📐 TAPAS

Overview

TAPAS (Torsion Angles from PACSY for Structure calculation; command Ts) predicts, from the protein sequence and your assigned chemical shifts alone:

• backbone torsions φ, ψ, ω and side-chain χ1, χ2;
• 3-state secondary structure (H helix, E strand, C coil) with probabilities and a calibrated reliability; a helix is always at least 3 residues and a strand at least 2, so a single residue is never reported as helix or strand;
• RCI-S2 order parameters (backbone dynamics);
• SASA (solvent accessible surface area) and hydrophobic exposure;
• a calibrated reliability and expected error for every predicted angle.

No structure, NOE data or local software install is needed: POKY sends the shifts to the POKY TAPAS web server and downloads the results. TAPAS first checks your chemical-shift referencing with LACS (the same algorithm used for the PACSY database) and corrects the CA/CB/C/HA shifts by the offsets it finds; the offsets are listed in the results. When LACS cannot estimate an offset for a nucleus (too few usable residues), or returns one outside the plausible range of referencing errors, the shifts of that nucleus are left uncorrected and the results page flags it, so check the referencing of such data.

Open it from Structure prediction → Torsion/SS/RCI/SASA prediction with TAPAS (POKY), or type Ts.

Running TAPAS (Ts)

1. Tick the conditions whose resonances should be used, and choose whether to include unused resonances (resonances not linked to any peak).
2. Optionally enter your name and e-mail; the server e-mails you when the job finishes.
3. Press Submit. POKY sends the sequence (user.seq) and the assigned shifts (user.list) and fills in the Job ID.
4. The job usually takes one to several minutes. The dialog waits in the background, so you can keep working in POKY. Stop ends the wait without cancelling the job; later, type or keep the job ID and press Check job to fetch the results.

What you get back

When the job finishes, POKY unpacks the results into <project folder>/TAPAS/<job ID>/, shows a per-residue summary (φ, ψ, their reliabilities, SS, SASA, S2) in the dialog and opens the interactive dashboard in your browser.

Reliability and abstention

Every angle comes with a reliability between 0 and 1 and an expected error in degrees. Both were calibrated on held-out proteins against the observed prediction error, so a higher reliability really means a smaller typical error. Residues whose φ or ψ reliability is below the threshold are marked abstain: TAPAS still reports its best guess, but it does not write a restraint for them. Flexible residues, termini and many glycines (whose φ is often genuinely two-state) are typical abstentions.

Secondary-structure boundaries and missing shifts. The ends of helices and strands are where secondary-structure calls are least certain. TAPAS predicts, for every residue, how likely it is to sit at such a boundary (ss_edge_prob in report.json) and how complete its backbone shifts are (shift_completeness). Residues at a predicted boundary or with few assigned shifts nearby (boundary_or_sparse) get a more cautious φ/ψ calibration, so their restraints are wider or withheld. The secondary-structure CONFIDENCE in tapas_predSS.tab is the calibrated probability that the call is correct; it is lower for residues with few shifts nearby and for close calls between strand and coil. An isolated β-bridge residue in a turn (for example ubiquitin Thr22 and Thr55) is reported as coil.

Dihedral restraints

tapas_restraints.aco (CYANA) and tapas_restraints.tbl (X-PLOR/CNS) contain φ/ψ restraints for the reliable residues only. Each restraint is centred on the prediction and its half-width is a multiple of the calibrated expected error; ranges that cross ±180° are written correctly. They can be used wherever TALOS-N's talos.aco/talos.tbl are used, for example in a PONDEROSA-C/S calculation.

TAPAS with TALOS-N and TALOS+ jobs

Every TALOS-N and TALOS+ job submitted from POKY (tl) also starts a TAPAS job on the same chemical shifts. The TALOS results arrive as before; a REMARK line near the top of each TALOS pred*.tab file gives the address of the TAPAS results page with a one-line explanation, so you can compare the two predictions.

TAPAS with I-PINE jobs

Every Enhanced I-PINE job (ep) also runs TAPAS on the backbone chemical shifts I-PINE has just assigned. This happens after the I-PINE results e-mail, and takes a few minutes. When it finishes, the I-PINE dashboard gains a TAPAS tab in the secondary-structure card and a TAPAS section that embeds the TAPAS results page. The TAPAS files are added to the I-PINE results under tapas_results/ (TALOS-format input, tapas_pred.tab/tapas_predSS.tab, CYANA/XPLOR restraints and report.json). You also receive a short TAPAS e-mail with the link to the full TAPAS page.

The I-PINE assignment engine itself also uses knowledge learned for TAPAS. Since October 2026 it scores candidate spin systems against a backbone chemical-shift distribution per residue type and secondary structure. That distribution was rebuilt from the same LACS-referenced BMRB/PDB data TAPAS was trained on, and replaces the older static table. On the 104 gradable proteins of the 106-protein ARTINA set it assigns more backbone residues correctly with no extra wrong ones (9887 vs 9872 correct, 26 wrong), and raises the ARTINA-graded side-chain fraction from 93.28 % to 93.58 %.

Accuracy and speed

On a held-out test set of 402 proteins (none used for training or model selection), over the residues TALOS-N predicts, TAPAS's mean absolute error was lower than TALOS-N 4.21's for φ (19.1° vs 21.5°) and ψ (24.8° vs 29.7°), and its 3-state secondary structure was correct for 89% of residues vs 80% for TALOS-N. On the residues TALOS-N itself rates as confident, TAPAS was also more accurate (φ 14.9° vs 16.2°, ψ 17.7° vs 19.4°; both differences significant). At the ends of helices and strands, where secondary structure is hardest, TAPAS's 3-state accuracy was 82% vs 75% for TALOS-N on the same residues. (TAPAS version 0.2.5.) TAPAS also predicts SASA, which TALOS-N does not. On the same inputs and computer, TAPAS took about 0.35 s per protein versus about 37 s for TALOS-N (roughly 100 times faster); a full web job takes longer because of queueing and the optional structural-prior step.