UNICORNE Documentation

A guide to using the Unified Conformational Ensemble database for exploring multiconformational protein structures.

1. View Modes

UNICORNE offers two ways to browse protein structures:

UniProt View (default)

A flat list of all UniProt entries. Each entry contains the aligned monomeric and oligomeric structure ensembles that map to that UniProt ID. Best for exploring a specific protein.

Fold/Domain View

Structures organized by fold and domain classifications. Navigate through Fold → Domain → UniProt → Structures. Best for exploring structural relationships across protein families.

Switch between views using the UniProt and Fold/Domain buttons at the top of the sidebar.

2. Searching

Type in the search bar and press Enter to search. You can search by:

• UniProt ID — e.g., P00323
• Entry name — e.g., FLAV_NITV2
• Protein name — e.g., Flavodoxin
• Structure ID — e.g., 1AKQ

In Fold/Domain view, you can also search by fold name or domain name.

3. Filters

Click the filter icon button to open the filter panel. Filters differ between views. Set your desired range, then press Enter or search to apply.

UniProt View Filters

• Conformations — Number of available structure conformations
• Min-TM Score — The minimum pairwise TM-score among all aligned structures (0–1). Lower values indicate greater structural diversity.
• Protein Length — Number of amino acid residues
• Avg Coverage — Average fraction of the sequence covered by aligned structures (0–1)

Fold/Domain View Filters

• Total Avg Angle — Average principal angle across all pairwise comparisons within the fold (degrees)
• Avg Seq Similarity — Average pairwise sequence identity among proteins in the fold (0–1)
• Direct UniProts — Number of UniProt entries directly assigned to the fold

All filters are double-sided range sliders. You can drag the endpoints or type exact values into the number fields.

4. UniProt Entries

Click on a UniProt entry to expand it. Inside each entry you will see:

• Entry Name and Protein Name — from the UniProt database
• Select All — Checks all structure checkboxes and starts the NMD animation in the normal mode viewer
• Clear — Unchecks all structures for this UniProt
• Share — Copies a direct URL to this UniProt entry to your clipboard
• Structure checkboxes — Check individual structures to load them. Clicking a checkbox loads the NMD animation for that UniProt in the normal mode viewer.
• Oligomers sublist — multi-chain assemblies for this UniProt, in a collapsible section (shown when available).
• Unaligned sublist — conformations with no aligned version on the site (see below), in a collapsible section.

Unaligned structures

Most conformations are structurally superposed onto a common reference so they overlay cleanly in the superposition viewer. A minority could not be aligned by the automated pipeline (for example very short chains, partial models, or atypical geometry). Rather than drop them, these are offered in a collapsible Unaligned sublist as their raw deposited coordinates — exactly as in the PDB, with no site alignment applied — so no conformation is lost.

To superimpose them yourself, check two or more unaligned structures to load them into the superposition viewer, then follow the step-by-step Aligning structures (TM-align) instructions below. TM-align is sequence-independent and often succeeds where the standard pipeline could not.

5. Normal Mode Viewer (Upper Panel)

Displays a single protein structure animated along its principal modes of conformational variation (PCA-based normal mode dynamics). Both 3D viewers are powered by Mol*.

Animation

• Play (▶) — The play button in the viewport toolbar animates the structure back and forth along the selected mode. Press again to pause.
• Mode selector — The dropdown at the top-left of the viewer chooses which principal component mode to animate. Each entry lists the mode's peak displacement in angstroms and the fraction of total variance it explains (e.g., “Mode 1 · 2.76 Å · 80.8%”). The animation amplitude matches the mode's √λ scaling, so higher-variance modes move further.

Displacement Colouring

By default the structure is coloured by per-residue displacement (RMSF): blue marks the most stable residues and red the most mobile, with the legend scaled in angstroms (Å) for the current mode. The scale is recomputed for whichever mode is selected. This displacement colouring stays the default even when you switch representation from the controls panel.

To use a different colour theme, see Changing the colour theme below — it works the same in both viewers.

Displacement Vectors

Click the arrow button (↗) in the viewport toolbar to toggle displacement vectors — cone-tipped arrows drawn from each residue showing the direction and magnitude of motion for the current mode. The button highlights teal while vectors are shown. Vectors follow the selected mode and clear automatically when you load a different protein.

Representation & Settings

Open the controls panel (the Structure Tools button in the viewport toolbar) to change the representation (cartoon, molecular surface, ball & stick, putty, …). Quick Styles gives one-click presets (Default / Cartoon / Spacefill / Surface). The toolbar also provides fullscreen (top), reset view, screenshot, and Mol*'s settings, including background colour.

Changing the colour theme

Colouring is handled by Mol*'s own controls, in both viewers:

1. Click Structure Tools in the viewport toolbar.
2. Scroll to Components and find the Polymer row.
3. Click the ⋯ (Actions) button on that row.
4. Choose Set Coloring, then pick a theme.

