ஓம் சரவணபவ

🧬 AChE Molecular Machine Learning

Structure-aware prediction of acetylcholinesterase inhibitory potency

---

📑 Contents

• 🔬 Project Overview
• 🎯 Objective
• 🧪 Dataset
• 🧹 Data Preparation
• 🧬 Molecular Representations
• 🧩 Scaffold-Aware Partitioning
• 🤖 Classical Machine Learning
• ⚙️ Hyperparameter Optimization
• 📏 Evaluation
• 📊 Experimental Design
• 🔬 Graph Neural Network
• 🛠️ Technology Stack
• 🚧 Current Status
• 🔭 Future Direction
• 📚 Technical Documentation
• 📖 Citation
• 👤 Author

---

🔬 Project Overview

This project develops a structure-aware molecular machine-learning pipeline for predicting the inhibitory potency of compounds against acetylcholinesterase (AChE).

The current implementation establishes a classical QSAR/ML benchmark using:

• ChEMBL bioactivity data
• Molecular standardization and quality control
• IC50 normalization and pIC50 transformation
• Morgan circular fingerprints
• RDKit molecular descriptors
• Scaffold-based molecular partitioning
• Grouped cross-validation
• Random Forest
• XGBoost
• LightGBM
• Optuna hyperparameter optimization
• Multi-metric evaluation and residual analysis
• Implemented GNNs

Research direction: move from fixed molecular feature representations toward increasingly structure-aware and scientifically constrained molecular learning.

---

🎯 Objective

Primary Task

Predict continuous pIC50 values from molecular structure.

$$ pIC_{50} = -\log_{10}(IC_{50}[M]) $$

The learning problem is formulated as:

Molecular Structure
        ↓
Molecular Representation
        ↓
Machine Learning Model
        ↓
Predicted pIC50

Core Questions

Question	Purpose
How informative are circular molecular fingerprints?	Evaluate local structural representation
What additional information do molecular descriptors provide?	Evaluate physicochemical representation
Does combining both representations improve prediction?	Compare complementary feature spaces
How does structural separation affect performance?	Test molecular generalization
How do classical models compare under the same representation?	Establish reliable baselines
Can graph representations improve molecular learning?	GNN extension

---

🧪 Dataset

Property	Configuration
Data source	ChEMBL
Target	Acetylcholinesterase
ChEMBL Target ID	CHEMBL220
Activity endpoint	IC50
Learning target	pIC50
Task	Continuous regression

Activity records are retrieved for the AChE target with available IC50 measurements.

Raw activity data are processed before being passed to the learning algorithms.

---

🧹 Data Preparation

The preprocessing pipeline converts raw bioactivity records into a consistent molecular-level regression dataset.

Processing flow

ChEMBL Activity Records
          ↓
Data Validation
          ↓
Salt Removal
          ↓
Molecular Identity / InChIKey
          ↓
Duplicate Handling
          ↓
IC50 Unit Normalization
          ↓
pIC50 Transformation
          ↓
Replicate Aggregation
          ↓
Activity Filtering
          ↓
Final Molecular Dataset

Molecular cleaning

The current pipeline performs:

• SMILES validation
• salt removal
• canonical molecular representation
• InChIKey generation
• duplicate handling
• activity-value validation

IC50 normalization

Supported source representations are normalized to a common concentration representation before pIC50 calculation.

$$ pIC_{50}=-\log_{10}(IC_{50}[M]) $$

Replicate measurements

Multiple measurements corresponding to the same molecular identity are grouped using InChIKey.

The current implementation calculates the median pIC50 and uses group-level median absolute deviation as a quality-control criterion.

This step is treated as part of the dataset definition, rather than as a modeling operation.

---

🧬 Molecular Representations

Two complementary engineered representations form the current classical ML feature space.

1. Morgan Fingerprints

Parameter	Value
Representation	Morgan circular fingerprint
Radius	2
Fingerprint size	1024 bits

Molecular Structure
       ↓
Circular Atom Environments
       ↓
Morgan Fingerprint
       ↓
1024-bit Vector

2. RDKit Molecular Descriptors

The current descriptor representation captures molecular physicochemical and structural properties including:

• Molecular weight
• MolLogP
• H-bond donors
• H-bond acceptors
• TPSA
• Partial-charge descriptors
• Heteroatom count
• Rotatable bonds
• Fraction Csp3
• Aromatic ring count
• Ring count
• Aliphatic nitrogen count
• Formal charge

Feature configurations

Configuration	Input
Morgan	1024-bit Morgan fingerprint
RDKit	Molecular descriptors
Combined	Morgan + RDKit descriptors

---

🧩 Scaffold-Aware Partitioning

A central component of the project is structure-aware evaluation.

Random molecular splits can distribute closely related compounds between training and test sets. This can make a model appear to generalize when it is benefiting from structurally similar compounds during training.

The pipeline therefore groups molecules according to their Bemis–Murcko scaffold and assigns scaffold groups to dataset partitions.

