{"success":true,"data":{"author":"Cho, Tsuboyama, Ovchinnikov, Rocklin et al. — Rocklin Lab, Northwestern University","bucketAccessType":"public","bucketListable":true,"bucketName":"scigantic-mgnify-stability","category":"Structural Biology","cloudProvider":"aws","createdAt":"2026-06-30T00:00:00.000Z","description":"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.","displayTitle":"MGnify Mega-Scale Protein Stability Dataset","doi":"10.5281/zenodo.19411305","fuseMountBrowsable":true,"id":"e5cd0cc0-c9c9-5d1c-b25b-09c3b1105be7","license":"CC-BY-4.0","objectCountEstimate":1,"probeStatus":"ok","referenceUrl":"https://www.biorxiv.org/content/10.64898/2026.05.19.726285v1","region":"us-east-1","schemaCard":{"accessHints":[{"language":"python","snippet":"# Memory-safe scan of the 2.2 GB table from the read-only FUSE mount.\nimport os, pandas as pd\nMNT = os.environ.get('SCIGANTIC_MOUNT_PATH', '/mnt/archive')\npath = f'{MNT}/230515_K50dG_dmsv4_dmsv5_dmsv7_concat260429.csv'\nuse = ['name', 'aa_seq', 'deltaG', 'deltaG_95CI', 'mgnify', 'split']\nchunks = pd.read_csv(path, usecols=use, chunksize=200_000)\nstable = pd.concat([c[(c['deltaG'] > 2) & (c['deltaG_95CI'] < 0.5)] for c in chunks])\nprint(len(stable), 'high-confidence stable domains')"},{"language":"python-train","snippet":"# Reproduce the published split + target (sequence -> ΔG regression).\nimport os, pandas as pd\nMNT = os.environ.get('SCIGANTIC_MOUNT_PATH', '/mnt/archive')\nidx = pd.read_csv(f'{MNT}/mgnify_training_index.csv')  # name, split, seq, dG, PDB_name\ntrain = idx[idx['split'] == 'train']\nval   = idx[idx['split'] == 'val']\ntest  = idx[idx['split'] == 'test']\nX_train, y_train = train['seq'], train['dG']   # featurize seq (ESM/SaProt/k-mers)\n# Held-out external benchmarks live under benchmarks/ (S1724, TED)."},{"language":"bash","snippet":"# Bucket is public — no credentials. Pull the headline table or just peek at it.\nB=s3://scigantic-mgnify-stability\naws s3 ls --no-sign-request --region us-east-1 $B/\naws s3 ls --no-sign-request --region us-east-1 $B/benchmarks/\n# stream just the header without downloading 2.2 GB:\naws s3 cp --no-sign-request --region us-east-1 \\\n  $B/mgnify_training_index.csv - | head -5"}],"caveats":["deltaG is the folding free energy (ΔG) in kcal/mol and is the headline label: MORE POSITIVE = MORE STABLE (folded state favored). It is NOT a ΔΔG mutational effect — these are absolute per-domain stabilities for ~1.8M distinct sequences.","ΔG is only measurable within the assay's dynamic window (roughly -5 to +5 kcal/mol). Sequences outside that window are CLAMPED: per-protease columns show sentinel values like -15, -5, -25 (see deltaG_t in the first sample row). Do not treat clamped values as real energies — filter on the CI columns / use the consensus deltaG, and consider dropping rows where deltaG sits at a clamp bound before training a regressor.","Two ΔG flavors exist. deltaG in the concat table is the RECALIBRATED value. train_dG_from_earlier_processing (and the dG column in mgnify_training_index.csv) is the EARLIER-processing value used as the training target for the published SaProtΔG / ESM3ΔG models. They differ slightly (e.g. -1.39 vs -0.91 for rocklin_batch2_0). Reproduce the paper with the training index; use deltaG for the best current estimate.","To reproduce the published model evaluation, train on mgnify_training_index.csv (it carries the exact split) and report on benchmarks/ (S1724 wild-types and TED domains are held out, not in train). The concat table's own split column is the broader assignment.","aa_seq is the domain; aa_seq_full prepends/appends a fixed SAGGS expression tag (linker). Model on aa_seq unless you specifically need the assayed construct. The same name keys the row across the concat table, the count tables, and the training index.","The count tables (220122_dmsv4*, 220331_dmsv5*) are RAW NGS read counts across 24 