saarantras1/scMPRAforge_models

scMPRA stratified negative binomial fits

Fitted parameters for every stratified negative binomial (NB) and zero-inflated negative binomial (ZINB) model reported in Modeling, calibration, and power analysis of single-cell massively parallel reporter assays. There are fifteen fits over three published scMPRA datasets: three canonical (one per dataset, the model the paper's analyses use) and twelve counterfactuals kept so the model-selection comparisons can be reproduced.

The parameter tables are plain parquet and need nothing but pandas. The orthos are the complete fitted objects, loadable with scMPRAforge. Bounds objects derived from these fits are included with scMPRAforge.

Files

file	rows	description
fitted_means.parquet	45,296	fitted mean and dispersion, one row per level
zero_inflation.parquet	6,016	fitted zero inflation, ZINB fits only
fits.csv	15	one row per fit: dataset, model family, zero handling, canonical flag
export_fits.py		the script that produced all three
orthos/<fit>.tar.gz	15	the complete fitted objects, one archive per fit

Schema

fitted_means.parquet

• fit — fit name, joins to fits.csv
• stratified_by — cell_type or cre
• stratum — the level of the stratifying axis this model was fit on
• level_axis / level — what the mean is indexed by, and which one
• mu — fitted mean
• theta — NB dispersion for this stratum

zero_inflation.parquet is the same minus theta, with zi in place of mu, and level_axis always rep_id.

Datasets

dataset	study	assay
shendure	Lalanne et al. 2024, Nat Methods 21:983, 10.1038/s41592-024-02260-3	scQers, developmental CREs in mouse embryoid bodies
cohen	Zhao et al. 2023, Nat Genet 55:346, 10.1038/s41588-022-01278-7	scMPRA, promoter variants in live mouse retinas
seelig	Yin et al. 2025, Cell Systems 16:101302, 10.1016/j.cels.2025.101302	scMPRA of designed enhancers in HepG2/K562

The lab names are kept as identifiers so these join to the analysis code. Make sure you cite the papers above in downstream analysis.

Loading

import pandas as pd

fits  = pd.read_csv("fits.csv")
means = pd.read_parquet("fitted_means.parquet")

canonical = set(fits.loc[fits.canonical, "fit"])
mu = means[means.fit.isin(canonical) & (means.stratified_by == "cell_type")]

Orthos

Each ortho holds its training data in training_data.scmpra/: the count table post filtering and negative-control pooling. ortho.load reads it from there, so an ortho loads and computes from any working directory.

Each fit is one archive, so download only the ones you need:

import tarfile
from huggingface_hub import hf_hub_download
from distributed import Client, LocalCluster
from scMPRAforge.core import ortho

name = "seelig_cm_moib_nb_phantom"
archive = hf_hub_download("saarantras1/scMPRAforge_models", f"orthos/{name}.tar.gz",
                          repo_type="dataset")
tarfile.open(archive).extractall("orthos", filter="data")

client = Client(LocalCluster())
o = ortho.load(client, "orthos", name)

This needs scMPRAforge 0.1.0.dev1 or later (pip install "scMPRAforge>=0.1.0.dev1").

Licence and source data

files	licence	reason
orthos/seelig_*.tar.gz	CC-BY-NC-4.0	contains Yin et al.'s count table, published under CC-BY-NC-4.0
everything else	CC-BY-4.0	our fits, and the shendure and cohen count tables, whose sources are CC-BY-4.0

The count tables are processed from the depositing studies' public data:

dataset	deposit	source terms
shendure	GEO GSE217686	paper is CC-BY-4.0
cohen	GEO GSE188639, Zenodo 14907846	Zenodo deposit is CC-BY-4.0
seelig	GEO GSE269037 (scMPRA), GSE269036 (bulk)	paper is CC-BY-NC-4.0

Citation

TODO on acceptance.