Kinase Inhibitor Activity Prediction

This repository contains the code and analysis for predicting kinase inhibitor activity using Kindel dataset (https://arxiv.org/pdf/2410.08938)

The project is structured in three parts: exploratory analysis, proof of concept (PoC), and deployment planning.

Structure of the repository

1. exploratory/

   • Exploratory analysis of different featurization methods and models.
   • Only data from random split 1 is used.
   • Splitting, evaluation, and metrics follow the Kindel paper.
   • From this analysis, the ChemBERTa+DNN model (ChemBERTa embeddings + 32-unit DNN head, fine-tuned) was selected for the PoC.

2. poc/

   • Code for the PoC model (ChemBERTa+DNN).
   • Includes scripts for training and prediction.
   • pixi.toml defines the environment dependencies.
   • Alternatively Docker image can be used.

3. deployment/

   • Contains information for deploying the ChemBERTaDNN model as a production-ready system.

Results and Discussion

• ChemBERTa+DNN outperformed all other models in the exploratory phase. Fine-tuning ChemBERTa embeddings gave the best performance overall.

• On test datasets, all models performed worse than on validation sets (highlighting generalization issues).

• The above conclusions are valid both for DDR1 and MAPK14 targets.

• Leveraging pre-trained transformer embeddings can be a promising approach, as it consistently improves performance compared to traditional featurization methods.

• However, overall predictive performance remains limited. The results indicate that the models, including ChemBERTa+DNN, still struggle to generalize to unseen compounds, especially in off-DNA test sets.

• Limitations:

  • No hyperparameter tuning or extensive architecture search was performed.
  • Limited to a subset of the dataset (random split 1).

• Opportunities for improvement:

  • Explore different architectures and hyperparameters.
  • Test alternative transformer models (e.g., MolFORMER) and Graph Neural Networks (GNNs).
  • Incorporate biological context (e.g., protein target information).
  • Evaluate across the full Kindel dataset.

Next Steps

• Deepen understanding of dataset properties, evaluation metrics, and biology/chemistry of the targets.
• Experiment with more architectures and featurizations.
• Enhance deployment pipeline.
• Extend experiments beyond split 1.