RosettaCommons/MegaScale

Mega-scale experimental analysis of protein folding stability in biology and design

The full MegaScale dataset contains 1,841,285 thermodynamic folding stability measurements using cDNA display proteolysis of natural and designed proteins. From these 776,298 high-quality folding stabilities (dataset2) cover all single amino acid variants and selected double mutants of 331 natural and 148 de novo designed protein domains 40–72 amino acids in length. Of these mutations, 607,839 have the wild-type ΔG is below 4.75 kcal mol^−1 (dataset3) allowing for the estimate of the ΔΔG of mutation. Of these

*** IMPORTANT! Please register your use of these data so that we (the Rocklin Lab) can continue to release new useful datasets!! This will take 10 seconds!! ***

Quickstart Usage

Install HuggingFace Datasets package

Each subset can be loaded into python using the Huggingface datasets library. First, from the command line install the datasets library

$ pip install datasets

Optionally set the cache directory, e.g.

$ HF_HOME=${HOME}/.cache/huggingface/
$ export HF_HOME

then, from within python load the datasets library

>>> import datasets

Load model datasets

To load one of the MegaScale model datasets (see available datasets below), use datasets.load_dataset(...):

>>> dataset_tag = "dataset3_single"
>>> dataset3_single = datasets.load_dataset(
  path = "RosettaCommons/MegaScale",
  name = dataset_tag,
  data_dir = dataset_tag)
Downloading readme: 100%|██████████████████████████████| 17.0k/17.0k [00:00<00:00, 290kB/s]
Downloading data: 100%|███████████████████████████████| 39.8M/39.8M [00:01<00:00, 36.9MB/s]
Downloading data: 100%|███████████████████████████████| 41.2M/41.2M [00:00<00:00, 57.3MB/s]
Downloading data: 100%|███████████████████████████████| 40.0M/40.0M [00:00<00:00, 43.9MB/s]
Downloading data: 100%|███████████████████████████████| 15.5M/15.5M [00:00<00:00, 26.8MB/s]
Downloading data: 100%|███████████████████████████████| 14.9M/14.9M [00:00<00:00, 29.4MB/s]
Generating train split: 100%|█████████| 1503063/1503063 [00:05<00:00, 262031.56 examples/s]
Generating test split: 100%|████████████| 169032/169032 [00:00<00:00, 264056.98 examples/s]
Generating val split: 100%|█████████████| 163968/163968 [00:00<00:00, 251806.22 examples/s]

