Machine learningOct 2026 · London AI × Science hackathon · Team of 2

PepShield

Peptide–HLA stability, predicted with a calibrated interval.

PyTorch models trained on 28,166 stability measurements across 75 HLA alleles. I rebuilt the evaluation after finding a 93% train/test leak, compared three ways to encode an allele, and shipped a live app with conformal 90% intervals.

Open the live app (opens in a new tab)Read the code (opens in a new tab)

sarah04menla--pepshield.modal.run

The live app: 10.3 h predicted, with its 90% interval.

RoleDeveloper 1: data, evaluation, models and uncertainty

TeamTeam of 2

WhenOctober 2026, one hackathon weekend

Built withPyTorch, Streamlit, Modal

Skills it shows

Leakage-aware evaluation

Found the same peptide in train and test for about 93% of random-split test rows, then built peptide-, similarity-clustered and allele-held-out splits.

Controlled experiments

Compared three allele encodings in one fixed network. Frozen ESM-2 lost to the allele-name baseline on unseen peptides (Spearman 0.54 vs 0.64).

Uncertainty quantification

A 5-model ensemble with split-conformal 90% intervals, reaching about 90% coverage on unseen peptides.

ML deployment

Served on Modal as a Streamlit app. On an unseen allele the spread stopped tracking error, so the app withholds the interval there.

Problem, build, result

1Problem

Whether a T cell ever sees a target depends on how long the peptide–HLA complex holds together. Lab measurement is slow, and models usually give one number with no error bar.

2What I built

One unchanged MLP for every experiment, so differences come from the features: allele name, ESM-2, and 34 groove-contact residues. Then an ensemble with conformal intervals, behind an API.

3Result

A deployed predictor with about 90% interval coverage on unseen peptides. Only the groove-contact encoding ranked an unseen allele above chance (Spearman 0.15, 95% CI 0.05–0.25).

28,166stability measurements

~93%train/test leak found and fixed

0.64Spearman on unseen peptides

~90%coverage of the 90% interval