mhcflurry 2.3.0rc1

MHC Binding Predictor

mhcflurry

MHC I ligand prediction package with competitive accuracy and a fast and documented implementation.

[!IMPORTANT] Version 2.3.0 keeps the same external API as 2.2.0 and ships substantial performance and tooling improvements for users training their own models or running large prediction workloads:

• Device-resident affinity training: Class1NeuralNetwork.fit() keeps peptides, alleles, targets, and the random-negative pool on the active torch device for the lifetime of one fit, eliminating per-batch host↔device copies.
• Multi-GPU prediction by default: mhcflurry-predict, mhcflurry-predict-scan, mhcflurry-calibrate-percentile-ranks, and the sweep eval script auto-discover visible GPUs and fan out across them.
• Orchestrator auto-tuning: mhcflurry-class1-train-pan-allele-models resolves --num-jobs, --max-workers-per-gpu, --dataloader-num-workers, and random_negative_pool_epochs from the box's hardware so the same recipe runs on a workstation, single-GPU node, or 8×A100 host. --dataloader-num-workers applies to streaming pretraining; affinity fine-tuning batches from device-resident tensors.
• torch.compile + TF32 + matmul-precision are first-class CLI flags on the train commands; the in-process Inductor cache is warmed by a single worker before the production pool launches.

If you are upgrading from 2.1.x or 2.2.x, simply pip install --upgrade mhcflurry. The published pre-trained models are unchanged and will be loaded automatically. Internal refactors (per-fit device-resident training tensors, torch-side peptide encodings) do not affect the public Python or CLI surface.

Earlier release: Version 2.2.0 was the first release to use PyTorch as its neural network backend, replacing TensorFlow/Keras. It introduced the Python 3.10+ and pandas >= 2.0 requirements and added Apple Silicon (MPS) support.

MHCflurry implements class I peptide/MHC binding affinity prediction. The current version provides pan-MHC I predictors supporting any MHC allele of known sequence. MHCflurry runs on Python 3.10+ using the PyTorch neural network library. It exposes command-line and Python library interfaces.

MHCflurry also includes two experimental predictors, an "antigen processing" predictor that attempts to model MHC allele-independent effects such as proteosomal cleavage and a "presentation" predictor that integrates processing predictions with binding affinity predictions to give a composite "presentation score." Both models are trained on mass spec-identified MHC ligands.

If you find MHCflurry useful in your research please cite:

T. O'Donnell, A. Rubinsteyn, U. Laserson. "MHCflurry 2.0: Improved pan-allele prediction of MHC I-presented peptides by incorporating antigen processing," Cell Systems, 2020. https://doi.org/10.1016/j.cels.2020.06.010

T. O'Donnell, A. Rubinsteyn, M. Bonsack, A. B. Riemer, U. Laserson, and J. Hammerbacher, "MHCflurry: Open-Source Class I MHC Binding Affinity Prediction," Cell Systems, 2018. https://doi.org/10.1016/j.cels.2018.05.014

Please file an issue if you have questions or encounter problems.

Have a bugfix or other contribution? We would love your help. See our contributing guidelines.

Try it now

You can generate MHCflurry predictions without any setup by running our Google colaboratory notebook.

Installation (pip)

Install the package:

$ pip install mhcflurry

Download our datasets and trained models:

$ mhcflurry-downloads fetch

You can now generate predictions:

$ mhcflurry-predict \
       --alleles HLA-A0201 HLA-A0301 \
       --peptides SIINFEKL SIINFEKD SIINFEKQ \
       --out /tmp/predictions.csv

Wrote: /tmp/predictions.csv

Or scan protein sequences for potential epitopes:

$ mhcflurry-predict-scan \
        --sequences MFVFLVLLPLVSSQCVNLTTRTQLPPAYTNSFTRGVYYPDKVFRSSVLHS \
        --alleles HLA-A*02:01 \
        --out /tmp/predictions.csv

Wrote: /tmp/predictions.csv

Unified mhcflurry parent command

Starting in 2.3.0 there is also a single mhcflurry command that dispatches to every subcommand:

$ mhcflurry predict \
        --alleles HLA-A0201 HLA-A0301 \
        --peptides SIINFEKL SIINFEKD SIINFEKQ \
        --out /tmp/predictions.csv

$ mhcflurry compare-models \
        --a results/new_run/ \
        --b public \
        --out results/comparison/

$ mhcflurry plot-model-comparison --input results/comparison/

Every historical command is reachable as a subcommand (mhcflurry-predict ↔ mhcflurry predict, mhcflurry-downloads ↔ mhcflurry downloads, mhcflurry-class1-train-pan-allele-models ↔ mhcflurry class1-train-pan-allele-models, etc.). Both forms run the same underlying entry point; the legacy mhcflurry-* scripts remain installed as compat shims and are not changing. mhcflurry --help lists every available subcommand.

The two new-in-2.3.0 model-comparison tools, compare-models and plot-model-comparison, only have the unified form.

See the documentation for more details.

Development and tests

From a checkout, source develop.sh to create and activate the editable environment:

$ source develop.sh

For quick feedback, run lint plus a focused unit subset:

$ ./lint.sh
$ pytest -q test/test_amino_acid.py test/test_random_negative_peptides.py

pytest test/ is the full test suite, not a fast unit-only loop. It includes small end-to-end training runs, command subprocess tests, public-model smoke tests that require cached MHCflurry download bundles, and speed/regression checks, so it can take many minutes. Use pytest -q test -m "not slow and not downloads" for the broad fast tier, and pytest -q test --durations=25 when auditing slow tests. See the testing documentation for the current test tiers.

Docker

You can also try the latest (GitHub master) version of MHCflurry using the Docker image hosted on Dockerhub by running:

$ docker run -p 9999:9999 --rm openvax/mhcflurry:latest

This will start a jupyter notebook server in an environment that has MHCflurry installed. Go to http://localhost:9999 in a browser to use it.

To build the Docker image yourself, from a checkout run:

$ docker build -t mhcflurry:latest .
$ docker run -p 9999:9999 --rm mhcflurry:latest

Predicted sequence motifs

Sequence logos for the binding motifs learned by MHCflurry BA are available here.

Common issues and fixes

Problems downloading data and models

Some users have reported HTTP connection issues when using mhcflurry-downloads fetch. As a workaround, you can download the data manually (e.g. using wget) and then use mhcflurry-downloads just to copy the data to the right place.

To do this, first get the URL(s) of the downloads you need using mhcflurry-downloads url:

$ mhcflurry-downloads url models_class1_presentation
https://github.com/openvax/mhcflurry/releases/download/1.6.0/models_class1_presentation.20200205.tar.bz2```

Then make a directory and download the needed files to this directory:

$ mkdir downloads
$ wget  --directory-prefix downloads https://github.com/openvax/mhcflurry/releases/download/1.6.0/models_class1_presentation.20200205.tar.bz2```

HTTP request sent, awaiting response... 200 OK
Length: 72616448 (69M) [application/octet-stream]
Saving to: 'downloads/models_class1_presentation.20200205.tar.bz2'

Now call mhcflurry-downloads fetch with the --already-downloaded-dir option to indicate that the downloads should be retrived from the specified directory:

$ mhcflurry-downloads fetch models_class1_presentation --already-downloaded-dir downloads