NatalijaStepurko/seq-structure-convergence

Residue-aligned protein embeddings: sequence and structure models, every layer

Nine model arms — sequence models (ESM-2 35M and 650M, CARP-38M), structure models (ProteinMPNN, ESM-IF1), two copies of ESM-1v that differ only by random seed, and untrained copies of ESM-2 35M and ProteinMPNN — each describing the same proteins at every layer, in two parts:

• Residue level: 10,045 residues from 35 whole protein domains, aligned row for row across all nine arms, with structural and functional labels per residue.
• Protein level: one vector per protein (the mean over its residues) for 4,894 protein domains, every arm and layer, with fold, kingdom and function labels per protein.

It is built as a test bed for measures of representation similarity (CKA, SVCCA, mutual k-NN and others). The seeded ESM-1v pair is a ceiling for how alike two models can be, the untrained copies are floors, and the sequence–structure pairs sit in between. Reference scores for twelve pairs are included, so an implementation can be checked against known values on identical inputs.

Produced by seq-structure-convergence, which studies whether a sequence model and a structure model learn the same biology; results.

Why it is useful

• Known answers at both ends. Measures of representation similarity are usually tested on pairs whose true relationship is unknown. Here the range is fixed: two copies of ESM-1v that differ only by random seed set the ceiling (seeded copies of one protein model are rarely released), untrained copies set the floor, and the sequence–structure pairs sit in between. A new measure can be checked for whether it orders these correctly.
• A genuinely cross-modal pair. The sequence models never see coordinates and the structure models never see the sequence, yet both describe the same residues, row for row. Pairs like this are hard to find outside proteins.
• Every layer. Where in a network a property appears, and where two networks agree, can be studied directly, without re-running any model.
• Widths from 128 to 1,280 dimensions. Several similarity measures drift with representation width even on unrelated data; the range here makes that effect testable.
• A shared training target to control for. All trained models output amino-acid identity, which by itself produces agreement. The residue table carries the amino acid, so methods for removing a shared target can be tested against the reference scores computed with and without it.
• Reference values to test code against. Implementations of CKA, SVCCA and nearest-neighbour overlap differ in details (centring, row caps, neighbour counts). reference_scores.csv gives 72 values on these exact inputs, each beside its score on scrambled data.
• Labels for probing. Secondary structure, solvent accessibility, burial, flexibility and binding, active and modification sites per residue; fold class, kingdom and function per protein.
• Protein-level comparisons at scale. The protein-level part covers 4,894 domains, enough for protein properties such as CATH fold class or enzyme class to be well represented, in 1.4 GB.
• It saves compute. Producing these tensors takes the model weights, three separate software environments (ESM-IF1 and CARP have conflicting dependencies) and several hours of computation.

Model arms

File (embeddings/)	Model	Reads	Layers	Dim	Size
esm2_35m.safetensors	ESM-2 35M (esm2_t12_35M_UR50D)	sequence	13: input embedding + 12 blocks	480	125 MB
esm2_650m.safetensors	ESM-2 650M (esm2_t33_650M_UR50D)	sequence	34: input embedding + 33 blocks	1280	874 MB
esm1v_seed1.safetensors	ESM-1v (esm1v_t33_650M_UR90S_1)	sequence	34	1280	874 MB
esm1v_seed2.safetensors	ESM-1v (esm1v_t33_650M_UR90S_2), same model, different seed	sequence	34	1280	874 MB
carp_38m.safetensors	CARP-38M (convolutional)	sequence	4: blocks 4, 8, 12, 16 of 16	1024	82 MB
proteinmpnn.safetensors	ProteinMPNN encoder (v_48_020)	backbone coordinates	3 encoder layers	128	8 MB
esmif1.safetensors	ESM-IF1 (esm_if1_gvp4_t16_142M_UR50)	backbone coordinates	1: encoder output	512	10 MB
esm2_35m_untrained.safetensors	ESM-2 35M architecture, random weights	sequence	13	480	125 MB
proteinmpnn_untrained.safetensors	ProteinMPNN encoder, random weights	backbone coordinates	3	128	8 MB

Sizes are for the residue-level files. The same nine files under embeddings_protein/ are together 1.4 GB (ESM-2 650M and each ESM-1v copy 426 MB; the rest under 62 MB each).

Each residue-level file holds one tensor, layers, float16, shape [n_layers, 10045, dim]; row k is the same residue in every file and in residues.csv. Each protein-level file holds layers of shape [n_layers, 4894, dim]; row k is the same protein in every file and in proteins.csv. The structure models never see the sequence and the sequence models never see coordinates; all trained models are trained to output amino-acid identity, which is a source of agreement between them.

