SSP: protein secondary structure

Protein chains with a secondary-structure label for every residue, written as strings ("CCHHHHEEE…"):

• Training data and classic test sets from TAPE (Rao et al., 2019): train, validation, cb513, ts115, casp12. TAPE took this data from NetSurfP-2.0 (Klausen et al., 2019). Every row and value is TAPE's; the integer labels are written as letters, primary is renamed sequence, disorder is renamed resolved, and protein_length (always equal to the sequence length) is dropped.
• CASP test sets, CASP12 to CASP16, labelled with DSSP by a procedure that reproduces NetSurfP-2.0's training labels on 99.95% of residues (see How the CASP test sets were built). Each round comes in two versions: built from the PDB entries CASP lists (rcsb, default) and from CASP's own target files (casp).

The task: given a chain's amino-acid sequence (sequence), predict each residue's 3-state (ss3) or 8-state (ss8) label.

Quick start

from datasets import load_dataset

ssp = load_dataset("hazemessam/SSP")             # version `rcsb` (default)
train, valid, test = ssp["train"], ssp["validation"], ssp["casp14"]
print(train[0]["sequence"][:40], train[0]["ss3"][:40])

tape = load_dataset("hazemessam/SSP", "tape")    # TAPE's data only: train, validation, cb513, ts115, casp12
casp = load_dataset("hazemessam/SSP", "casp")    # CASP test sets built from CASP's own target files

Versions (subsets) and splits

All three versions have the same columns and the same train, validation, cb513 and ts115 (the files in tape/). They differ in the CASP splits:

	rcsb (default)	casp	tape
Splits	train, validation, cb513, ts115, casp12 … casp16	train, validation, cb513, ts115, casp12 … casp16	train, validation, cb513, ts115, casp12
CASP structures	the PDB entry CASP lists for the target, current version from RCSB	CASP's public target files	TAPE's CASP12
CASP row	one PDB chain (lists the CASP targets it contains)	one CASP target	one PDB chain
CASP sequence	the PDB entity's full canonical sequence	CASP's target sequence	TAPE's
CASP labels	mkdssp on the chain alone	mkdssp on CASP's target structure	TAPE's

split	version	chains	residues	with structure (resolved)	sequence length
train	TAPE (all versions)	8,678	2,221,507	93.5%	20 to 1,632 (median 219)
validation	TAPE (all versions)	2,170	556,931	93.7%	20 to 1,228 (median 220)
cb513	TAPE (all versions)	513	144,360	97.6%	20 to 874 (median 238)
ts115	TAPE (all versions)	115	29,704	92.6%	43 to 1,111 (median 233)
casp12	rcsb	41	12,082	90.4%	75 to 1,494 (median 228)
casp12	casp	40	12,094	87.0%	75 to 670 (median 261)
casp12	tape	21	7,256	88.5%	76 to 1,494 (median 246)
casp13	rcsb	41	12,007	92.0%	41 to 756 (median 252)
casp13	casp	20	5,629	95.4%	41 to 718 (median 262)
casp14	rcsb	17	5,986	89.5%	74 to 2,194 (median 186)
casp14	casp	34	10,688	85.4%	74 to 2,180 (median 196)
casp15	rcsb	58	24,332	83.2%	48 to 3,364 (median 296)
casp15	casp	30	11,695	86.3%	114 to 1,029 (median 375)
casp16	rcsb	24	12,679	83.7%	115 to 1,770 (median 331)
casp16	casp	61	27,550	81.2%	115 to 1,770 (median 363)

Labels

ss3	meaning	train residues with structure
H	helix (G, H, I)	799,086
E	strand (B, E)	475,456
C	coil / other (S, T, C)	802,292

ss8	meaning	train residues with structure
G	3-10 helix	80,325
H	alpha helix	707,583
I	pi helix	11,178
B	beta bridge	21,215
E	beta strand	454,241
S	bend	169,046
T	turn	227,423
C	coil / other	405,823

• ss8 has the eight DSSP states. ss3 is ss8 reduced with TAPE's rule (G, H, I → H; B, E → E; S, T, C → C); validate_ssp.py checks this on every residue of every split.
• ss8_pp (rcsb and casp only) is the output of mkdssp version 4.6.1, which has a ninth state, P (polyproline II). In ss8, a P residue is written T if DSSP's turn columns place it inside a 3-, 4- or 5-turn, S if DSSP's bend column marks it, and C otherwise. In this build 2,368 residues are P: 0 became T, 0 became S, 2,368 became C.

