Talea

Sequence-only discovery of elongated protein-solenoid architecture from protein-language-model attention.

Talea ranks proteins for solenoid-like structural repetition directly from amino-acid sequence. It requires no experimental structure, predicted structure, multiple-sequence alignment, or database search at inference time. Alongside its primary protein-level score, Talea reports candidate repeat intervals, periods, and registered units from the same model pass.

The name talea refers to the repeating rhythmic pattern in medieval isorhythm: a recurring temporal unit that may persist beneath a changing melodic surface. That is a pleasingly close analogy for divergent protein sequences preserving an underlying repeated geometry.

At a glance

Input	Protein FASTA
Primary output	Continuous score for elongated/open solenoid-like architecture
Localization output	Candidate interval, period, and registered repeat units
Structure required at inference	No
Alignment required	No
Maximum frozen length	2,046 aa
Accepted residues	The 20 canonical amino acids only
Current status	Research software; ranking model, not a curated annotation caller

Note

Talea deliberately has no hard positive/negative threshold. Its primary output is a ranking score for discovery and follow-up, not a definitive structural or biological annotation.

Why Talea?

Protein solenoids are built from repeated structural units, but the sequence identity between units can be weak or effectively absent. Sequence-repeat finders therefore miss many structurally repetitive proteins, while structure-based tools first require an experimental model or a successful structure prediction. The latter can be expensive at metagenomic scale and can be unreliable for unusual, long, oligomeric, or context-dependent proteins.

Protein-language-model attention offers another view. Repeated structural relationships can produce periodic, translated patterns in attention space even when conventional sequence alignment is uninformative. Talea reads both that local geometry and the protein's global language-model representation. It is intended as a fast triage layer: identify the sequences most worth folding, inspecting, or testing with structure-aware methods.

“Sequence-only” describes inference. Structural annotations were used during model development and evaluation.

How Talea works

flowchart LR
    S["Protein sequence"] --> E["Pinned ESM++ forward pass"]
    E --> A["Layers 13 and 35<br/>APC-corrected attention"]
    E --> H["Layers 13 and 35<br/>pooled hidden states"]

    A --> R["Residue-state<br/>attention U-Net"]
    H --> G["Global PCA +<br/>logistic head"]

    R --> ZR["teacher_state_z"]
    G --> ZG["global_z"]
    ZR --> T["Talea discovery score"]
    ZG --> T

    A --> L["Auxiliary topology decoder"]
    E --> P["Layer 35 residue states<br/>helix/strand/loop probes"]
    L --> U["Phase-SSM<br/>unit placement"]
    P --> U
    U --> O["Interval · period · units"]

Loading

One ESM++ pass supplies both branches:

1. Residue-state branch. Attention from zero-based transformer layers 13 and 35 is symmetrized, average-product corrected, rectified, and averaged across heads. A compact two-dimensional U-Net reads the two attention channels plus a valid pair mask in 128 residue tiles. Predictions are reconstructed from each tile's central 64-residue diagonal. The mean elongated-state probability is logit-transformed and standardized against frozen development controls.
2. Global branch. Residue-wise means and population standard deviations of hidden states from the same two layers are passed through a frozen standardization, whitened 64-component PCA, and sample-weighted logistic head. Its probability is likewise logit-transformed and standardized against frozen controls.
3. Auxiliary topology branch. Multi-scale attention spectra propose repeat periods and intervals. Talea then registers translated units and evaluates terminal closure and plausible boundaries. These outputs support localization and interpretation but do not enter the primary v2 rank.
4. Unit placement. Within the registered extent, Talea 0.2.0 places repeat units with a phase-aware state-space decoder. Three linear probes on the layer-35 hidden states give per-residue helix, strand, and loop probabilities. Demodulating each track at the registered period gives every residue a vote for its position within a unit. A hidden semi-Markov model then decodes units of locally varying length (0.6–1.6 × period) whose phase agrees with the votes and whose length is supported by attention stacking. Scores, the wedge interval, the period, closure, and the boundary scan are computed before this step and are exactly those of v0.1.0.

