Structure →
potency.

Start with the enzyme, not the algorithm.

Acetylcholinesterase is a serine hydrolase responsible for the rapid hydrolysis of acetylcholine. Its catalytic activity regulates the lifetime of acetylcholine in cholinergic signalling, making AChE an established molecular target for modulation of cholinergic signalling.

8,793 observations. One target. Many experimental origins.

The raw data are literature-derived ChEMBL activity records associated with human acetylcholinesterase, CHEMBL220. The dataset should therefore be understood as an aggregation of measurements across assays and sources, not as one homogeneous experiment.

Every discarded row has a reason.

Scroll horizontally through the actual curation sequence. The endpoint is not simply “the cleaned dataset”; it is a documented chain of transformations.

The same molecule, three ways to tell a model what matters.

Representation is not an implementation detail. It determines what structural information is even available to the learner.

The test set was separated by chemistry, not just by row number.

Bemis–Murcko scaffolds were used to keep scaffold groups together. The saved partition contains 3,656 training molecules and 915 held-out test molecules, with zero scaffold overlap.

3,656training molecules

915test molecules

1,191train scaffolds

915test scaffolds

Scaffold allocation saved with the project.

Three ensemble families. Nine classical experiments.

Random Forest, XGBoost and LightGBM were each tested with Morgan, descriptors and combined features. Hyperparameters were selected with the saved Optuna workflow under scaffold-grouped validation.

Model comparison across the original saved classical experiments.

XGBoost · combined representation.

Random Forest · combined representation.

Then the molecule stops being a bit-vector.

The GNN stage represents the molecule explicitly as a graph. Seven categorical atom features and thirteen edge channels are processed through three message-passing steps, then combined with the same 14 global descriptors used by the classical descriptor experiments.

Saved GNN diagnostics.

0.6908 pIC50 MAE

The saved GNN reached RMSE 0.9093, R² 0.3535 and Pearson r 0.6225 on the scaffold-held-out test partition. Best validation MAE was 0.6507 at epoch 48; training stopped at epoch 68.

Terminology matters: this is a graph-plus-descriptor regressor, not a graph-only model.

A model is more than one number.

The repository stores per-model predictions, residuals and diagnostic plots. The figures below make the behaviour of each classical model visible rather than reducing the study to a leaderboard.

Statistical model diagnostics.

XGBoost · Morgan.

XGBoost · descriptors.

LightGBM · combined.

LightGBM · Morgan.

LightGBM · descriptors.

Random Forest · Morgan.

Random Forest · descriptors.

A negative GNN result can still be a useful result.

A dedicated clean audit reserves 548 validation molecules from the original training portion while retaining the same 915-molecule test partition. This makes the classical/GNN comparison more directly comparable.

The uncomfortable parts stay on the page.

The project includes a forensic implementation audit. These findings are part of the research record, not something to hide behind a polished interface.

What the study actually demonstrates.

Where the evidence stops.

A repository you can interrogate.

The research record contains source modules, data snapshots, metrics, predictions, Optuna trials, audit tables, model checkpoints and diagnostic figures.

The next experiment should answer a harder question.

These are directions, not claims of completed work.