Which residues to score

• resolved: 1 = the residue has structure; 0 = it has none, and is labelled C, as in TAPE (so the labels have only 3 or 8 classes).
  • TAPE's splits: TAPE's disorder values. In all five splits every residue with 0 is labelled coil. On TAPE's CASP12, 6,417 of the 6,420 residues with 1 have a DSSP label in today's PDB entry, and 833 of the 836 residues with 0 have none.
  • rcsb and casp: 1 exactly where mkdssp gives the residue a label. Every residue with 1 has N, CA, C and O coordinates. Of the 115,681 residues that have all four, 1 has no label: a D-histidine (DHI, a HETATM record) in 7utd-A.
• valid_mask (TAPE's splits only): TAPE's valid_mask values. It is 0 only in cb513: 60,024 residues, 57,059 of them with structure. In 511 of the 513 cb513 rows, valid_mask = 1 on one continuous stretch of the chain; in 2 rows on no residue. The 513 rows have 434 distinct ids; rows sharing an id have the same sequence and labels, and differ only in valid_mask, whose stretches do not overlap.
• TAPE's own training and evaluation code (tape/datasets.py, class SecondaryStructureDataset, commit 6d345c2 (github.com/songlab-cal/tape)) reads only primary and ss3, so it scores every residue, including residues without structure (as coil). Scoring only resolved == 1 (and valid_mask == 1) gives different numbers; state which one you use.

def accuracy(pred: str, row: dict, labels: str = "ss3"):
    """Correct and scored residues of one row, over residues with structure (and valid_mask == 1 where present)."""
    mask = row["resolved"]
    if row["valid_mask"] is not None:
        mask = [a & b for a, b in zip(mask, row["valid_mask"])]
    hits = sum(p == t for p, t, m in zip(pred, row[labels], mask) if m)
    return hits, sum(mask)

Columns

column	type	filled in	content
id	string	all	rcsb/tape: PDB chain, e.g. 5uvn-A; casp: CASP target, e.g. T1024
sequence	string	all	input: amino-acid sequence, one letter per residue (for X, see Modified residues)
ss3, ss8	string	all	labels: one letter per residue (tables above)
ss8_pp	string	rcsb, casp	DSSP 4 labels including P
resolved	list of 0/1	all	1 = residue has structure (see above)
valid_mask	list of 0/1	TAPE splits	TAPE's valid_mask (see above)
rsa	list of float	TAPE splits	TAPE's rsa (relative solvent accessibility): 0 to 1 on residues with structure; 0 on residues without
asa_max	list of float	TAPE splits	TAPE's asa_max: 1 to 240.5; -1 at 390 residues, all X
phi, psi	list of float	TAPE splits	TAPE's backbone dihedral angles: -180 to 180, or 360 (17,186 of 17,209 phi = 360 are at the chain's first residue or after a residue without structure; 17,186 of 17,209 psi = 360 at the last residue or before one without structure); 0 on residues without structure
interface	list of 0/1	TAPE splits	TAPE's interface (1 on 319,807 residues, none of them without structure)
evolutionary	list of [30 floats]	TAPE splits	TAPE's evolutionary: 30 values per residue, 0 to 1
n_xyz, ca_xyz, c_xyz, o_xyz, cb_xyz	list of [x, y, z] or null	rcsb, casp	coordinates (Å) of the N, CA, C, O and CB atoms from the structure used; null where the atom is absent (e.g. CB of glycine) or the residue is missing
casp_targets	list of string	CASP splits	CASP targets this row covers
target_match	list of string	CASP splits	per target: byte-identical, contained (one stretch of sequence), pieces (several separate stretches in order) or differs
target_identity	list of float	CASP splits	per target: fraction of the CASP sequence's residues identical in sequence (local alignment, BLOSUM62)
target_regions	list of [[start, end], …]	CASP splits	per target: 1-based, inclusive positions in sequence (null for differs)
pdb_id, chain	string	all	PDB entry and author chain (casp: the PDB code(s) CASP lists, if any)
method, resolution, release_date, pdb_revision_date	string / float	rcsb	from the PDB entry used
source_files	list of string	all	the file(s) the row was built from
in_tape	bool	all	the row's chain (casp: the target's CASP-listed chain) is in TAPE's data

Columns a split can't fill are null.

How the CASP test sets were built

1. Targets. CASP's target list for each round, canceled targets removed. rcsb: the PDB code(s) CASP's target list gives for the target (no sequence search). casp: every target structure in CASP's public target archive.
2. Structure. rcsb: the current mmCIF entry from RCSB, and in it the protein chain whose sequence has the highest identity to the CASP target sequence; hydrogens, alternative conformations (all but the first), ligands and water removed. casp: CASP's target file (CASP15: the domain files of a target merged back into one chain by residue number; CASP16: the first version of targets released in several versions).
3. Labels. mkdssp version 4.6.1 on the chain alone (not the whole entry).
4. Duplicates. Within a round, rows with identical sequences are kept once (the others are listed in logs/).

