protein-molecule-binding-challenge

There are 3 Jupyter Notebooks:

• Data_Exploration_Preparation
• ML_Pipeline
• Run_Test_Set

Data_Exploration_Preparation.ipynb

Table of Contents

• Import and inspect dataset

• Protein Embedding

  • Embed Uniprot IDs using ProtT5 model (prottrans_t5_xl_u50)

• Molecule Embedding

  • Get SMILES representation from PubChem
  • Convert SMILES to Embedding
    • RDKit for Morgan Fingerprints
    • Alternative option, using BERT model (not used in current solution)

• Preparing data for Machine Learning Pipeline

  • Normalizing and saving embedded data

ML_Pipeline.ipynb

LGBM model gets trained on 80% of the data and performs as follows on the held out 20%:

• Best number of estimators: 215
• Mean Squared Error: 391223160770.29596
• R^2 Score: 0.1777882974932672

Run_Test_Set.ipynb

A new test set can be loaded and predictions can be made using the trained model.

Conclusion and Remarks:

An R² score of 0.177 is relatively low but indicates that the model is capturing some signal from the data. Further optimization and exploration are necessary to improve performance.

Notes

Cross-validation should be conducted to validate and generalize this performance. Additional analysis can determine whether using estimated or non-estimated KIBA scores impacts the model's performance.

Potential Next Steps to Improve Performance

Feature Engineering

• Incorporate additional features such as Kd, Ki, and IC50 scores.

Enhanced Embeddings

• Experiment with more informative embeddings, such as:
  • PubChem10M_SMILES_BERT
  • DeepChem
  • AlphaFold embeddings
  • ProtTrans

Explore Advanced Models

Implement deep learning techniques, including:

• Graph Neural Networks (GNNs) for representing molecular and protein structures.
• End-to-end models for drug-target binding prediction, such as DeepAffinity.

By incorporating these strategies, the model's performance can likely be enhanced, leading to more robust predictions.