MGnify Mega-Scale Protein Stability Dataset

Provider
Cho, Tsuboyama, Ovchinnikov, Rocklin et al. — Rocklin Lab, Northwestern University
License
CC-BY-4.0
Records
~1.8M protein domains (one row each) in the concat table, spanning 200,000+ MGnify sequence families plus de novo designs
Format
CSV (in sampled files)
Stored at
Amazon S3, s3://scigantic-mgnify-stability (us-east-1)
Access
Free and public. No account with the data provider is needed.
DOI
10.5281/zenodo.19411305
Category
Structural Biology
Source
https://www.biorxiv.org/content/10.64898/2026.05.19.726285v1

About MGnify Mega-Scale Protein Stability Dataset

A mega-scale experimental measurement of protein folding stability for roughly 1.8 million small protein domains (about 60-80 amino acids), generated by the Rocklin Lab at Northwestern using cDNA display proteolysis, a cell-free assay that resolves folding stability for nearly a million domains in a single experiment. The domains are drawn primarily from the MGnify metagenomic database and span more than 200,000 sequence families, greatly expanding the sequence diversity of folding-stability data beyond the lab's 2023 release. The core table reports recalibrated folding free energies (dG, in kcal/mol) per domain across three measurement campaigns (DMSv4, DMSv5, DMSv7), accompanied by the raw NGS count tables, the MGnify training index, and a benchmarks set (S1724 wild-type proteins, curated ThermoMutDB stability measurements, and filtered TED structural domains) used to evaluate predictors. Yehlin Cho used these data to train SaProtdG and ESM3dG, models that predict absolute folding stability for small domains to about 0.8 kcal/mol RMSE. Published by Cho, Tsuboyama, Ovchinnikov, Rocklin et al. (bioRxiv 2026); redistributed from the authors' Zenodo deposit (10.5281/zenodo.19411306) under CC-BY-4.0. Hosted by Scigantic for direct mounting.

How to load MGnify Mega-Scale Protein Stability Dataset in Python

# MGnify mega-scale protein stability — start from the concat ΔG table.
import os, pandas as pd
MNT = os.environ.get('SCIGANTIC_MOUNT_PATH', '/mnt/archive')  # also symlinked at ~/archive

# 2.2 GB file: load only the columns you need.
cols = ['name', 'aa_seq', 'deltaG', 'deltaG_95CI', 'mgnify', 'dm_design', 'lib', 'split']
df = pd.read_csv(f'{MNT}/230515_K50dG_dmsv4_dmsv5_dmsv7_concat260429.csv', usecols=cols)
print(df.shape)                       # ~1.8M rows
print(df['deltaG'].describe())        # folding free energy (kcal/mol); higher = more stable
print('from MGnify:', int(df['mgnify'].sum()), ' de novo designs:', int(df['dm_design'].sum()))

# Keep confidently-measured rows (drop clamped / wide-CI):
clean = df[(df['deltaG'].between(-5, 5)) & (df['deltaG_95CI'] < 1.0)]
print(clean.shape)

# To reproduce the published train/val/test split + target:
idx = pd.read_csv(f'{MNT}/mgnify_training_index.csv')  # name, split, seq, dG, PDB_name
print(idx['split'].value_counts())

Written for a Scigantic notebook, where /mnt/archive is the mounted dataset.

Data structure

File types in sampled files: CSV

Fields

namestring — unique domain id (e.g. rocklin_batch2_0); join key to the count tables and training index
aa_seqstring — the protein domain sequence (~60-80 aa). Model on this.
aa_seq_fullstring — full expressed construct = aa_seq flanked by SAGGS linker/tag residues. Use aa_seq unless you need the exact…
dna_seqstring — codon-optimized DNA encoding aa_seq_full
deltaGfloat (kcal/mol) — THE recalibrated folding free energy and the main label. More positive = more stable. Combined…
deltaG_95CI_high / deltaG_95CI_low / deltaG_95CIfloat — 95% confidence bounds and CI width on deltaG
deltaG_t / deltaG_cfloat (kcal/mol) — protease-specific ΔG from trypsin (t) and chymotrypsin (c) separately; deltaG is the consensus of…
deltaG_t_95CI* / deltaG_c_95CI*float — 95% CI bounds/width for the per-protease ΔG
log10_K50_t / log10_K50_cfloat — log10 of the protease concentration at half-maximal cleavage (K50), the raw proteolysis readout per enzyme; ΔG…
log10_K50_t_95CI* / log10_K50_c_95CI*float — 95% CI bounds/width on the K50 fits
fitting_error_t / fitting_error_c / rmsd_t / rmsd_cfloat — goodness-of-fit diagnostics for the trypsin/chymotrypsin titration curves
log10_K50unfolded_t / log10_K50unfolded_c / log10_A0_K50float — fitted unfolded-state K50 references / amplitude used in the ΔG calculation
libstring — source library (e.g. dms4, dms5); which experimental campaign the row came from
mgnifybool — True if the domain is a natural sequence drawn from the MGnify metagenomic database
dm_designbool — True if the domain is a de novo designed protein rather than a natural sequence
splitstring — train/val/test assignment (may be empty for rows not used in the published training)
train_name / train_seqstring — id / sequence as used in training (empty when the row is not in the train set)
train_dG_from_earlier_processingfloat — the ΔG value from the earlier processing pipeline used as the training target; differs slightly from the…

Files

S3 bucket s3://scigantic-mgnify-stability (us-east-1, anonymous read), mounted read-only. Contents:
  230515_K50dG_dmsv4_dmsv5_dmsv7_concat260429.csv   — THE dataset: ~1.8M domains, one row each, recalibrated folding free energy ΔG (2.2 GB)
  mgnify_training_index.csv               — clean train/val/test index (name, split, seq, dG, PDB_name) used to train the published models
  dmsv4_filtered_train_splits.csv         — DMSv4 rows after unstructured-terminus filtering (name, split, aa_seq, deltaG)
  220122_dmsv4_count_table.csv            — raw DMSv4 NGS counts: 24 trypsin (v4_T01..24) + 24 chymotrypsin (v4_C01..24) bins
  220331_dmsv5_count_table.csv            — raw DMSv5 NGS counts (v5_T01..24 / v5_C01..24)
  benchmarks/                             — held-out evaluation sets (S1724 wild-types, ThermoMutDB, TED domains)
  scripts_to_calculate_K50_and_dG.zip     — the pipeline that turns the count tables into K50/ΔG
  README.md                               — the authors' file-by-file notes
Start from the concat table (ΔG already computed). The count tables are only for re-deriving ΔG from raw NGS. Per-domain predicted structures (PDB/CIF) referenced by PDB_name are NOT in this mount — they live in the authors' Google Drive (see README); this mount holds sequences + measurements.

Example file: 230515_K50dG_dmsv4_dmsv5_dmsv7_concat260429.csv (read columns name, aa_seq, deltaG, mgnify, split)

Analyze this dataset without downloading it

On Scigantic, MGnify Mega-Scale Protein Stability Dataset opens in a hosted JupyterLab notebook with the data already mounted, so you can read files where they are stored instead of copying them first. Each notebook includes an AI assistant that can see the mounted files and write and run analysis code. Open the dataset page to launch a notebook, or browse the full catalog.

This page is also available as Markdown and as JSON.

Related Structural Biology datasets