# PAULY — Full Reference

PAULY is a predictive-modelling company for environmental chemistry, operated by
Pauly Predictions GmbH, Essener Straße 100, 04357 Leipzig, Germany (Amtsgericht
Leipzig, HRB 45535). It builds machine-learning models that predict the
physico-chemical properties, environmental fate, and toxicity of chemical
substances, so that data exists before laboratory testing is done. The site is
published in English (/en/) and German (/de/).

## What PAULY is for
- **Regulatory data gaps.** Generating robust physicochemical properties to fill
  data gaps in REACH evaluations and compliance reporting.
- **Analytical chemistry.** Predicting retention times and ion-mobility data to
  eliminate false positives in suspect-screening workflows and to shorten method
  development.
- **Property prediction methods.** QSPR, graph-convolutional neural networks,
  group-contribution approaches, k-nearest-neighbour models, and LSER solute
  descriptors.
- **Environmental fate and toxicity.** Predicting distribution, bioaccumulation,
  and adverse effects across more than 219 million registered chemicals.

## Founders
PAULY was founded by Dr. Nadin Ulrich, Anton Anders, and Karsten Voigt.
Dr. Ulrich has spent over ten years at the Helmholtz Centre for Environmental Research (UFZ) combining experimental insight with computational modelling.

## Toolboxes

### Risk assessment of chemicals
Generate robust physicochemical properties to fill data gaps in your REACH evaluations. We provide the pure, high-fidelity calculation data you need to build your compliance reports.
Peer-reviewed basis:
  - "Prediction of the water solubility by a graph convolutional-based neural network on a highly curated dataset.", J. Cheminformatics (2025) — https://doi.org/10.1186/s13321-025-01000-9
  - "Prediction of melting points of chemicals with a data augmentation-based neural network approach.", ACS Omega (2025) — https://doi.org/10.1021/acsomega.5c00205
  - "Prediction of solute descriptors for linear solvation energy relationships using k-nearest neighbours, group contributions, and graph-convolutional neural networks", ACS Environ. Au (2026) — https://doi.org/10.1021/acsenvironau.6c00063

