A Chemprop v2 multi-component MPNN model that predicts 7 spectroscopic properties of organic chromophores from molecular structure (SMILES) and solvent.

Model Description

Architecture: MulticomponentMPNN (depth=4, hidden=400, FFN=400, dropout=0.15)
Parameters: 1.1M
Framework: Chemprop 2.2.1 (PyTorch Lightning)
Inputs: Chromophore SMILES + Solvent SMILES
Training data: 20,502 chromophore-solvent pairs from Scientific Data publication
Training: 100 epochs on NVIDIA A100 GPU, masked loss for missing values

Predicted Properties

Property	MAE	R²	Test Samples
Absorption max (nm)	15.40	0.9503	1,739
Emission max (nm)	18.31	0.9212	1,847
Quantum yield	0.1232	0.6754	1,377
abs FWHM (cm⁻¹)	462.2	0.8247	664
emi FWHM (cm⁻¹)	383.6	0.7576	1,091
log(ε/mol⁻¹ dm³ cm⁻¹)	0.1403	0.8582	817
Lifetime (ns)	4.84	0.0834	685

Note: Lifetime prediction has very low R² and should not be relied upon.

Usage

Installation

pip install chemprop>=2.0

Prediction

bash

chemprop predict \
  -i input.csv \
  --model-paths best.pt \
  -o predictions.csv \
  -s Chromophore Solvent

Input CSV format

csv

Chromophore,Solvent
CCN(CC)c1ccc2c(C)cc(=O)oc2c1,CCO
Nc1ccc2c(C(F)(F)F)cc(=O)oc2c1,CC#N

Python API

python

import torch
from chemprop.models import MPNN
from chemprop.data import MoleculeDatapoint, MoleculeDataset, MulticomponentDataset, build_dataloader
from chemprop.featurizers import SimpleMoleculeMolGraphFeaturizer

# Load model
model_data = torch.load("best.pt", map_location="cpu", weights_only=False)
model = MPNN(
    model_data["hyper_parameters"]["message_passing"],
    model_data["hyper_parameters"]["agg"],
    model_data["hyper_parameters"]["predictor"]
)
model.load_state_dict(model_data["state_dict"])
model.eval()

Training Details

Dataset: DB for chromophore_Sci_Data_rev03.csv (20,836 raw entries, 20,502 after cleaning)
Data cleaning: Removed 314 duplicates, 17 entries with negative Stokes shift, 20 entries with invalid solvent
Missing values: Handled via masked loss (Chemprop native support)
Split: Random 80/10/10 (train: 16,401 / val: 2,050 / test: 2,051)
Optimizer: Adam with warmup (5 epochs) + cosine decay
Hardware: NVIDIA A100-SXM4-40GB, RAIDEN HPC Cluster (RIKEN)

Limitations

Quantum yield predictions have moderate accuracy (MAE=0.12, R²=0.68) due to inherent measurement noise and limited structural information
Lifetime predictions are unreliable (R²<0.1) - this property requires quantum mechanical calculations beyond SMILES
Best performance on common solvents (DCM, acetonitrile, toluene); rare solvents may have higher error
Training data is from literature compilations with varying experimental conditions

Citation

If you use this model, please cite:

bibtex

@article{joung2020experimental,
  title={Experimental database of optical properties of organic compounds},
  author={Joung, Joonyoung F and Han, Minhi and Jeong, Minseok and Park, Sungnam},
  journal={Scientific Data},
  volume={7},
  pages={295},
  year={2020}
}

@article{heid2024chemprop,
  title={Chemprop: A Machine Learning Package for Chemical Property Prediction},
  author={Heid, Esther and others},
  journal={Journal of Chemical Information and Modeling},
  volume={64},
  number={1},
  pages={9--17},
  year={2024}
}

HuggingTaro/chromophore-spectral-predictor

Chromophore Spectral Property Predictor (7 Properties)