trypsin (T) and 24 chymotrypsin (C) protease concentrations. They exist to re-derive K50/ΔG via scripts_to_calculate_K50_and_dG.zip. Almost all users want the concat ΔG table instead — do not confuse counts with stabilities.","Per-domain structures (the PDB/CIF files behind PDB_name) are NOT in this mount. The structure-aware models (SaProtΔG, ESM3ΔG) fold each sequence first; the predicted structures live in the authors' Google Drive (linked in README.md). This archive provides sequences + measurements; generate structures with ESMFold/ColabFold or pull them from the Drive folder if you need them.","This is a Scigantic mirror of the authors' CC-BY-4.0 Zenodo deposit (10.5281/zenodo.19411306). Attribution required: cite Cho, Tsuboyama, Ovchinnikov, Rocklin et al., bioRxiv 2026. The full deposit also includes figures.zip and a dmsv7_analysis folder (paper-reproduction artifacts) that were intentionally NOT mirrored here to keep the mount analysis-focused; fetch them from Zenodo if needed.","The concat table is 2.2 GB. Reading it whole needs several GB of RAM — prefer usecols= to load just name/aa_seq/deltaG/mgnify/split, or read in chunks. For quick experiments start from mgnify_training_index.csv (~138 MB)."],"columns":[{"name":"name","type":"string — unique domain id (e.g. rocklin_batch2_0); join key to the count tables and training index"},{"name":"aa_seq","type":"string — the protein domain sequence (~60-80 aa). Model on this."},{"name":"aa_seq_full","type":"string — full expressed construct = aa_seq flanked by SAGGS linker/tag residues. Use aa_seq unless you need the exact assayed construct."},{"name":"dna_seq","type":"string — codon-optimized DNA encoding aa_seq_full"},{"name":"deltaG","type":"float (kcal/mol) — THE recalibrated folding free energy and the main label. More positive = more stable. Combined trypsin+chymotrypsin estimate."},{"name":"deltaG_95CI_high / deltaG_95CI_low / deltaG_95CI","type":"float — 95% confidence bounds and CI width on deltaG"},{"name":"deltaG_t / deltaG_c","type":"float (kcal/mol) — protease-specific ΔG from trypsin (t) and chymotrypsin (c) separately; deltaG is the consensus of the two"},{"name":"deltaG_t_95CI* / deltaG_c_95CI*","type":"float — 95% CI bounds/width for the per-protease ΔG"},{"name":"log10_K50_t / log10_K50_c","type":"float — log10 of the protease concentration at half-maximal cleavage (K50), the raw proteolysis readout per enzyme; ΔG is derived from these"},{"name":"log10_K50_t_95CI* / log10_K50_c_95CI*","type":"float — 95% CI bounds/width on the K50 fits"},{"name":"fitting_error_t / fitting_error_c / rmsd_t / rmsd_c","type":"float — goodness-of-fit diagnostics for the trypsin/chymotrypsin titration curves"},{"name":"log10_K50unfolded_t / log10_K50unfolded_c / log10_A0_K50","type":"float — fitted unfolded-state K50 references / amplitude used in the ΔG calculation"},{"name":"lib","type":"string — source library (e.g. dms4, dms5); which experimental campaign the row came from"},{"name":"mgnify","type":"bool — True if the domain is a natural sequence drawn from the MGnify metagenomic database"},{"name":"dm_design","type":"bool — True if the domain is a de novo designed protein rather than a natural sequence"},{"name":"split","type":"string — train/val/test assignment (may be empty for rows not used in the published training)"},{"name":"train_name / train_seq","type":"string — id / sequence as used in training (empty when the row is not in the train set)"},{"name":"train_dG_from_earlier_processing","type":"float — the ΔG value from the earlier processing pipeline used as the training target; differs slightly from the recalibrated deltaG. mgnify_training_index.csv.dG is this value."