Prediction models:
  - Octanol-Water Partition Coefficient log Kow (endpoint logKow_GNN, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/risikobewertung/logKow_GNN/
      Model type: Neuronal network including 2 hidden layers with 128 and 64 neurons, trained over 40 epochs
      Basis data: curated dataset of Mansouri et al. (initially 14,050 chemicals) - containing 13,889 chemicals
      Training set: 10,000 chemicals
  - Water solubility log Sw (endpoint logSw_GNN, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/risikobewertung/logSw_GNN/
      Model type: Neural network - Graph Convolutional Neural Network, 2 layers (64, 128 neurons)
      Basis data: Basic data set: curated data set from Sorkun et al. with additional data from various compilations - a total of 9,800 chemicals
      Training set: 6,860 data points
  - Bioconcentration factor log BCF (endpoint bcf_direct_gnn, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net (direct prediction)
      https://i-am-pauly.com/en/packages/risikobewertung/bcf_direct_gnn/
      Model type: Neuronal net including 2 hidden layers with 32 neurons per layer, early stopping was applied for training
      Basis data: curated dataset of the Bioconcentration Factor Database
      Training set: 738 chemicals
  - Octanol-water partition coefficient log Kow (endpoint logKow_qspr, Group-contribution approach GCA)
      Group contribution Approach
      https://i-am-pauly.com/en/packages/risikobewertung/logKow_qspr/
  - Water solubility log Sw (endpoint logSw_qspr, Group-contribution approach GCA)
      Group Contribution Approach
      https://i-am-pauly.com/en/packages/risikobewertung/logSw_qspr/
  - Bioconcentration factor log BCF (endpoint logBCF_Gnn, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net (log Kow) & determination according to Veith et al.
      https://i-am-pauly.com/en/packages/risikobewertung/logBCF_Gnn/
      Model type: The log BCF is calculated according to Veith et al. based on the predicted log Kow (graph-convolutional neural net). The range of log Kow, for which the model can be applied is between 0 and 6.
      Basis data: The linear relationship is as follows: log BCF = 0.85 log Kow - 0.70.
      Training set: Veith et al. developed their model based on 55 chemicals for the species Pimephales promelas.
  - Octanol-Water Partition Coefficient log Kow (endpoint logKow_knn, k-nearest neighbors kNN)
      K-Nearest Neighbors Approach
      https://i-am-pauly.com/en/packages/risikobewertung/logKow_knn/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts
      Basis data: curated dataset of Mansouri et al. (initially 14,050 chemicals) - containing 13,889 chemicals
      Training set: 12,500 Chemicals
  - Acute Fish Toxicity log LC50 (endpoint lc50_fish_GNN, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/risikobewertung/lc50_fish_GNN/
      Model type: Neural Network including 2 hidden layers with 16 neurons located in each layer, trained by the early stopping approach. The unit of the LC50 values is mol/L (given as log value).
      Basis data: Dataset of Schüürmann et al. - contains 691 chemicals
      Training set: 539 Chemicals
  - Bioconcentration factor log BCF (endpoint bcf_qspr, Group-contribution approach GCA)
      Group contribution approach (log Kow) & determination according to Veith et al.
      https://i-am-pauly.com/en/packages/risikobewertung/bcf_qspr/
  - Acute toxicity Daphnia Magna log LC50 (endpoint lc50_daphnia_GNN, Graph-convolutional neural network GNN)
      Graph Convolutional Neural Net
      https://i-am-pauly.com/en/packages/risikobewertung/lc50_daphnia_GNN/
      Model type: Neural Network including 2 hidden layers with 32 and 64 neurons located in each layer, trained by the early stopping approach. The unit of the LC50 values is mol/L (given as log value).
      Basis data: Dataset of Kühne et al. - contains 1,355 chemicals
      Training set: 1,083 chemicals
  - Air-Water Partition Coefficient log Kaw (endpoint kaw_Gnn, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/risikobewertung/kaw_Gnn/
      Model type: Neural net including 2 hidden layers with 16 and 32 neurons, early stopping was applied
      Basis data: extended dataset of Ebert et al. - 3267 chemicals
      Training set: 2,286 chemicals
  - Acute toxicity fish log LC50 (endpoint lc50_qspr, Group-contribution approach GCA)
      Group contribution approach
      https://i-am-pauly.com/en/packages/risikobewertung/lc50_qspr/
  - Octanol-Air Partition Coefficient log Koa (endpoint koa_Gnn, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/risikobewertung/koa_Gnn/
      Model type: Neuronal network including 2 hidden layers with 16 and 32 neurons, trained by the application of an early stopping approach
      Basis data: curated dataset of Ebert et al. - containing 2,732 chemicals
      Training set: 1,911 chemicals
  - OC-Water Partition Coefficient log Koc (endpoint koc_Gnn, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/risikobewertung/koc_Gnn/
      Model type: Neural Network - Graph-Convolutional Neural Network, 2 layers (16, 32 neurons)
      Basis data: EPA (2014) EPI Suite curated by Mansouri et al.
      Training set: 507 datapoints
  - Melting Point MP (endpoint mp_GNN, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/risikobewertung/mp_GNN/
      Model type: Consensus graph convolutional neural net to predict melting points of organic chemicals, 2 hidden layer (64, 128 neurons), unit °C
      Basis data: curated Bradley dataset containing 19,995 chemicals
      Training set: 13,989 data
  - Boiling point (endpoint bp_GNN, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/risikobewertung/bp_GNN/
      Model type: Consensus graph convolutional neural net to predict melting points of organic chemicals, 2 hidden layers (64, 128 neurons), unit °C
      Basis data: dataset of EPA (2014) EPI Suite curated by Mansouri et al.
      Training set: 4,226 dataponints
  - Melting Point MP (endpoint mp_KNN, k-nearest neighbors kNN)
      K-Nearest Neighbors Approach
      https://i-am-pauly.com/en/packages/risikobewertung/mp_KNN/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts, unit °C
      Basis data: curated dataset of Bradley et al.
      Training set: 17,659 datapoints
  - OC-Water Partition Coefficient log Koc (endpoint koc_KNN, k-nearest neighbors kNN)
      K-Nearest Neighbors Ansatz
      https://i-am-pauly.com/en/packages/risikobewertung/koc_KNN/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts
      Basis data: EPA (2014) EPI Suite curated by Mansouri et al.
      Training set: 655 chemicals
  - Octanol-Air Partition Coefficient log Koa (endpoint koa_KNN, k-nearest neighbors kNN)
      K-Nearest Neighbors Approach
      https://i-am-pauly.com/en/packages/risikobewertung/koa_KNN/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts
      Basis data: curated dataset of Ebert et al. - containing 2,732 chemicals
      Training set: 2,458 chemicals
  - Air-Water Partition Coefficient log Kaw (endpoint kaw_KNN, k-nearest neighbors kNN)
      K-Nearest Neighbors Approach
      https://i-am-pauly.com/en/packages/risikobewertung/kaw_KNN/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts
      Basis data: extended dataset of Ebert et al. - 3267 chemicals
      Training set: 2,940 chemicals
  - Bioconcentration Factor log BCF (endpoint bcf_KNN, k-nearest neighbors kNN)
      K-Nearest Neighbors Approach
      https://i-am-pauly.com/en/packages/risikobewertung/bcf_KNN/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts
      Basis data: curated dataset of the Bioconcentration Factor Database
      Training set: 950 datapoints
  - Water Solubility log Sw (endpoint sw_KNN, k-nearest neighbors kNN)
      K-Nearest Neighbors Approach
      https://i-am-pauly.com/en/packages/risikobewertung/sw_KNN/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts
      Basis data: curated data set from Sorkun et al. with additional data from various compilations - a total of 9,800 chemicals
      Training set: 8,818 chemicals
  - Boiling Point BP (endpoint bp_KNN, k-nearest neighbors kNN)
      K-Nearest Neighbors Approach
      https://i-am-pauly.com/en/packages/risikobewertung/bp_KNN/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts, unit °C
      Basis data: dataset of EPA (2014) EPI Suite curated by Mansouri et al.
      Training set: 4,890 datapoints
  - Acute Toxicity Daphnia Magna log LC50 (endpoint lc50_daphnie_KNN, k-nearest neighbors kNN)
      K-Nearest Neighbors Approach
      https://i-am-pauly.com/en/packages/risikobewertung/lc50_daphnie_KNN/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts. Note that the unit of LC50 is [mol/L] (given as log value).
      Basis data: Dataset of Kühne et al. - contains 1,355 chemicals
      Training set: 1,219 datapoints
  - Acute Fish Toxicity log LC50 (endpoint lc50_fish_KNN, k-nearest neighbors kNN)
      K-Nearest Neighbors Approach
      https://i-am-pauly.com/en/packages/risikobewertung/lc50_fish_KNN/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts. Note that the unit of LC50 is [mol/L] (given as log value).
      Basis data: Dataset of Schüürmann et al. - contains 691 chemicals
      Training set: 621 datapoints
  - Algae Growth Inhibition log EC50 (endpoint ec50_algae_KNN, k-nearest neighbors kNN)
      K-Nearest Neighbors Aproach
      https://i-am-pauly.com/en/packages/risikobewertung/ec50_algae_KNN/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts. Note that the unit of EC50 is [mol/L] (given as log value).
      Basis data: inhouse UFZ dataset
      Training set: 44 datapoints
  - Algae Growth Inhibition log EC50 (endpoint ec50_algea_GNN, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/risikobewertung/ec50_algea_GNN/

