BIOPREDICT

A QSAR Machine Learning Framework for Bioactivity Prediction of Query Compounds
Cheminformatics · QSAR modelling · Random Forest · PubChem fingerprints · ChEMBL

Abstract

BIOPREDICT is an end-to-end quantitative structure–activity relationship (QSAR) modelling framework for predicting the bioactivity (pIC50) of small molecules against a target of interest. The pipeline queries bioactivity data directly from the ChEMBL database, performs exploratory data analysis with rigorous statistical validation, computes PubChem fingerprint descriptors using PaDEL-Descriptor, and trains a Random Forest regression model to predict pIC50 values for query compounds. The framework includes a comprehensive model comparison module using LazyPredict to benchmark the Random Forest against the full suite of available scikit-learn regressors. The trained model is exported as a serialised pickle object for downstream deployment and prediction on novel compounds.

The current implementation demonstrates the pipeline against a Receptor Tyrosine Kinase (RTK) target retrieved from ChEMBL, with IC50-based bioactivity classification and pIC50 prediction.

Key Features
Pipeline Overview
Repository Structure
Data Description
Bioactivity Classification Thresholds
Descriptors and Feature Engineering
Machine Learning Model
Statistical Analysis
Output Files
Installation and Dependencies
Usage
Results
Author

Key Features

Feature	Description
Automated data retrieval	Queries IC50 bioactivity data from ChEMBL via REST API
Activity classification	Labels compounds as active / inactive / intermediate based on IC50 thresholds
Lipinski descriptor calculation	MW, LogP, NumHDonors, NumHAcceptors computed via RDKit
PubChem fingerprints	881-bit PubChem fingerprint descriptors via PaDEL-Descriptor
Statistical validation	Mann-Whitney U test comparing active vs inactive compound properties
Low-variance feature removal	VarianceThreshold-based feature selection before model training
Random Forest QSAR model	Trained on pIC50 with 80/20 train/test split
Comprehensive model comparison	LazyPredict benchmarks 40+ regression algorithms side-by-side
Model serialisation	Trained model exported as `.pkl` for reuse and deployment
Diagnostic visualisations	Scatter plots (experimental vs predicted pIC50), box plots, R² and RMSE bar charts

Pipeline Overview

The workflow is implemented as six sequential Jupyter notebooks:

ChEMBL Database
      │
      ▼
┌─────────────────────────────────────────────────────────────────────┐
│  Notebook 1 — Data Collection                                       │
│  • Query ChEMBL by target name                                      │
│  • Retrieve IC50 bioactivity records                                │
│  • Remove duplicates and missing values                             │
│  • Classify compounds: active / inactive / intermediate             │
│  Output: 1_bioactivity_data_raw.csv → 3_bioactivity_data_curated.csv│
└──────────────────────────┬──────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────────────────┐
│  Notebook 2 — Exploratory Data Analysis                             │
│  • Lipinski descriptors (RDKit): MW, LogP, HBD, HBA                │
│  • IC50 → pIC50 conversion (cap at 100 μM)                         │
│  • Distribution plots, scatter plots, box plots                     │
│  • Mann-Whitney U test: active vs inactive (5 descriptors)          │
│  Output: 4_bioactivity_data_3class_pIC50.csv                        │
│          5_bioactivity_data_2class_pIC50.csv                        │
│          mannwhitneyu_*.csv                                         │
└──────────────────────────┬──────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────────────────┐
│  Notebook 3 — Descriptor Calculation                                │
│  • Export SMILES to molecule.smi                                    │
│  • PaDEL-Descriptor: PubChem fingerprints (881 bits, via padel.sh)  │
│  • Combine fingerprints with pIC50 labels                           │
│  Output: descriptors_output.csv                                     │
│          6_bioactivity_data_3class_pIC50_pubchem_fp.csv             │
└──────────────────────────┬──────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────────────────┐
│  Notebook 4 — Model Building                                        │
│  • VarianceThreshold feature selection (threshold = 0.16)           │
│  • 80/20 train/test split                                           │
│  • Random Forest Regressor (n_estimators=100)                       │
│  • R² evaluation on test set                                        │
│  • Experimental vs predicted pIC50 scatter plot                     │
└──────────────────────────┬──────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────────────────┐
│  Notebook 5 — Model Deployment                                      │
│  • Final model: Random Forest (n_estimators=500, random_state=42)   │
│  • VarianceThreshold feature selection (threshold = 0.1)            │
│  • Trained on full dataset                                          │
│  • MSE and R² evaluation                                            │
│  • Serialised model export: RTK_model.pkl                           │
│  • descriptor_list.csv: selected feature set for prediction         │
└──────────────────────────┬──────────────────────────────────────────┘
                           │
                           ▼
┌─────────────────────────────────────────────────────────────────────┐
│  Notebook 6 — Model Comparison                                      │
│  • LazyPredict: benchmarks 40+ sklearn regressors automatically     │
│  • Reports R², RMSE, and runtime for each algorithm                 │
│  • Bar chart visualisations (R², RMSE, computation time)            │
│  Output: model comparison PDFs                                      │
└─────────────────────────────────────────────────────────────────────┘