Molecules
    ↓
Murcko Scaffold
    ↓
Scaffold Groups
    ↓
Training / Test Partition

Molecules sharing a scaffold are kept within the same partition.

---

🤖 Classical Machine Learning

The current benchmark consists of three tree-based ensemble models.

Model	Role
Random Forest	Non-parametric ensemble baseline
XGBoost	Gradient-boosted tree model
LightGBM	Gradient-boosted tree model

Each model is evaluated across the same molecular representation configurations:

                 Morgan     RDKit     Combined
                    │          │          │
Random Forest       ✓          ✓          ✓
XGBoost             ✓          ✓          ✓
LightGBM            ✓          ✓          ✓

This provides a controlled comparison between model family and molecular representation.

---

⚙️ Hyperparameter Optimization

Hyperparameter selection is performed using Optuna.

Training Data
      ↓
Scaffold-Grouped Cross-Validation
      ↓
Model Evaluation
      ↓
Validation MAE
      ↓
Optuna Trials
      ↓
Best Hyperparameters
      ↓
Final Training

The current implementation uses 3-fold GroupKFold, with scaffold identity used as the grouping variable.

The optimization objective is based on mean absolute error.

---

📏 Evaluation

Metric	Interpretation
MAE	Mean absolute prediction error
RMSE	Error with stronger penalty for large deviations
R²	Variance explained relative to the target mean
Pearson r	Linear association
Spearman ρ	Rank-order association

Residuals are calculated as:

$$ e_i = y_i - \hat{y}_i $$

This enables analysis of:

• systematic prediction bias
• high-error compounds
• activity-range dependence
• residual distributions
• prediction calibration
• model failure regions

---

📊 Experimental Design

Representation comparison

Morgan
   │
   ├── Random Forest
   ├── XGBoost
   └── LightGBM

RDKit Descriptors
   │
   ├── Random Forest
   ├── XGBoost
   └── LightGBM

Morgan + RDKit
   │
   ├── Random Forest
   ├── XGBoost
   └── LightGBM

Validation hierarchy

                Full Dataset
                     │
                     ▼
          Scaffold-Based Split
              ┌──────┴──────┐
              ▼             ▼
           Training        Test
              │
              ▼
       Grouped 3-Fold CV
              │
              ▼
       Optuna Optimization
              │
              ▼
       Final Model Training
              │
              ▼
        Held-Out Evaluation

The held-out test partition is reserved for final evaluation rather than hyperparameter selection.

---

Graph Neural Network

Classical models receive fixed fingerprints or descriptor vectors. The GNN represents each molecule as a chemical graph and learns atom embeddings by passing bond-conditioned messages between neighboring atoms. In this study the learned graph representation is combined with molecular descriptors, so the evaluated model is a graph-plus-descriptor model, not a graph-only model.

The GNN uses the same scaffold-held-out test partition as the matched classical comparison described in Experimental Design. Its implementation and saved outputs are linked under Reproducibility Artifacts.

Molecular Graph

For a molecule, the graph is $G=(V,E)$, where $V$ is the set of atoms and $E$ is the set of bonds. The implementation encodes each atom with seven categorical features:

Atom feature	Encoded information
Atomic number	Element identity
Degree	Number of directly bonded atoms
Formal charge	Charge, clipped to a supported range and shifted to a nonnegative category
Hybridization	Supported RDKit hybridization states
Aromaticity	Whether the atom is aromatic
Hydrogen count	Number of attached hydrogens
Ring membership	Whether the atom belongs to a ring

Each bond has 13 edge channels: four one-hot bond types (single, double, triple, aromatic), three bond properties (conjugated, aromatic, ring), and six one-hot stereochemical states. Every bond is added in both directions so messages can travel between either pair of neighboring atoms.

Message Passing

Each of the seven atom categories is mapped through its own embedding layer. Those embeddings are concatenated and projected into a 128-dimensional hidden state. At each of three message-passing steps, the source atom state is combined with the 13 bond channels and transformed into a message. Incoming messages are summed and divided by the destination atom's degree, then passed to a GRU-based update. A residual connection, dropout, and layer normalization produce the next atom state.

Atom categories ──> embeddings ──> node projection
                                           │
Bond channels ──> edge-conditioned messages
                                           │
                   degree-normalized aggregation
                                           │
                      GRU update + residual
                                           │
                    dropout + layer normalization
                                           │
                          repeat 3 times

Graph Readout and Prediction

After message passing, node states are pooled per molecule using both mean and max pooling. The two graph-level vectors are concatenated with 14 RDKit descriptors. Descriptor imputation and standardization are fitted on the training partition only; target values are also standardized for training and predictions are returned to the original pIC50 scale.

