Bio lab · GMU Young Scholars

Peptide self-assembly from sequence

Some short peptides assemble into ordered nanostructures on their own. Most don’t. I built three pipelines that predict which, from the amino-acid sequence alone.

NextGen Science: Machine Learning & Bioinformatics, George Mason University Young Scholars Research Program · Oct to Dec 2025 · with Christopher Lockhart · Orange Data Mining, Python

Illustration. Some chains find each other and stack, most drift past.

ResultCross-validated

All three above AUC 0.99

Pipeline	Best model	CV AUC	Other metrics
Physicochemical descriptors	AdaBoost	0.996	CA 0.979 · F1 0.979 · MCC 0.958
k-mer bag of words	Neural net	0.993	not in table
Engineered descriptors	not in table	> 0.99	not in table

Every pipeline reached a cross-validated AUC above 0.99 with at least one model. Oversampling ran before the split, so treat these as upper bounds. “Not in table” marks numbers the results table doesn’t give.

LimitsRead with the table

Why 0.996 is an upper bound

Oversampling ran before the split. That can leak duplicates across folds and inflate the scores, so I read them as an upper bound, not a clean estimate of how the models generalize.

The paper’s abstract gives a higher figure for the k-mer pipeline than its results table supports. I use the table. Its best is 0.996.

MethodThree representations

Three ways to turn a sequence into numbers

PIPELINE 1

Physicochemical descriptors

Amino-acid physicochemical properties as features.

PIPELINE 2

k-mer bag of words

The NLP-inspired one, above. Counts of every length-k window.

PIPELINE 3

Engineered descriptors

Derived sequence descriptors built by hand.

The classes were moderately imbalanced, so I used stratified oversampling. Every pipeline was scored with stratified 5-fold cross-validation across logistic regression, an RBF-kernel SVM, gradient boosting, AdaBoost and a neural net.

Orange Data Mining workflow: CSV import, class imbalance plot, data samplers, concatenate, corpus, text preprocessing, bag of words, learners, Test and Score, confusion matrix, ROC analysis and predictions — **Fig. 1**Pipeline 2 as an Orange workflow, from CSV import through class balancing and text encoding to Test and Score and ROC analysis.

Try itk-mer counting

k-mer counting

PaperPublished Jan 1, 2026

Paper

Predicting Peptide Self-Assembly Using Machine Learning: A Comparative Study of Sequence-Based Feature Representations

Daniel Tsen and Christopher Lockhart (George Mason University). GMU Young Scholars Research Program, Jan 1, 2026.

doi:10.13021/MARS/15267 ↗