Skip to content

Known gaps

What mhctools does not do yet, and where the work is tracked. This is the one page that carries issue numbers; the rest of the documentation describes what exists. It is reviewed at release time.

Cleavage validation and coverage

  • Cathepsin S/L/B and AEP (legumain) prediction for novel sequences. The built-in panel has no transferable model; the gap is shown explicitly in batch coverage reports. Experimental observations can be imported now through reference_panels. #470
  • ProsperousPlus adapter, blocked on the open-license requirement. #281
  • Serum and extracellular peptidases beyond the current candidates: activated blood and inflammation-associated proteases (thrombin, plasmin, kallikrein, elastase, cathepsin G, proteinase 3), other exopeptidases and compartment-specific enzymes. Curate activation, inhibitors, tissue exposure and assay conditions before choosing a predictor; a generic Arg/Lys or hydrophobic-residue scan would not distinguish these enzymes. #278
  • MMP substrate predictions (CleaveNet), which need their own native endpoint and assay validation rather than conversion to per-bond probabilities. #476
  • Remaining intracellular peptidases, including LAP3 and BLMH, and the unsettled THOP1 epitope-substrate disagreement in the primary reports. TPP2 and NPEPPS also need exact substrate observations so their permissive rules gain reproduction records. #334
  • Conflicting IRAP precursor sequence in Georgiadou 2010 (DIRSSVQNKL in the text and Table I, DIRSSQVNKL in the Figure 3F caption) blocks curating that record. #332

Benchmarks and validation

  • Held-out, assay-specific benchmarking of the shipped cleavage models, including observed non-cleavages, terminal modifications, homologous-sequence leakage checks and abstention. Purified-enzyme turnover and disappearance of intact peptide in serum are separate endpoints. Reconstructing the Pepsickle paper's processed validation partition belongs here too. #291
  • Half-life endpoints: auditing pepADMET endpoints, overlap with the Tan model, and missing local inference artifacts. #294

Downstream integration

Cleavage overlays, source-table reconciliation and ranking belong to other projects; mhctools provides the evidence contract.

Maintenance

  • Making the SMM subset audit independent of zlib compression differences. #465