Academic project · 2025–2026

Antimicrobial-resistance prediction in Neisseria gonorrhoeae

Evaluating genomic predictors of resistance, with grouped validation and an explicit account of what internal results can establish.

Python / Pandas / scikit-learn

Profile-group bootstrap intervals and descriptive continent-level balanced accuracy for the AMR models.
Original project figure. Five hundred profile-group bootstrap resamples and descriptive geographic subgroups from the same out-of-fold predictions; not leave-one-continent-out validation.
Open full-size figure ↗

The question

Can genomic unitig features distinguish resistant from susceptible Neisseria gonorrhoeae isolates, and how sensitive are the results to evaluation choices?

This Bioinformatics Computing II course project examined azithromycin and ciprofloxacin resistance. The repository preserves the original submission alongside a reproducibility-focused revision.

Data and approach

I reviewed 3,786 metadata records and matched labelled isolates to antibiotic-specific genomic matrices. Identical complete unitig profiles stay together in five-fold cross-validation. Variance filtering and chi-square feature selection are fitted inside each training fold.

I compared logistic regression and random forest models with a prevalence baseline, then examined sampling variability with profile-group bootstrap intervals and descriptive geographic subgroups.

Results in context

The repository reports the following internal out-of-fold estimates for class-weighted logistic regression:

Antibiotic Labelled isolates Balanced accuracy
Azithromycin 3,478 0.949
Ciprofloxacin 3,088 0.969

These figures describe this internal evaluation. They are not independent validation results. I did not train a cefixime classifier because only five isolates were labelled resistant.

Biological interpretation

BLAST matches for selected unitigs included regions associated with penA and gyrA. These alignments add biological context; they do not establish that a sequence causes resistance.

What I would test next

Independent data and validation that includes upstream feature discovery would provide a stronger test of generalisation. Geographic subgroup differences also warrant evaluation on held-out populations rather than treating descriptive subgroups as external validation.

Evidence & source material