v15 — Full Backbone + Confidence Prediction

Each protein is represented as a chain of residues. v15 inherits the same input format as v14, tracking 4 backbone atoms per residue:

• ids — (B, L) token sequence
• true_coords — (B, L, 3) Cα coordinates
• true_backbone — (B, L, 4, 3) [N, CA, C, O] atoms
• coord_mask — (B, L) per-residue mask
• backbone_mask — (B, L, 4) per-atom mask

A pre-trained encoder converts amino acid sequences into two representations. Its weights are not updated during training.

• Pre-trained ContactClassifier (weights frozen)
• Output: single (B, L, 128) — per-residue embeddings
• Output: contact_probs (B, L, L) — pairwise contact map

Builds a rich pairwise representation by combining several signals, then refines it with triangle multiplicative updates.

• SingleToPairV13 projects single → pair representation
• LogScaledRPE — 128 bins, max distance 512 residues
• OuterProductMean — d_opm=32
• Contact Map Conditioning — gated, d=128
• 8× EnhancedPairBlock: TriMul (out + in) + AxialBlock
• Output: pair (B, L, L, 128) — also feeds pAE head

• Normalize Cα coordinates (center + scale)
• Sample diffusion timestep t ~ U(0, T)
• Add Gaussian noise to Cα positions: x_t = α_t x_0 + σ_t ε
• Note: Noise applied only to Cα — full backbone is derived from frames

Each residue gets a local coordinate frame (rotation + translation) built from N, CA, C atoms.

• e1 = norm(CA->C)
• e3 = norm((N-CA) x e1)
• e2 = e3 x e1
• R = [e1 | e2 | e3], t = CA position

Self-Conditioning (50%): With 50% probability, run denoiser once to get x̂₀, then feed as additional input for the real forward pass.

IPA is a special attention mechanism that operates in 3D space, making it SE(3)-equivariant. Three attention channels are merged: scalar QKV, point-based geometric attention, and pair bias.

BackboneAtomHead — predicting all 4 atoms:

After IPA predicts CA positions and frames, the BackboneAtomHead places N, C, O atoms relative to each CA using ideal geometry + learned corrections.

• MLP: Linear(512→256) → ReLU → Linear(256→256) → ReLU → Linear(256→9)
• Output 9 values = 3 atoms × 3 coords (dN, dC, dO)
• pos_atom = t + R @ (ideal_offset + delta)
• Zero-init: starts at ideal geometry

Output: backbone_pred (B, L, 4, 3) and single representation h (B, L, 512) — feeds pLDDT head

pLDDT (predicted Local Distance Difference Test) measures per-residue accuracy by checking whether local CA-CA distances are preserved within distance thresholds (0.5, 1.0, 2.0, 4.0 Å). The head learns to predict its own accuracy — forcing the trunk to represent uncertainty.

Architecture:

LDDT definition:

For each residue, LDDT checks all CA-CA pairs within a 15Å radius and computes the fraction of distances preserved within thresholds:

• Ranges from 0 (completely wrong) to 1 (perfect)
• The 50-bin distribution is trained with cross-entropy on binned ground-truth LDDT
• Loss weight: w = 0.2
• At inference: pLDDT = ∑ bin_center × softmax(logits) gives a scalar per residue

Why it matters: The head forces the trunk's single representation to encode uncertainty. Loops and disordered regions get low pLDDT, well-packed cores get high pLDDT — the model learns what it doesn't know.

For each pair (i, j), pAE predicts the error of residue j's position when the structure is aligned using residue i's local frame. This tells us which parts of the structure are well-determined relative to each other — critical for multi-domain proteins where domains may be individually correct but their relative orientation is uncertain.

Architecture:

The L×L aligned error matrix:

• Input: pair representation (B, L, L, 128) from the pair stack
• MLP: Linear(128→64) → ReLU → Linear(64→64)
• 64 bins over [0, 32]Å aligned error range
• Loss: cross-entropy on binned aligned error, w = 0.2
• At inference: pAE_ij = ∑ bin_center × softmax(logits)

Key insight: pAE is not symmetric — aligning on residue i and scoring residue j is different from the reverse. Block-diagonal structure in the pAE matrix reveals domain boundaries: within-domain pairs have low error, cross-domain pairs have high error when the relative orientation is uncertain.

• AuxPairStack — lightweight pair refinement
• OrdinalDistanceHead — ordinal distance bins
• RgPredictor — radius of gyration prediction

Two new loss terms for the confidence prediction heads. These are not included in the structural metric.

Ground truth LDDT is computed from predicted vs. true CA-CA distances, then binned into 50 uniform bins over [0, 1]. Ground truth aligned error is computed per-pair by aligning on frame i and measuring distance of residue j, then binned into 64 bins over [0, 32]Å.

Total losses in v15: 9 (CA) + 4 (BB) + 2 (confidence) = 15 terms

Initialization: loaded from v14 run3 best weights v15 NEW

All existing modules (encoder, pair stack, IPA denoiser, backbone atom head, auxiliary heads) are initialized from the best v14 run3 checkpoint. The two new confidence heads (pLDDT, pAE) are randomly initialized.

Per-module Learning Rates (AdamW):

LR Schedule:

CFG: 10% dropout (classifier-free guidance)

EMA: 0.999 exponential moving average