Maryam-Taherzadeh/agentic-ai-mdm2-p53

Agentic AI Drug Discovery Workflow for MDM2–p53 Inhibitor Prioritization

Live App

🔗 Try the app on Hugging Face Spaces

Overview

This project is an end-to-end machine learning and agentic AI workflow for MDM2–p53 inhibitor prioritization.

The workflow combines ChEMBL bioactivity data collection, RDKit molecular descriptor generation, machine learning model development, pIC50 prediction, Lipinski drug-likeness evaluation, rule-based candidate recommendation, and deployment as an interactive Streamlit application.

The current deployed version supports live pIC50 prediction directly from SMILES using RDKit molecular descriptors and an optimized XGBoost regression model.

The final goal is to build a decision-support system for early-stage AI-driven drug discovery that can later integrate classical machine learning, graph neural networks, SMILES-based transformer models, ADMET prediction, docking, and cloud-based deployment.

Featured Project Summary

Agentic-AI-Drug-Discovery-MDM2-P53 is an end-to-end agentic AI drug discovery workflow for MDM2–p53 inhibitor prioritization. The project includes ChEMBL bioactivity processing, RDKit descriptor generation, live pIC50 prediction from SMILES using XGBoost, Lipinski drug-likeness evaluation, rule-based candidate recommendation, and a deployed Streamlit app on Hugging Face Spaces.

Project Pipeline

This project is organized into 7 main steps.

Part 01 — ChEMBL Bioactivity Data Collection

Bioactivity data for MDM2 inhibitors was collected from ChEMBL.

The dataset includes compounds with reported IC50 values against the MDM2 target. These values were transformed into pIC50 values for regression-based machine learning modeling.

Part 02 — RDKit Feature Generation

Molecular features were generated using RDKit.

These features include chemically meaningful descriptors such as:

• —Molecular weight
• —LogP
• —Hydrogen bond donors
• —Hydrogen bond acceptors
• —TPSA
• —Rotatable bonds
• —Ring counts
• —Aromatic ring counts

These RDKit descriptors enable direct prediction from SMILES in the deployed app.

Part 03 — PaDEL Descriptor Generation

PaDEL descriptors and molecular fingerprints were generated to numerically represent molecular structures.

These descriptors were used in the original model development and comparison workflow.

Part 04 — Random Forest Baseline Model

A Random Forest regression model was trained as the baseline model for predicting pIC50 values.

This model served as the first benchmark for evaluating predictive performance.

Part 05 — Model Comparison

Multiple classical machine learning models were compared, including:

• —Random Forest
• —Support Vector Regression
• —XGBoost

The goal was to identify the strongest model for pIC50 prediction.

Part 06 — XGBoost Optimization

XGBoost was selected as the final optimized classical machine learning model.

The optimized PaDEL-based model was evaluated using standard regression metrics.

Original PaDEL-XGBoost model performance:

Part 07 — Agentic AI Workflow and Streamlit App

An agentic AI-style decision-support workflow was developed for MDM2–p53 inhibitor prioritization.

The workflow:

1. 1.Accepts a candidate molecule as a SMILES string
2. 2.Generates RDKit molecular descriptors
3. 3.Predicts pIC50 using a trained XGBoost regression model
4. 4.Classifies the compound as Active, Intermediate, or Inactive
5. 5.Evaluates Lipinski drug-likeness
6. 6.Generates a rule-based next-step recommendation

The deployed app uses RDKit descriptors because they allow live prediction directly from SMILES.

Live RDKit-XGBoost Deployment Model

To support interactive deployment, a second deployment-friendly model was trained using RDKit molecular descriptors.

This enables the Streamlit app to perform live pIC50 prediction directly from a user-provided SMILES string.

Deployment model performance:

Streamlit App Features

The deployed Streamlit app includes:

• —SMILES input
• —Example candidate dropdown
• —Live RDKit descriptor generation
• —Live pIC50 prediction using XGBoost
• —Activity classification
• —Lipinski drug-likeness evaluation
• —Rule-based recommendation
• —Model performance summary
• —Clean, light-themed interface
• —Hugging Face Spaces deployment

Example App Output

For a candidate molecule, the app returns:

Current Technology Stack

• —Python
• —RDKit
• —Pandas
• —NumPy
• —Scikit-learn
• —XGBoost
• —Joblib
• —Streamlit
• —Hugging Face Spaces

Planned Version 3 — Multi-Model AI Drug Discovery Workflow

Future versions will extend the current RDKit-XGBoost app into a multi-model AI workflow.

Planned models:

• —Random Forest
• —Support Vector Regression
• —XGBoost
• —Graph Neural Network
• —ChemBERTa
• —Multi-model consensus scoring

The planned model selection interface will allow users to choose:

RDKit-XGBoost
Random Forest
GNN
ChemBERTa
Consensus Model