The frozen primary score is the unweighted mean of the two independently standardized protein views:

Talea_discovery = 0.5 × teacher_state_z + 0.5 × global_z

No coefficient in this fusion was fit to the evaluation cohort. The 50:50 formulation was nevertheless selected after that cohort had been opened; see What “post-opening” means here.

Why layers 13 and 35?

Layer indices are zero-based transformer-block indices. All 36 ESM++ layers were compared under the same attention-harmonic formulation on a canonical development cohort of 127 positives and 126 length-matched controls:

• layer 13 had the highest development AUROC: 0.663, with AP 0.658;
• layer 35 had the highest development AP: 0.688, with AUROC 0.651.

The pair therefore preserves the best layer under each pre-existing metric without a combinatorial layer-pair search. The choice was made before the 452-protein comparator cohort existed and remained fixed thereafter. Blindly averaging the final eight layers performed substantially worse on the homology-disjoint confirmation cohort.

Installation

Requirements

• Python 3.10 or newer
• PyTorch
• Transformers 4.57 or newer
• CUDA or Apple MPS strongly recommended

Clone and install in an isolated environment:

git clone https://github.com/jwestrob/Talea.git
cd Talea

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -e .

The small Talea heads and calibrations are included in the package. The pinned Synthyra/ESMplusplus_large checkpoint is not vendored. Its primary weight file is approximately 2.3 GB and is downloaded into the normal Hugging Face cache on first use.

Talea pins the ESM++ repository revision in its deployment contract. Loading that model currently requires --trust-remote-code. After the exact revision has been cached, --local-files-only prevents further network access.

Command-line use

Apple Silicon

talea \
  --input proteins.faa \
  --output predictions.jsonl \
  --device mps \
  --trust-remote-code

CUDA

talea \
  --input proteins.faa \
  --output predictions.jsonl \
  --device cuda \
  --trust-remote-code

Automatic device selection

talea \
  -i proteins.faa \
  -o predictions.jsonl \
  --device auto \
  --trust-remote-code

auto selects CUDA, then MPS, then CPU. CPU inference is supported but is generally impractical for large screens.

Talea writes:

• one JSON object per protein to the requested JSONL file; and
• a sibling <output>.audit.json containing exact input, output, model, environment, and artifact checksums.

CLI options

Option	Meaning	Default
-i, --input	Protein FASTA	required
-o, --output	Prediction JSONL	required
--device	auto, cpu, mps, cuda, or an explicit Torch device	auto
--dtype	auto, fp16, fp32, or bf16	auto
--maximum-batch	Maximum proteins per padded ESM++ batch	64
--square-budget	Upper bound on batch size × padded sequence length²	4,000,000
--postprocess-workers	CPU workers for topology decoding	all available CPUs (process mode); 1 (thread mode)
--postprocess-mode	process to scale across CPU cores, or thread to stay in one process	process
--capture-mode	manual (numerically identical to the frozen run) or sdpa (faster, small numerical drift)	manual
--cache-clear-interval	Release accelerator caches every N batches; 0 never does	0
--deterministic	Pin cuDNN to deterministic convolution algorithms	off
--unit-placement	phase-ssm, or comb for the v0.1.0 equal-period units	phase-ssm
--long-proteins	reject proteins above 2,046 aa, or score them as overlapping windows	reject
--window-length	Residues per window in window mode	2,000
--window-stride	Residues between window starts in window mode	1,000
--trust-remote-code	Permit the pinned ESM++ model implementation to load	off
--local-files-only	Use only the local Hugging Face cache	off
--overwrite	Replace an existing output and audit	off
--quiet	Suppress progress messages	off
--version	Print Talea's version	—

Run talea --help for the live interface.

