ChemPredictor

Table of Contents

1. Overview
2. Objectives
3. Key Features
4. Usage
   1. Installation and Setup
   2. How to Use
5. Dependencies
6. File Structure
7. Hosted Version
8. How it Works

Overview

• ChemPredictor is a Streamlit Web Interface that uses an artificial neural network (ANN) trained using TensorFlow for quantitative structure-property relationship (QSPR) analysis of molecules to predict 9 thermodynamic properties.
• Users can either enter the common name of the compound or draw the molecule using an interactive sketch tool and obtain predicted properties in real time.
• The project aims to provide a user-friendly interface for chemists, researchers, and students to analyze and predict the properties of organic molecules.
• The integration of Google's AI language model, Gemini Pro, allows users to obtain additional information about the predicted properties or the compound itself. (API Key required)

Objectives

• Predict Molecular Properties: Utilize an ANN model to predict 9 thermodynamic properties of organic molecules based on their molecular structure.

• User-Friendly Interface: Develop a Streamlit WebUI that allows users to input the common name, SMILES string, or draw the molecule for property prediction.

• Gemini LLM Integration: Integrate Google's AI language model, Gemini Pro, to provide additional information about the predicted properties or the compound itself. (API Key required) Table 1 - Predicted properties of the QM9 dataset

  No.	Property	Unit	Description
  1	μ	D	Dipole moment
  2	α	a³	Isotropic polarizability
  3	homo	Ha	Energy of HOMO
  4	lumo	Ha	Energy of LUMO
  5	gap	Ha	Gap (lumo − homo)
  6	U	Ha	Internal energy at 298.15 K
  7	H	Ha	Enthalpy at 298.15 K
  8	G	Ha	Free energy at 298.15 K
  9	Cv	cal/mol K	Heat capacity at 298.15 K

  • Dipole moment (µ): Measurement of polarity of a molecule.
  • Electronic polarizability (α): Tendency of non-polar molecules to shift their electron clouds relative to their nuclei.
  • Energy of HOMO: Energy of the highest occupied Molecular Orbital.
  • Energy of LUMO: Energy of the lowest unoccupied Molecular Orbital.
  • Band Gap Energy: Energy of LUMO – Energy of HOMO.
  • Internal energy of atomization (U): Energy required to break a molecule into separate atoms.
  • Enthalpy of atomization (H): Amount of enthalpy change when a compound's bonds are broken, and the component atoms are separated into single atoms.
  • Free energy of atomization (G): Extra energy needed to break up a molecule into separate atoms.
  • Heat capacity (Cv): Amount of heat required to increase the temperature of the molecule by one degree.

Key Features

• Gemini LLM Integration:

• Multiple Input Options: ChemPredictor supports various input methods for user convenience.

  • SMILES Input: Predict properties by entering the SMILES string of the compound directly.
  • Common Name Input: Input the common name of the compound (e.g., "Aspirin") to predict its properties.
  • CSV File Upload: Upload a CSV file containing SMILES strings to predict properties for multiple molecules.
  • Molecule Drawing: Use an interactive drawing board to draw the molecule for property prediction.

• Similar Molecules: Shows similar molecules based on the predicted properties, providing insights into chemical similarity.

• 3D Molecular Visualization: Shows molecular structures in three dimensions, providing an insight into the compound's spatial arrangement."

Usage

Installation and Setup

To run the ChemPredictor WebUI, follow these steps:

1. Clone the repository:

   git clone https://github.com/CubeStar1/ChemPredictorv2.git
   cd ChemPredictorv2

2. Create a virtual environment:

   python -m venv venv

3. Activate the virtual environment:

   • On Windows:

     .\venv\Scripts\activate

   • On Unix or MacOS:

     source venv/bin/activate

4. Install dependencies:

   pip install -r requirements.txt
   

How to Use

1. Run the Streamlit app:

    streamlit run app.py

2. Open the WebUI in your browser:

   • Open your web browser and navigate to http://localhost:8501 to use the ChemPredictor WebUI.

3. Predict Molecular Properties:

   Choose one of the following ways to predict molecular properties:

   • Enter Common Name : Input the common name of the compound (e.g., "Aspirin") to predict its properties.

   • Enter SMILES String:

     • Enter the SMILES string of the compound in the provided input field.

   • Upload SMILES CSV File:

     • Upload a CSV file containing a set of SMILES strings for bulk prediction.

   • Draw Molecule:

     • Utilize the interactive drawing board to draw the molecule for prediction.

4. Click on the Predict Button:

   • Once the input is provided (SMILES string, CSV file, or drawn molecule), click on the "Predict" button.

5. View Predicted Values:

   • The predicted values will be displayed on the right side of the screen.

Dependencies

Dataset

• QM9 Dataset: The QM9 dataset is used for training the ANN model. It contains 133,885 stable small organic molecules with up to nine heavy atoms (C, N, O, F). The dataset includes 9 thermodynamic properties for each molecule.

Data Analysis and Exploration

• Libraries Used: NumPy, Pandas, Matplotlib
• Overview: Data analysis and exploration are performed using NumPy, Pandas, and Matplotlib to understand the dataset's structure and properties.

Data Preparation

• Libraries Used: RDKit, Mordred Molecular Descriptors
• Overview: RDKit is used to convert the SMILES strings into molecular structures, and Mordred is used to calculate 1826 molecular descriptors for each molecule.

Machine Learning

• Libraries Used: TensorFlow with Keras, Scikit-learn
• Overview: An ANN model is trained using TensorFlow with Keras to predict the 9 thermodynamic properties of organic molecules. The model is trained on the QM9 dataset and evaluated using Scikit-learn.

Model Deployment and Prediction

• Libraries Used: Streamlit
• Overview: Streamlit is used to make a user-friendly Web Interface. User input is accepted via the WebUI, preprocessed, and predictions are made using the deployed ANN model.
• Gemini LLM Integration: Google's AI language model, Gemini Pro, is integrated to provide additional information about the predicted properties or the compound itself. (API Key required)

File Structure

project-root/
├── app.py                       # Streamlit WebUI
│
├── scripts/
│   ├── image_handling.py        # Image handling functions
│   ├── predict_property         # Property prediction functions
│   ├── project_overview_page.py # Project overview page
│   ├── utils.py                 # Project utility functions
│
├── utilities/
│   ├── assets/                 # Images
│   ├── dataset/                # QM9 dataset
│   ├── descriptors/            # Mordred descriptor files
│   ├── models/                 # Trained model files
│   ├── scalers/                # Pickled MinMaxScaler file
│
├── images/                     # Image storage 
│   
├── requirements.txt
├── packages.txt
│
└── README.md

Hosted Version

For a quick demo, you can also access the hosted version of ChemPredictor at https://chempredictor.streamlit.app/

How it Works