### Migration Models
Longer description coming soon.
Peer-reviewed basis:
  - "Predicting partitioning from low density polyethylene to blood and adipose tissue by linear solvation energy relationship models", J. Biomed. Mater. Res. Part B (2023) — https://doi.org/10.1002/jbm.b.35304

Prediction models:
  - PE/Water Partition Coefficient (endpoint lser_pe_water, Graph-convolutional neural network GNN)
      Prediction of log K PE/water
      https://i-am-pauly.com/en/packages/migrationsmodellierung/lser_pe_water/
  - Octanol-Water Partition Coefficient log Kow (endpoint logKow_GNN, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/migrationsmodellierung/logKow_GNN/
      Model type: Neuronal network including 2 hidden layers with 128 and 64 neurons, trained over 40 epochs
      Basis data: curated dataset of Mansouri et al. (initially 14,050 chemicals) - containing 13,889 chemicals
      Training set: 10,000 chemicals
  - Octanol-water partition coefficient log Kow (endpoint logKow_qspr, Group-contribution approach GCA)
      Group contribution Approach
      https://i-am-pauly.com/en/packages/migrationsmodellierung/logKow_qspr/

### LSER Toolbox
Predict highly accurate solute descriptors using our advanced machine learning models, then seamlessly calculate partition coefficients, retention times, and diverse physicochemical properties based on established phase equations.
Peer-reviewed basis:
  - "Prediction of solute descriptors for linear solvation energy relationships using k-nearest neighbours, group contributions, and graph-convolutional neural networks", ACS Environ. Au (2026) — https://doi.org/10.1021/acsenvironau.6c00063

