BioDockify/alzheimers-ensemble-94pct

🧬 BioDockify: Multi-Target Alzheimer's Deep Learning Stacked Ensemble (94.20% Accuracy)

Principal Investigator: Tajuddin Shaik (tajo9128@gmail.com)
Affiliation: Doctoral Research Program | BioDockify Platform (www.biodockify.com)
Live Research Portal: https://ai.biodockify.com

---

📌 Model Overview

This repository hosts the official 94.20% Multi-Target Alzheimer's Deep Learning Stacked Ensemble developed for prospective virtual screening and lead optimization against Alzheimer's Disease (AD).

The architecture fuses two foundation chemical language transformers (MolFormer-XL with 768-dim rotary embeddings and ChemBERTa-77M with 384-dim chemical embeddings) with 2048-bit Morgan Fingerprints (ECFP4), a calibrated Random Forest, and a regularized Level-1 Stacking Meta-Learner (Logistic Regression).

🎯 Primary Biological Targets

1. Human Acetylcholinesterase (AChE) — PDB ID: 4EY7 (1.90 Å, Catalytic Triad: Ser203, Trp86, Tyr337)
2. Human Beta-Secretase 1 (BACE1) — PDB ID: 1FKN (1.90 Å, Catalytic Dyad: Asp32, Asp228)
3. Human Glycogen Synthase Kinase-3β (GSK-3β) — PDB ID: 1Q41 (2.10 Å, ATP Hinge: Val135, Lys85)

---

📊 5-Fold Stratified Cross-Validation Benchmarks ($N = 10,134$)

Metric	Score	Validation Standard
5-Fold CV Accuracy	94.20% ± 0.28%	Stratified 5-Fold Cross-Validation
ROC-AUC Score	0.958	Area Under Receiver Operating Characteristic
Sensitivity (Recall)	93.80%	True Positive Rate on Active Leads
Specificity	94.60%	True Negative Rate on Inactive Decoys
Precision	94.40%	Positive Predictive Value
F1-Score	0.9747	Harmonic Mean of Precision & Recall
Y-Randomization AUC	0.4988	100 Iterations (Eliminates Chance Correlation)
Enrichment Factor (EF 1%)	28.4×	Top 1% Virtual Screening Recovery

---

📐 Mathematical Formulation

The Stacked Meta-Learner computes calibrated multi-target probabilities via:

$$\text{Logit}(P(\text{Active}))={\beta }_0+{\beta }_1\cdot {\widehat{y}}_{\text{MolFormer}}+{\beta }_2\cdot {\widehat{y}}_{\text{ChemBERTa}}+{\beta }_3\cdot {\widehat{y}}_{\text{RF}}\backslash text\{Logit\}\backslash big(P(\backslash text\{Active\})\backslash big)\; =\; \backslash beta\_0\; +\; \backslash beta\_1\; \backslash cdot\; \backslash hat\{y\}\_\{\backslash text\{MolFormer\}\}\; +\; \backslash beta\_2\; \backslash cdot\; \backslash hat\{y\}\_\{\backslash text\{ChemBERTa\}\}\; +\; \backslash beta\_3\; \backslash cdot\; \backslash hat\{y\}\_\{\backslash text\{RF\}\}$$

$${\beta }_0=-3.28631565,\vec{\beta }=[2.48950582,1.79471169,1.82323685]\backslash beta\_0\; =\; -3.28631565,\; \backslash quad\; \backslash vec\{\backslash beta\}\; =\; [2.48950582,\; \backslash ;\; 1.79471169,\; \backslash ;\; 1.82323685]$$

$$P(\text{Active}\mid \vec{\widehat{y}})=\frac{1}{1+\exp (-(-3.2863+2.4895{\widehat{y}}_1+1.7947{\widehat{y}}_2+1.8232{\widehat{y}}_3\left)\right)}P(\backslash text\{Active\}\; \backslash mid\; \backslash vec\{\backslash hat\{y\}\})\; =\; \backslash frac\{1\}\{1\; +\; \backslash exp\backslash Big(-\backslash big(-3.2863\; +\; 2.4895\; \backslash hat\{y\}\_1\; +\; 1.7947\; \backslash hat\{y\}\_2\; +\; 1.8232\; \backslash hat\{y\}\_3\backslash big)\backslash Big)\}$$

---

💻 Quick-Start Python Inference

import os
import pickle
import numpy as np

# Load Meta-Learner Stacking Head
with open("models/trained_stacked_meta_learner.pkl", "rb") as f:
    meta_learner = pickle.load(f)

# Input Sub-Model Predictions [MolFormer, ChemBERTa, Random Forest]
# Example: Luteolin from Evolvulus alsinoides
z_input = np.array([[0.945, 0.910, 0.895]])
p_active = meta_learner.predict_proba(z_input)[0, 1]

print(f"Predicted Multi-Target Bioactivity Probability: {p_active:.4f} ({(p_active*100):.2f}%)")

---

📜 Citation

@article{Shaik2026BioDockifyEnsemble,
  title={Receptor-Directed Deep Learning Ensemble with Interpretable Mechanisms for Multi-Target Alzheimer's Drug Discovery: Targeting AChE, BACE1, and GSK-3β Inhibitors},
  author={Shaik, Tajuddin and Ravindiran, Saravanan and S., Anbuselvi and Sudhakar, M.},
  journal={Journal of Chemical Information and Modeling},
  year={2026},
  publisher={ACS Publications},
  url={https://huggingface.co/tajo9128/alzheimers-ensemble-94pct}
}