Motivation and Scope¶
In addition to physics-based modeling (docking and molecular dynamics), a small exploratory machine learning (ML) analysis was performed to gain methodological exposure to peptide–protein interaction modeling. Given the extremely limited dataset size, this analysis is not intended for prediction, validation, or hypothesis testing, but purely as a qualitative exploration of baseline ML behavior.
This chapter presents that exploratory machine learning (ML) exercise, required dataset, feature preparation, training stratezy and limitation of models.
Dataset and Feature Preparation¶
The dataset consists of a very small number of peptides investigated in the protein–peptide molecular mimicry study. Each peptide was described using simple, interpretable features:
Peptide length
Peptide sequence
Binding energy
H-bond details
saltbridges (analysed during analysis in gromacs by gmx_mpi)
Average hydrophobicity, computed using the Kyte–Doolittle scale
Hydrophobicity values were calculated as the mean residue hydropathy across each peptide sequence and used as a physicochemical descriptor.
Machine Learning Models¶
Several standard supervised classification models were explored:
Decision Tree
Random Forest
Support Vector Machine (SVM)
XGBoost
The goal was to observe how different model families behave under extreme data sparsity rather than to optimize performance.
Training Strategy¶
The dataset was split into training and test subsets using a simple train–test split. Models were trained using default hyperparameters to avoid overfitting through manual tuning.
Model performance was assessed using accuracy and confusion matrix visualization.
Results and Observations¶
Due to the very small dataset size:
Model accuracies varied widely and were unstable
Strong class imbalance effects were observed
Some models failed to correctly identify one of the classes
Confusion matrix visualizations illustrate this unstable classification behavior, which is expected under such data constraints.
Trained model objects were saved using the joblib format for transparency
and reproducibility, rather than reuse.
Limitations¶
The machine learning results presented here cannot be considered statistically meaningful or generalizable. The analysis is limited by:
Extremely small sample size
Simplistic feature representation
Absence of independent validation data
As a result, ML outputs are interpreted qualitatively only and are not used to support or refute conclusions drawn from docking or MD simulations.
Summary¶
This exploratory ML analysis serves as a methodological complement to the physics-based workflow, demonstrating familiarity with basic machine learning pipelines while respecting the limitations imposed by the data.