}],"format":"v2-hybrid","generatedAt":"2026-10-28T16:48:24.380Z","layout":{"example":"230515_K50dG_dmsv4_dmsv5_dmsv7_concat260429.csv  (read columns name, aa_seq, deltaG, mgnify, split)","pattern":"S3 bucket s3://scigantic-mgnify-stability (us-east-1, anonymous read), mounted read-only. Contents:\n  230515_K50dG_dmsv4_dmsv5_dmsv7_concat260429.csv   — THE dataset: ~1.8M domains, one row each, recalibrated folding free energy ΔG (2.2 GB)\n  mgnify_training_index.csv               — clean train/val/test index (name, split, seq, dG, PDB_name) used to train the published models\n  dmsv4_filtered_train_splits.csv         — DMSv4 rows after unstructured-terminus filtering (name, split, aa_seq, deltaG)\n  220122_dmsv4_count_table.csv            — raw DMSv4 NGS counts: 24 trypsin (v4_T01..24) + 24 chymotrypsin (v4_C01..24) bins\n  220331_dmsv5_count_table.csv            — raw DMSv5 NGS counts (v5_T01..24 / v5_C01..24)\n  benchmarks/                             — held-out evaluation sets (S1724 wild-types, ThermoMutDB, TED domains)\n  scripts_to_calculate_K50_and_dG.zip     — the pipeline that turns the count tables into K50/ΔG\n  README.md                               — the authors' file-by-file notes\nStart 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."},"mountPath":"/mnt/archive","rowCount":"~1.8M protein domains (one row each) in the concat table, spanning 200,000+ MGnify sequence families plus de novo designs","sampleHeaders":[{"bytes":"PDB_chain,PDBfasta,paper,gabetrimmed,notes\r\n1AJ3_A,AKLNESHRLHQFFRDMDDEESWIKEKKLLVSSEDYGRDLTGVQNLRKKHKRLEAELAAHEPAIQGVLDTGKKLSDDNTIGKEEIQQRLAQFVDHWKELKQLAAARGQLE,ESHRLHQFFRDMDDEESWIKEKKLLVSSEDYGRDLTGVQNLRKKHKRLEAELAAHEPAIQGVLDTGKKLSDDNTIGKEEIQQRLAQFVDHWKELKQLAAARGQRLEESLEYQQFVA,ESHRLHQFFRDMDDEESWIKEKKLLVSSEDYGRDLTGVQNLRKKHKRLEAELAAHEPAIQGVLDTGKKLSDDNTIGKEEIQQRLAQFVDHWKELKQLAAARGQRLEESLEYQQFVA,\"according to DOI 10.1016/j.jmb.2004.09.037, this study of R16 used 4aa+106aa+6aa, as described in doi 10.1021/bi0025","encoding":"utf8","key":"benchmarks/S1724_PDB_sequences.csv","truncatedAt":512},{"bytes":"DOI,weight,blosum62,measure,pos,pdbs_template,year,protein,PDBs_template,id,aro,PDB_wild,reference,ca_depth,mut_count,neg,pdb_mutant,sul,relative_bfactor,mutation_type,source,mutation_code,swissprot,mutation_based,ph,neu,ddg,method,phi,psi,sst,rsa,effect,mutated_chain,res_depth,temperature,acc,don,dtm,pam250,length,uniprot,PMID,PIR_ID,clean_DOI,dg_wt,dg,paper_seq,GJR_trim_seq,paper_seq_muts\n10.1016/j.jmb.2004.09.023,285474.16,-3.0,Fluorescence,,,2004.0,\"Spectrin alpha chain, non-erythrocytic 1\",,7260.0,,1AJ","encoding":"utf8","key":"benchmarks/S1724_thermomutdb_cleaned_withseq.csv","truncatedAt":512}],"sampleObjects":["230515_K50dG_dmsv4_dmsv5_dmsv7_concat260429.csv  (read this first — ΔG is already computed)","mgnify_training_index.csv  (name, split, seq, dG, PDB_name — the published train/val/test split)","benchmarks/S1724_thermomutdb_cleaned_withseq.csv  (curated ThermoMutDB measurements with sequences)","benchmarks/TED_under100aa_filtered_20percent_5k.csv  (TED domains <100 aa, 20% identity, 5k/bin)","220122_dmsv4_count_table.csv  (raw NGS counts; only needed to re-derive ΔG)"],"sampleRows":[{"aa_seq":"DEEPAAKNNETKAEKFIRLGEYRMNKAIDAIGRLEHLANRSAYEYTAEQVEAMFGALENKVADVKAKFNT","deltaG":-1.39,"deltaG_c":-0.13,"deltaG_t":-15,"dm_design":false,"lib":"dms4","mgnify":true,"name":"rocklin_batch2_0","split":"train"},{"aa_seq":"ISGDYAADQEAITKLLQGVSAFIADAGPAVEAGYLPLEAAKAMIMAAVRRFKLGREVEDALDMIG","deltaG":1.72,"deltaG_c":1.87,"deltaG_t":1.45,"dm_design":false,"lib":"dms4","mgnify":true,"name":"rocklin_batch2_550891","split":""}],"sidecars":[{"bytes":"# SaProtΔG / ESM3ΔG — Supplementary Data\n\nPredicted and experimental values accompanying the paper:\n**\"Accurate protein stability prediction for small domains using mega-scale experiments\"**\n\nEach CSV contains identifiers (name/pdb/ted_id), experimental values,\nand one `predicted_<model>` column per model evaluated.