Checks behind these choices (scripts in scripts/):

check	result	script
labels from this procedure vs NetSurfP-2.0's training labels (200 random train chains, seed 0, today's PDB entries)	99.95% (8-state), 99.96% (3-state) of 48,889 residues; 191 of 200 chains identical	check_netsurfp_train.py
mkdssp on the chain alone vs on the whole entry, against TAPE's CASP12 labels	99.94% vs 99.77% (8-state) of 6,417 residues	compare_ss_tools.py
TAPE's CASP12 chains vs the PDB entries CASP lists	21 of 21 are the CASP-listed entry and chain	build_ssp.py (in_tape), validate_ssp.py
rcsb CASP12 labels vs TAPE's	99.94% of 6,417 residues with structure in both	validate_ssp.py
CASP's archive files vs TAPE's CASP12 (16 of TAPE's 21 targets are in the archive)	labels: 91.92% (8-state), 94.69% (3-state) of 4,121 residues; CA RMSD to the PDB entry after superposition: 9 targets 0.00-0.06 Å, 2 targets 0.35-0.55 Å, 5 targets 1.5-5.6 Å	compare_casp12_archive.py, casp12_superpose.py

The rcsb version uses the PDB entries because they reproduce TAPE's CASP12 labels (99.94%) and CASP's archive files do not (91.92%); the table above shows that the archive files differ in coordinates from the PDB entries for 7 of the 16 targets.

round	version	rows	CASP target vs PDB chain (target, row) pairs	left out (reason: count)
CASP12	rcsb	45 of 45 CASP-listed targets, in 41 rows	byte-identical 26, contained 7, differs 12	none
CASP12	casp	40 targets	—	none
CASP13	rcsb	41 of 41 CASP-listed targets, in 41 rows	byte-identical 24, contained 11, differs 5, pieces 1	none
CASP13	casp	20 targets	—	none
CASP14	rcsb	26 of 26 CASP-listed targets, in 17 rows	byte-identical 10, contained 11, differs 3, pieces 2	none
CASP14	casp	34 targets	—	none
CASP15	rcsb	59 of 65 CASP-listed targets, in 58 rows	byte-identical 37, contained 7, differs 14, pieces 1	PDB entry 8pko not available from RCSB (2); same sequence as another row (4)
CASP15	casp	30 targets	—	none
CASP16	rcsb	43 of 44 CASP-listed targets, in 24 rows	byte-identical 26, contained 7, differs 10	same sequence as another row (1)
CASP16	casp	61 targets	—	no CASP sequence (1)

Modified residues

In every TAPE chain compared, TAPE writes every non-standard residue as X. rcsb sequences are the PDB canonical sequence, which writes the parent amino acid where there is one (selenomethionine MSE as M).

How this was found (scripts/check_modified_residues.py): each TAPE chain was compared position by position with the canonical sequence of its current PDB entry, and the PDB residue name was read at every position.

chains	chains compared	non-standard residues	TAPE's letter there	residues
TAPE's CASP12	20	44	X at all 44	MSE 44
TAPE's train (entries of the 200-chain check above)	238	308	X at all 308	MSE 296, FME 2, ACE 2, CSO 2, 6V1 2, CSD 1, CME 1, PCA 1, TOX 1

(One CASP12 chain, 5uvn-A, was not compared: its current PDB entry is 453 residues long, TAPE's sequence 455.)

Notes

• Sequence overlap between the CASP rows and train/validation was not checked. TAPE's CASP12 has 21 chains; in rcsb they cover 25 of the 45 CASP12 targets with a CASP-listed PDB code. in_tape marks them in every version.
• rcsb sequences are the full canonical sequence of the PDB entity, so they can be longer than the CASP target. Of the 214 (target, row) pairs in rcsb: byte-identical 123, contained 43, differs 44, pieces 4. Use target_regions to cut out a target. For differs, target_identity ranges from 0.32 to 1.00.
• CB513 ids are not unique (see valid_mask).

Verification

scripts/validate_ssp.py ran on 2026-11-14; all 28 checks passed (results in logs/validation.json):

• every file has the same columns and types;
• per row: every per-residue column has the sequence's length; label alphabets; ss3 derived from ss8, and ss8 from ss8_pp; residues without structure labelled C; residues with structure have N, CA, C and O coordinates (rcsb, casp);
• TAPE's splits equal TAPE's JSON files field by field, same rows in the same order;
• rcsb CASP12 contains TAPE's 21 chains; labels agree with TAPE on 99.94% of 6,417 residues;
• every CASP-listed target is in a row or listed in logs/ with the reason it was left out.