Python API

from talea.artifacts import load_deployment
from talea.inference import TaleaPredictor
from talea.io import read_fasta


def main():
    deployment = load_deployment()
    records = read_fasta(
        "proteins.faa",
        maximum_length=deployment["input_contract"]["maximum_residues"],
    )
    predictor = TaleaPredictor(
        device="cuda",
        dtype="fp16",
        trust_remote_code=True,
    )
    for prediction in predictor.iter_predictions(records, progress=True):
        print(
            prediction["record_id"],
            prediction["talea_discovery_score"],
            prediction["wedge_call"],
        )


if __name__ == "__main__":
    main()

Postprocessing runs in worker processes, and each worker re-imports the calling script (Python's rule for spawned processes), so keep the entry point under if __name__ == "__main__": as above. postprocess_mode="thread" runs in one process. Every new worker re-imports the calling script, including torch if the script imports Talea at top level, which takes seconds on network storage (13 s per worker on biotite's login node). The predictor therefore starts its worker pool on the first call and keeps it, so scoring proteins one call at a time pays this once; predictor.close() (or with TaleaPredictor(...) as predictor:) stops it.

TaleaPredictor(unit_placement="comb") reproduces v0.1.0 unit boundaries. Load one TaleaPredictor and reuse it. Starting multiple processes on the same GPU duplicates ESM++ and its attention buffers.

Input contract

Talea validates the complete FASTA before loading ESM++.

• Only ACDEFGHIKLMNPQRSTVWY is accepted.
• A single terminal * is removed as a FASTA serialization marker.
• Internal stops, repeated terminal stops, ambiguity codes, and all other noncanonical characters are errors.
• Residues are never masked, imputed, or replaced.
• FASTA record identifiers—the first whitespace-delimited header token—must be unique.
• Empty sequences are errors.
• The frozen v2 maximum is 2,046 residues.
• Sequences longer than the evaluation maximum of 1,732 residues are accepted through 2,046 aa but explicitly marked as length extrapolations.

If any record violates the contract, the run stops before model inference.

Predictions are emitted in length-sorted order to reduce padded-attention memory. Join outputs back to inputs by record_id, not line number.

Output interpretation

All residue coordinates are zero-based and half-open: start is included and end is excluded.

Field	Interpretation
record_id	FASTA identifier
sequence_sha256	SHA256 of the exact normalized amino-acid sequence
talea_discovery_score	Primary stable v2 ranking score
primary_score_field	Explicit declaration of the field to rank
teacher_state_probability	Mean predicted elongated-state probability
teacher_state_z	Control-standardized logit of the state probability
global_probability	Frozen global-head probability
global_z	Control-standardized logit of the global probability
topology_development_control_z	Auxiliary attention-topology score
status	ok when a wedge was localized; otherwise no_wedge_call
candidate_periods	Period hypotheses in residues, including plausible aliases
wedge_call	Highest-scoring candidate interval, period, and wedge score
secondary_wedge_calls	Further wedge calls after the primary one; those the residue-state track supports are registered into units as the primary call is (see Repeat regions)
repeat_regions	Every repeat region: the registered primary region plus supported further calls, each with period, units, wedge score and mean residue-state probability
registration.units	Repeat-unit boundaries in the registered extent
registration.unit_placement	phase_ssm, or comb_fallback when fewer than three units decode
registration.comb_units	The v0.1.0 equal-period units, retained for comparison
closure	Terminal-to-adjacent attention summary for the equal-period units
boundary_scan	Alternative plausible boundary-pair diagnostics
frozen_spectrum	Layer-specific harmonic attention summaries
length_scope	Evaluation-length boundary and extrapolation flag
evidence_status	Explicitly marks the result as an unverified model output

Two additional fields deserve care:

• talea_teacher_score is retained for historical schema compatibility. It is not the primary v2 ranking field.
• predicted_teacher_mean_class_probability contains auxiliary soft model outputs inherited from the structural teacher. These are diagnostic model probabilities, not curated structural class assignments.

To make a simple ranked TSV:

jq -r \
  '[.record_id, .length, .talea_discovery_score, .teacher_state_z, .global_z,
    (.wedge_call.start // ""), (.wedge_call.end // ""),
    (.wedge_call.period // "")] | @tsv' \
  predictions.jsonl |
sort -t $'\t' -k3,3gr > ranked_talea.tsv

Scaling and hardware

Transformer attention is quadratic in sequence length. Talea therefore length-sorts inputs and constrains batches by batch_size × padded_length², rather than by record count alone.

