ChemXploreML

Appearance

Machine Learning Training ​

The Machine Learning Training module is the core of ChemXploreML, providing a comprehensive suite of tools for building, optimizing, and evaluating predictive models for molecular properties. The interface is organized into two main tabs: ML Model for training and ML Prediction for using a trained model.

ML Model Tab ​

This tab contains a multi-panel interface that guides you through the entire model training workflow, from data selection to advanced hyperparameter tuning and model evaluation.

ML Training Overview

Workflow Overview ​

  1. Select Input Data: Choose the molecular embeddings and the corresponding data file.
  2. Configure Training: Set up the train/test split, cross-validation, and other basic parameters.
  3. Choose a Model: Select a machine learning algorithm and configure its parameters.
  4. Set Advanced Options: Optionally, configure advanced features like data cleaning, hyperparameter tuning, and model interpretability analysis.
  5. Save and Run: Specify a location to save the trained model and its results, then start the training process.
  6. Analyze Results: Review the performance metrics, plots, and analysis results to evaluate your model.

1. Training File Panel ​

This panel is for selecting the input data for model training.

  • Embeddings File: Choose the .npy file containing the molecular embeddings you generated previously.
  • Training Data File: Select the corresponding data file (e.g., CSV) that contains the target property values (Y-column).

2. Control Panel ​

Here, you configure the fundamental aspects of the training and validation process.

  • Train/Test Split: Set the percentage of data to be used for the test set.
  • Cross-Validation: Enable k-fold cross-validation and specify the number of folds. This is highly recommended for robust model evaluation.
  • Hyperparameter Tuning: Choose a strategy for optimizing your model's hyperparameters:
    • None: Train the model with the manually specified parameters.
    • Grid Search: Exhaustively search over a specified subset of the hyperparameter space.
    • Randomized Grid Search: Samples a fixed number of parameter settings from the specified distributions.
    • Halving Grid Search: An efficient method that successively prunes underperforming parameter combinations.
    • Optuna: A powerful Bayesian optimization framework that intelligently searches for the best hyperparameters. You can monitor Optuna's progress using the built-in dashboard.

3. Model Panel ​

Select the machine learning algorithm and configure its parameters.

  • Model: Choose from a wide range of models, including:
    • Linear Models: Linear Regression, Ridge, Lasso, ElasticNet
    • Support Vector Machines: SVR
    • Neighbors: KNN
    • Gaussian Process: GPR
    • Ensemble Methods: Random Forest (RFR), Gradient Boosting (GBR)
    • Advanced Gradient Boosting: XGBoost, LightGBM, CatBoost
  • Parameters: For each model, you can manually set its parameters or define a search space for hyperparameter tuning.

4. More Options Panel ​

This panel contains advanced features for improving your model and gaining deeper insights.

  • Y-Transformation and Scaling: Apply transformations (e.g., log, Box-Cox, Yeo-Johnson) and scaling (e.g., StandardScaler, MinMaxScaler) to the target variable. This can be very effective for models that assume a normal distribution of the target.
  • Data Augmentation:
    • Bootstrap: Create a larger training set by resampling the existing data with replacement.
    • Noise Injection: Add random noise to the target variable in the bootstrapped samples to improve model robustness.
  • Cleanlab: Use the Cleanlab algorithm to automatically detect and remove mislabeled data points from your dataset before training.
  • Learning Curve: Generate a learning curve plot to diagnose whether your model is suffering from high bias or high variance.
  • SHAP Analysis: Compute SHAP (SHapley Additive exPlanations) values to understand the importance of each feature in the model's predictions.

5. Save Model and Run ​

  • Save Pretrained Model: Specify a name and location to save your trained model (.pkl file) and all associated results (plots, metrics, etc.).
  • Begin Training: Click this button to start the training process. You can monitor the progress in the application's console.

6. Results Panel ​

After training is complete, this panel will populate with a comprehensive summary of your model's performance.

  • Metrics: View key metrics like R², MSE, RMSE, and MAE for both the training and test sets.
  • Plots: Analyze the parity plot (predicted vs. actual values) and other visualizations.
  • Applicability Domain: Review the leverage and Mahalanobis distance plots to assess whether your model is making reliable predictions.

ML Prediction Tab ​

Once you have a trained model, you can use the ML Prediction tab to predict the properties of new, unseen molecules.

  1. Load Model: Select the .pkl file of your saved model.
  2. Provide Input: Load a file containing the SMILES strings of the molecules you want to predict.
  3. Run Prediction: The application will generate embeddings for the new molecules and use the loaded model to predict the target property.
  4. View and Save Results: The predictions will be displayed and can be saved to a file.