file	rows	MB	SHA-256
casp/casp12.parquet	40	0.8	68bf65529d901fbd6cae0c78b68a3fb3cf171a31eaeb51a42afecf3b8d77b0ba
casp/casp13.parquet	20	0.4	c210f35a62da293bd372747307ede2f659abc7251f9b1c6dac2148dd2fad31c8
casp/casp14.parquet	34	0.6	55849523ddfb4a86523ec6ce77d76332f0a3e346c00a0ff386a3450ce0fa01ed
casp/casp15.parquet	30	0.8	d8285e921bd0e2c3e3a5607a2ad248d222edc7e697031f106c20377a404a7198
casp/casp16.parquet	61	2.1	b841ef6f28d0498545a177a96468ff0877abe687967aeab85da5313df34e1322
rcsb/casp12.parquet	41	0.9	562665a4e04a35fdc45f4e5dbed9b30ac4932f10539e84045e57d99803e76171
rcsb/casp13.parquet	41	0.9	c45aa8f39a94944a7198aa24d77d9980bdfd220b6dfb48425b2a29d01dd666be
rcsb/casp14.parquet	17	0.4	65df8ca882579ad6e08278b72ad5ef519ca9165f8bf8b9016cbe680f8cefd271
rcsb/casp15.parquet	58	1.9	16de4dc6f7e92bcdd27321dd9af212c01f64a9beadb6607f9a5460b56d7b2c18
rcsb/casp16.parquet	24	0.8	db8fae1cfda1b029b8a666d253b0ce4469560e7ee12f945472e45a333ff608ca
tape/casp12.parquet	21	0.4	8be8d62661610f1a9eb4c896b95ca54698e776afa48655992e1e839fd35a49c3
tape/cb513.parquet	513	7.5	b144bf97c4f1a0c8afb872b4c4a780b0f6510945fe8f6d192b58ee1b44abfb2e
tape/train.parquet	8,678	111.2	45fd4b950c6f8a732824ee6481bc8ef9fc608f553a31fa17879464a60b0142b9
tape/ts115.parquet	115	1.5	172cf937cd31f77ab84c02466f315ba777eab56982ab9c7bdc565c75cde5a018
tape/validation.parquet	2,170	27.9	a230d85f8b18fb203bdd6b552202fb9238c32d41a157caf44003d17942ef9f35

scripts/ holds the code that built and checked this repository and the analyses cited above; they read the source downloads from data/original/ (paths in each script). logs/ records, per version and split, every target left out and why, how P residues were converted, and the DSSP version.

Sources, licenses and citation

• TAPE's splits: TAPE's download http://s3.amazonaws.com/songlabdata/proteindata/data_raw_pytorch/secondary_structure.tar.gz; TAPE's repository is BSD-3-Clause.
• rcsb: mmCIF files from the RCSB PDB (https://www.rcsb.org); PDB data are CC0 1.0.
• casp: target structures and sequences from the Protein Structure Prediction Center (https://predictioncenter.org).
• Labels: DSSP (mkdssp version 4.6.1, https://github.com/PDB-REDO/dssp).

Please cite TAPE and NetSurfP-2.0 for the training data and the classic test sets, and CASP, the PDB and DSSP for the CASP test sets:

@inproceedings{rao2019tape,
  title     = {Evaluating Protein Transfer Learning with TAPE},
  author    = {Rao, Roshan and Bhattacharya, Nicholas and Thomas, Neil and Duan, Yan and Chen, Xi and Canny, John and Abbeel, Pieter and Song, Yun S.},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2019}
}

@article{klausen2019netsurfp,
  title   = {NetSurfP-2.0: Improved prediction of protein structural features by integrated deep learning},
  author  = {Klausen, Michael Schantz and Jespersen, Martin Closter and Nielsen, Henrik and Jensen, Kamilla Kjaergaard and Jurtz, Vanessa Isabell and Soenderby, Casper Kaae and Sommer, Morten Otto Alexander and Winther, Ole and Nielsen, Morten and Petersen, Bent and others},
  journal = {Proteins: Structure, Function, and Bioinformatics},
  year    = {2019}
}

@article{kabsch1983dssp,
  title   = {Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features},
  author  = {Kabsch, Wolfgang and Sander, Christian},
  journal = {Biopolymers},
  volume  = {22},
  number  = {12},
  pages   = {2577--2637},
  year    = {1983}
}

@article{berman2000pdb,
  title   = {The Protein Data Bank},
  author  = {Berman, Helen M. and Westbrook, John and Feng, Zukang and Gilliland, Gary and Bhat, T. N. and Weissig, Helge and Shindyalov, Ilya N. and Bourne, Philip E.},
  journal = {Nucleic Acids Research},
  volume  = {28},
  number  = {1},
  pages   = {235--242},
  year    = {2000}
}

For CASP, cite the CASP assessment paper of the round you use.

Changes

Created on 2026-11-14.