The combined vector is passed through a multilayer regression head with ReLU activations and dropout to predict pIC50. This makes the role of the two input sources explicit:

Molecular graph ──> message passing ──> mean + max pooled graph vector ──┐
                                                                                  ├─> regression head ──> pIC50
14 RDKit descriptors ──> train-fitted imputation / scaling ────────────┘

Architecture and Training

Component	Configuration
Atom features	7 categorical channels
Bond features	13 channels; each bond is bidirectional
Global molecular descriptors	14 RDKit descriptors
Hidden size	128
Message-passing steps	3
Graph pooling	Mean + max
Node update	GRU cell with residual connection
Regularization	Dropout 0.15 + LayerNorm
Regression loss	Smooth L1
Optimizer	AdamW; learning rate 0.001; weight decay 0.0001
Batch size	64
Gradient clipping	5.0
Early stopping	Validation MAE; patience 20 epochs
Random seed	50
Hardware	CPU

The scaffold-disjoint partitions contain 3,108 training, 548 validation, and 915 test molecules. The best validation MAE was 0.6507 at epoch 48; early stopping ended training after 68 epochs. No molecules were rejected during graph construction for this run.

Matched Results

The table reports held-out metrics from the matched audit. Every model is evaluated on the same 915-molecule test partition; the test set is excluded from hyperparameter selection and training.

Model	Representation	Test MAE	RMSE	R²	Pearson r
XGBoost	Morgan + RDKit	0.6352	0.8347	0.4553	0.6795
XGBoost	Morgan	0.6510	0.8519	0.4326	0.6648
Random Forest	Morgan + RDKit	0.6764	0.8507	0.4341	0.6644
Random Forest	Morgan	0.6852	0.8631	0.4175	0.6506
GNN	Graph + RDKit	0.6908	0.9093	0.3535	0.6225
XGBoost	RDKit	0.6969	0.8892	0.3818	0.6188
Random Forest	RDKit	0.7169	0.8978	0.3698	0.6140

On this split, the best classical model's MAE is 0.0556 pIC50 lower than the GNN's. This is a result for this dataset, scaffold split, graph encoding, descriptor augmentation, and training configuration; it is not evidence that GNNs are generally inferior. The audit recommends repeated scaffold splits or an external test set for a more stable estimate.

Reproducibility Artifacts

• GNN implementation
• Architecture figure
• Matched classical/GNN audit report
• GNN run metrics
• Training history
• Held-out test predictions
• All GNN predictions

🛠️ Technology Stack

Category	Technologies
Language	Python
Bioactivity	ChEMBL
Cheminformatics	RDKit
Data	NumPy · Pandas · DuckDB
Classical ML	scikit-learn · XGBoost · LightGBM
Optimization	Optuna
Statistics	SciPy
Visualization	Matplotlib · Seaborn
Experiment Tracking	MLflow
Graph ML	PyTorch · PyTorch Geometric

---

🚧 Current Status

Component	Status
ChEMBL data retrieval	✅
AChE / CHEMBL220 dataset	✅
Molecular cleaning	✅
IC50 normalization	✅
pIC50 transformation	✅
Replicate handling	✅
Morgan fingerprints	✅
RDKit descriptors	✅
Scaffold grouping	✅
Scaffold-aware split	✅
Random Forest	✅
XGBoost	✅
LightGBM	✅
Optuna optimization	✅
Grouped CV	✅
Multi-metric evaluation	✅
Residual analysis	✅
Dataset / validation audit	✅
GNN	✅

---

🔭 Future Direction

The project progresses through increasingly structured molecular representations:

Engineered Features
        ↓
Classical Molecular ML
        ↓
Molecular Graphs
        ↓
Graph Neural Networks
        ↓
Physics-Informed GNNs
        ↓
Structure-aware Molecular Machine Learning

The longer-term direction is to connect:

molecular structure → biological activity → mechanistic understanding

rather than optimizing predictive performance in isolation.

---

📚 Technical Documentation

A separate technical report contains the detailed methodology, data-processing decisions, feature definitions, model configurations, hyperparameter spaces, validation analysis, GNN methodology, and experimental results.

The complete research record and archived software release are preserved through Zenodo.

---

📖 Citation

If you use this repository, its methodology, implementation, or derived results in academic or scientific work, please cite the archived software release:

Venkataramanan, D. (2026).
AChE Molecular Machine Learning: Structure-Aware Prediction of Acetylcholinesterase Inhibitory Potency.
Zenodo.
https://doi.org/10.5281/zenodo.23191332

BibTeX

@software{venkataramanan_ache_qsar_2026,
  author       = {Venkataramanan, Darshan},
  title        = {AChE Molecular Machine Learning: Structure-Aware Prediction of Acetylcholinesterase Inhibitory Potency},
  year         = {2026},
  version      = {1.0.1},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.23191332},
  url          = {https://doi.org/10.5281/zenodo.23191332}
}

---

👤 Author

தர்ஷன் வெங்கட்டரமணன்

கணக்கீட்டு உயிரியல் • வேதியியல் தகவல் புள்ளியியல் • மூலக்கூறு இயந்திரக் கற்றல் • கணக்கீட்டு மருந்து கண்டுபிடிப்பு

a.k.a

Darshan Venkataramanan

Computational Biology · Cheminformatics · Molecular Machine Learning · Computational Drug Discovery

Building from molecular representations toward scientifically grounded learning.

---