Prediction models:
  - LSER Solute Descriptors (endpoint absolv_GNN, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Network
      https://i-am-pauly.com/en/packages/lser-substanzdeskriptoren/absolv_GNN/
      Model type: For the solute descriptors A, B, E, S, L individual neural nets were trained with 2 hidden layers including 32 and 64 neurons.
      Basis data: For training of the models the curated dataset of Ulrich et al. was used, the number of chemicals varies for the different solute descriptors.
      Training set: Number of training set chemicals varies for each descriptor.
  - LSER Solute Descriptors (endpoint absolv_qspr, Group-contribution approach GCA)
      Group Contribution Approach
      https://i-am-pauly.com/en/packages/lser-substanzdeskriptoren/absolv_qspr/
  - LSER Solute descriptors (endpoint absolv_knn, k-nearest neighbors kNN)
      K-Nearest Neighbors Approach
      https://i-am-pauly.com/en/packages/lser-substanzdeskriptoren/absolv_knn/
      Model type: k-nearest-neighbors model based on the Jaccard metric in the space of atom-centered fragment (ACF) counts
      Basis data: curated Abraham Absolv dataset
      Training set: 5,728 datapoints for the S descriptor. Note that the number of datapoints varies for the different solute descriptors.

### Toolbox Analytical Chemistry
Generate highly accurate retention time and ion mobility data to eliminate false positives in your suspect screening workflows. Our predictive models streamline your method development before you even step into the lab.
Peer-reviewed basis:
  - "Prediction of solute descriptors for linear solvation energy relationships using k-nearest neighbours, group contributions, and graph-convolutional neural networks", ACS Environ. Au (2026) — https://doi.org/10.1021/acsenvironau.6c00063
  - "Linear solvation energy relationships as classifiers in non-target analysis - a capillary liquid chromatography approach",J. Chromatogr. A (2011) — https://doi.org/10.1016/j.chroma.2011.09.031
  - "Prediction of gas chromatographic retention indices as classifier in non-target analysis of environmental samples",J. Chromatogr. A (2013) — https://doi.org/10.1016/j.chroma.2013.02.037

Prediction models:
  - Ion mobility CCS for [M+H]+ Ions (endpoint css_GNN, Graph-convolutional neural network GNN)
      Graph-convolutional Neural Net
      https://i-am-pauly.com/en/packages/analytische-chemie/css_GNN/
      Model type: graph convolutional neural net to predict CCS values of organic [M+H]+ ions, 2 hidden layer (32, 32 neurons)
      Basis data: METLIN CCS database
      Training set: 134,800 datapoints
  - Ion Mobility CCS for [M+Na]+ Ions (endpoint ccsna_GNN, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/analytische-chemie/ccsna_GNN/
      Model type: graph convolutional neural net to predict CCS values of organic [M+Na]+ ions, 2 hidden layer (16, 32 neurons)
      Basis data: Metlin CCS database
      Training set: 85,700 datapoints
  - Ion Mobility CCS for [M-H]- Ions (endpoint ccsm_GNN, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net
      https://i-am-pauly.com/en/packages/analytische-chemie/ccsm_GNN/
      Model type: graph convolutional neural net to predict CCS values of organic [M-H]- ions, 2 hidden layer (16, 32 neurons)
      Basis data: Metlin CCS database
      Training set: 109,600 datapoints
  - RT Prediction C18 Positive Mode (endpoint rtC18pos_Gnn, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net - UFZ Method
      https://i-am-pauly.com/en/packages/analytische-chemie/rtC18pos_Gnn/
      Model type: graph convolutional neural net to predict retention times on a C18 column in positive mode, retention times are given in min, UFZ inhouse model, 2 hidden layers (16, 32 neurons)
      Basis data: UFZ inhouse dataset
      Training set: 474 datapoints
  - RT Prediction C18 Negative Mode (endpoint rtC18neg_Gnn, Graph-convolutional neural network GNN)
      Graph-Convolutional Neural Net - UFZ Method
      https://i-am-pauly.com/en/packages/analytische-chemie/rtC18neg_Gnn/
      Model type: graph convolutional neural net to predict retention times on a C18 column in negative mode, retention times are given in min, UFZ inhouse model, 2 hidden layers (16, 32 neurons)
      Basis data: UFZ inhouse dataset
      Training set: 282 datapoints

## Modelling approaches
- Graph-Convolutional Neural Network (GNN) — Learns directly from a molecule's structural graph — powerful, but data-hungry. (https://i-am-pauly.com/en/knowledge-hub/gnn/)
- Group-Contribution Approach (GCA) — Builds a prediction from known molecular fragments — transparent, but limited to familiar structures. (https://i-am-pauly.com/en/knowledge-hub/GCA/)
- k-Nearest Neighbors (kNN) — Predicts from the most similar known molecules — robust, but only as good as its nearest neighbour. (https://i-am-pauly.com/en/knowledge-hub/knn/)

## Contact
- Website: https://i-am-pauly.com
- Software: https://app.pauly-predictions.de/
- Email: info@pauly-predictions.de
- Languages: English, German