LiteFold/MegaScale-Tsuboyama2023

MegaScale Tsuboyama 2023 Stability

Advances in DNA sequencing and machine learning are providing insights into protein sequences and structures on an enormous scale. However, the energetics driving folding are invisible in these structures and remain largely unknown. The hidden thermodynamics of folding can drive disease, shape protein evolution and guide protein engineering, and new approaches are needed to reveal these thermodynamics for every sequence and structure. Here we present cDNA display proteolysis, a method for measuring thermodynamic folding stability for up to 900,000 protein domains in a one-week experiment. From 1.8 million measurements in total, we curated a set of around 776,000 high-quality folding stabilities covering 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. Using this extensive dataset, we quantified (1) environmental factors influencing amino acid fitness, (2) thermodynamic couplings (including unexpected interactions) between protein sites, and (3) the global divergence between evolutionary amino acid usage and protein folding stability. We also examined how our approach could identify stability determinants in designed proteins and evaluate design methods. The cDNA display proteolysis method is fast, accurate and uniquely scalable, and promises to reveal the quantitative rules for how amino acid sequences encode folding stability.

Splits

Rows are assigned with a deterministic hash split: sha256(record_id) % 10, where bucket 0 is test and buckets 1-9 are train.

Split	Rows
train	269,451
test	29,820
total	299,271

Source Tables

Record type	Rows
double_mutants	269,516
site_features	17,093
natural_sites	5,974
gemme_natural_sites	5,396
external_delta_g	1,292

Columns

The table starts with normalized columns for browsing and model input:

• record_id: stable SHA-256 row identifier
• dataset_id, source_file, source_table, source_row_index: source provenance
• table_name, figure, record_type: source table metadata
• protein_id, mutation_name, sequence, sequence_length: protein and mutation identifiers when present
• position, position_1, position_2, wt_aa, mut_aa: site and residue fields when present
• delta_g, reconstructed_delta_g, thermodynamic_delta_g, previous_delta_g, target, score_value: normalized score fields when present
• raw_row_json: complete original upstream row as JSON
• split_bucket: deterministic split bucket

All original source fields are also preserved as snake_case string columns. See metadata/column_mapping.parquet for the original field-name mapping and metadata/source_tables.parquet for per-table row counts.

Usage

from datasets import load_dataset

ds = load_dataset("LiteFold/MegaScale-Tsuboyama2023")
print(ds)
print(ds["train"][0])

Load selected columns:

from datasets import load_dataset

cols = [
    "record_id",
    "record_type",
    "protein_id",
    "mutation_name",
    "sequence",
    "delta_g",
    "score_value",
]
train = load_dataset("LiteFold/MegaScale-Tsuboyama2023", split="train", columns=cols)

Filter to double mutants:

from datasets import load_dataset

train = load_dataset("LiteFold/MegaScale-Tsuboyama2023", split="train")
double_mutants = train.filter(lambda row: row["record_type"] == "double_mutants")

Read a raw source field from the JSON payload:

import json
from datasets import load_dataset

row = next(iter(load_dataset("LiteFold/MegaScale-Tsuboyama2023", split="train", streaming=True)))
raw = json.loads(row["raw_row_json"])
print(raw.keys())

Metadata tables can be loaded directly:

from datasets import load_dataset

source_tables = load_dataset(
    "parquet",
    data_files="hf://datasets/LiteFold/MegaScale-Tsuboyama2023/metadata/source_tables.parquet",
    split="train",
)
column_mapping = load_dataset(
    "parquet",
    data_files="hf://datasets/LiteFold/MegaScale-Tsuboyama2023/metadata/column_mapping.parquet",
    split="train",
)

Preparation

The normalization script used for this upload is included at scripts/prepare_megascale_tsuboyama2023_dataset.py.

Citation

@article{tsuboyama2023megascale,
  title     = {Mega-scale experimental analysis of protein folding stability in biology and design},
  author    = {Tsuboyama, Kotaro and Dauparas, Justas and Chen, Jonathan and Laine, Elodie and Mohseni Behbahani, Yasser and Weinstein, Jonathan J. and Mangan, Niall M. and Ovchinnikov, Sergey and Rocklin, Gabriel J.},
  journal   = {Nature},
  volume    = {620},
  number    = {7973},
  pages     = {434--444},
  year      = {2023},
  publisher = {Nature Publishing Group},
  doi       = {10.1038/s41586-023-06328-6}
}