Skip to content

Prediction Model Builder

Jump to bottom
Shoichi Ishida edited this page Feb 21, 2025 · 2 revisions

This tool offers a GUI application that allows users to build their own prediction models using an automated machine learning tool (the FLAML).

The application accepts two input types: a CSV file for building general prediction models and a UniProt ID for constructing quantitative structure-activity relationship (QSAR) models from the ChEMBL database.

1. Upload CSV File

When a user uploads a CSV file, it is rendered as a table on the interface to enable the user to verify its content. To build prediction models, the user must select the names of two columns: one containing molecule structures in the SMILES format and another containing the values you want to predict.

Initially, the user adjusts the data proportion for the test dataset.

Estimator Selection

Following this, the user has the option to select the use auto for default ML estimators or manually choose ML estimators from a list option, including

  • random forest
  • LightGBM
  • eXtreme Gradient Boosting (XGBoost)
  • Categorical Boosting (CatBoost)
  • Extremely Randomized Trees (Extra Trees)
  • Logistic Regression with L1 or L2 regularization
  • k-nearest neighbors.

In the FLAML, the list of the default estimators are defined as follows: LightGBM, random forest, XGBoost, CatBoost, Extra Trees.

Task Selection

Prediction Model Builder currently supports only regression task.

Metric Selection

The metric is used to evaluate the performance of the constructed model during training, with the option to utilize automatic metric selection.

If the automatic metric selection is chosen, the FLAML uses

  • $1-r^2$ as the minimizing error metric, where $r^2$ is the coefficient of determination.

It is recommended that this checkbox is left checked.

Budget Time

A time budget can be set to manage the amount of time that computational resources are dedicated to performing the AutoML search process. (Default is set as 60 sec)

Target Value Standardization

If the values of the target variable are not normalized, the user should employ the standardization function in this application to standardize the target values, ensuring that reward functions can be appropriately designed.

Once all the settings are finalized, this tool runs the AutoML process to find quality models and saves the model that performs best on the test dataset for use in a reward function.

2. Fetch Data from ChEMBL Database Using UniProt ID

When a user inputs a UniProt ID, the application fetches and processes data from the ChEMBL database using the ChEMBL webresource client package.

The application verifies the existence of records for the specified UniProt ID and then allows users to refine the dataset by

  • deduplicating molecules based on pChEMBL values
  • retaining records according to their specific assay types
  • filtering out records with certain assay descriptions and activity types.

Subsequently, the workflow is the same as that of the process described above after uploading a CSV file.

Clone this wiki locally