• Lower --square-budget if a GPU runs out of memory.
• Lower --maximum-batch when screening many short proteins.
• Postprocessing (CPU-only topology work) runs in a process pool on every CPU the job may use, including a Slurm allocation's CPU set. Workers do not load additional ESM++ copies. --postprocess-workers caps the pool; --postprocess-mode thread keeps the work in one process.
• Batches are pipelined: each batch's postprocessing runs in the pool while the next batch is on the device, and rows keep their order and values.
• ESM++ runs with selective capture: Talea retains only the two attention layers it reads (288 MB rather than 5.06 GB at 2,046 aa). Under the default --capture-mode manual, output is byte-identical to the full-capture path.
• Attention matrices, residue embeddings, and residue-state tracks are reduced in memory and are not retained on disk.
• For reproducibility and memory efficiency, run one Talea process per GPU.

The 2,046-aa ceiling is part of the frozen deployment contract, not a fundamental limit of the method. Longer proteins are scored only on request, as overlapping windows (next section); Talea never silently tiles or truncates.

Long proteins

--long-proteins window (Python: TaleaPredictor.iter_windowed_predictions) scores each protein longer than 2,046 aa as windows of --window-length residues (2,000) starting every --window-stride residues (1,000), plus one window ending at the C-terminus, so every window has the same length and residues more than 1,000 positions from either end are read twice. The 2,000/1,000 grid matches the chunks used to fold the Omnitrophota giant proteins, so window k is chunk k. Each window is scored exactly as a standalone protein. Proteins within the contract are scored whole and their rows are unchanged.

A windowed protein yields one row: the best window's row (highest discovery score) with every residue position moved into the full-protein frame, plus windowing, which lists every window's start, end, scores, period, wedge interval and registered extent in full-protein coordinates.

Validation (2026-10-27; development data only):

• Exact windows. On the 19 standard-alphabet parents of the 35 curated giant-protein tube chunks, the 34 windows at chunk positions are the curated chunk sequences, and their scores match the frozen chunk scores within 0.0008.
• Tube localization. Within every parent, the curated tube windows outrank the parent's other windows (within-parent AUROC 0.78 on average; median tube window at the 19th percentile). The other windows are unlabeled, so this is a lower bound.
• Whole versus windowed. 300 DPANN proteins of 1,600–2,046 aa (the 150 top-scored and 150 random), scored whole and as forced 1,000/500 windows, which give less context than deployment: the best-window score has rank correlation 0.93 with the whole-protein score, 81 of the top 100 are shared, and for 85 of the top 100 the period agrees within 10%. The best window's locus overlaps the whole-protein locus for 67 of the top 100. For 28 more, another window calls the whole-protein locus: these proteins carry repeats of the same period in several regions, and the highest-scoring window holds a different one. Read the row's locus as the strongest window's repeat region; repeat_regions gathers the regions of every window (see Repeat regions). Rescored with repeat_regions (scores and primary calls identical for all 300), the whole-protein locus lies in one of the windowed protein's regions for all 100 of the top 100, and 99% of the regions found when the protein is scored whole are among the windowed regions (2.24 regions per protein whole, 2.68 windowed).

Read window scores as a ranking. Like whole proteins above 1,000 aa, 2,000-aa windows score high in absolute terms: 99% of the 809 windows of the tube parents exceed the CATH-control cutoff that classifies 95% of CATH controls as negative. Compare windows with windows, and rank within a protein to localize. Windows put several 2,000-aa sequences in one batch only when --square-budget allows it (about 4,008,000 per window); raise it on large GPUs so their postprocessing runs in parallel.

Repeat regions

repeat_regions lists every repeat region Talea finds in a protein, sorted by position. The wedge returns the single best period and interval, so a protein with two repeat regions, such as a solenoid split by an insertion or two repeat domains, was previously reported by one of them. Talea now continues the search: it sets the diagonal values of every residue pair touching the primary call's interval to that diagonal's median and searches the wedge again, stopping at the first call whose wedge score falls below 12 or below half the primary call's. A further call becomes a region when the U-Net residue-state track averages at least 0.2 over it. Each region extends half a period past its wedge interval, as the registration seed does; a region of the same period (within 15%) that overlaps or nearly touches another merges into it.

