A machine-learning model for predicting whether a chemical compound is likely to be active or inactive against EGFR based on its molecular structure represented as a SMILES string.

The model uses RDKit molecular descriptors and Morgan molecular fingerprints as input features and a Random Forest classifier for binary classification.

Research use only: This model is intended for educational, research, and compound-prioritization purposes. It is not a clinical diagnostic tool and must not be used to make medical, therapeutic, regulatory, or patient-care decisions.

Model Overview

Property	Value
Target	EGFR
ChEMBL Target ID	CHEMBL203
Assay endpoint	IC50
Activity unit	nM
Task	Binary classification
Active threshold	IC50 ≤ 1000 nM
Inactive threshold	IC50 > 1000 nM
Model	Random Forest Classifier
Molecular descriptors	8
Morgan fingerprint	2048 bits
Total features	2056

Objective

The objective of this project is to develop an end-to-end machine-learning pipeline that takes a compound SMILES string and predicts:

ACTIVE
INACTIVE

against the EGFR target.

The project combines:

Chemical data curation
Cheminformatics
Molecular descriptors
Molecular fingerprints
Machine learning
Model evaluation
Scaffold-based validation
Interactive Streamlit deployment

Input

The model accepts a valid chemical SMILES representation.

Example:

CCO

or:

CCOC1=CC=CC=C1

The SMILES is converted into an RDKit molecular representation before feature generation.

Feature Engineering

Each molecule is represented using two feature groups.

Molecular descriptors

The following eight descriptors are calculated using RDKit:

Molecular Weight
LogP
Hydrogen Bond Donors
Hydrogen Bond Acceptors
Rotatable Bonds
Topological Polar Surface Area
Ring Count
Heavy Atom Count

Morgan fingerprint

A Morgan circular fingerprint is generated using:

text

Radius: 2
Fingerprint size: 2048 bits

Therefore:

text

8 descriptors
+
2048 Morgan fingerprint bits
=
2056 features

Activity Definition

The curated dataset uses an IC50 threshold of:

1000 nM

Classification:

text

IC50 ≤ 1000 nM
        ↓
     ACTIVE

text

IC50 > 1000 nM
        ↓
    INACTIVE

This threshold is a project-level classification rule and should not be interpreted as a universal biological definition of EGFR activity.

Dataset

The final curated dataset contains:

13,577 compounds

Class distribution:

Class	Compounds	Fraction
ACTIVE	9,142	67.33%
INACTIVE	4,435	32.67%
Total	13,577	100%

The data curation pipeline includes:

Removal of invalid measurements
IC50 filtering
Unit normalization
Missing-value handling
SMILES validation
Duplicate measurement removal
Compound-level aggregation
Activity classification

Model Comparison

Three machine-learning models were evaluated.

Model	Accuracy	Precision	Recall	F1	ROC-AUC	PR-AUC
Logistic Regression	0.8384	0.8913	0.8655	0.8782	0.8776	0.9193
Random Forest	0.8921	0.9364	0.9010	0.9184	0.9512	0.9743
XGBoost	0.8840	0.9079	0.9213	0.9145	0.9410	0.9690

The Random Forest model was selected based primarily on ROC-AUC.

Random-Split Evaluation

The selected Random Forest model achieved:

Metric	Score
Accuracy	0.8921
Precision	0.9364
Recall	0.9010
F1	0.9184
ROC-AUC	0.9512
PR-AUC	0.9743

Five-fold cross-validation produced:

text

Mean ROC-AUC: 0.9478
Std ROC-AUC : 0.0054

Scaffold-Split Evaluation

A Bemis-Murcko scaffold split was additionally performed to evaluate generalization to chemically different molecular scaffolds.

Results:

Metric	Score
Accuracy	0.8401
Precision	0.8760
Recall	0.8884
F1	0.8822
ROC-AUC	0.9118
PR-AUC	0.9551

The scaffold split is more challenging than a conventional random split because structurally related compounds are less likely to appear across both training and test sets.

Generalization

The difference between random-split and scaffold-split ROC-AUC was:

text

Random split ROC-AUC   = 0.9512
Scaffold split ROC-AUC = 0.9118

Generalization gap     = 0.0394

The scaffold evaluation provides a more conservative estimate of how the model may perform on chemically distinct compounds.

Model Files

This repository contains:

text

bioactivity_model.joblib
model_metadata.json
README.md

`bioactivity_model.joblib`

Serialized scikit-learn Random Forest model.