Files

File	Contents
embeddings/*.safetensors	residue level: the nine arms above, 10,045 residues
embeddings_protein/*.safetensors	protein level: the nine arms, one mean vector per protein, 4,894 proteins
proteins.csv	one row per protein, in protein-level tensor order: PDB ID, chain, length, CATH code and levels, UniProt accession, kingdom, enzyme class, protein class, localisation
residues.csv	one row per residue, in tensor order: chain, position, amino acid, secondary structure (helix / sheet / coil), relative solvent accessibility, buried (RSA < 0.25) or exposed, per-chain-normalised B-factor, binding / active / PTM site flags
chains.csv	one row per chain: PDB ID, chain, length, CATH code and levels, UniProt accession, kingdom, enzyme class, protein class, localisation
reference_scores.csv	CKA, SVCCA and mutual k-NN for twelve pairs on these residues, raw and with the per-amino-acid mean subtracted, each beside a scrambled-residue null
arms.json, arms_protein.json	layer count, dimension and size per arm, for each part

Loading

from huggingface_hub import snapshot_download
from safetensors.numpy import load_file
import pandas as pd

path = snapshot_download("NatalijaStepurko/seq-structure-convergence", repo_type="dataset",
                         allow_patterns=["residues.csv", "embeddings/esm2_35m.safetensors",
                                         "embeddings/proteinmpnn.safetensors"])
seq = load_file(f"{path}/embeddings/esm2_35m.safetensors")["layers"]     # (13, 10045, 480)
struct = load_file(f"{path}/embeddings/proteinmpnn.safetensors")["layers"]  # (3, 10045, 128)
residues = pd.read_csv(f"{path}/residues.csv")

# protein level: one vector per protein, e.g. the last layer of ESM-2 35M against CATH class
path = snapshot_download("NatalijaStepurko/seq-structure-convergence", repo_type="dataset",
                         allow_patterns=["proteins.csv", "embeddings_protein/esm2_35m.safetensors"])
prot = load_file(f"{path}/embeddings_protein/esm2_35m.safetensors")["layers"]  # (13, 4894, 480)
proteins = pd.read_csv(f"{path}/proteins.csv")
X, y = prot[-1], proteins["cath_class"]

The three 650M-parameter arms are 874 MB each; allow_patterns fetches only what you need.

How the proteins were chosen

From the 4,898 non-redundant CATH S35 domains with experimental structures used in the parent study, chains were taken in a random order (seed 42), keeping only chains that every arm describes in full, until the total reached 10,000 residues. Chains are whole, so estimates can be resampled by chain: residues within a chain are correlated, and resampling them independently understates the spread.

The protein-level part keeps the 4,894 of the 4,898 domains that all nine arms describe in full. Each vector is the mean over the protein's residues, in float32, stored as float16 — the pooling the parent study's protein-level probes use.

Reference scores

reference_scores.csv has, per pair, mode and measure: the layer pair used (layer_a, layer_b, indices into the tensors), the score, and the score with one model's rows shuffled (scrambled), which is what the measure reports for unrelated data. The layer pair is the one at which the parent study found that pair's peak; partial subtracts each residue's per-amino-acid mean from both models before measuring.

These values depend on the sample. CKA, SVCCA and mutual k-NN are all biased upward at smaller sample sizes, and unequally, so they are comparable within this dataset and not against numbers computed on a different number of residues. The implementations in the parent repository (src/ssc/metrics.py) evaluate SVCCA on at most 5,000 rows and mutual k-NN (k = 10) on at most 4,000; the reference values use those defaults.

Limits

• The residue-level part has 35 proteins: enough for residue-level comparisons. Protein-level questions belong on the protein-level part, which covers 4,894.
• Mean pooling discards where in the protein a signal sits; the protein-level vectors suit protein properties, not residue ones.
• The structure models are inverse-folding encoders; no structure-prediction model (such as AlphaFold) is included.
• CARP is stored at four of its sixteen blocks.

Provenance

Built by scripts/07_hf_dataset.py in the parent repository (--protein-level for the protein-level part) from the per-residue embeddings written by scripts/02_extract.py.

Licence and attribution

CC BY 4.0. The models are MIT-licensed (fair-esm, ProteinMPNN, sequence-models). The labels come from CATH (CC BY 4.0), UniProt (CC BY 4.0), RCSB PDB (CC0) and SIFTS (EMBL-EBI); when you use them, credit CATH (Sillitoe et al., 2021), UniProt (The UniProt Consortium, 2023) and SIFTS (Dana et al., 2019).

Citation

@misc{stepurko2026seqstructure,
  author = {Stepurko, Natalija},
  title  = {Residue-aligned protein embeddings from sequence and structure models},
  year   = {2026},
  url    = {https://huggingface.co/datasets/NatalijaStepurko/seq-structure-convergence}
}