Themes are grouped by category — Atom Property, Chain Property, Residue Property, Symmetry, Validation, and Miscellaneous (Uniform, Illustrative, …). Common choices are sequence position (rainbow, N→C), secondary structure, chain, and uniform.

Hover

Hover over any residue to highlight it and show an info label with the chain, residue name, and residue number.

6. Superposition Viewer (Lower Panel)

Overlays multiple selected structures for visual comparison, also rendered with Mol*. Structures load automatically as you check them — there is no separate Visualize button. Unchecking a structure hides it without reloading the others.

Appearance

• Each structure is drawn as a cartoon in its own colour.
• To change the colour theme, see Changing the colour theme.
• Show 2 furthest (min TM) — a button in the stats sidebar loads only the two structures with the lowest pairwise TM-score (the greatest structural difference).

Overlaying structures (TM-align)

Structures loaded from the normal list are already superposed. Structures from the Unaligned sublist are raw deposited coordinates, so they arrive in unrelated frames and will look scattered until you overlay them.

Click the Overlay button (the stacked-layers icon at the bottom of the viewport toolbar). Every loaded structure is TM-aligned onto the first one, and the viewer reports the result — for example “Overlaid 5 onto the first — mean RMSD 2.20 Å, TM 0.815”.

Hover

Hover over any residue to see an info label; the structure's PDB ID is included so the overlaid models can be told apart.

Fullscreen

Use the fullscreen button — the first button in the viewport toolbar — to expand either viewer.

7. Sequence Panel

When at least one structure is loaded into the superposition viewer, a Sequence panel appears at the bottom of the right sidebar. It lists each loaded structure's sequence, aligned by residue number so the same column corresponds to the same UniProt position across all superposed structures.

• Each row is labeled with the structure ID in its model colour; gaps where a structure does not cover a residue show as ·.
• Sequence wraps in fixed-width blocks that match the panel's current width — resize the sidebar to see more or fewer residues per line.

8. Resizable Layout

Both sidebars are draggable. Hover over the divider between the protein list and the viewer area, or between the viewer area and the stats sidebar — the cursor will change to a resize cursor. Drag to set your preferred widths.

• The 3D viewers automatically resize as you drag, so the structure stays centered without clipping.
• The sequence panel re-flows its blocks to fit the new sidebar width.
• Widths are saved in your browser's local storage and persist across reloads.
• Keyboard: focus the divider with Tab and use ←/→ to nudge by 16 px at a time.

9. Statistics Panels

Statistics appear in the right sidebar when you select a UniProt, fold, or domain. Click the ⓘ button next to any statistic for a detailed explanation.

UniProt Statistics (flat view)

• Length — Number of amino acid residues
• Avg Coverage — Average fraction of the sequence covered by aligned structures
• Max GLOCON Score — Highest principal angle, indicating maximum conformational diversity
• Max GLOCON IDs — The two structures with the maximum principal angle difference
• Min TM Score — Lowest pairwise TM-score (greatest structural divergence)
• Min TM IDs — The two structures with the minimum TM-score

Fold Statistics (Fold/Domain view)

• Total Average — Average principal angle across all pairwise comparisons
• Min Average — Most conformationally conserved UniProt
• Avg Min — Average of each UniProt's minimum principal angle
• Min Min — Lowest principal angle observed in the fold
• Avg Seq Similarity — Average pairwise sequence identity

Domain Statistics (Fold/Domain view)

• Total Average — Average principal angle within the domain group
• Min Average — Most conformationally conserved UniProt in the domain
• Avg Min — Average of each UniProt's minimum principal angle
• Min Min — Lowest principal angle in the domain

10. Global Controls

• Download PDBs — Downloads all checked structures as individual PDB files
• Select All — Checks all visible structure checkboxes
• Clear All — Unchecks everything, stops animations, clears both viewers, and resets the Show dropdown
• Theme Toggle — Switch between dark mode (default) and light mode. Your preference is saved.
• Leaderboard — Opens the structure-prediction leaderboard (see section 12).
• Run — Opens /run, where you can generate ensembles for your own sequence (see section 13).

11. Icons & Indicators

• ▶ Play — Appears next to UniProt entries that have normal mode animation data available. In Fold/Domain view, play icons reflect availability for that specific fold context.
• Conformation count — Shown next to each UniProt: (N confs) when every conformation has an aligned version, or (X aligned / Y confs) when some are unaligned (X of the Y total conformations have aligned versions; the rest are in the Unaligned sublist).
• Green dot — Status indicator at the bottom of the page showing the application is connected and data is loaded.

12. Benchmark Leaderboard

The leaderboard at /leaderboard evaluates structure-prediction tools on four benchmark datasets, selected with the Dataset dropdown:

• UNICORNE_BENCH — 854 protein ensembles, each with multiple experimentally-determined ground-truth conformations. The headline benchmark, and the only one with afsample3.
• UNICORNE_V3 — 1950 ensembles, the broadest set.
• UNICORNE_OC24 — A curated 24-protein subset, each with 2–5 representative GT conformations drawn from a complete-linkage clustering of available PDB structures (see benchmarking-dataset-OC24-complete-linkage.csv). Smaller, faster to iterate against, used for focused method development.
• TRANSPORTERS — 14 membrane-transporter ensembles, each with 2 ground-truth conformations. A hard, domain-specific set.

