Python, PyTorch · immunopeptidomics · not yet peer reviewed
Can a 1D convolutional network tell experimentally observed HLA class-I peptides from matched decoys using sequence alone, and how does it compare with published presentation predictors on identical rows? Built on the public PRIDE dataset PXD024871, preregistered before any model was trained, and frozen with a checksum-verified record.
The CNN reached an average precision of 0.7551 on held-out units. Four independently built predictors scored within 0.0046 of each other, which suggests the CNN's lead may partly reflect how this cohort's decoys were constructed. A second cohort is needed to settle it.
Are bacterial HLA ligands in tumour immunopeptidomes real?
Python, Docker, Colab · MSc thesis · public data only
An error-controlled reanalysis of public tumour immunopeptidomes, with entrapment, human-homology and benign-tissue controls, auditing a public atlas of bacteria-derived HLA epitopes.
Re-searching the raw files behind one dataset at 1% FDR found 20 bacterial peptides where the atlas attributes 6,363 — 143 to 176 times fewer after adjusting for coverage.
Repository currently private; available on request.
Mechanism-resolved prediction of cancer mutation effects in SH3 domains
Python, PyTorch, ESM-2 · protein machine learning
Does a cancer missense mutation in a human SH3 domain destabilise the fold, or leave it intact and change binding? Every result is rebuilt from raw data, starting with the COSMIC atlas, using protein language models and measured folding energies.
Folding effects are predictable (median Spearman ρ 0.839 on 15 held-out domains) and track residue burial, not the binding interface. Binding effects are not yet predictable from current data, and the shortfall is quantified.
Repository currently private; available on request.