and the dataset is loaded as a datasets.arrow_dataset.Dataset

>>> dataset3_single
DatasetDict({
    train: Dataset({
        features: ['name', 'dna_seq', 'log10_K50_t', 'log10_K50_t_95CI_high', 'log10_K50_t_95CI_low', 'log10_K50_t_95CI', 'fitting_error_t', 'log10_K50unfolded_t', 'deltaG_t', 'deltaG_t_95CI_high', 'deltaG_t_95CI_low', 'deltaG_t_95CI', 'log10_K50_c', 'log10_K50_c_95CI_high', 'log10_K50_c_95CI_low', 'log10_K50_c_95CI', 'fitting_error_c', 'log10_K50unfolded_c', 'deltaG_c', 'deltaG_c_95CI_high', 'deltaG_c_95CI_low', 'deltaG_c_95CI', 'deltaG', 'deltaG_95CI_high', 'deltaG_95CI_low', 'deltaG_95CI', 'aa_seq_full', 'aa_seq', 'mut_type', 'WT_name', 'WT_cluster', 'log10_K50_trypsin_ML', 'log10_K50_chymotrypsin_ML', 'dG_ML', 'ddG_ML', 'Stabilizing_mut', 'pair_name', 'split_name'],
        num_rows: 1503063
    })
    test: Dataset({
        features: ['name', 'dna_seq', 'log10_K50_t', 'log10_K50_t_95CI_high', 'log10_K50_t_95CI_low', 'log10_K50_t_95CI', 'fitting_error_t', 'log10_K50unfolded_t', 'deltaG_t', 'deltaG_t_95CI_high', 'deltaG_t_95CI_low', 'deltaG_t_95CI', 'log10_K50_c', 'log10_K50_c_95CI_high', 'log10_K50_c_95CI_low', 'log10_K50_c_95CI', 'fitting_error_c', 'log10_K50unfolded_c', 'deltaG_c', 'deltaG_c_95CI_high', 'deltaG_c_95CI_low', 'deltaG_c_95CI', 'deltaG', 'deltaG_95CI_high', 'deltaG_95CI_low', 'deltaG_95CI', 'aa_seq_full', 'aa_seq', 'mut_type', 'WT_name', 'WT_cluster', 'log10_K50_trypsin_ML', 'log10_K50_chymotrypsin_ML', 'dG_ML', 'ddG_ML', 'Stabilizing_mut', 'pair_name', 'split_name'],
        num_rows: 169032
    })
    val: Dataset({
        features: ['name', 'dna_seq', 'log10_K50_t', 'log10_K50_t_95CI_high', 'log10_K50_t_95CI_low', 'log10_K50_t_95CI', 'fitting_error_t', 'log10_K50unfolded_t', 'deltaG_t', 'deltaG_t_95CI_high', 'deltaG_t_95CI_low', 'deltaG_t_95CI', 'log10_K50_c', 'log10_K50_c_95CI_high', 'log10_K50_c_95CI_low', 'log10_K50_c_95CI', 'fitting_error_c', 'log10_K50unfolded_c', 'deltaG_c', 'deltaG_c_95CI_high', 'deltaG_c_95CI_low', 'deltaG_c_95CI', 'deltaG', 'deltaG_95CI_high', 'deltaG_95CI_low', 'deltaG_95CI', 'aa_seq_full', 'aa_seq', 'mut_type', 'WT_name', 'WT_cluster', 'log10_K50_trypsin_ML', 'log10_K50_chymotrypsin_ML', 'dG_ML', 'ddG_ML', 'Stabilizing_mut', 'pair_name', 'split_name'],
        num_rows: 163968
    })
})

which is a column oriented format that can be accessed directly, written to disk as a parquet file or converted in to a pandas.DataFrame, e.g.

>>> dataset3_single['train'].data.column('name')
>>> dataset3_single['train'].to_parquet("dataset3_single_train.parquet")
>>> dataset3_single.to_pandas()[[WT_name', 'mut_type', 'dG_ML', 'ddG_ML']]
                        WT_name mut_type               dG_ML               ddG_ML
0        r10_437_TrROS_Hall.pdb      E1Q  2.9212264903176783   0.2949200736672686
1        r10_437_TrROS_Hall.pdb      E1Q  2.9212264903176783   0.2949200736672686
2        r10_437_TrROS_Hall.pdb      E1Q  2.9212264903176783   0.2949200736672686
3        r10_437_TrROS_Hall.pdb      E1Q  2.9212264903176783   0.2949200736672686
4        r10_437_TrROS_Hall.pdb      E1Q  2.9212264903176783   0.2949200736672686
...                         ...      ...                 ...                  ...
1503058       HEEH_rd3_0055.pdb     L43C   1.629862324762064  0.07132877903687329
1503059       HEEH_rd3_0055.pdb     L43C   1.629862324762064  0.07132877903687329
1503060       HEEH_rd3_0055.pdb     L43C   1.629862324762064  0.07132877903687329
1503061       HEEH_rd3_0055.pdb     L43C   1.629862324762064  0.07132877903687329
1503062       HEEH_rd3_0055.pdb     L43C   1.629862324762064  0.07132877903687329

Overview of Datasets

dataset1: The whole dataset 1,841,285 stability measurements

• All mutations in G0-G11 (see below)

dataset2: The curated a set of 776,298 high-quality folding stabilities covers

• All mutations in G0 + G1 (see below)
• all single amino acid variants and selected double mutants of 331 natural and 148 de novo designed protein domains 40–72 amino acids in length
• comprehensive double mutations at 559 site pairs spread across 190 domains (a total of 210,118 double mutants)
• 36 different 3-residue networks
  • all possible single and double substitutions in both the wild-type background and the background in which the third amino acid was replaced by alanine
  • (400 mutants × 3 pairs × 2 backgrounds ≈ 2,400 mutants in total for each triplet)

dataset3: Curated set of 325,132 ΔG measurements at 17,093 sites in 365 domains

• All mutations in G0
• All mutations in dataset2 where the wild-type ΔG is below 4.75 kcal mol^−1 (dataset3) allowing for the estimate of the ΔΔG of mutation.

