MHCflurry 2.3.0

MHCflurry 2.3.0

MHCflurry 2.3.0 provides updated presentation models, reproducible training and
model comparison workflows, and shared percentile calibration across the
PyTorch predictors. Python 3.10 or newer is required.

Install and upgrade

pip install --upgrade "mhcflurry==2.3.0"
mhcflurry downloads fetch models_class1_presentation

Pinning the version also handles installations that used a higher-numbered
development build. Existing 2.2.x model directories remain loadable. Software
and model downloads have separate provenance; a software upgrade alone does
not replace a user-specified model directory.

Models and evaluation

The new full presentation model uses the 2023 training-data snapshot and the
release training recipe. Its with-flank processing component
is an equal mixture of short-flank and cleavage-boundary networks. The
separate long-flank processing ensemble is not the presentation component.

Evaluate complete presentation predictions separately from affinity and
processing diagnostics. Full-model improvements do not imply improvements in
every component. Comparisons must use identical positive/negative rows and
shared overlap exclusions; external-predictor training overlap may remain
unknown. View the model comparison (PDF).

The PDF includes all comparison figures: full presentation AP, PPV@N and AUROC
with paired patient-bootstrap intervals versus MHCflurry 2.1.5/2.2.0/2.2.1,
NetMHCpan 4.0/4.1/4.2 BA and EL, and MixMHCpred 3.0. It leads with peptide,
MHC and N/C flanks, followed by the no-flank comparison. All models use the
same 181,131 retained rows; the external tools receive no flank inputs.
Separate affinity/processing diagnostics and training-overlap limitations
are included. Tables and source data (.tar.gz)
are available for reanalysis.

Prediction and calibration

• Prediction progress is written to stderr so redirected stdout contains valid CSV.
• The mhcflurry command groups prediction, download, training and evaluation
  commands. Existing standalone mhcflurry-* entry points remain supported.
• CPU, CUDA and Apple Silicon execution use PyTorch. Automatic worker and
  prediction-batch planning respects caller overrides and device capacity.
• Affinity, processing and presentation share percentile-calibration methods.
  New compact mappings use independent background data; existing histogram
  calibrations retain their original semantics when loaded.
• Model manifests, allele pseudosequences, saved calibration and prediction
  weights travel together. Calibration changes do not alter raw predictions.

Training and reproducibility

• Streaming pretraining validation now reverses concentration inequalities
  when converting to the decreasing regression-target scale. This fixes future
  validation/retry decisions; it does not modify the selected released weights.
• Release workflows record source, data, configuration and artifact checksums;
  validate holdout exclusions; and preserve resumable training stages.
• Processing preparation uses matched negatives with unique assignments,
  bounded pool expansion and deterministic replay.
• Affinity and processing training can retain checkpoint alternatives for
  development. Inference downloads contain the selected model, while the
  training archive preserves alternative states and selection evidence.
• Evaluation supports saved predictions, paired uncertainty estimates and
  external NetMHCpan/MixMHCpred scores on explicitly shared rows.

See the training recipe,
evaluation guide, and
release workflow for the supported commands.

Download files

• Presentation model (includes affinity and processing)
• Standalone affinity models
• Standalone processing models
• All comparison figures (PDF) (PDF checksum)
• Comparison tables and source data (.tar.gz)
• Model and data SHA256 checksums

The model information files preserve the original training software/source
identifiers. The stable code and download release is 2.3.0.

The comparison PDF includes a separate 2.3.0-pre — 2020 versus 2023 training data component appendix. The 2023 prerelease arm has no full presentation model, and its scored processing architecture differs from the 2020 arm; this is not a controlled data-only ablation. The main comparison covers the released full predictors on identical rows. The revised archive contains public labels and self-contained rendering inputs; original IDs remain in audit provenance.

Revised comparison archive SHA256.