Regions locate Talea's repeat evidence; read them for proteins that rank high on the discovery score, since every protein with a wedge call has a primary region. The first region (source: primary) is the registered extent, and registration and every score describe it alone; further regions (source: secondary) carry a period, extent, wedge score and mean residue-state probability. Every region carries units: the primary region the registration's units, a further region the units of its own wedge call, registered and placed exactly as the primary call's are (only further calls that pass the region gate are registered). A merge adds the incoming units whose midpoints fall outside the region's units, a trimmed region keeps the units whose midpoints fall inside it, and units are clipped where they meet and to their region. In window mode, repeat_regions is the union of every window's supported regions in full-protein coordinates, with the window each came from: the best window's regions, plus the other windows' further regions and those of their primary regions that pass the same gate as a further call (mean residue-state probability at least 0.2, wedge score at least 12), since every window has a primary region whether or not it holds a repeat. Each window's own regions are listed under windowing.windows.

Validation (2026-10-28; Talea's development data plus a new cohort):

• Proteins with several curated regions. RepeatsDB class 3 chains with two to four curated regions were excluded from the repeat-extraction benchmark by its one-region rule; 455 of them (≤2,046 aa, canonical, exactly mapped, sequences outside the benchmark; 98% AlphaFold DB, mostly α solenoids such as MIF4G, cullin, armadillo and TPR domains) form this cohort. Against the primary region alone, curated regions covered at least half by a reported region rise from 60.0% to 76.7% (+0.167 [+0.129, +0.210]), and regions covered by a region of their own from 45.2% to 59.0% (+0.138 [+0.104, +0.177]). Residue precision holds (0.790 vs 0.783; −0.006 [−0.017, +0.005]) and residue F1 rises from 0.653 to 0.748. Intervals are paired bootstrap 95% CIs over 176 Pfam-architecture groups. The acceptance rule was chosen on this cohort; choosing it on one half of the Pfam groups and scoring the other half selects the same rule in both directions, with held-out coverage 0.629 → 0.789 and 0.573 → 0.738.

• One curated region per protein. On the 2,286 development records, residue F1 is unchanged (0.811 vs 0.814; +0.003 [−0.003, +0.010] over 350 homology components), recall rises (0.970 → 0.992; the further search finds the curated region where the primary call missed it), and the share of proteins with a region outside the curated one rises from 3.4% to 4.7% (+0.014 [+0.007, +0.022]). Proteins carry 1.05 regions on average here and 1.47 in the multi-region cohort.

• Units in every region. Scored with the benchmark's array scorer (unit boundaries within 5 residues; curated units of every curated region), units in the further regions raise boundary F1 on the multi-region cohort from 0.289 to 0.318 (+0.028 [+0.017, +0.041], 176 Pfam groups) against further regions that carry only their ends, with matched-unit IoU 0.653 → 0.663 and unit-count error 6.25 → 5.43 (−0.82 [−1.63, −0.12]). Curated regions found by a further region get boundary F1 0.316 and matched-unit IoU 0.631 there (primary-region matches 0.395 and 0.674), at a median unit count equal to the curated one. On the 2,286 single-region development records boundary F1 holds (0.537 vs 0.537; −0.0005 [−0.0045, +0.0029], 350 components) and unit-count error rises by 0.20 units (+0.20 [+0.11, +0.31]), the units of the extra regions that 4.7% of those proteins carry.

The further search adds one wedge search per protein when nothing passes and one more per region found (2.4 per protein in the multi-region cohort). The wedge search computes each diagonal's interval sums once per search and scores every period of a width together, with bit-identical results, so postprocessing with the further search takes less CPU time than v0.2.0 postprocessing without it: 15.6 s against 18.3 s, single-threaded, for 240 Omnitrophota proteins (96-2,046 aa, median 274) whose attention was captured once. Every other field of those 240 rows equals v0.2.0's. Long proteins run more further searches (6.4 further calls per protein at 1,800-2,046 aa, the length of every window in window mode): 24 such proteins take 12.5 s against 9.9 s for v0.2.0 (22.8 s on the previous search). Registering the units of supported further calls (567 of 1,081 calls in the multi-region cohort) adds no measurable time on the 240 natural proteins and 8% at window length.

