If the same genetic variant survives two computational shortlists, have they selected the same object for a later experiment? A variant can give rise to different predicted peptides, short protein fragments, paired with different HLA molecules, which are involved in displaying peptides to the immune system. Agreement on the variant alone leaves those choices unresolved.
I used public cancer-cell-line data to compare variant-only agreement with agreement on the complete variant–peptide–HLA combination. The reproducibility question was whether repeated analyses selected the same combinations, not merely whether familiar variant names remained in the list.
Defining the object being compared
The study used public SEQC2 data from the HCC1395 cancer cell line and its matched HCC1395BL normal line. Its 63 analytical conditions are not 63 patients or independent tumors. They come from 21 sequencing libraries distributed across seven sites, with three alignment programs applied to each library. Results from the same library share an input, so I retained that structure in the comparisons.
Each entry records the variant’s position and base change, the exact peptide sequence, and the HLA allele. I distinguished a shared variant or peptide from a shared complete combination. Agreement was measured as the number of shared entries divided by the number of distinct entries across both lists, called Jaccard similarity. Changing what counts as one entry can change how similar the same two lists appear.
Changing HLA conditions before selecting the winners
Public HLA reports for this cell line came from different analytical contexts. I did not designate one as the correct genotype. Instead, I examined how selection changed when each reported panel was used.
Simply deleting entries from an already selected list would miss combinations that lost under the original panel but would win under the revised one. I therefore re-evaluated the full set of peptide combinations under the relevant HLA conditions before selecting winners. In the released analysis, strong variant-level agreement did not carry through to the exact peptide–HLA combinations.
Comparisons between analytical conditions also preserved the site, library, and aligner structure. They describe variation within this dataset and computational design; they do not establish that a particular site or program is biologically more accurate.
What the public release lets a reader check
On 23 August 2026, I released v1.0.0 with the analysis code, frozen configuration, result files, and tests. Checksums identify file contents rather than filenames, allowing a reader to detect whether the material has changed. The reproducibility archive includes a hash manifest for that purpose.
Auditing the bundled results is different from regenerating every calculation from upstream inputs. A complete upstream run also needs the identified public source data and MHCflurry model files. Those third-party payloads are not all redistributed in the ZIP. Author-generated code is licensed under MIT, and author-generated documentation and derived artifacts under CC BY 4.0; the providers’ terms still apply to third-party materials.
Repeated selection does not establish that a peptide is displayed on a cell or elicits an immune response. No biological candidate was nominated in this study. Agreement between computational lists remains a separate question from experimental efficacy or safety.
GitHub v1.0.0 code and reproducibility archive
This DOI identifies the software release. I submitted the manuscript to Computational Biology and Chemistry; it was in editorial processing as of 3 September 2026, with journal acceptance still unconfirmed.