ChEMBL ML Predictor (Streamlit)

Streamlit web app that fetches ChEMBL bioactivity data for a target (via UniProt ID), computes molecular descriptors, and trains a simple predictive model (regression, with an optional classification fallback).

What it does

• Input: a UniProt ID (e.g. Q9UNA0).
• Fetches:
  • HGNC gene symbol (via genenames.org)
  • ChEMBL target, assays, and activities (via ChEMBL API)
• Builds a dataset with:
  • pchembl_value as the main label
  • RDKit physicochemical descriptors (e.g. MW, LogP, HBD/HBA, TPSA, rings, rotatable bonds)
  • RDKit fragment descriptors (fr_*)
  • Optional ErG-style features using FixedPharmacophoreAnalyzer from erg_calc_fragments_topo.py
• Trains:
  • Regression: Random Forest (default) or XGBoost (if installed)
  • Optional fallback: binary classification (Active if pChEMBL ≥ 6.5) with optional SMOTE
• Exports:
  • Pickled model + preprocessing objects
  • Predictions CSV
  • Top-features JSON
  • Full dataset CSV

Requirements

Core:

• Python 3.9+
• streamlit, pandas, numpy, requests, scikit-learn, matplotlib, seaborn

Optional:

• RDKit (required to compute descriptors; app will stop without it)
• xgboost (enables XGBoost option)
• optuna (enables hyperparameter search)
• imbalanced-learn (enables SMOTE for classification fallback)

Install

Example (pip):

pip install streamlit pandas numpy requests scikit-learn matplotlib seaborn

RDKit installation depends on your platform (common options are conda-forge or a prebuilt wheel).

Run

From this folder:

streamlit run .\chembl_ml_app.py

Read a generated PKL from output

The analysis pipeline in erg_eda_enhanced_with_optuna.py writes per-target pickle files under the run output folder, for example:

output/erg_analysis_inhibition_activation_YYYYMMDD_HHMMSS/models/inhibition/AKT1_results.pkl

You can load one of those files with:

import pickle
from pathlib import Path

pkl_path = Path("output/erg_analysis_inhibition_activation_YYYYMMDD_HHMMSS/models/inhibition/AKT1_results.pkl")

with pkl_path.open("rb") as handle:
  results = pickle.load(handle)

print(results["target"])
print(results["n_samples"])
print(results["regression"].keys())
print(results["classification"].keys())

Typical top-level keys in the loaded object are:

• target
• threshold
• n_samples
• regression
• classification
• plots

Example access to a saved regression model result:

rf_regression = results["regression"].get("random_forest")

if rf_regression:
  print("R2:", rf_regression["r2"])
  print("RMSE:", rf_regression["rmse"])
  print("Top features:", rf_regression["top_features"][:5])

Notes / troubleshooting

• ChEMBL API calls are rate-limited in the app; large targets may take time.
• If you see runtime errors when starting the app, check that the selected Python environment has the required packages installed.
• The ErG analyzer is loaded from erg_calc_fragments_topo.py and is optional.