End to end on an H200 with 24 CPUs (process workers, pipelined batches; two rounds in mirrored order), Talea 0.3.0 scores 2,000 natural Omnitrophota proteins at 15,795 residues/s against 14,746 for v0.2.0, and 120 proteins of 1,800-2,046 aa at 5,941 against 6,072 (v0.2.0's own runs on that set ranged from 5,506 to 6,109 residues/s across two benchmark jobs on this shared node).

Benchmark

The principal comparator cohort contains 452 canonical proteins: 226 elongated/open-repeat positives and 226 reviewed alternative-repeat controls. It contains one representative per sensitive homology group, and resampling is performed over those groups rather than treating related sequences as independent. Prospectively, comparator groups represented in the RepeatsDB-derived residue-state training branch were removed, along with exact sequence and PDB identities to either residue-state source. Remote homology against the second residue-state source was not jointly recomputed before the cohort was opened.

SOLeNNoID received chain-specific experimental coordinates. Talea received amino-acid sequence only.

Evaluation	Talea AP	Talea AUROC	Talea normalized pAUROC at FPR 0.1
Original sealed Talea score	0.991	0.988	0.953
Stable discovery v2	0.996	0.995	0.976

The SOLeNNoID mean-total comparator reached AP 0.984, AUROC 0.983, and normalized pAUROC 0.896. Mean-total probability was fixed before lockbox inference as the stronger comparator summary observed in prior development; the published majority-class fraction is retained in the benchmark bundle as a secondary summary. SOLeNNoID was run from official repository commit 87c13fefdd6acd7cb43831979d89fd11e437716f.

These labels come from the same RepeatsDB elongated-open-repeat ontology used to define Talea's target during development. The cohort is therefore a favorable or “home-field” task for Talea relative to a comparator developed under a different solenoid definition. It supports competitiveness on this task, not universal superiority across repeat ontologies.

For the original sealed Talea score, Talea minus SOLeNNoID was +0.007 AP with a 95% homology-group paired interval of [-0.003, +0.019], and +0.005 AUROC with an interval of [-0.007, +0.017]. Talea had the higher point estimates, but the predeclared conclusive-win rule required the lower AP bound to exceed zero and therefore failed.

For stable discovery v2, 20,000 label-stratified homology-group paired bootstrap replicates gave Talea-minus-SOLeNNoID intervals entirely above zero:

• AP: [+0.00317, +0.02426]
• AUROC: [+0.00182, +0.02283]
• normalized pAUROC: [+0.02561, +0.15189]

A post-opening sensitivity analysis reweighted the frozen state and global outputs without refitting. RepeatsDB AP remained between 0.994 and 0.996 for state weights from 0.30 to 0.70; on CATH S20 the same interval produced AP 0.799–0.810. The equal fusion lies on a broad plateau rather than at a narrow RepeatsDB-specific optimum.

A score-blind post-opening MMseqs2 audit compared the complete comparator with the full Talea training sequence union: 457 residue-state records and the exact 1,200 global-head records. Qualifying links required at least 20% identity, E-value at most 1e-3, and at least 50% coverage of the shorter sequence. The union linked 141 comparator proteins. After excluding them, 151 positives and 160 controls remained; sealed Talea, released v2, and SOLeNNoID reached AP 0.989, 0.994, and 0.990. The paired Talea-minus-SOLeNNoID intervals were [-0.010, +0.007] for the sealed score and [-0.002, +0.010] for released v2. Both cross zero. This sensitivity repairs the scope of the original firewall; it does not turn the known cohort into new validation.

A separate score-blind repair recomputed homology between the two residue-state training sources. Sixty-one cross-source links collapsed the 457 records into 315 intact validation components. Corrected out-of-fold AP was 0.960 without the structural teacher and 0.959 with it; the paired difference interval was [-0.014, +0.013]. Thus the residue-state signal survives the repaired firewall, but the teacher is not claimed to improve out-of-fold ranking. The released full model was not refit or altered by this audit.

