Naturalness Prediction

Enter SMILES string to get the prediction from the fine-tuned ChemBERTa model.
Examples

Naturalness Predictor Using ChemBERTa

This project aims to predict the naturalness of chemical compounds using a fine-tuned ChemBERTa foundation model. The work serves as a preliminary proof-of-concept study with promising results.

Dataset

The dataset was compiled from two major sources:

  1. Natural Products Atlas: Contains information on natural products.
  2. ChEMBL: A comprehensive database of small molecules.

The dataset consists of the following:

  • Each entry in the two data sources was matched using the InChI Key.
  • Out of approximately 1.5 million molecules in ChEMBL, 5,298 were marked as natural, with the remaining marked as unnatural. Note that this dataset is not 100% accurate but follows a common strategy where negative data is expected to dominate.

Methodology

  1. Data Preparation:

    • All natural products in the annotated ChEMBL dataset were used for further analysis.
    • Unnatural molecules were randomly sampled to make the negative data 1.5x the size of the positive data.

  2. Model Training:

    • 80% of the dataset was used to fine-tune the ChemBERTa foundation model.
    • The remaining 20% of the dataset was reserved for performance evaluation.

  3. Results:

    • The prediction achieved an accuracy of 93.88%. This was accomplished using a simple fine-tuning strategy, which required only a few minutes on a personal computer.

Future Work

This project is just a starting point and will benefit from further refinement and testing for a more robust evaluation of the strategy. Future work will focus on enhancing data accuracy and exploring additional features that can improve prediction capability.

🔗 Links