Repository Structure

BIOPREDICT/
│
├── 1-Data-Collection.ipynb                      # ChEMBL query, IC50 retrieval, activity labelling
├── 2-Exploratory-Data-Analysis.ipynb            # Lipinski descriptors, pIC50, EDA, Mann-Whitney U
├── 3-Descriptor-Calculation.ipynb               # PaDEL PubChem fingerprints
├── 4-Model_Building.ipynb                       # Random Forest model training and evaluation
├── 5-Model-deployment.ipynb                     # Final model, pickle export, deployment
├── 6-Model-Comparison.ipynb                     # LazyPredict multi-algorithm benchmark
│
├── padel.sh                                     # Shell script to run PaDEL-Descriptor via Java
│
├── 1_bioactivity_data_raw.csv                   # Raw IC50 records from ChEMBL
├── 2_bioactivity_data_preprocessed.csv          # Deduplicated: ChEMBL ID, SMILES, IC50
├── 3_bioactivity_data_curated.csv               # + bioactivity class labels
├── 4_bioactivity_data_3class_pIC50.csv          # + Lipinski descriptors + pIC50 (3 classes)
├── 5_bioactivity_data_2class_pIC50.csv          # Active / inactive only (intermediates removed)
├── 6_bioactivity_data_3class_pIC50_pubchem_fp.csv # PubChem fingerprints + pIC50
│
├── molecule.smi                                 # SMILES input file for PaDEL-Descriptor
├── descriptors_output.csv                       # Full 881-bit PubChem fingerprint matrix
├── descriptor_list.csv                          # Variance-filtered feature subset used for prediction
│
├── RTK_model.pkl                                # Serialised trained Random Forest model
│
├── mannwhitneyu_pIC50.csv                       # Mann-Whitney U: pIC50 (active vs inactive)
├── mannwhitneyu_MW.csv                          # Mann-Whitney U: Molecular Weight
├── mannwhitneyu_LogP.csv                        # Mann-Whitney U: LogP
├── mannwhitneyu_NumHDonors.csv                  # Mann-Whitney U: H-bond donors
├── mannwhitneyu_NumHAcceptors.csv               # Mann-Whitney U: H-bond acceptors
│
└── User_Manual.pdf                              # User manual

Data Description

Bioactivity data is retrieved from the ChEMBL database using the chembl_webresource_client Python library.

Parameter	Value
Target	Receptor Tyrosine Kinase (RTK)
Activity type	IC50 (half-maximal inhibitory concentration)
Data source	ChEMBL database
Query field	`standard_type = "IC50"`
Columns retained	`molecule_chembl_id`, `canonical_smiles`, `standard_value`
Preprocessing	Rows with missing `standard_value` or `canonical_smiles` removed; duplicate SMILES dropped

Bioactivity Classification Thresholds

Compounds are classified into three bioactivity categories based on IC50 values (in nM):

Class	IC50 threshold	Interpretation
Active	≤ 1,000 nM	Potent inhibitors
Intermediate	1,000 – 10,000 nM	Moderate activity; excluded from binary model training
Inactive	≥ 10,000 nM	Non-inhibitors

IC50 values are capped at 100,000,000 nM (100 μM) before conversion to pIC50 to prevent infinite or extreme values.

