lamm-mit/silkome-masp-idv-grouped

Silkome MaSp Idv Grouped

lamm-mit/silkome-masp-idv-grouped is an idv-grouped MaSp-only Silkome dataset for predicting dragline fiber-level mechanical properties from the set of major ampullate spidroin sequences associated with the same measured fiber/property identifier.

This dataset is derived from lamm-mit/silkome-masp. Instead of placing one protein sequence in each row, this dataset groups sequence rows with the same idv into one sample:

{MaSp sequences for one idv} -> {toughness, E, strength, strain}

This formulation is designed for ESMC embedding models followed by a set-level aggregation model.

Dataset Summary

• Rows / unique idv groups: 233
• Splits:
  • full: 233 grouped samples
  • train: 198 grouped samples
  • test: 35 grouped samples
• Split method: inherited benchmark_split from source dataset
• Train/test idv overlap: 0
• Train/test 4-property tuple overlap: 0
• Source rows before filtering/grouping: 1033
• MaSp sequence rows represented: 1033
• Sequences per idv:
  • min: 1
  • median: 4.0
  • mean: 4.43
  • max: 13
• Maximum X-linker concatenated sequence length: 5896 residues/tokens

Category counts across all grouped sequence lists:

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

Representation

The recommended model input is:

• sequences: list of amino-acid sequences for the idv
• sequence_categories: category for each sequence, e.g. MaSp1, MaSp2, MiSp, AcSp
• sequence_lengths: length of each sequence
• category_counts_json or the n_* count columns for coarse composition features

For ESMC-style workflows, embed each sequence independently and then aggregate the sequence embeddings with a permutation-invariant model such as mean pooling, attention pooling, DeepSets, or a set transformer. A useful architecture is:

ESMC(sequence_i) + learned_category_embedding(category_i) -> set aggregator -> property head

Example:

from datasets import load_dataset

ds = load_dataset("lamm-mit/silkome-masp-idv-grouped", split="train")
row = ds[0]

for seq, category in zip(row["sequences"], row["sequence_categories"]):
    # embed seq with ESMC independently
    # optionally concatenate/add a learned category embedding
    ...

target = row["targets"]  # [toughness, E, strength, strain]

The dataset also includes convenience serializations:

• sequence_fasta: multi-FASTA text for inspection/export
• sequence_concat_x25: sequences concatenated with 25 X residues between proteins
• sequence: alias of sequence_concat_x25 for compatibility with simple sequence-only pipelines

The concatenated string is only a baseline convenience. It is not the primary biological representation, and it can be too long for some ESMC settings. Prefer sequences whenever possible.

Columns

Grouping and sequence columns:

• idv
• n_sequences
• sequences
• sequence_lengths
• sequence_hashes
• sequence_categories
• unique_sequence_categories
• sequence_subcategories
• source_row_indices
• source_splits
• benchmark_split
• sequence_fasta
• sequence_concat_x25
• sequence_concat_length
• sequence

Composition and metadata columns:

• family, genus, species
• sex
• ncbi
• category_counts_json
• category count columns: n_acsp, n_agsp1, n_agsp2, n_ampullate_spidroin, n_crsp, n_cysp, n_flag, n_masp, n_masp1, n_masp2, n_masp2b, n_masp3, n_masp3b, n_misp, n_pflag, n_putative_spidroin, n_pysp, n_spidroin
• has_masp1, has_masp2, has_masp3
• total_sequence_length, mean_sequence_length, max_sequence_length

Target columns:

• toughness
• E
• strength
• strain
• targets: [toughness, E, strength, strain]
• toughnessSD, ESD, strengthSD, strainSD
• target_sds: [toughnessSD, ESD, strengthSD, strainSD]
• normalized versions where available: toughnessNorm, ENorm, strengthNorm, strainNorm, and the corresponding SD-normalized columns

Intended Use

This dataset is intended for:

• fiber-level mechanical property prediction from grouped Silkome sequence sets
• ESMC embedding plus set aggregation benchmarks
• analysis of sequence category composition and sequence-property relationships
• leakage-safer train/test evaluation than random row-level sequence splits

Important Caveats

The targets are fiber-level measurements, not direct labels for each protein. This dataset is MaSp-only. It is the most direct grouped formulation for dragline mechanics, but it still does not prove that the listed proteins alone determine the measured fiber properties.

The original Spider Silkome study measured dragline fiber properties and associated them with silkome sequence data. Experimental, environmental, spinning, structural, and post-translational factors can also influence fiber mechanics.

Provenance

The grouped dataset was created from lamm-mit/silkome-masp by grouping rows by idv. The train/test splits are provided through the benchmark_split column and exposed as Hugging Face splits.

Citation

Please cite the original Spider Silkome resource:

@article{arakawa2022spider,
  title = {1000 spider silkomes: Linking sequences to silk physical properties},
  author = {Arakawa, Kazuharu and Kono, Nobuaki and Malay, Ali D. and Tateishi, Ayaka and Ifuku, Nao and Masunaga, Hiroyasu and Sato, Ryota and Tsuchiya, Kousuke and Ohtoshi, Rintaro and Pedrazzoli, Daniel and others},
  journal = {Science Advances},
  volume = {8},
  number = {41},
  pages = {eabo6043},
  year = {2022},
  doi = {10.1126/sciadv.abo6043}
}

Please also cite SilkomeGPT if you use this dataset in the context of SilkomeGPT-style modeling:

@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}
}