`model_metadata.json`

Contains model configuration and reproducibility metadata including:

Target information
Activity threshold
Feature configuration
Molecular descriptors
Morgan fingerprint configuration
Validation metrics
Dataset statistics
Software versions

Example Inference

A compound can be processed using the project's prediction wrapper.

Example:

python

from src.predict import BioactivityPredictor

predictor = BioactivityPredictor(
    "models/bioactivity_model.joblib"
)

result = predictor.predict(
    "CCOC1=CC=CC=C1"
)

print(result["prediction"])
print(result["active_probability"])

Example output:

text

ACTIVE
0.874

The probability represents the model's estimated probability for the predicted class and should not be interpreted as a calibrated probability of biological efficacy.

Application

The model is integrated into an interactive Streamlit application.

The application supports:

SMILES input
SMILES validation
Molecular property calculation
ACTIVE/INACTIVE prediction
Prediction probability
CSV batch prediction
Molecular descriptors
Chemical structure visualization

Technology Stack

Programming

Python 3.14

Cheminformatics

RDKit

Data Science

Pandas
NumPy
SciPy

Machine Learning

Scikit-learn
XGBoost
Joblib

Visualization

Matplotlib
Seaborn

Application

Streamlit

Development

JupyterLab
Pytest
Ruff
Git
GitHub
Hugging Face Hub

Project Architecture

text

SMILES
   │
   ▼
RDKit Molecular Parsing
   │
   ├───────────────┐
   │               │
   ▼               ▼
Molecular       Morgan
Descriptors     Fingerprint
   │               │
   └───────┬───────┘
           ▼
      2056 Features
           │
           ▼
    Random Forest
           │
           ▼
 ┌─────────┴─────────┐
 │                   │
 ▼                   ▼
ACTIVE            INACTIVE
 │
 ▼
Probability

Limitations

This model has several important limitations.

1. Target-specific model

The model was trained for:

EGFR / CHEMBL203

It should not automatically be assumed to generalize to other biological targets.

2. Endpoint-specific model

The training endpoint is:

IC50

The model should not be interpreted as a predictor for other assay endpoints such as:

EC50
Ki
Kd
cellular viability
toxicity
clinical response

without appropriate retraining and validation.

3. Dataset bias

Bioactivity databases contain heterogeneous experimental measurements generated using different:

assay protocols
experimental conditions
laboratories
compound series
measurement practices

Therefore, the model may inherit biases present in the underlying data.

4. Chemical space limitations

Performance can decrease for molecules substantially different from the training distribution.

The scaffold-split ROC-AUC of approximately 0.912 demonstrates that performance is lower under a more challenging chemical split than under a random split.

5. Probability calibration

The predicted probability is a machine-learning model output and has not been established as a clinically calibrated probability of biological activity.

6. No clinical interpretation

The model does not predict:

drug approval
therapeutic efficacy
patient response
toxicity
pharmacokinetics
pharmacodynamics
clinical safety

Intended Use

Appropriate uses include:

Educational cheminformatics projects
Machine-learning research
Molecular data analysis
Computational compound prioritization
Benchmarking molecular fingerprints
Exploring structure-activity relationships
Prototyping drug-discovery workflows

Out-of-Scope Use

This model should not be used as the sole basis for:

Clinical decisions
Drug prescribing
Patient treatment
Regulatory submissions
Safety-critical decisions
Confirmation of biological activity
Replacement of laboratory experiments

Predictions should be experimentally validated before drawing biological conclusions.

Reproducibility

The project includes:

Data curation scripts
Feature generation
Model training
Model evaluation
Scaffold evaluation
Automated tests
Model metadata
Evaluation reports

The software environment is documented in the project requirements files.

Repository

Source code:

https://github.com/sintu7781/drug-bioactivity-predictor

Hugging Face model:

https://huggingface.co/prime7781/drug-bioactivity-predictor

Disclaimer

This project is a research and educational machine-learning application.

Predictions are computational estimates and do not constitute experimental evidence of biological activity, medical advice, or clinical recommendations.

Any compound prioritization based on this model should be followed by appropriate experimental validation.

Future Improvements

Potential future improvements include:

Probability calibration
External validation on independent datasets
Larger multi-target datasets
Additional molecular fingerprints
Graph neural networks
Molecular transformers
Explainable AI
Applicability-domain estimation
Uncertainty quantification
Active learning
Multi-task bioactivity prediction
Experimental validation