A score-blind structural-exposure audit compared every comparator structure with the complete union of the 457 residue-state/teacher structures, 1,200 global-head training structures, and all 348 published SOLeNNoID development structures before classifier scores were joined. Five positive comparator chains exactly match PDB-and-chain identifiers in the SOLeNNoID archive:

Maximum Foldseek TM score	Positive/control	Sealed Talea AP	v2 AP	SOLeNNoID AP
< 0.50	22/19	0.975	0.987	0.979
< 0.60	33/34	0.948	0.980	0.951
< 0.70	67/63	0.973	0.985	0.961

Paired intervals crossed zero at every structural threshold. The rankings are therefore robust after removing structurally close evaluation proteins, but the reduced cohorts do not establish a structurally disjoint Talea advantage.

A separate post-opening subtype diagnostic found strong released-score ranking for all four RepeatsDB elongated-open topologies. The smallest subgroup, beta-hairpin repeats, contained only six proteins and remains imprecisely estimated. Five-fold out-of-fold baselines reached AP 0.565 from length, 0.622 from SEG low-complexity fraction, 0.797 from amino-acid composition, and 0.798 from composition plus length, compared with 0.996 for released v2 on the same cohort.

What “post-opening” means here

Important

The 452-protein cohort was genuinely sealed and score-blind for the original Talea evaluation. Stable discovery v2 was formulated only after that cohort had been opened. The underlying state and global models, their calibrations, and the selected ESM++ layers remained frozen, and the v2 fusion is a transparent 50:50 mean rather than a coefficient fit to these labels. Nevertheless, the formulation decision occurred with the cohort's results available. The v2 comparison is therefore post-opening secondary evidence, not an independent held-out estimate.

In practical terms:

• The original sealed result is the confirmatory benchmark result.
• The stable v2 result shows that the released formulation performs very strongly on this known cohort and supports its use for discovery.
• The v2 result should not be described as an independently confirmed win.
• Confirming v2 requires a genuinely new cohort whose labels and comparator scores remain unseen until the v2 score is frozen.

This distinction is statistical, not semantic: unchanged components and non-fitted equal weights reduce the opportunity for overfitting, but they do not make an opened evaluation set new again.

CATH S20 transfer diagnostic

The exact released score was also reconstructed without refitting on 819 CATH S20 domain sequences: 91 positive domains and 728 length-matched controls. The positive set is narrow, with 80 of 91 records classified as alpha horseshoes.

Cohort	Talea AP	SOLeNNoID AP	Paired Talea minus SOLeNNoID 95% interval
All CATH S20 domains	0.810	0.754	[-0.011, 0.123]
Exact all-training-unions TM < 0.60	0.440	0.419	[-0.139, 0.180]

The exact structural filter covers the 457-structure Talea state/teacher union, the 348-structure published SOLeNNoID development archive, and the 1,200-structure released Talea global-head collection.

CATH labels and earlier comparator results had already been opened during method development. This is a post-opening, no-refit transfer diagnostic, not independent validation. The point estimates are competitive, the paired intervals do not distinguish the methods, and the lower absolute performance under structural and ontology shift is an important scope limit.

Repeat-unit placement

Unit placement was evaluated on a RepeatsDB class 3 repeat-extraction benchmark built by Kinsey Long: solenoid entries from PDB and AlphaFold DB with curated unit boundaries, split by homology component into development (2,286 records) and test (571 records, 466 unique sequences) partitions. Every method choice used development records. The test partition was scored once, under a plan whose SHA256 was recorded before any test label was read. A predicted boundary counts as correct within ±5 residues of a curated one; PDB and AlphaFold DB records carry equal weight, and intervals are paired bootstrap 95% CIs over homology components.

Test partition (571 proteins)	Boundary F1
Equal-period comb (v0.1.0)	0.375
Phase votes on an equal-period comb	0.488
Phase-SSM (v0.2.0)	0.549

Phase-SSM minus comb: +0.173 [+0.124, +0.223]; the state-space decode contributes +0.061 [+0.036, +0.087] over the same votes on a comb. Region F1 is 0.90 for both, since both use the same extent. Per protein, by RepeatsDB subclass (comb → phase-SSM): α-solenoids (3.3, n = 394) 0.42 → 0.59; α/β-solenoids (3.2, n = 88) 0.37 → 0.64; β-solenoids (3.1, n = 71) 0.44 → 0.49; subclass 3.4 (n = 18) 0.49 → 0.43.