IC50 → pIC50 conversion:

pIC50 = −log₁₀(IC50 in M)

Descriptors and Feature Engineering

Lipinski Descriptors (RDKit)

Computed in Notebook 2 for exploratory analysis:

Descriptor	Definition
`MW`	Molecular weight (Da)
`LogP`	Octanol–water partition coefficient
`NumHDonors`	Number of hydrogen bond donors
`NumHAcceptors`	Number of hydrogen bond acceptors

PubChem Fingerprints (PaDEL-Descriptor)

Computed in Notebook 3 via padel.sh:

Fingerprint type: PubChem fingerprints (PubchemFingerprinter.xml)
Dimensionality: 881 binary bits per compound
Preprocessing: Salt removal (-removesalt), nitro group standardisation (-standardizenitro)
Tool: PaDEL-Descriptor (Java, ≥ JRE 1.8)

# padel.sh — executed internally by Notebook 3
java -Xms1G -Xmx1G -Djava.awt.headless=true \
  -jar ./PaDEL-Descriptor/PaDEL-Descriptor.jar \
  -removesalt -standardizenitro -fingerprints \
  -descriptortypes ./PaDEL-Descriptor/PubchemFingerprinter.xml \
  -dir ./ -file descriptors_output.csv

Feature Selection

Low-variance features are removed using sklearn.feature_selection.VarianceThreshold:

Notebook	Threshold	Use
Notebook 4 (training)	0.16 (≈ 80/20 variance)	Model building and test evaluation
Notebook 5 (deployment)	0.10	Final production model and `descriptor_list.csv`

The selected feature names are saved to descriptor_list.csv and must be used consistently when predicting on new compounds.

Machine Learning Model

Algorithm

Random Forest Regressor (sklearn.ensemble.RandomForestRegressor)

Parameter	Notebook 4 (exploration)	Notebook 5 (deployment)
`n_estimators`	100	500
`random_state`	—	42
Training set	80%	Full dataset
Test set	20%	Full dataset (training evaluation)

Prediction target

pIC50 — continuous regression output. Higher pIC50 indicates greater predicted potency.

Model file

The final trained model is serialised using Python's pickle module:

import pickle
model = pickle.load(open('RTK_model.pkl', 'rb'))
predictions = model.predict(X_new)

Note: Input features for prediction must match the descriptor columns in descriptor_list.csv and be computed using the same PaDEL PubChem fingerprint settings as the training data.

Statistical Analysis

Mann-Whitney U Test

Non-parametric statistical comparison between active and inactive compound classes is performed in Notebook 2 for five descriptors. Results are saved individually as CSV files:

File	Descriptor tested	Null hypothesis
`mannwhitneyu_pIC50.csv`	pIC50	Active and inactive have the same pIC50 distribution
`mannwhitneyu_MW.csv`	Molecular weight	Same MW distribution
`mannwhitneyu_LogP.csv`	LogP	Same LogP distribution
`mannwhitneyu_NumHDonors.csv`	H-bond donors	Same HBD distribution
`mannwhitneyu_NumHAcceptors.csv`	H-bond acceptors	Same HBA distribution

Significance threshold: α = 0.05. Results report the test statistic, p-value, and interpretation (reject / fail to reject H₀).

Output Files

File	Description
`1_bioactivity_data_raw.csv`	Raw ChEMBL IC50 records
`2_bioactivity_data_preprocessed.csv`	Cleaned: ChEMBL ID, SMILES, IC50
`3_bioactivity_data_curated.csv`	+ activity class labels (active/inactive/intermediate)
`4_bioactivity_data_3class_pIC50.csv`	+ Lipinski descriptors + pIC50 (3 classes)
`5_bioactivity_data_2class_pIC50.csv`	Active/inactive only; intermediates removed
`6_bioactivity_data_3class_pIC50_pubchem_fp.csv`	PubChem fingerprints (881-bit) + pIC50
`descriptors_output.csv`	Full PaDEL output before feature selection
`descriptor_list.csv`	Variance-filtered feature names for deployment
`molecule.smi`	SMILES file used as PaDEL input
`RTK_model.pkl`	Serialised Random Forest model (pickle)
`mannwhitneyu_*.csv`	Mann-Whitney U test results (5 files)

Installation and Dependencies

Python packages

Install all required packages via pip:

pip install chembl_webresource_client rdkit pandas numpy scikit-learn seaborn matplotlib lazypredict

Or using conda:

conda install -c conda-forge rdkit
pip install chembl_webresource_client lazypredict scikit-learn seaborn matplotlib pandas numpy

Package	Purpose
`chembl_webresource_client`	ChEMBL database query
`rdkit`	Lipinski descriptor calculation, SMILES processing
`pandas` / `numpy`	Data manipulation
`scikit-learn`	Random Forest, VarianceThreshold, train/test split, metrics
`seaborn` / `matplotlib`	Visualisation
`lazypredict`	Multi-algorithm model comparison (Notebook 6)
`pickle`	Model serialisation

PaDEL-Descriptor

PaDEL-Descriptor requires Java Runtime Environment (JRE) ≥ 1.8.

