lamm-mit/silkome-masp

Silkome MaSp

lamm-mit/silkome-masp is the major ampullate spidroin (MaSp) sequence-property subset used for the SilkomeGPT study:

Wei Lu, David L. Kaplan, and Markus J. Buehler, "Generative Modeling, Design, and Analysis of Spider Silk Protein Sequences for Enhanced Mechanical Properties", Advanced Functional Materials 34, 2311324 (2024).

The dataset is curated from lamm-mit/silkome-full by selecting rows whose category1 is one of:

MaSp, MaSp1, MaSp2, MaSp2B, MaSp3, MaSp3B

This reproduces the paper's MaSp-focused sequence-to-fiber-property table: protein-level spidroin sequences paired with fiber-level mechanical measurements.

Dataset Summary

• Rows: 1,033
• Unique sequences: 1,028
• Unique idv fiber/property identifiers: 233
• Splits:
  • full: 1,033 rows, the complete paper-style MaSp subset
  • train: 895 rows, benchmark training split
  • test: 138 rows, benchmark test split
• Benchmark split method:
  • deterministic random split with seed 0
  • grouped by idv, so all sequences associated with the same fiber/property measurement stay in only one split
  • test fraction: 15% of idv groups, giving 35 test groups out of 233 total groups
  • zero idv overlap between train and test
  • zero overlap in 4-property tuples (toughness, E, strength, strain) between train and test
  • zero overlap in 8-property tuples (toughness, toughnessSD, E, ESD, strength, strengthSD, strain, strainSD) between train and test
• Source split provenance remains available in source_split:
  • source train: 912 rows in full
  • source test: 121 rows in full
• Sequence length range:
  • minimum: 115 amino acids
  • median: 378 amino acids
  • maximum: 1,854 amino acids

Category counts:

category1	rows
MaSp1	349
MaSp2	331
MaSp	223
MaSp3B	47
MaSp3	43
MaSp2B	40

Columns

Sequence and metadata columns:

• idv: fiber/property identifier from the source Silkome data
• sequence: amino acid sequence
• length: sequence length
• family, genus, species
• category1, category2
• ncbi
• sex
• source_split: original split in lamm-mit/silkome-full
• benchmark_split: deterministic idv-grouped split used for this dataset's train/test

Core fiber-level mechanical properties:

• toughness
• toughnessSD
• E
• ESD
• strength
• strengthSD
• strain
• strainSD

The eight core values have no missing entries in this subset.

Normalized property columns from the source dataset:

• toughnessNorm
• toughnessSDNorm
• ENorm
• ESDNorm
• strengthNorm
• strengthSDNorm
• strainNorm
• strainSDNorm

These normalized columns correspond to the 8D property vector used for conditional SilkomeGPT generation:

[toughness, SD toughness, E, SD E, strength, SD strength, strain, SD strain]

The dataset also preserves additional available Silkome material-property columns, including 1weightloss, 5weightloss, 10weightloss, cryst, birefri, dia, WC, and SuperContra, plus their normalized variants where available.

Usage

from datasets import load_dataset

full = load_dataset("lamm-mit/silkome-masp", split="full")
train = load_dataset("lamm-mit/silkome-masp", split="train")
test = load_dataset("lamm-mit/silkome-masp", split="test")

print(full, train, test)

row = train[0]
sequence = row["sequence"]
target = [
    row["toughness"],
    row["toughnessSD"],
    row["E"],
    row["ESD"],
    row["strength"],
    row["strengthSD"],
    row["strain"],
    row["strainSD"],
]

For a normalized SilkomeGPT-style target:

target_norm = [
    row["toughnessNorm"],
    row["toughnessSDNorm"],
    row["ENorm"],
    row["ESDNorm"],
    row["strengthNorm"],
    row["strengthSDNorm"],
    row["strainNorm"],
    row["strainSDNorm"],
]

Intended Use

This dataset is intended for research on:

• spider silk protein language modeling
• sequence-to-property prediction
• conditional sequence generation
• protein/material design workflows
• analysis of MaSp sequence motifs and fiber-level mechanical measurements

Important Caveats

The mechanical targets are fiber-level measurements, while each row contains a protein-level spidroin sequence. Multiple spidroin sequences can share the same measured fiber-property tuple because spider silk fibers are complex multicomponent materials. This dataset is therefore best understood as a weakly supervised sequence-to-fiber-property resource, not as evidence that a single sequence alone causally determines a fiber's mechanical behavior.

For benchmark work, avoid leakage between examples that share identical or near-identical property measurements or identical fiber IDs. The provided train/test split is grouped by idv for this reason. The full split is provided for paper-style fine-tuning or generation workflows that use the full MaSp sequence-property subset.

Provenance

This dataset is derived from the Spider Silkome resource via lamm-mit/silkome-full. The full split was constructed by combining the source train and test splits, retaining the original split in source_split, and filtering to MaSp categories listed above. The benchmark train/test splits are then constructed from full using the deterministic idv-grouped split described above.

Splitting Guidance

For sequence-to-property prediction, a plain random row split is usually too optimistic because many rows can share the same fiber-level measurement. Prefer one of:

1. Use the provided split: train/test in this repository. This is grouped by idv.
2. Make a new grouped split by idv if you need a different train/test fraction.
3. Make a stricter grouped split by property tuple if your model-selection protocol must prevent identical target vectors from appearing in both train and test.

Example:

import numpy as np
from datasets import load_dataset

ds = load_dataset("lamm-mit/silkome-masp", split="full")
df = ds.to_pandas()

rng = np.random.default_rng(0)
groups = np.array(sorted(df["idv"].dropna().unique()))
rng.shuffle(groups)

n_test = round(0.15 * len(groups))
test_groups = set(groups[:n_test])

test_df = df[df["idv"].isin(test_groups)]
train_df = df[~df["idv"].isin(test_groups)]

Citation

Please cite the SilkomeGPT paper if you use this dataset:

@article{lu2024generative,
  title = {Generative Modeling, Design, and Analysis of Spider Silk Protein Sequences for Enhanced Mechanical Properties},
  author = {Lu, Wei and Kaplan, David L. and Buehler, Markus J.},
  journal = {Advanced Functional Materials},
  volume = {34},
  number = {11},
  pages = {2311324},
  year = {2024},
  doi = {10.1002/adfm.202311324},
  publisher = {Wiley}
}

Also cite the original Spider Silkome resource associated with the source data where appropriate.