dataset3_single: The single point mutations in dataset3

• Using the train/val/test splits defined in ThermoMPNN (Dieckhaus, et al., 2024)

dataset3_single_cv: The single point mutations in dataset3

• Using the 5-fold cross validation splits (train_[0-4]/val_[0-4]/test_[0-4]) defined in ThermoMPNN (Dieckhaus, et al., 2024)

AlphaFold_model_PDBs: AlphaFold predicted structures of wildtype domains (even if structures exist in the Protein Databank)

Target Selection

Targets consist of natural, designed, and destabilized wild-type

983 natural targets were selected from the all monomeric proteins in the protein databank having 30–100 amino acid length range that met the following criteria:

• Conisted of more than a single helix
• Did not contain other molecules (for example, proteins, nucleic acids or metals)
• Were not annotated to have DNAse, RNAse, or protease inhibition activity
• Had at most four cysteins
• Were not sequence redundant (amino acid sequence distance <2) with another selected sequence These were then processed by
• AlphaFold was used to predict the structure (including those that had solved structures in the PDB), which was used to trim amino acids from the N- and C termini that had a low number of contacts with any other residues.
• selected domains with up to 72 amino acids after excluding N- or C-terminal flexible loops

designed targets were selected from

• previous Rosetta designs with ααα, αββα, βαββ, and ββαββ topologies (40 to 43 amino acids)
• new ββαα proteins designed using Rosetta (47 amino acids)
• new domains designed by trRosetta hallucination (46 to 69 amino acids)

121 destabilized wild-type backgrounds targets were also included.

Library construction

The cDNA proteolysis screening was conducted as four giant synthetic DNA oligonucleotide libraries and obtained K50 values for 2,520,337 sequences, 1,841,285 of these measurements are included here:

• Library 1:
  • ~250 designed proteins and ~50 natural proteins that are shorter than 45 amino acids
  • padding by Gly, Ala and Ser amino acids so that all sequences have 44 amino acids
  • ~244,000 sequences Purchased from Agilent Technologies, length 230 nt.
• Library 2:
  • ~350 natural proteins that have PDB structures that are in a monomer state and have 72 or less amino acids after removing N and C-terminal linkers
  • padding by Gly, Ala and Ser amino acids so that all sequences have 72 amino acids
  • ~650,000 sequences
  • also includes scramble sequences to construct unfolded state model.
  • Purchased from Twist Bioscience, length 250 nt.
• Library 3:
  • ~150 designed proteins
  • comprehensive deletion and Gly or Ala insertion of all wild-type proteins included in Library 1 and Libary 2
  • amino acid sequences for comprehensive double mutant analysis on polar amino acid pairs
  • ~840,000 sequences
  • Purchased from Twist Bioscience, length 250 nt.
• Library 4:
  • Amino acid sequences for exhaustive double mutant analysis on amino acid pairs located in close proximity
  • overlapped sequences to calibrate effective protease concentration and to check consistency between libraries
  • ~900,000 sequences
  • Purchased from Twist Bioscience, length 300 nt.

Bayesian Stability Analysis

Each target was analyzed and given a single quality category score G0-G11, which were then sorted into one of three datasets. The quality scores are

• G0: Good (wild-type ΔG values below 4.75 kcal mol^−1)
• G1: Good but WT outside dynamic range
• G2: Too much missing data
• G3: WT dG is too low
• G4: WT dG is inconsistent
• G5: Poor trypsin vs. chymotrypsin correlation
• G6: Poor trypsin vs. chymotrypsin slope
• G7: Too many stabilizing mutants
• G8: Multiple cysteins (probably folded properly)
• G9: Multiple cysteins (probably misfolded)
• G10: Poor T-C intercept
• G11: Probably cleaved in folded state(s)

ThermoMPNN splits

ThermoMPNN is a message passing neural network that predicts protein ΔΔG of mutation based on ProteinMPNN (Dauparas et al., 2022). ThermoMPNN uses in part data from the MegaScale dataset. From the mutations in dataset2, 272,712 mutations across 298 proteins were curated that were single point mutants, reliable, and where the baseline is wildtype.