP

SMILES language models for natural-product-likeness scoring (CLaM-NP Score).

This repository hosts up to three causal language models, one per subfolder:

Subfolder	Training data	Role
natural	COCONUT	P(x | natural)
synthetic	ZINC22	P(x | synthetic)
general	ChEMBL	P(x | general) — stabilizer

Tokenizer: kohbanye/SmilesTokenizer_PubChem_1M.

Training data

The three models are trained on a joint scaffold split: the scaffolds of all three datasets are pooled and clustered together (ECFP4 Tanimoto >= 0.6), and each cluster goes to a single split, so no scaffold is shared between the training data of one dataset and the evaluation data of another. The split is built by scripts/joint_split.py.

• natural: COCONUT, August 2025 release (coconut_csv-08-2025.zip)
• synthetic: 700,000 compounds drawn at random from ZINC 22. A random draw cannot be repeated, so the sample is deposited here as data/zinc22_sample.smi (one SMILES per line, as standardized for training).
• general: ChEMBL 35

Usage

from clamnp import CLaMNPScorer

scorer = CLaMNPScorer.from_pretrained()          # this repo
print(scorer.score("CC(=O)Oc1ccccc1C(=O)O"))         # CLaM-NP Score

Or load a single model directly:

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("kohbanye/clamnp", subfolder="natural")