Preserve biological context
Generate candidate 9-mers around multiple mutation classes without silently padding, truncating, or shifting the peptide away from the true altered sequence.
Graph AI / precision oncology
Personalized cancer vaccine design.
An end-to-end pipeline that turns patient variants into biologically grounded peptide candidates, graph representations, binding estimates, and a ranked shortlist.
Interactive architecture
From mutation record to ranked candidate.
Hover a node to isolate its dependencies, or switch focus to inspect the inputs, transformation core, and output boundary.
Case study
The system behind the diagram
Connect sequence and structure
TCGA records are matched to UniProt or reconstructed through Ensembl CDS, encoded as peptide graphs, scored by a GNN, and aligned to learned motifs in log-domain Sinkhorn.
Make the whole path inspectable
The public repository includes preprocessing, candidate generation, graph construction, training, prediction, tests, CSV outputs, and a generated report.
Pipeline detail
Four places where correctness matters
Normalize variants
Separate missense, frameshift, and in-frame mutation handling instead of forcing every event through one shortcut.
Reconstruct sequence
Recover wild-type protein context and simulate altered coding sequence where the mutation changes the downstream frame.
Learn representations
Attach biophysical features to residue nodes and use graph message passing to estimate peptide–MHC binding.
Rank candidates
Combine binding estimates and learned motif alignment into auditable candidate outputs rather than an opaque terminal score.
Evidence boundary
A computational design pipeline, not clinical validation.
The repository demonstrates data processing, model construction, training, and candidate reporting. It does not establish clinical efficacy, and the rankings should be read as research outputs requiring experimental validation.