Tools scored: alphaflow_pdb, afsample2, bioemu, cfrandom, esmdiff, esmflow_pdb, speach_af, plus afsample3 on UNICORNE_BENCH. Coverage varies slightly by dataset — a blank cell means that tool has no predictions for that ensemble.

The table

Each row is one ensemble; each tool column shows the mean of the per-GT best TM-score — for every ground-truth conformation, the closest predicted structure's TM is taken, and those values are averaged. Higher is better. Click a column header to sort. Use the search bar to filter by UniProt ID or protein name.

• GT — Number of ground-truth conformations in the ensemble.
• len — Reference sequence length.
• Pair TM — Minimum pairwise TM among the GT conformations. Lower means the ensemble is more conformationally diverse, which makes it harder to predict.
• Per-tool columns — Mean of best-per-GT TM. Color-coded green (≥ 0.8), amber (0.6–0.8), red (< 0.6).

Graph view

Toggle to Graph for a scatter plot of min pairwise TM across GT conformations (x — ensemble diversity) vs. worst-GT best-pred TM (y — tool score on the hardest conformation). Each point is one ensemble; one trace per tool. Click any point to load that ensemble's viewer. Filter by AF2 training-cutoff status (in / out / all) to compare tools on truly out-of-distribution targets.

3D viewer

Click a row (or graph point) to open the superposition viewer (rendered with Mol*). The viewer has two modes:

• GTs — Superpose every ground-truth conformation in the ensemble in grayscale. Toggle individual GTs on / off in the legend.
• Pred — Compare predictions against the ground truth. GT chips at the top act as multi-select toggles: clicking adds (or removes) that GT from the superposition; the most-recently-clicked GT is "active" and gets an outer outline. Each visible tool's best prediction for every selected GT is overlaid at the active opacity (slider) and in its tool color.

• Show GT — in Pred mode, once at least one prediction is visible, this checkbox hides the ground truth so you can look at the predictions on their own. Hiding it also makes the predictions fully opaque.
• Pred opacity slider controls the prediction transparency.
• Settings (gear icon) — Choose Color scheme (Per-tool, Rainbow, Secondary structure), Background, and Reset zoom.
• Fullscreen (⛶) and double-click on the viewer toggle fullscreen.

Metrics + Sequence

Below the viewer: a metrics table lists each tool's TM against the active GT, the full tool×tool pairwise-TM matrix at that GT (i.e., TM between every pair of tools' best-pred structures), and RMSD. The header reads Pairwise TM at GT: <name> — in Pred mode when multiple GTs are selected, this matrix follows the last-selected GT, so you always know which conformation the numbers correspond to. Eye toggles per row hide / show individual tools in the viewer. To the right of the viewer, the reference vs. GT-chain sequence alignment with mismatches in lowercase — hover any residue to highlight it in the alignment.

The inter-tool TM values come from data/inputs/<dataset>/<tool>-inter-tool-TM-scores.csv: one CSV per tool, one row per (uniprot, GT_conf), columns being the other six tools' TM-scores. Values are max(TM_chain1, TM_chain2, TM_avg) from a single TMalign call per pair.

13. Running tools on your own sequence

/run submits your own sequence to the same ensemble-generation tools the leaderboard scores, on lab GPU workstations. Jobs run one at a time in a first-in-first-out queue; the page shows live progress and the finished structures are downloadable as a ZIP.

Submitting

• Sequence — a single amino-acid sequence, up to 600 residues. You can also paste several sequences to submit them as a batch, which queues one job per sequence.
• Tool — pick one of the tools below.
• Conformations — how many structures to generate, from the tool's minimum up to 100. The minimum is shown under the field, because some tools cannot produce fewer.

Available tools

Tool	Min. conformations	MSA	Notes
BioEMU	1	no	Sequence → backbone ensemble, emulating equilibrium fluctuations.
ESMFlow	1	no	ESMFold-based flow model.
ESMDiff	1	no	Diffusion model over ESMFold.
AlphaFlow	1	yes	AlphaFold-based flow model.
SPEACH_AF	1	yes	Systematic MSA mutation to surface alternative states.
AFsample2	5	yes	MSA subsampling for conformational diversity.
CFrandom	25	yes	ColabFold with random seeds + variable MSA depth.

Watching a job

• Each job shows its state (queued → running → done), live progress, and an ETA estimated from the sequence length and the number of conformations requested.
• Stop ends a running job immediately; Cancel removes one that is still queued. In a batch you can cancel the whole batch, or expand it and cancel individual sequences.
• The jobs list scrolls, so earlier runs stay reachable.

Results

A finished job offers a Download button giving every generated conformation as a ZIP, and the structures can be previewed in the built-in viewer. Outputs are kept in object storage under a per-job directory.

---

UNICORNE — MIT License — Affiliation withheld for double-blind review