Protein Structure
The 3D models, confidence signals, structural impacts, and alignment logic that project splicing events into physical space.
Every isoform in SPLISOFORMS is anchored to a three-dimensional protein model generated using AlphaFold 3 (AF3). These models, together with their per-residue confidence signals, are the substrate on which every downstream panel (domains, structural impact, disorder, PTM accessibility, interfaces) is computed. All canonical and novel isoforms are modelled under the same uniform AF3 workflow to ensure fully cross-comparable structures.
Model Confidence — pLDDT & PAE
Every predicted structure carries two AlphaFold confidence signals that gate the rest of the analysis stack — low-confidence regions are explicitly flagged rather than analysed as if reliable:
pLDDT — per-residue confidence
A score in [0, 100] stored in the CA B-factor column of the mmCIF. Higher = better. We follow standard bands:
- ≥ 90 — very high confidence (atomic detail trustworthy)
- 70–90 — confident backbone, plausible side chains
- 50–70 — low confidence; backbone may be roughly right but side chains drift
- < 50 — very low confidence; often disordered or modelled as a guess
The PTM, DSSP, and structural-landscape panels apply pLDDT-confidence gates so they mark low-confidence regions asundetermined /low_confidence rather than emitting structural state changes that ride on noise.
PAE — predicted aligned error
A residue × residue matrix where entry (i, j) is the model's expected position error of residue j when the structure is aligned on residue i. Low PAE between two residues = "their relative position is confident" = rigid pair.
We summarise the matrix as rigid blocks: contiguous regions where the block-internal mean PAE is below 6 Å. The block decomposition powers the structural-landscape panel and gates the Gold-Medal neoantigen tier. The greedy algorithm can over-fragment large rigid regions when edge residues bump above threshold; an adjacent-block merge rule (gap ≤ 3 residues, combined-mean tolerance ≤ +1) mitigates this.
Structural Impact (DSSP)
For every isoform structure we extract the N/CA/C/O backbone coordinates and run PyDSSP (a vectorized implementation of the DSSP algorithm) to assign per-residue secondary structures. By pairing canonical and novel residues via the splice-aware aligner, we track precise fold consequences of splicing. The resulting per-residue structural_change is one of the following:
DSSP assignment (Helix/Sheet/Coil) is unchanged between novel and canonical at this aligned position.
Confident structural transitions: both canonical and novel pLDDT ≥ 70. Six pairwise transitions between Helix, Sheet, and Coil are tracked.
At least one of canonical / novel pLDDT < 70. Side-chain placement is unreliable there, so the structural transition direction cannot be trusted.
The novel residue has no canonical counterpart in the splice-aware alignment — splicing introduced new sequence at this position.
Disorder & Functional Binding (IUPred3 / ANCHOR2)
Splicing frequently rewires intrinsically disordered regions (IDRs) which act as flexible hubs for protein-protein interactions. For each isoform, we predict per-residue intrinsic disorder and binding propensity using the IUPred3 and ANCHOR2 algorithms:
IUPred3 Disorder Propensity
Estimates the probability of a residue lacking a well-defined 3D structure. The IUPred3 model captures local sequence features and energy potentials.
Scores ≥ 0.5 mark residues that are highly likely to be flexible/disordered in vivo. Splicing within these regions often alters linear motifs without destabilizing a structured domain.
ANCHOR2 Functional Binding
Predicts protein binding regions located within intrinsically disordered sequences. ANCHOR2 evaluates the thermodynamic probability of an unstructured region undergoing a disorder-to-order transition upon binding a partner.
Scores ≥ 0.5 indicate a likely binding-coupled IDR. Splicing out these residues can cleanly ablate a specific protein interaction.
Structure Viewer
Every gene page includes an interactive Structure Viewer (powered by Mol*) that overlays multiple isoform structures simultaneously onto the canonical reference.
Superposition
Isoforms are superposed onto the canonical using only shared-exon residues identified by the splice-aware alignment. This prevents bogus superpositions across novel exon sequences.
RMSD
Root-Mean-Square Deviation (Å) is computed for each overlaid isoform and shown in the legend panel. Lower RMSD indicates a more structurally similar isoform.
Color Coding
The canonical reference is rendered in slate-silver. Each novel isoform receives a distinct color from a fixed palette (blue, red, emerald, amber, violet…).
Multi-Select
Up to several isoforms can be overlaid at once. Use the Mapped Transcripts table checkboxes (or Select All) to add/remove isoforms from the 3D view.
Sequence Alignment Panel (MSA)
The gene page also displays a multi-sequence alignment (MSA) panel comparing all selected isoforms to the canonical. Alignment is performed by a custom splice-aware algorithm rather than standard Needleman-Wunsch, ensuring biologically correct coloring:
Green columns
Novel exon insertions — sequence present only in the isoform, not in canonical.
Pink dashes
Skipped canonical exons — regions present in canonical but absent in the isoform.
Amber columns
Short substitutions (≤ 3 aa) — likely SNPs or minor junction differences, not splice events.