README.md

metadata

license: cc-by-4.0
library_name: mpac
tags:
  - biology
  - genomics
  - dna
  - mpra
  - cis-regulatory
  - variant-effect-prediction
pipeline_tag: other

MPAC

MPAC (Malinois with Parallel Aggregated Cross-validation) predicts cis-regulatory activity of 200 bp human sequences in K562, HepG2 and SK-N-SH, and the allelic skew caused by non-coding variants.

Identifying non-coding variant effects at scale via machine learning models of cis-regulatory reporter assays.

This repository holds the 110 published checkpoints, converted to safetensors from the Zenodo deposit with no retraining or modification.

Usage

from modeling_mpac import MPACEnsemble

# Loads the ten models that held chr7 out of training
ensemble = MPACEnsemble.from_pretrained("saarantras1/MPAC", chromosome=7, device="cuda")

preds = ensemble.predict(["ACGT" * 50], device="cuda")   # (n, 3): K562, HepG2, SKNSH

Use from_pretrained and predict rather than loading a checkpoint or calling the model directly: they select the ensemble that did not train on your query's chromosome, and they add the MPRA vector context and average over both strands. Skipping either step returns plausible-looking but wrong numbers instead of an error.

predict follows vcf_predict.py from the upstream code base, which generated the published predictions: the reverse strand is the reverse complement of the 200 bp insert placed back in the forward-orientation vector, matching the assay, rather than a reverse complement of the whole 600 bp construct.

MPAC covers autosomes only; from_pretrained raises on chrX, chrY and anything else with no held-out fold.

For allelic skew, pass matched reference and alternate contexts. Each variant is tiled into eighteen 200 bp windows at stride 10 and averaged, reproducing the scheme behind the published predictions.

out = ensemble.predict_skew(ref_contexts, alt_contexts, device="cuda")
out["skew"]   # (n, 3), alt minus ref

Command-line tools for MPAC can be found at: Reilly-Lab-Yale/coda_mpac.

Which provide greater control of prediction parameters including (but not limited to): window number step size strand reduction insert/full plasmid reverse complement window aggregation (average, max, min, etc.)

An example prediction for SNVs from a VCF can be found in the methods of the publication and below:

python vcf_predict.py --artifact_path {10X $MODEL} \ CHROMOSOME HOLDOUT MODELS TO ENSEMBLE --vcf_file ${VCF} \ VARIANTS OF INTEREST --fasta_file ${FASTA} \ REFERENCE GENOME --output ${OUTPUT} \ OUTPUT PATH --relative_start 9 \ START WINDOW --relative_end 180 \ END WINDOW --step_size 10 \ N WINDOWS --strand_reduction mean \ REDUCTION METHOD OF FWD/REV STRAND PREDICTIONS --window_reduction mean \ REDUCTION METHOD OF PREDICTION WINDOWS

And include specialized prediction methods for: small indels (<= 10bp recommended) (vcf_predict_indel.py) haplotypes (vcf_predict_haplotype.py)

Which can be run in the same way as vcf_predict.py but with modifications to handle sequence padding (vcf_predict_indel.py) or windowing (vcf_predict_haplotype.py)

Citation

@article{butts2025mpac,
  title   = {Identifying non-coding variant effects at scale via machine learning
             models of cis-regulatory reporter assays},
  author  = {Butts, John C. and Rong, Stephen and Gosai, Sager J. and
             Castro, Rodrigo I. and Noon, Mackenzie and Adeniran, Kehinde and
             Ghosh, Rohit and Sabeti, Pardis C. and Tewhey, Ryan and Reilly, Steven K.},
  journal = {bioRxiv},
  year    = {2025},
  doi     = {10.1101/2025.04.16.648420}
}

License

CC-BY-4.0, matching the Zenodo deposit. modeling_mpac.py derives from the MIT licensed model code in sjgosai/boda2 and retains that notice.