\n\n---\n\n## Training / Raw Datasets\n\n| File | Description |\n|------|-------------|\n| `230515_K50dG_dmsv4_dmsv5_dmsv7_concat260429.csv` | Full MGnify Stability dataset with recalibrated ΔG values (dmsv4, dmsv5, dmsv7 combined) |\n| `220122_dmsv4_count_table.csv` | DMSv4 raw NGS count table |\n| `220331_dmsv5_count_table.csv` | DMSv5 raw NGS count table |\n| `mgnify_training_index.csv` | MGnify sequence index used for training |\n| `dmsv4_filtered_train_splits.csv` | DMSv4 training split after unstructured terminus filtering |\n| `dmsv7_analysis.zip` | Complete DMSv7 analysis folder (notebooks, scripts, intermediate results) |\n\n---\n\n## Benchmark Datasets\n\nAll benchmark evaluation datasets are bundled in **`benchmarks.zip`**.\n\n| File (inside benchmarks.zip) | Description |\n|------------------------------|-------------|\n| `S1724_PDB_sequences.csv` | S1724 benchmark: PDB sequence data for 35 wild-type proteins |\n| `S1724_thermomutdb_cleaned_withseq.csv` | S1724 benchmark: curated ThermoMutDB stability measurements with sequences |\n| `TED_under100aa_filtered_20percent_5k.csv` | TED domains <100 aa, filtered to 20% identity, 5k cap per temperature bin |\n| `TED_100to200aa_filtered_1point3mill.csv` | TED domains 100–200 aa, ~1.3M sequences |\n\n---\n\n## Prediction Results\n\nAll per-figure prediction CSVs are bundled in **`figures.zip`**.\n\n| File (inside figures.zip) | Figure | Description |\n|---------------------------|--------|-------------|\n| `fig1_2__mgnify_stability_predictions.csv` | 1 & 2 | MGnify test set stability predictions |\n| `fig3__megascale_point_mutants_predictions.csv` | 3 | Megascale ΔΔG point mutant predictions |\n| `fig3__mgnify_insdel_predictions.csv` | 3 | MGnify insertion/deletion predictions |\n| `fig3__megascale_insdel_predictions.csv` | 3 | Megascale insertion/deletion predictions |\n| `fig4__thermomut_predictions.csv` | 4 | ThermoMut (S1724) benchmark predictions |\n| `fig5__TED_temperature_predictions.csv` | 5 | TED domains with organism growth temperature predictions |\n| `fig6__dmsv2_denovo_predictions.csv` | 6 | De novo designs from Cho et al. (2025) |\n| `fig6__nanobody_predictions.csv` | 6 | Nanobody apparent melting temperature predictions |\n| `fig6__rosetta_dark_matter_designs_predictions.csv` | 6 | Rosetta de novo design predictions |\n| `fig6__rfdiffusion_binders_predictions.csv` | 6 | RFdiffusion binder design predictions |","encoding":"utf8","key":"README.md","name":"README.md","size":2701,"truncated":false}],"sourceObjectKey":"benchmarks/S1724_PDB_sequences.csv","starterCell":"# MGnify mega-scale protein stability — start from the concat ΔG table.\nimport os, pandas as pd\nMNT = os.environ.get('SCIGANTIC_MOUNT_PATH', '/mnt/archive')  # also symlinked at ~/archive\n\n# 2.2 GB file: load only the columns you need.\ncols = ['name', 'aa_seq', 'deltaG', 'deltaG_95CI', 'mgnify', 'dm_design', 'lib', 'split']\ndf = pd.read_csv(f'{MNT}/230515_K50dG_dmsv4_dmsv5_dmsv7_concat260429.csv', usecols=cols)\nprint(df.shape)                       # ~1.8M rows\nprint(df['deltaG'].describe())        # folding free energy (kcal/mol); higher = more stable\nprint('from MGnify:', int(df['mgnify'].sum()), ' de novo designs:', int(df['dm_design'].sum()))\n\n# Keep confidently-measured rows (drop clamped / wide-CI):\nclean = df[(df['deltaG'].between(-5, 5)) & (df['deltaG_95CI'] < 1.0)]\nprint(clean.shape)\n\n# To reproduce the published train/val/test split + target:\nidx = pd.read_csv(f'{MNT}/mgnify_training_index.csv')  # name, split, seq, dG, PDB_name\nprint(idx['split'].value_counts())","totalObjects":"The concat ΔG table (2.2 GB) + 2 raw NGS count tables (DMSv4 ~486 MB, DMSv5 ~513 MB) + the training index (~138 MB) + the DMSv4 filtered splits (~60 MB) + benchmarks/ (S1724, ThermoMutDB, TED) + the K50/ΔG pipeline zip + README."},"summary":"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…","title":"MGnify Mega-Scale Protein Stability Dataset","updatedAt":"2026-10-28T16:48:24.380Z","visibility":"public","websiteUrl":"https://doi.org/10.5281/zenodo.19411306"}}