Download PaDEL-Descriptor and place the PaDEL-Descriptor/ directory in the repository root.

Install Java:

# Ubuntu / Debian
sudo apt install -y default-jre

# macOS (Homebrew)
brew install --cask temurin

Verify Java installation:
```
java -version
```

Update the Java path in padel.sh if your JRE is not at the default location:

# Example path — adjust to your system
/home/rashid/Downloads/jre-8u333-linux-x64/jre1.8.0_333/bin/java \
  -Xms1G -Xmx1G ...

Usage

Run the notebooks in sequential order:

jupyter notebook

Step	Notebook	Action
1	`1-Data-Collection.ipynb`	Query ChEMBL, retrieve IC50, classify compounds
2	`2-Exploratory-Data-Analysis.ipynb`	Lipinski EDA, pIC50 conversion, Mann-Whitney U test
3	`3-Descriptor-Calculation.ipynb`	Run PaDEL fingerprints via `padel.sh`
4	`4-Model_Building.ipynb`	Train and evaluate Random Forest on 80/20 split
5	`5-Model-deployment.ipynb`	Train final model, export `RTK_model.pkl`
6	`6-Model-Comparison.ipynb`	Benchmark all sklearn regressors with LazyPredict

Adapting to a different target

To run BIOPREDICT against a different ChEMBL target:

In 1-Data-Collection.ipynb, modify the target search query:
```
target_query = target.search('YOUR_TARGET_NAME')
```
Select the appropriate target from the results dataframe and update the index accordingly.
Re-run all notebooks in order. All intermediate CSV files will be regenerated automatically.

Predicting on new compounds

To predict pIC50 for new query compounds using the trained model:

Prepare a SMILES file of your query compounds.
Run PaDEL-Descriptor using padel.sh to generate PubChem fingerprints.
Load descriptor_list.csv to align the feature columns:

import pickle, pandas as pd

# Load model and descriptor list
model = pickle.load(open('RTK_model.pkl', 'rb'))
desc_list = pd.read_csv('descriptor_list.csv')

# Load your query fingerprints and align columns
query_fp = pd.read_csv('your_query_descriptors.csv')
query_fp = query_fp[desc_list.columns]

# Predict pIC50
predictions = model.predict(query_fp)
print(predictions)

Results

Model performance summary

Metric	Value
Algorithm	Random Forest Regressor
Fingerprint type	PubChem (881 bits)
Feature selection	VarianceThreshold (threshold = 0.1)
Training set	Full dataset (Notebook 5)
Evaluation metric	R², MSE

Detailed performance metrics and comparison plots are generated within Notebooks 4, 5, and 6.

Author

Rashid Hussain, PhD, RSci, MRSC
Postdoctoral Researcher in Computational Pathology
Humanitas Research Hospital (IRCCS), Milan, Italy
rashid.bioinfo@gmail.com · https://rashid-bioinfo.github.io

QSAR · Cheminformatics · Machine Learning for Drug Discovery

Navigation Menu

BIOPREDICT

Abstract

Table of Contents

Key Features

Pipeline Overview

Repository Structure

Data Description

Bioactivity Classification Thresholds

Descriptors and Feature Engineering

Lipinski Descriptors (RDKit)

PubChem Fingerprints (PaDEL-Descriptor)

Feature Selection

Machine Learning Model

Algorithm

Prediction target

Model file

Statistical Analysis

Mann-Whitney U Test

Output Files

Installation and Dependencies

Python packages

PaDEL-Descriptor

Usage

Adapting to a different target

Predicting on new compounds

Results

Model performance summary

Author

FilesExpand file tree

README.md

Latest commit

History

README.md

File metadata and controls

BIOPREDICT

Abstract

Table of Contents

Key Features

Pipeline Overview

Repository Structure

Data Description

Bioactivity Classification Thresholds

Descriptors and Feature Engineering

Lipinski Descriptors (RDKit)

PubChem Fingerprints (PaDEL-Descriptor)

Feature Selection

Machine Learning Model

Algorithm

Prediction target

Model file

Statistical Analysis

Mann-Whitney U Test

Output Files

Installation and Dependencies

Python packages

PaDEL-Descriptor

Usage

Adapting to a different target

Predicting on new compounds

Results

Model performance summary

Author