Ian Holmes suggested the state-space formulation. Kinsey Long built the repeat-extraction benchmark and the helix/strand/loop probes.

Limitations

• Talea ranks candidates; it does not make curated structural or biological calls.
• No universal score threshold has been calibrated for arbitrary sequence collections. Score distributions can shift with proteome composition.
• Average precision on the balanced comparator does not predict precision at the lower and unknown prevalence of elongated open repeats in natural proteomes.
• The current comparator contains 226 positives and 226 alternative-repeat controls, one per sensitive homology group. A new independent cohort is still needed to confirm stable v2 because this cohort had already been opened when v2 was formulated.
• Candidate periods can have harmonic or subharmonic aliases. Treat the reported period and unit registration as hypotheses for inspection. Unit placement inherits the period and extent, so an aliased period yields units at the aliased length.
• Unit placement gains are largest for α and α/β solenoids, where helix and loop probabilities oscillate strongly within each unit.
• Units of further regions match curated unit counts (median ratio 1.0) and place unit boundaries less precisely than the primary region's (boundary F1 0.32 vs 0.39 on matched curated regions of the multi-region cohort).
• The 1,733–2,046-aa range is a declared length extrapolation. Proteins above 2,046 aa are scored only as 2,000-aa windows (--long-proteins window), which see at most 2,000 residues of context.
• Runtime and memory scale quadratically with sequence length.
• “Sequence-only” applies to inference; the residue-state head was developed with structural supervision.
• A high score is evidence for the architecture Talea learned, not proof of monomeric state, oligomeric context, membrane association, or biological function.

Reproducibility

The released tool is intentionally inference-only. It contains:

• runtime source under talea/;
• a frozen 132-KB attention head;
• a frozen 2.4-MB global head;
• an 87-KB unit-placement asset (probes, phase coefficients, decoder settings);
• calibration and deployment contracts;
• checksum verification for every bundled artifact;
• focused unit and integration tests; and
• a toy FASTA smoke input.

It does not contain training pipelines, source datasets, attention-matrix caches, comparator installations, or generated benchmark result trees.

Key frozen provenance:

Artifact	SHA256
Deployment contract	7aa5b7f320323b34d2db27b0df488732d1fcec39d72e2d0f5b774ebec1edea2e
Residue-state teacher head	130840931f4f8be768f2789cbe8a5db8bc2611fd5605fc797f1023156d4c625e
Global head	8701dad34e3392a0f84884e88d0bf3e9741a75903882ec7658bbb8b2fb2161de
Unit-placement asset	fecba553557eef59dae2b0b1b12982c21563c9d26c73c29c079638f685724595
Pinned ESM++ revision	f813401638b3fddab09748aec1ad2bf537aa4208
External ESM++ weight file	4aff3f8c5de68c4d3e3824eb2c478e4a47355d3f849f3c745e5c8a5ee6cff851

The complete bundled manifest is talea/assets/SHA256SUMS, and the machine-readable deployment contract is talea/assets/deployment.json.

The standalone refactor was checked against the original production runner on a fresh ESM++ pass. The primary score, both component scores, auxiliary topology score, candidate periods, wedge interval, and registered units were identical.

Run the test suite:

python -m pip install -e '.[test]'
pytest

Name and historical identifiers

Talea was originally developed under the name Pangu. Historical machine-readable keys such as frozen_pangu_development_control_z remain in some auxiliary output and contracts where renaming them would break artifact provenance. New public interfaces use the Talea name.

Citation

A Talea methods manuscript is in preparation. Until a DOI is available, please cite the repository URL together with the version or Git commit used, and retain the generated audit file for exact model provenance.

Contributing

Issues and focused pull requests are welcome. Changes to model arithmetic, layers, calibrations, or bundled artifacts should be treated as a new model version and evaluated on a cohort not used to choose the change.